1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation -----===//
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 defines the interfaces that AArch64 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "aarch64-isel"
17 #include "AArch64ISelLowering.h"
18 #include "AArch64MachineFunctionInfo.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "Utils/AArch64BaseInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/Support/MathExtras.h"
33 static TargetLoweringObjectFile *createTLOF(AArch64TargetMachine &TM) {
34 assert (TM.getSubtarget<AArch64Subtarget>().isTargetELF() &&
35 "unknown subtarget type");
36 return new AArch64ElfTargetObjectFile();
39 AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
40 : TargetLowering(TM, createTLOF(TM)), Itins(TM.getInstrItineraryData()) {
42 const AArch64Subtarget *Subtarget = &TM.getSubtarget<AArch64Subtarget>();
44 // SIMD compares set the entire lane's bits to 1
45 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
47 // Scalar register <-> type mapping
48 addRegisterClass(MVT::i32, &AArch64::GPR32RegClass);
49 addRegisterClass(MVT::i64, &AArch64::GPR64RegClass);
51 if (Subtarget->hasFPARMv8()) {
52 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
53 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
54 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
55 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
58 if (Subtarget->hasNEON()) {
60 addRegisterClass(MVT::v1i8, &AArch64::FPR8RegClass);
61 addRegisterClass(MVT::v1i16, &AArch64::FPR16RegClass);
62 addRegisterClass(MVT::v1i32, &AArch64::FPR32RegClass);
63 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
64 addRegisterClass(MVT::v1f64, &AArch64::FPR64RegClass);
65 addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
66 addRegisterClass(MVT::v4i16, &AArch64::FPR64RegClass);
67 addRegisterClass(MVT::v2i32, &AArch64::FPR64RegClass);
68 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
69 addRegisterClass(MVT::v2f32, &AArch64::FPR64RegClass);
70 addRegisterClass(MVT::v16i8, &AArch64::FPR128RegClass);
71 addRegisterClass(MVT::v8i16, &AArch64::FPR128RegClass);
72 addRegisterClass(MVT::v4i32, &AArch64::FPR128RegClass);
73 addRegisterClass(MVT::v2i64, &AArch64::FPR128RegClass);
74 addRegisterClass(MVT::v4f32, &AArch64::FPR128RegClass);
75 addRegisterClass(MVT::v2f64, &AArch64::FPR128RegClass);
78 computeRegisterProperties();
80 // We combine OR nodes for bitfield and NEON BSL operations.
81 setTargetDAGCombine(ISD::OR);
83 setTargetDAGCombine(ISD::AND);
84 setTargetDAGCombine(ISD::SRA);
85 setTargetDAGCombine(ISD::SRL);
86 setTargetDAGCombine(ISD::SHL);
88 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
89 setTargetDAGCombine(ISD::INTRINSIC_VOID);
90 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
92 // AArch64 does not have i1 loads, or much of anything for i1 really.
93 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
94 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
95 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
97 setStackPointerRegisterToSaveRestore(AArch64::XSP);
98 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
99 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
100 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
102 // We'll lower globals to wrappers for selection.
103 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
104 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
106 // A64 instructions have the comparison predicate attached to the user of the
107 // result, but having a separate comparison is valuable for matching.
108 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
109 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
110 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
111 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
113 setOperationAction(ISD::SELECT, MVT::i32, Custom);
114 setOperationAction(ISD::SELECT, MVT::i64, Custom);
115 setOperationAction(ISD::SELECT, MVT::f32, Custom);
116 setOperationAction(ISD::SELECT, MVT::f64, Custom);
118 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
119 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
120 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
121 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
123 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
125 setOperationAction(ISD::SETCC, MVT::i32, Custom);
126 setOperationAction(ISD::SETCC, MVT::i64, Custom);
127 setOperationAction(ISD::SETCC, MVT::f32, Custom);
128 setOperationAction(ISD::SETCC, MVT::f64, Custom);
130 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
131 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
132 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
134 setOperationAction(ISD::VASTART, MVT::Other, Custom);
135 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
136 setOperationAction(ISD::VAEND, MVT::Other, Expand);
137 setOperationAction(ISD::VAARG, MVT::Other, Expand);
139 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
140 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
142 setOperationAction(ISD::ROTL, MVT::i32, Expand);
143 setOperationAction(ISD::ROTL, MVT::i64, Expand);
145 setOperationAction(ISD::UREM, MVT::i32, Expand);
146 setOperationAction(ISD::UREM, MVT::i64, Expand);
147 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
148 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
150 setOperationAction(ISD::SREM, MVT::i32, Expand);
151 setOperationAction(ISD::SREM, MVT::i64, Expand);
152 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
153 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
155 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
156 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
157 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
158 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
160 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
161 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
163 // Legal floating-point operations.
164 setOperationAction(ISD::FABS, MVT::f32, Legal);
165 setOperationAction(ISD::FABS, MVT::f64, Legal);
167 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
168 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
170 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
171 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
173 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
174 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
176 setOperationAction(ISD::FNEG, MVT::f32, Legal);
177 setOperationAction(ISD::FNEG, MVT::f64, Legal);
179 setOperationAction(ISD::FRINT, MVT::f32, Legal);
180 setOperationAction(ISD::FRINT, MVT::f64, Legal);
182 setOperationAction(ISD::FSQRT, MVT::f32, Legal);
183 setOperationAction(ISD::FSQRT, MVT::f64, Legal);
185 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
186 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
188 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
189 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
190 setOperationAction(ISD::ConstantFP, MVT::f128, Legal);
192 // Illegal floating-point operations.
193 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
194 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
196 setOperationAction(ISD::FCOS, MVT::f32, Expand);
197 setOperationAction(ISD::FCOS, MVT::f64, Expand);
199 setOperationAction(ISD::FEXP, MVT::f32, Expand);
200 setOperationAction(ISD::FEXP, MVT::f64, Expand);
202 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
203 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
205 setOperationAction(ISD::FLOG, MVT::f32, Expand);
206 setOperationAction(ISD::FLOG, MVT::f64, Expand);
208 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
209 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
211 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
212 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
214 setOperationAction(ISD::FPOW, MVT::f32, Expand);
215 setOperationAction(ISD::FPOW, MVT::f64, Expand);
217 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
218 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
220 setOperationAction(ISD::FREM, MVT::f32, Expand);
221 setOperationAction(ISD::FREM, MVT::f64, Expand);
223 setOperationAction(ISD::FSIN, MVT::f32, Expand);
224 setOperationAction(ISD::FSIN, MVT::f64, Expand);
226 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
227 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
229 // Virtually no operation on f128 is legal, but LLVM can't expand them when
230 // there's a valid register class, so we need custom operations in most cases.
231 setOperationAction(ISD::FABS, MVT::f128, Expand);
232 setOperationAction(ISD::FADD, MVT::f128, Custom);
233 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
234 setOperationAction(ISD::FCOS, MVT::f128, Expand);
235 setOperationAction(ISD::FDIV, MVT::f128, Custom);
236 setOperationAction(ISD::FMA, MVT::f128, Expand);
237 setOperationAction(ISD::FMUL, MVT::f128, Custom);
238 setOperationAction(ISD::FNEG, MVT::f128, Expand);
239 setOperationAction(ISD::FP_EXTEND, MVT::f128, Expand);
240 setOperationAction(ISD::FP_ROUND, MVT::f128, Expand);
241 setOperationAction(ISD::FPOW, MVT::f128, Expand);
242 setOperationAction(ISD::FREM, MVT::f128, Expand);
243 setOperationAction(ISD::FRINT, MVT::f128, Expand);
244 setOperationAction(ISD::FSIN, MVT::f128, Expand);
245 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
246 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
247 setOperationAction(ISD::FSUB, MVT::f128, Custom);
248 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
249 setOperationAction(ISD::SETCC, MVT::f128, Custom);
250 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
251 setOperationAction(ISD::SELECT, MVT::f128, Expand);
252 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
253 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
255 // Lowering for many of the conversions is actually specified by the non-f128
256 // type. The LowerXXX function will be trivial when f128 isn't involved.
257 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
258 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
259 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
260 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
261 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
262 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
263 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
264 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
265 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
266 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
267 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
268 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
269 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
270 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
272 // This prevents LLVM trying to compress double constants into a floating
273 // constant-pool entry and trying to load from there. It's of doubtful benefit
274 // for A64: we'd need LDR followed by FCVT, I believe.
275 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
276 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
277 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
279 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
280 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
281 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
282 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
283 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
284 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
286 setExceptionPointerRegister(AArch64::X0);
287 setExceptionSelectorRegister(AArch64::X1);
289 if (Subtarget->hasNEON()) {
290 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i8, Expand);
291 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Expand);
292 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Expand);
293 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v1i64, Expand);
294 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v16i8, Expand);
295 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i16, Expand);
296 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Expand);
297 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Expand);
299 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i8, Custom);
300 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
301 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
302 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i16, Custom);
303 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
304 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
305 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i32, Custom);
306 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
307 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
308 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
309 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
310 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
311 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
312 setOperationAction(ISD::BUILD_VECTOR, MVT::v1f64, Custom);
313 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
315 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
316 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
317 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
318 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i16, Custom);
319 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
320 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i32, Custom);
321 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
322 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
323 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f32, Custom);
324 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
325 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1f64, Custom);
326 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
328 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i32, Legal);
329 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Legal);
330 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Legal);
331 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Legal);
332 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Legal);
333 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Legal);
334 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Legal);
336 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i8, Custom);
337 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i16, Custom);
338 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
339 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
340 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
342 setOperationAction(ISD::SETCC, MVT::v8i8, Custom);
343 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
344 setOperationAction(ISD::SETCC, MVT::v4i16, Custom);
345 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
346 setOperationAction(ISD::SETCC, MVT::v2i32, Custom);
347 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
348 setOperationAction(ISD::SETCC, MVT::v1i64, Custom);
349 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
350 setOperationAction(ISD::SETCC, MVT::v2f32, Custom);
351 setOperationAction(ISD::SETCC, MVT::v4f32, Custom);
352 setOperationAction(ISD::SETCC, MVT::v1f64, Custom);
353 setOperationAction(ISD::SETCC, MVT::v2f64, Custom);
355 setOperationAction(ISD::FFLOOR, MVT::v2f32, Legal);
356 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
357 setOperationAction(ISD::FFLOOR, MVT::v1f64, Legal);
358 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
360 setOperationAction(ISD::FCEIL, MVT::v2f32, Legal);
361 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
362 setOperationAction(ISD::FCEIL, MVT::v1f64, Legal);
363 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
365 setOperationAction(ISD::FTRUNC, MVT::v2f32, Legal);
366 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
367 setOperationAction(ISD::FTRUNC, MVT::v1f64, Legal);
368 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
370 setOperationAction(ISD::FRINT, MVT::v2f32, Legal);
371 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
372 setOperationAction(ISD::FRINT, MVT::v1f64, Legal);
373 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
375 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Legal);
376 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
377 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Legal);
378 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
380 setOperationAction(ISD::FROUND, MVT::v2f32, Legal);
381 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
382 setOperationAction(ISD::FROUND, MVT::v1f64, Legal);
383 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
385 setOperationAction(ISD::SINT_TO_FP, MVT::v1i8, Custom);
386 setOperationAction(ISD::SINT_TO_FP, MVT::v1i16, Custom);
387 setOperationAction(ISD::SINT_TO_FP, MVT::v1i32, Custom);
388 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
389 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
390 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
392 setOperationAction(ISD::UINT_TO_FP, MVT::v1i8, Custom);
393 setOperationAction(ISD::UINT_TO_FP, MVT::v1i16, Custom);
394 setOperationAction(ISD::UINT_TO_FP, MVT::v1i32, Custom);
395 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
396 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
397 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
399 setOperationAction(ISD::FP_TO_SINT, MVT::v1i8, Custom);
400 setOperationAction(ISD::FP_TO_SINT, MVT::v1i16, Custom);
401 setOperationAction(ISD::FP_TO_SINT, MVT::v1i32, Custom);
402 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
403 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
404 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Custom);
406 setOperationAction(ISD::FP_TO_UINT, MVT::v1i8, Custom);
407 setOperationAction(ISD::FP_TO_UINT, MVT::v1i16, Custom);
408 setOperationAction(ISD::FP_TO_UINT, MVT::v1i32, Custom);
409 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
410 setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom);
411 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Custom);
413 // Neon does not support vector divide/remainder operations except
414 // floating-point divide.
415 setOperationAction(ISD::SDIV, MVT::v1i8, Expand);
416 setOperationAction(ISD::SDIV, MVT::v8i8, Expand);
417 setOperationAction(ISD::SDIV, MVT::v16i8, Expand);
418 setOperationAction(ISD::SDIV, MVT::v1i16, Expand);
419 setOperationAction(ISD::SDIV, MVT::v4i16, Expand);
420 setOperationAction(ISD::SDIV, MVT::v8i16, Expand);
421 setOperationAction(ISD::SDIV, MVT::v1i32, Expand);
422 setOperationAction(ISD::SDIV, MVT::v2i32, Expand);
423 setOperationAction(ISD::SDIV, MVT::v4i32, Expand);
424 setOperationAction(ISD::SDIV, MVT::v1i64, Expand);
425 setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
427 setOperationAction(ISD::UDIV, MVT::v1i8, Expand);
428 setOperationAction(ISD::UDIV, MVT::v8i8, Expand);
429 setOperationAction(ISD::UDIV, MVT::v16i8, Expand);
430 setOperationAction(ISD::UDIV, MVT::v1i16, Expand);
431 setOperationAction(ISD::UDIV, MVT::v4i16, Expand);
432 setOperationAction(ISD::UDIV, MVT::v8i16, Expand);
433 setOperationAction(ISD::UDIV, MVT::v1i32, Expand);
434 setOperationAction(ISD::UDIV, MVT::v2i32, Expand);
435 setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
436 setOperationAction(ISD::UDIV, MVT::v1i64, Expand);
437 setOperationAction(ISD::UDIV, MVT::v2i64, Expand);
439 setOperationAction(ISD::SREM, MVT::v1i8, Expand);
440 setOperationAction(ISD::SREM, MVT::v8i8, Expand);
441 setOperationAction(ISD::SREM, MVT::v16i8, Expand);
442 setOperationAction(ISD::SREM, MVT::v1i16, Expand);
443 setOperationAction(ISD::SREM, MVT::v4i16, Expand);
444 setOperationAction(ISD::SREM, MVT::v8i16, Expand);
445 setOperationAction(ISD::SREM, MVT::v1i32, Expand);
446 setOperationAction(ISD::SREM, MVT::v2i32, Expand);
447 setOperationAction(ISD::SREM, MVT::v4i32, Expand);
448 setOperationAction(ISD::SREM, MVT::v1i64, Expand);
449 setOperationAction(ISD::SREM, MVT::v2i64, Expand);
451 setOperationAction(ISD::UREM, MVT::v1i8, Expand);
452 setOperationAction(ISD::UREM, MVT::v8i8, Expand);
453 setOperationAction(ISD::UREM, MVT::v16i8, Expand);
454 setOperationAction(ISD::UREM, MVT::v1i16, Expand);
455 setOperationAction(ISD::UREM, MVT::v4i16, Expand);
456 setOperationAction(ISD::UREM, MVT::v8i16, Expand);
457 setOperationAction(ISD::UREM, MVT::v1i32, Expand);
458 setOperationAction(ISD::UREM, MVT::v2i32, Expand);
459 setOperationAction(ISD::UREM, MVT::v4i32, Expand);
460 setOperationAction(ISD::UREM, MVT::v1i64, Expand);
461 setOperationAction(ISD::UREM, MVT::v2i64, Expand);
463 setOperationAction(ISD::FREM, MVT::v2f32, Expand);
464 setOperationAction(ISD::FREM, MVT::v4f32, Expand);
465 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
466 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
468 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
469 setOperationAction(ISD::SELECT, MVT::v16i8, Expand);
470 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
471 setOperationAction(ISD::SELECT, MVT::v8i16, Expand);
472 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
473 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
474 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
475 setOperationAction(ISD::SELECT, MVT::v2i64, Expand);
476 setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
477 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
478 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
479 setOperationAction(ISD::SELECT, MVT::v2f64, Expand);
481 setOperationAction(ISD::SELECT_CC, MVT::v8i8, Custom);
482 setOperationAction(ISD::SELECT_CC, MVT::v16i8, Custom);
483 setOperationAction(ISD::SELECT_CC, MVT::v4i16, Custom);
484 setOperationAction(ISD::SELECT_CC, MVT::v8i16, Custom);
485 setOperationAction(ISD::SELECT_CC, MVT::v2i32, Custom);
486 setOperationAction(ISD::SELECT_CC, MVT::v4i32, Custom);
487 setOperationAction(ISD::SELECT_CC, MVT::v1i64, Custom);
488 setOperationAction(ISD::SELECT_CC, MVT::v2i64, Custom);
489 setOperationAction(ISD::SELECT_CC, MVT::v2f32, Custom);
490 setOperationAction(ISD::SELECT_CC, MVT::v4f32, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Custom);
492 setOperationAction(ISD::SELECT_CC, MVT::v2f64, Custom);
494 // Vector ExtLoad and TruncStore are expanded.
495 for (unsigned I = MVT::FIRST_VECTOR_VALUETYPE;
496 I <= MVT::LAST_VECTOR_VALUETYPE; ++I) {
497 MVT VT = (MVT::SimpleValueType) I;
498 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
499 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
500 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
501 for (unsigned II = MVT::FIRST_VECTOR_VALUETYPE;
502 II <= MVT::LAST_VECTOR_VALUETYPE; ++II) {
503 MVT VT1 = (MVT::SimpleValueType) II;
504 // A TruncStore has two vector types of the same number of elements
505 // and different element sizes.
506 if (VT.getVectorNumElements() == VT1.getVectorNumElements() &&
507 VT.getVectorElementType().getSizeInBits()
508 > VT1.getVectorElementType().getSizeInBits())
509 setTruncStoreAction(VT, VT1, Expand);
513 // There is no v1i64/v2i64 multiply, expand v1i64/v2i64 to GPR i64 multiply.
514 // FIXME: For a v2i64 multiply, we copy VPR to GPR and do 2 i64 multiplies,
515 // and then copy back to VPR. This solution may be optimized by Following 3
516 // NEON instructions:
517 // pmull v2.1q, v0.1d, v1.1d
518 // pmull2 v3.1q, v0.2d, v1.2d
519 // ins v2.d[1], v3.d[0]
520 // As currently we can't verify the correctness of such assumption, we can
521 // do such optimization in the future.
522 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
523 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
525 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
526 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
527 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
528 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
529 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
530 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
531 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
532 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
533 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
536 setTargetDAGCombine(ISD::SETCC);
537 setTargetDAGCombine(ISD::SIGN_EXTEND);
538 setTargetDAGCombine(ISD::VSELECT);
541 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
542 // It's reasonably important that this value matches the "natural" legal
543 // promotion from i1 for scalar types. Otherwise LegalizeTypes can get itself
544 // in a twist (e.g. inserting an any_extend which then becomes i64 -> i64).
545 if (!VT.isVector()) return MVT::i32;
546 return VT.changeVectorElementTypeToInteger();
549 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
552 static const unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
553 AArch64::LDXR_word, AArch64::LDXR_dword};
554 static const unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
555 AArch64::LDAXR_word, AArch64::LDAXR_dword};
556 static const unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
557 AArch64::STXR_word, AArch64::STXR_dword};
558 static const unsigned StoreRels[] = {AArch64::STLXR_byte,AArch64::STLXR_hword,
559 AArch64::STLXR_word, AArch64::STLXR_dword};
561 const unsigned *LoadOps, *StoreOps;
562 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
567 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
568 StoreOps = StoreRels;
570 StoreOps = StoreBares;
572 assert(isPowerOf2_32(Size) && Size <= 8 &&
573 "unsupported size for atomic binary op!");
575 LdrOpc = LoadOps[Log2_32(Size)];
576 StrOpc = StoreOps[Log2_32(Size)];
579 // FIXME: AArch64::DTripleRegClass and AArch64::QTripleRegClass don't really
580 // have value type mapped, and they are both being defined as MVT::untyped.
581 // Without knowing the MVT type, MachineLICM::getRegisterClassIDAndCost
582 // would fail to figure out the register pressure correctly.
583 std::pair<const TargetRegisterClass*, uint8_t>
584 AArch64TargetLowering::findRepresentativeClass(MVT VT) const{
585 const TargetRegisterClass *RRC = 0;
587 switch (VT.SimpleTy) {
589 return TargetLowering::findRepresentativeClass(VT);
591 RRC = &AArch64::QPairRegClass;
595 RRC = &AArch64::QQuadRegClass;
599 return std::make_pair(RRC, Cost);
603 AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
605 unsigned BinOpcode) const {
606 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
607 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
609 const BasicBlock *LLVM_BB = BB->getBasicBlock();
610 MachineFunction *MF = BB->getParent();
611 MachineFunction::iterator It = BB;
614 unsigned dest = MI->getOperand(0).getReg();
615 unsigned ptr = MI->getOperand(1).getReg();
616 unsigned incr = MI->getOperand(2).getReg();
617 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
618 DebugLoc dl = MI->getDebugLoc();
620 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
622 unsigned ldrOpc, strOpc;
623 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
625 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
626 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
627 MF->insert(It, loopMBB);
628 MF->insert(It, exitMBB);
630 // Transfer the remainder of BB and its successor edges to exitMBB.
631 exitMBB->splice(exitMBB->begin(), BB,
632 llvm::next(MachineBasicBlock::iterator(MI)),
634 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
636 const TargetRegisterClass *TRC
637 = Size == 8 ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
638 unsigned scratch = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
642 // fallthrough --> loopMBB
643 BB->addSuccessor(loopMBB);
647 // <binop> scratch, dest, incr
648 // stxr stxr_status, scratch, ptr
649 // cbnz stxr_status, loopMBB
650 // fallthrough --> exitMBB
652 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
654 // All arithmetic operations we'll be creating are designed to take an extra
655 // shift or extend operand, which we can conveniently set to zero.
657 // Operand order needs to go the other way for NAND.
658 if (BinOpcode == AArch64::BICwww_lsl || BinOpcode == AArch64::BICxxx_lsl)
659 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
660 .addReg(incr).addReg(dest).addImm(0);
662 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
663 .addReg(dest).addReg(incr).addImm(0);
666 // From the stxr, the register is GPR32; from the cmp it's GPR32wsp
667 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
668 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
670 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(scratch).addReg(ptr);
671 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
672 .addReg(stxr_status).addMBB(loopMBB);
674 BB->addSuccessor(loopMBB);
675 BB->addSuccessor(exitMBB);
681 MI->eraseFromParent(); // The instruction is gone now.
687 AArch64TargetLowering::emitAtomicBinaryMinMax(MachineInstr *MI,
688 MachineBasicBlock *BB,
691 A64CC::CondCodes Cond) const {
692 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
694 const BasicBlock *LLVM_BB = BB->getBasicBlock();
695 MachineFunction *MF = BB->getParent();
696 MachineFunction::iterator It = BB;
699 unsigned dest = MI->getOperand(0).getReg();
700 unsigned ptr = MI->getOperand(1).getReg();
701 unsigned incr = MI->getOperand(2).getReg();
702 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
704 unsigned oldval = dest;
705 DebugLoc dl = MI->getDebugLoc();
707 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
708 const TargetRegisterClass *TRC, *TRCsp;
710 TRC = &AArch64::GPR64RegClass;
711 TRCsp = &AArch64::GPR64xspRegClass;
713 TRC = &AArch64::GPR32RegClass;
714 TRCsp = &AArch64::GPR32wspRegClass;
717 unsigned ldrOpc, strOpc;
718 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
720 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
721 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
722 MF->insert(It, loopMBB);
723 MF->insert(It, exitMBB);
725 // Transfer the remainder of BB and its successor edges to exitMBB.
726 exitMBB->splice(exitMBB->begin(), BB,
727 llvm::next(MachineBasicBlock::iterator(MI)),
729 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
731 unsigned scratch = MRI.createVirtualRegister(TRC);
732 MRI.constrainRegClass(scratch, TRCsp);
736 // fallthrough --> loopMBB
737 BB->addSuccessor(loopMBB);
741 // cmp incr, dest (, sign extend if necessary)
742 // csel scratch, dest, incr, cond
743 // stxr stxr_status, scratch, ptr
744 // cbnz stxr_status, loopMBB
745 // fallthrough --> exitMBB
747 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
749 // Build compare and cmov instructions.
750 MRI.constrainRegClass(incr, TRCsp);
751 BuildMI(BB, dl, TII->get(CmpOp))
752 .addReg(incr).addReg(oldval).addImm(0);
754 BuildMI(BB, dl, TII->get(Size == 8 ? AArch64::CSELxxxc : AArch64::CSELwwwc),
756 .addReg(oldval).addReg(incr).addImm(Cond);
758 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
759 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
761 BuildMI(BB, dl, TII->get(strOpc), stxr_status)
762 .addReg(scratch).addReg(ptr);
763 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
764 .addReg(stxr_status).addMBB(loopMBB);
766 BB->addSuccessor(loopMBB);
767 BB->addSuccessor(exitMBB);
773 MI->eraseFromParent(); // The instruction is gone now.
779 AArch64TargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
780 MachineBasicBlock *BB,
781 unsigned Size) const {
782 unsigned dest = MI->getOperand(0).getReg();
783 unsigned ptr = MI->getOperand(1).getReg();
784 unsigned oldval = MI->getOperand(2).getReg();
785 unsigned newval = MI->getOperand(3).getReg();
786 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
787 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
788 DebugLoc dl = MI->getDebugLoc();
790 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
791 const TargetRegisterClass *TRCsp;
792 TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
794 unsigned ldrOpc, strOpc;
795 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
797 MachineFunction *MF = BB->getParent();
798 const BasicBlock *LLVM_BB = BB->getBasicBlock();
799 MachineFunction::iterator It = BB;
800 ++It; // insert the new blocks after the current block
802 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
803 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
804 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
805 MF->insert(It, loop1MBB);
806 MF->insert(It, loop2MBB);
807 MF->insert(It, exitMBB);
809 // Transfer the remainder of BB and its successor edges to exitMBB.
810 exitMBB->splice(exitMBB->begin(), BB,
811 llvm::next(MachineBasicBlock::iterator(MI)),
813 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
817 // fallthrough --> loop1MBB
818 BB->addSuccessor(loop1MBB);
825 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
827 unsigned CmpOp = Size == 8 ? AArch64::CMPxx_lsl : AArch64::CMPww_lsl;
828 MRI.constrainRegClass(dest, TRCsp);
829 BuildMI(BB, dl, TII->get(CmpOp))
830 .addReg(dest).addReg(oldval).addImm(0);
831 BuildMI(BB, dl, TII->get(AArch64::Bcc))
832 .addImm(A64CC::NE).addMBB(exitMBB);
833 BB->addSuccessor(loop2MBB);
834 BB->addSuccessor(exitMBB);
837 // strex stxr_status, newval, [ptr]
838 // cbnz stxr_status, loop1MBB
840 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
841 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
843 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(newval).addReg(ptr);
844 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
845 .addReg(stxr_status).addMBB(loop1MBB);
846 BB->addSuccessor(loop1MBB);
847 BB->addSuccessor(exitMBB);
853 MI->eraseFromParent(); // The instruction is gone now.
859 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
860 MachineBasicBlock *MBB) const {
861 // We materialise the F128CSEL pseudo-instruction using conditional branches
862 // and loads, giving an instruciton sequence like:
871 // Using virtual registers would probably not be beneficial since COPY
872 // instructions are expensive for f128 (there's no actual instruction to
875 // An alternative would be to do an integer-CSEL on some address. E.g.:
880 // csel x0, x0, x1, ne
883 // It's unclear which approach is actually optimal.
884 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
885 MachineFunction *MF = MBB->getParent();
886 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
887 DebugLoc DL = MI->getDebugLoc();
888 MachineFunction::iterator It = MBB;
891 unsigned DestReg = MI->getOperand(0).getReg();
892 unsigned IfTrueReg = MI->getOperand(1).getReg();
893 unsigned IfFalseReg = MI->getOperand(2).getReg();
894 unsigned CondCode = MI->getOperand(3).getImm();
895 bool NZCVKilled = MI->getOperand(4).isKill();
897 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
898 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
899 MF->insert(It, TrueBB);
900 MF->insert(It, EndBB);
902 // Transfer rest of current basic-block to EndBB
903 EndBB->splice(EndBB->begin(), MBB,
904 llvm::next(MachineBasicBlock::iterator(MI)),
906 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
908 // We need somewhere to store the f128 value needed.
909 int ScratchFI = MF->getFrameInfo()->CreateSpillStackObject(16, 16);
911 // [... start of incoming MBB ...]
912 // str qIFFALSE, [sp]
915 BuildMI(MBB, DL, TII->get(AArch64::LSFP128_STR))
917 .addFrameIndex(ScratchFI)
919 BuildMI(MBB, DL, TII->get(AArch64::Bcc))
922 BuildMI(MBB, DL, TII->get(AArch64::Bimm))
924 MBB->addSuccessor(TrueBB);
925 MBB->addSuccessor(EndBB);
928 // NZCV is live-through TrueBB.
929 TrueBB->addLiveIn(AArch64::NZCV);
930 EndBB->addLiveIn(AArch64::NZCV);
935 BuildMI(TrueBB, DL, TII->get(AArch64::LSFP128_STR))
937 .addFrameIndex(ScratchFI)
940 // Note: fallthrough. We can rely on LLVM adding a branch if it reorders the
942 TrueBB->addSuccessor(EndBB);
946 // [... rest of incoming MBB ...]
947 MachineInstr *StartOfEnd = EndBB->begin();
948 BuildMI(*EndBB, StartOfEnd, DL, TII->get(AArch64::LSFP128_LDR), DestReg)
949 .addFrameIndex(ScratchFI)
952 MI->eraseFromParent();
957 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
958 MachineBasicBlock *MBB) const {
959 switch (MI->getOpcode()) {
960 default: llvm_unreachable("Unhandled instruction with custom inserter");
961 case AArch64::F128CSEL:
962 return EmitF128CSEL(MI, MBB);
963 case AArch64::ATOMIC_LOAD_ADD_I8:
964 return emitAtomicBinary(MI, MBB, 1, AArch64::ADDwww_lsl);
965 case AArch64::ATOMIC_LOAD_ADD_I16:
966 return emitAtomicBinary(MI, MBB, 2, AArch64::ADDwww_lsl);
967 case AArch64::ATOMIC_LOAD_ADD_I32:
968 return emitAtomicBinary(MI, MBB, 4, AArch64::ADDwww_lsl);
969 case AArch64::ATOMIC_LOAD_ADD_I64:
970 return emitAtomicBinary(MI, MBB, 8, AArch64::ADDxxx_lsl);
972 case AArch64::ATOMIC_LOAD_SUB_I8:
973 return emitAtomicBinary(MI, MBB, 1, AArch64::SUBwww_lsl);
974 case AArch64::ATOMIC_LOAD_SUB_I16:
975 return emitAtomicBinary(MI, MBB, 2, AArch64::SUBwww_lsl);
976 case AArch64::ATOMIC_LOAD_SUB_I32:
977 return emitAtomicBinary(MI, MBB, 4, AArch64::SUBwww_lsl);
978 case AArch64::ATOMIC_LOAD_SUB_I64:
979 return emitAtomicBinary(MI, MBB, 8, AArch64::SUBxxx_lsl);
981 case AArch64::ATOMIC_LOAD_AND_I8:
982 return emitAtomicBinary(MI, MBB, 1, AArch64::ANDwww_lsl);
983 case AArch64::ATOMIC_LOAD_AND_I16:
984 return emitAtomicBinary(MI, MBB, 2, AArch64::ANDwww_lsl);
985 case AArch64::ATOMIC_LOAD_AND_I32:
986 return emitAtomicBinary(MI, MBB, 4, AArch64::ANDwww_lsl);
987 case AArch64::ATOMIC_LOAD_AND_I64:
988 return emitAtomicBinary(MI, MBB, 8, AArch64::ANDxxx_lsl);
990 case AArch64::ATOMIC_LOAD_OR_I8:
991 return emitAtomicBinary(MI, MBB, 1, AArch64::ORRwww_lsl);
992 case AArch64::ATOMIC_LOAD_OR_I16:
993 return emitAtomicBinary(MI, MBB, 2, AArch64::ORRwww_lsl);
994 case AArch64::ATOMIC_LOAD_OR_I32:
995 return emitAtomicBinary(MI, MBB, 4, AArch64::ORRwww_lsl);
996 case AArch64::ATOMIC_LOAD_OR_I64:
997 return emitAtomicBinary(MI, MBB, 8, AArch64::ORRxxx_lsl);
999 case AArch64::ATOMIC_LOAD_XOR_I8:
1000 return emitAtomicBinary(MI, MBB, 1, AArch64::EORwww_lsl);
1001 case AArch64::ATOMIC_LOAD_XOR_I16:
1002 return emitAtomicBinary(MI, MBB, 2, AArch64::EORwww_lsl);
1003 case AArch64::ATOMIC_LOAD_XOR_I32:
1004 return emitAtomicBinary(MI, MBB, 4, AArch64::EORwww_lsl);
1005 case AArch64::ATOMIC_LOAD_XOR_I64:
1006 return emitAtomicBinary(MI, MBB, 8, AArch64::EORxxx_lsl);
1008 case AArch64::ATOMIC_LOAD_NAND_I8:
1009 return emitAtomicBinary(MI, MBB, 1, AArch64::BICwww_lsl);
1010 case AArch64::ATOMIC_LOAD_NAND_I16:
1011 return emitAtomicBinary(MI, MBB, 2, AArch64::BICwww_lsl);
1012 case AArch64::ATOMIC_LOAD_NAND_I32:
1013 return emitAtomicBinary(MI, MBB, 4, AArch64::BICwww_lsl);
1014 case AArch64::ATOMIC_LOAD_NAND_I64:
1015 return emitAtomicBinary(MI, MBB, 8, AArch64::BICxxx_lsl);
1017 case AArch64::ATOMIC_LOAD_MIN_I8:
1018 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::GT);
1019 case AArch64::ATOMIC_LOAD_MIN_I16:
1020 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::GT);
1021 case AArch64::ATOMIC_LOAD_MIN_I32:
1022 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::GT);
1023 case AArch64::ATOMIC_LOAD_MIN_I64:
1024 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::GT);
1026 case AArch64::ATOMIC_LOAD_MAX_I8:
1027 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::LT);
1028 case AArch64::ATOMIC_LOAD_MAX_I16:
1029 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::LT);
1030 case AArch64::ATOMIC_LOAD_MAX_I32:
1031 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LT);
1032 case AArch64::ATOMIC_LOAD_MAX_I64:
1033 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LT);
1035 case AArch64::ATOMIC_LOAD_UMIN_I8:
1036 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::HI);
1037 case AArch64::ATOMIC_LOAD_UMIN_I16:
1038 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::HI);
1039 case AArch64::ATOMIC_LOAD_UMIN_I32:
1040 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::HI);
1041 case AArch64::ATOMIC_LOAD_UMIN_I64:
1042 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::HI);
1044 case AArch64::ATOMIC_LOAD_UMAX_I8:
1045 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::LO);
1046 case AArch64::ATOMIC_LOAD_UMAX_I16:
1047 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::LO);
1048 case AArch64::ATOMIC_LOAD_UMAX_I32:
1049 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LO);
1050 case AArch64::ATOMIC_LOAD_UMAX_I64:
1051 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LO);
1053 case AArch64::ATOMIC_SWAP_I8:
1054 return emitAtomicBinary(MI, MBB, 1, 0);
1055 case AArch64::ATOMIC_SWAP_I16:
1056 return emitAtomicBinary(MI, MBB, 2, 0);
1057 case AArch64::ATOMIC_SWAP_I32:
1058 return emitAtomicBinary(MI, MBB, 4, 0);
1059 case AArch64::ATOMIC_SWAP_I64:
1060 return emitAtomicBinary(MI, MBB, 8, 0);
1062 case AArch64::ATOMIC_CMP_SWAP_I8:
1063 return emitAtomicCmpSwap(MI, MBB, 1);
1064 case AArch64::ATOMIC_CMP_SWAP_I16:
1065 return emitAtomicCmpSwap(MI, MBB, 2);
1066 case AArch64::ATOMIC_CMP_SWAP_I32:
1067 return emitAtomicCmpSwap(MI, MBB, 4);
1068 case AArch64::ATOMIC_CMP_SWAP_I64:
1069 return emitAtomicCmpSwap(MI, MBB, 8);
1074 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
1076 case AArch64ISD::BR_CC: return "AArch64ISD::BR_CC";
1077 case AArch64ISD::Call: return "AArch64ISD::Call";
1078 case AArch64ISD::FPMOV: return "AArch64ISD::FPMOV";
1079 case AArch64ISD::GOTLoad: return "AArch64ISD::GOTLoad";
1080 case AArch64ISD::BFI: return "AArch64ISD::BFI";
1081 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
1082 case AArch64ISD::Ret: return "AArch64ISD::Ret";
1083 case AArch64ISD::SBFX: return "AArch64ISD::SBFX";
1084 case AArch64ISD::SELECT_CC: return "AArch64ISD::SELECT_CC";
1085 case AArch64ISD::SETCC: return "AArch64ISD::SETCC";
1086 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
1087 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
1088 case AArch64ISD::TLSDESCCALL: return "AArch64ISD::TLSDESCCALL";
1089 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
1090 case AArch64ISD::WrapperSmall: return "AArch64ISD::WrapperSmall";
1092 case AArch64ISD::NEON_MOVIMM:
1093 return "AArch64ISD::NEON_MOVIMM";
1094 case AArch64ISD::NEON_MVNIMM:
1095 return "AArch64ISD::NEON_MVNIMM";
1096 case AArch64ISD::NEON_FMOVIMM:
1097 return "AArch64ISD::NEON_FMOVIMM";
1098 case AArch64ISD::NEON_CMP:
1099 return "AArch64ISD::NEON_CMP";
1100 case AArch64ISD::NEON_CMPZ:
1101 return "AArch64ISD::NEON_CMPZ";
1102 case AArch64ISD::NEON_TST:
1103 return "AArch64ISD::NEON_TST";
1104 case AArch64ISD::NEON_QSHLs:
1105 return "AArch64ISD::NEON_QSHLs";
1106 case AArch64ISD::NEON_QSHLu:
1107 return "AArch64ISD::NEON_QSHLu";
1108 case AArch64ISD::NEON_VDUP:
1109 return "AArch64ISD::NEON_VDUP";
1110 case AArch64ISD::NEON_VDUPLANE:
1111 return "AArch64ISD::NEON_VDUPLANE";
1112 case AArch64ISD::NEON_REV16:
1113 return "AArch64ISD::NEON_REV16";
1114 case AArch64ISD::NEON_REV32:
1115 return "AArch64ISD::NEON_REV32";
1116 case AArch64ISD::NEON_REV64:
1117 return "AArch64ISD::NEON_REV64";
1118 case AArch64ISD::NEON_UZP1:
1119 return "AArch64ISD::NEON_UZP1";
1120 case AArch64ISD::NEON_UZP2:
1121 return "AArch64ISD::NEON_UZP2";
1122 case AArch64ISD::NEON_ZIP1:
1123 return "AArch64ISD::NEON_ZIP1";
1124 case AArch64ISD::NEON_ZIP2:
1125 return "AArch64ISD::NEON_ZIP2";
1126 case AArch64ISD::NEON_TRN1:
1127 return "AArch64ISD::NEON_TRN1";
1128 case AArch64ISD::NEON_TRN2:
1129 return "AArch64ISD::NEON_TRN2";
1130 case AArch64ISD::NEON_LD1_UPD:
1131 return "AArch64ISD::NEON_LD1_UPD";
1132 case AArch64ISD::NEON_LD2_UPD:
1133 return "AArch64ISD::NEON_LD2_UPD";
1134 case AArch64ISD::NEON_LD3_UPD:
1135 return "AArch64ISD::NEON_LD3_UPD";
1136 case AArch64ISD::NEON_LD4_UPD:
1137 return "AArch64ISD::NEON_LD4_UPD";
1138 case AArch64ISD::NEON_ST1_UPD:
1139 return "AArch64ISD::NEON_ST1_UPD";
1140 case AArch64ISD::NEON_ST2_UPD:
1141 return "AArch64ISD::NEON_ST2_UPD";
1142 case AArch64ISD::NEON_ST3_UPD:
1143 return "AArch64ISD::NEON_ST3_UPD";
1144 case AArch64ISD::NEON_ST4_UPD:
1145 return "AArch64ISD::NEON_ST4_UPD";
1146 case AArch64ISD::NEON_LD1x2_UPD:
1147 return "AArch64ISD::NEON_LD1x2_UPD";
1148 case AArch64ISD::NEON_LD1x3_UPD:
1149 return "AArch64ISD::NEON_LD1x3_UPD";
1150 case AArch64ISD::NEON_LD1x4_UPD:
1151 return "AArch64ISD::NEON_LD1x4_UPD";
1152 case AArch64ISD::NEON_ST1x2_UPD:
1153 return "AArch64ISD::NEON_ST1x2_UPD";
1154 case AArch64ISD::NEON_ST1x3_UPD:
1155 return "AArch64ISD::NEON_ST1x3_UPD";
1156 case AArch64ISD::NEON_ST1x4_UPD:
1157 return "AArch64ISD::NEON_ST1x4_UPD";
1158 case AArch64ISD::NEON_LD2DUP:
1159 return "AArch64ISD::NEON_LD2DUP";
1160 case AArch64ISD::NEON_LD3DUP:
1161 return "AArch64ISD::NEON_LD3DUP";
1162 case AArch64ISD::NEON_LD4DUP:
1163 return "AArch64ISD::NEON_LD4DUP";
1164 case AArch64ISD::NEON_LD2DUP_UPD:
1165 return "AArch64ISD::NEON_LD2DUP_UPD";
1166 case AArch64ISD::NEON_LD3DUP_UPD:
1167 return "AArch64ISD::NEON_LD3DUP_UPD";
1168 case AArch64ISD::NEON_LD4DUP_UPD:
1169 return "AArch64ISD::NEON_LD4DUP_UPD";
1170 case AArch64ISD::NEON_LD2LN_UPD:
1171 return "AArch64ISD::NEON_LD2LN_UPD";
1172 case AArch64ISD::NEON_LD3LN_UPD:
1173 return "AArch64ISD::NEON_LD3LN_UPD";
1174 case AArch64ISD::NEON_LD4LN_UPD:
1175 return "AArch64ISD::NEON_LD4LN_UPD";
1176 case AArch64ISD::NEON_ST2LN_UPD:
1177 return "AArch64ISD::NEON_ST2LN_UPD";
1178 case AArch64ISD::NEON_ST3LN_UPD:
1179 return "AArch64ISD::NEON_ST3LN_UPD";
1180 case AArch64ISD::NEON_ST4LN_UPD:
1181 return "AArch64ISD::NEON_ST4LN_UPD";
1182 case AArch64ISD::NEON_VEXTRACT:
1183 return "AArch64ISD::NEON_VEXTRACT";
1189 static const uint16_t AArch64FPRArgRegs[] = {
1190 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
1191 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7
1193 static const unsigned NumFPRArgRegs = llvm::array_lengthof(AArch64FPRArgRegs);
1195 static const uint16_t AArch64ArgRegs[] = {
1196 AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3,
1197 AArch64::X4, AArch64::X5, AArch64::X6, AArch64::X7
1199 static const unsigned NumArgRegs = llvm::array_lengthof(AArch64ArgRegs);
1201 static bool CC_AArch64NoMoreRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
1202 CCValAssign::LocInfo LocInfo,
1203 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1204 // Mark all remaining general purpose registers as allocated. We don't
1205 // backtrack: if (for example) an i128 gets put on the stack, no subsequent
1206 // i64 will go in registers (C.11).
1207 for (unsigned i = 0; i < NumArgRegs; ++i)
1208 State.AllocateReg(AArch64ArgRegs[i]);
1213 #include "AArch64GenCallingConv.inc"
1215 CCAssignFn *AArch64TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1218 default: llvm_unreachable("Unsupported calling convention");
1219 case CallingConv::Fast:
1220 case CallingConv::C:
1226 AArch64TargetLowering::SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG,
1227 SDLoc DL, SDValue &Chain) const {
1228 MachineFunction &MF = DAG.getMachineFunction();
1229 MachineFrameInfo *MFI = MF.getFrameInfo();
1230 AArch64MachineFunctionInfo *FuncInfo
1231 = MF.getInfo<AArch64MachineFunctionInfo>();
1233 SmallVector<SDValue, 8> MemOps;
1235 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(AArch64ArgRegs,
1237 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(AArch64FPRArgRegs,
1240 unsigned GPRSaveSize = 8 * (NumArgRegs - FirstVariadicGPR);
1242 if (GPRSaveSize != 0) {
1243 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1245 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1247 for (unsigned i = FirstVariadicGPR; i < NumArgRegs; ++i) {
1248 unsigned VReg = MF.addLiveIn(AArch64ArgRegs[i], &AArch64::GPR64RegClass);
1249 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1250 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1251 MachinePointerInfo::getStack(i * 8),
1253 MemOps.push_back(Store);
1254 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1255 DAG.getConstant(8, getPointerTy()));
1259 if (getSubtarget()->hasFPARMv8()) {
1260 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1262 // According to the AArch64 Procedure Call Standard, section B.1/B.3, we
1263 // can omit a register save area if we know we'll never use registers of
1265 if (FPRSaveSize != 0) {
1266 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1268 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1270 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1271 unsigned VReg = MF.addLiveIn(AArch64FPRArgRegs[i],
1272 &AArch64::FPR128RegClass);
1273 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1274 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1275 MachinePointerInfo::getStack(i * 16),
1277 MemOps.push_back(Store);
1278 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1279 DAG.getConstant(16, getPointerTy()));
1282 FuncInfo->setVariadicFPRIdx(FPRIdx);
1283 FuncInfo->setVariadicFPRSize(FPRSaveSize);
1286 unsigned StackOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), 8);
1287 int StackIdx = MFI->CreateFixedObject(8, StackOffset, true);
1289 FuncInfo->setVariadicStackIdx(StackIdx);
1290 FuncInfo->setVariadicGPRIdx(GPRIdx);
1291 FuncInfo->setVariadicGPRSize(GPRSaveSize);
1293 if (!MemOps.empty()) {
1294 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
1301 AArch64TargetLowering::LowerFormalArguments(SDValue Chain,
1302 CallingConv::ID CallConv, bool isVarArg,
1303 const SmallVectorImpl<ISD::InputArg> &Ins,
1304 SDLoc dl, SelectionDAG &DAG,
1305 SmallVectorImpl<SDValue> &InVals) const {
1306 MachineFunction &MF = DAG.getMachineFunction();
1307 AArch64MachineFunctionInfo *FuncInfo
1308 = MF.getInfo<AArch64MachineFunctionInfo>();
1309 MachineFrameInfo *MFI = MF.getFrameInfo();
1310 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1312 SmallVector<CCValAssign, 16> ArgLocs;
1313 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1314 getTargetMachine(), ArgLocs, *DAG.getContext());
1315 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1317 SmallVector<SDValue, 16> ArgValues;
1320 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1321 CCValAssign &VA = ArgLocs[i];
1322 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1324 if (Flags.isByVal()) {
1325 // Byval is used for small structs and HFAs in the PCS, but the system
1326 // should work in a non-compliant manner for larger structs.
1327 EVT PtrTy = getPointerTy();
1328 int Size = Flags.getByValSize();
1329 unsigned NumRegs = (Size + 7) / 8;
1331 unsigned FrameIdx = MFI->CreateFixedObject(8 * NumRegs,
1332 VA.getLocMemOffset(),
1334 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1335 InVals.push_back(FrameIdxN);
1338 } else if (VA.isRegLoc()) {
1339 MVT RegVT = VA.getLocVT();
1340 const TargetRegisterClass *RC = getRegClassFor(RegVT);
1341 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1343 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1344 } else { // VA.isRegLoc()
1345 assert(VA.isMemLoc());
1347 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
1348 VA.getLocMemOffset(), true);
1350 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1351 ArgValue = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1352 MachinePointerInfo::getFixedStack(FI),
1353 false, false, false, 0);
1358 switch (VA.getLocInfo()) {
1359 default: llvm_unreachable("Unknown loc info!");
1360 case CCValAssign::Full: break;
1361 case CCValAssign::BCvt:
1362 ArgValue = DAG.getNode(ISD::BITCAST,dl, VA.getValVT(), ArgValue);
1364 case CCValAssign::SExt:
1365 case CCValAssign::ZExt:
1366 case CCValAssign::AExt:
1367 case CCValAssign::FPExt: {
1368 unsigned DestSize = VA.getValVT().getSizeInBits();
1369 unsigned DestSubReg;
1372 case 8: DestSubReg = AArch64::sub_8; break;
1373 case 16: DestSubReg = AArch64::sub_16; break;
1374 case 32: DestSubReg = AArch64::sub_32; break;
1375 case 64: DestSubReg = AArch64::sub_64; break;
1376 default: llvm_unreachable("Unexpected argument promotion");
1379 ArgValue = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
1380 VA.getValVT(), ArgValue,
1381 DAG.getTargetConstant(DestSubReg, MVT::i32)),
1387 InVals.push_back(ArgValue);
1391 SaveVarArgRegisters(CCInfo, DAG, dl, Chain);
1393 unsigned StackArgSize = CCInfo.getNextStackOffset();
1394 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
1395 // This is a non-standard ABI so by fiat I say we're allowed to make full
1396 // use of the stack area to be popped, which must be aligned to 16 bytes in
1398 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
1400 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
1401 // a multiple of 16.
1402 FuncInfo->setArgumentStackToRestore(StackArgSize);
1404 // This realignment carries over to the available bytes below. Our own
1405 // callers will guarantee the space is free by giving an aligned value to
1408 // Even if we're not expected to free up the space, it's useful to know how
1409 // much is there while considering tail calls (because we can reuse it).
1410 FuncInfo->setBytesInStackArgArea(StackArgSize);
1416 AArch64TargetLowering::LowerReturn(SDValue Chain,
1417 CallingConv::ID CallConv, bool isVarArg,
1418 const SmallVectorImpl<ISD::OutputArg> &Outs,
1419 const SmallVectorImpl<SDValue> &OutVals,
1420 SDLoc dl, SelectionDAG &DAG) const {
1421 // CCValAssign - represent the assignment of the return value to a location.
1422 SmallVector<CCValAssign, 16> RVLocs;
1424 // CCState - Info about the registers and stack slots.
1425 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1426 getTargetMachine(), RVLocs, *DAG.getContext());
1428 // Analyze outgoing return values.
1429 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv));
1432 SmallVector<SDValue, 4> RetOps(1, Chain);
1434 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1435 // PCS: "If the type, T, of the result of a function is such that
1436 // void func(T arg) would require that arg be passed as a value in a
1437 // register (or set of registers) according to the rules in 5.4, then the
1438 // result is returned in the same registers as would be used for such an
1441 // Otherwise, the caller shall reserve a block of memory of sufficient
1442 // size and alignment to hold the result. The address of the memory block
1443 // shall be passed as an additional argument to the function in x8."
1445 // This is implemented in two places. The register-return values are dealt
1446 // with here, more complex returns are passed as an sret parameter, which
1447 // means we don't have to worry about it during actual return.
1448 CCValAssign &VA = RVLocs[i];
1449 assert(VA.isRegLoc() && "Only register-returns should be created by PCS");
1452 SDValue Arg = OutVals[i];
1454 // There's no convenient note in the ABI about this as there is for normal
1455 // arguments, but it says return values are passed in the same registers as
1456 // an argument would be. I believe that includes the comments about
1457 // unspecified higher bits, putting the burden of widening on the *caller*
1458 // for return values.
1459 switch (VA.getLocInfo()) {
1460 default: llvm_unreachable("Unknown loc info");
1461 case CCValAssign::Full: break;
1462 case CCValAssign::SExt:
1463 case CCValAssign::ZExt:
1464 case CCValAssign::AExt:
1465 // Floating-point values should only be extended when they're going into
1466 // memory, which can't happen here so an integer extend is acceptable.
1467 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1469 case CCValAssign::BCvt:
1470 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1474 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1475 Flag = Chain.getValue(1);
1476 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1479 RetOps[0] = Chain; // Update chain.
1481 // Add the flag if we have it.
1483 RetOps.push_back(Flag);
1485 return DAG.getNode(AArch64ISD::Ret, dl, MVT::Other,
1486 &RetOps[0], RetOps.size());
1489 unsigned AArch64TargetLowering::getByValTypeAlignment(Type *Ty) const {
1490 // This is a new backend. For anything more precise than this a FE should
1491 // set an explicit alignment.
1496 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1497 SmallVectorImpl<SDValue> &InVals) const {
1498 SelectionDAG &DAG = CLI.DAG;
1500 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1501 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1502 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1503 SDValue Chain = CLI.Chain;
1504 SDValue Callee = CLI.Callee;
1505 bool &IsTailCall = CLI.IsTailCall;
1506 CallingConv::ID CallConv = CLI.CallConv;
1507 bool IsVarArg = CLI.IsVarArg;
1509 MachineFunction &MF = DAG.getMachineFunction();
1510 AArch64MachineFunctionInfo *FuncInfo
1511 = MF.getInfo<AArch64MachineFunctionInfo>();
1512 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1513 bool IsStructRet = !Outs.empty() && Outs[0].Flags.isSRet();
1514 bool IsSibCall = false;
1517 IsTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1518 IsVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1519 Outs, OutVals, Ins, DAG);
1521 // A sibling call is one where we're under the usual C ABI and not planning
1522 // to change that but can still do a tail call:
1523 if (!TailCallOpt && IsTailCall)
1527 SmallVector<CCValAssign, 16> ArgLocs;
1528 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1529 getTargetMachine(), ArgLocs, *DAG.getContext());
1530 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1532 // On AArch64 (and all other architectures I'm aware of) the most this has to
1533 // do is adjust the stack pointer.
1534 unsigned NumBytes = RoundUpToAlignment(CCInfo.getNextStackOffset(), 16);
1536 // Since we're not changing the ABI to make this a tail call, the memory
1537 // operands are already available in the caller's incoming argument space.
1541 // FPDiff is the byte offset of the call's argument area from the callee's.
1542 // Stores to callee stack arguments will be placed in FixedStackSlots offset
1543 // by this amount for a tail call. In a sibling call it must be 0 because the
1544 // caller will deallocate the entire stack and the callee still expects its
1545 // arguments to begin at SP+0. Completely unused for non-tail calls.
1548 if (IsTailCall && !IsSibCall) {
1549 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
1551 // FPDiff will be negative if this tail call requires more space than we
1552 // would automatically have in our incoming argument space. Positive if we
1553 // can actually shrink the stack.
1554 FPDiff = NumReusableBytes - NumBytes;
1556 // The stack pointer must be 16-byte aligned at all times it's used for a
1557 // memory operation, which in practice means at *all* times and in
1558 // particular across call boundaries. Therefore our own arguments started at
1559 // a 16-byte aligned SP and the delta applied for the tail call should
1560 // satisfy the same constraint.
1561 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
1565 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1568 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, AArch64::XSP,
1571 SmallVector<SDValue, 8> MemOpChains;
1572 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1575 CCValAssign &VA = ArgLocs[i];
1576 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1577 SDValue Arg = OutVals[i];
1579 // Callee does the actual widening, so all extensions just use an implicit
1580 // definition of the rest of the Loc. Aesthetically, this would be nicer as
1581 // an ANY_EXTEND, but that isn't valid for floating-point types and this
1582 // alternative works on integer types too.
1583 switch (VA.getLocInfo()) {
1584 default: llvm_unreachable("Unknown loc info!");
1585 case CCValAssign::Full: break;
1586 case CCValAssign::SExt:
1587 case CCValAssign::ZExt:
1588 case CCValAssign::AExt:
1589 case CCValAssign::FPExt: {
1590 unsigned SrcSize = VA.getValVT().getSizeInBits();
1594 case 8: SrcSubReg = AArch64::sub_8; break;
1595 case 16: SrcSubReg = AArch64::sub_16; break;
1596 case 32: SrcSubReg = AArch64::sub_32; break;
1597 case 64: SrcSubReg = AArch64::sub_64; break;
1598 default: llvm_unreachable("Unexpected argument promotion");
1601 Arg = SDValue(DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1603 DAG.getUNDEF(VA.getLocVT()),
1605 DAG.getTargetConstant(SrcSubReg, MVT::i32)),
1610 case CCValAssign::BCvt:
1611 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1615 if (VA.isRegLoc()) {
1616 // A normal register (sub-) argument. For now we just note it down because
1617 // we want to copy things into registers as late as possible to avoid
1618 // register-pressure (and possibly worse).
1619 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1623 assert(VA.isMemLoc() && "unexpected argument location");
1626 MachinePointerInfo DstInfo;
1628 uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize() :
1629 VA.getLocVT().getSizeInBits();
1630 OpSize = (OpSize + 7) / 8;
1631 int32_t Offset = VA.getLocMemOffset() + FPDiff;
1632 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
1634 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
1635 DstInfo = MachinePointerInfo::getFixedStack(FI);
1637 // Make sure any stack arguments overlapping with where we're storing are
1638 // loaded before this eventual operation. Otherwise they'll be clobbered.
1639 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
1641 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
1643 DstAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1644 DstInfo = MachinePointerInfo::getStack(VA.getLocMemOffset());
1647 if (Flags.isByVal()) {
1648 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i64);
1649 SDValue Cpy = DAG.getMemcpy(Chain, dl, DstAddr, Arg, SizeNode,
1650 Flags.getByValAlign(),
1651 /*isVolatile = */ false,
1652 /*alwaysInline = */ false,
1653 DstInfo, MachinePointerInfo(0));
1654 MemOpChains.push_back(Cpy);
1656 // Normal stack argument, put it where it's needed.
1657 SDValue Store = DAG.getStore(Chain, dl, Arg, DstAddr, DstInfo,
1659 MemOpChains.push_back(Store);
1663 // The loads and stores generated above shouldn't clash with each
1664 // other. Combining them with this TokenFactor notes that fact for the rest of
1666 if (!MemOpChains.empty())
1667 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1668 &MemOpChains[0], MemOpChains.size());
1670 // Most of the rest of the instructions need to be glued together; we don't
1671 // want assignments to actual registers used by a call to be rearranged by a
1672 // well-meaning scheduler.
1675 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1676 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1677 RegsToPass[i].second, InFlag);
1678 InFlag = Chain.getValue(1);
1681 // The linker is responsible for inserting veneers when necessary to put a
1682 // function call destination in range, so we don't need to bother with a
1684 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1685 const GlobalValue *GV = G->getGlobal();
1686 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1687 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1688 const char *Sym = S->getSymbol();
1689 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1692 // We don't usually want to end the call-sequence here because we would tidy
1693 // the frame up *after* the call, however in the ABI-changing tail-call case
1694 // we've carefully laid out the parameters so that when sp is reset they'll be
1695 // in the correct location.
1696 if (IsTailCall && !IsSibCall) {
1697 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1698 DAG.getIntPtrConstant(0, true), InFlag, dl);
1699 InFlag = Chain.getValue(1);
1702 // We produce the following DAG scheme for the actual call instruction:
1703 // (AArch64Call Chain, Callee, reg1, ..., regn, preserveMask, inflag?
1705 // Most arguments aren't going to be used and just keep the values live as
1706 // far as LLVM is concerned. It's expected to be selected as simply "bl
1707 // callee" (for a direct, non-tail call).
1708 std::vector<SDValue> Ops;
1709 Ops.push_back(Chain);
1710 Ops.push_back(Callee);
1713 // Each tail call may have to adjust the stack by a different amount, so
1714 // this information must travel along with the operation for eventual
1715 // consumption by emitEpilogue.
1716 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
1719 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1720 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1721 RegsToPass[i].second.getValueType()));
1724 // Add a register mask operand representing the call-preserved registers. This
1725 // is used later in codegen to constrain register-allocation.
1726 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1727 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1728 assert(Mask && "Missing call preserved mask for calling convention");
1729 Ops.push_back(DAG.getRegisterMask(Mask));
1731 // If we needed glue, put it in as the last argument.
1732 if (InFlag.getNode())
1733 Ops.push_back(InFlag);
1735 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1738 return DAG.getNode(AArch64ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1741 Chain = DAG.getNode(AArch64ISD::Call, dl, NodeTys, &Ops[0], Ops.size());
1742 InFlag = Chain.getValue(1);
1744 // Now we can reclaim the stack, just as well do it before working out where
1745 // our return value is.
1747 uint64_t CalleePopBytes
1748 = DoesCalleeRestoreStack(CallConv, TailCallOpt) ? NumBytes : 0;
1750 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1751 DAG.getIntPtrConstant(CalleePopBytes, true),
1753 InFlag = Chain.getValue(1);
1756 return LowerCallResult(Chain, InFlag, CallConv,
1757 IsVarArg, Ins, dl, DAG, InVals);
1761 AArch64TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1762 CallingConv::ID CallConv, bool IsVarArg,
1763 const SmallVectorImpl<ISD::InputArg> &Ins,
1764 SDLoc dl, SelectionDAG &DAG,
1765 SmallVectorImpl<SDValue> &InVals) const {
1766 // Assign locations to each value returned by this call.
1767 SmallVector<CCValAssign, 16> RVLocs;
1768 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1769 getTargetMachine(), RVLocs, *DAG.getContext());
1770 CCInfo.AnalyzeCallResult(Ins, CCAssignFnForNode(CallConv));
1772 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1773 CCValAssign VA = RVLocs[i];
1775 // Return values that are too big to fit into registers should use an sret
1776 // pointer, so this can be a lot simpler than the main argument code.
1777 assert(VA.isRegLoc() && "Memory locations not expected for call return");
1779 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1781 Chain = Val.getValue(1);
1782 InFlag = Val.getValue(2);
1784 switch (VA.getLocInfo()) {
1785 default: llvm_unreachable("Unknown loc info!");
1786 case CCValAssign::Full: break;
1787 case CCValAssign::BCvt:
1788 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1790 case CCValAssign::ZExt:
1791 case CCValAssign::SExt:
1792 case CCValAssign::AExt:
1793 // Floating-point arguments only get extended/truncated if they're going
1794 // in memory, so using the integer operation is acceptable here.
1795 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
1799 InVals.push_back(Val);
1806 AArch64TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1807 CallingConv::ID CalleeCC,
1809 bool IsCalleeStructRet,
1810 bool IsCallerStructRet,
1811 const SmallVectorImpl<ISD::OutputArg> &Outs,
1812 const SmallVectorImpl<SDValue> &OutVals,
1813 const SmallVectorImpl<ISD::InputArg> &Ins,
1814 SelectionDAG& DAG) const {
1816 // For CallingConv::C this function knows whether the ABI needs
1817 // changing. That's not true for other conventions so they will have to opt in
1819 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1822 const MachineFunction &MF = DAG.getMachineFunction();
1823 const Function *CallerF = MF.getFunction();
1824 CallingConv::ID CallerCC = CallerF->getCallingConv();
1825 bool CCMatch = CallerCC == CalleeCC;
1827 // Byval parameters hand the function a pointer directly into the stack area
1828 // we want to reuse during a tail call. Working around this *is* possible (see
1829 // X86) but less efficient and uglier in LowerCall.
1830 for (Function::const_arg_iterator i = CallerF->arg_begin(),
1831 e = CallerF->arg_end(); i != e; ++i)
1832 if (i->hasByValAttr())
1835 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
1836 if (IsTailCallConvention(CalleeCC) && CCMatch)
1841 // Now we search for cases where we can use a tail call without changing the
1842 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
1845 // I want anyone implementing a new calling convention to think long and hard
1846 // about this assert.
1847 assert((!IsVarArg || CalleeCC == CallingConv::C)
1848 && "Unexpected variadic calling convention");
1850 if (IsVarArg && !Outs.empty()) {
1851 // At least two cases here: if caller is fastcc then we can't have any
1852 // memory arguments (we'd be expected to clean up the stack afterwards). If
1853 // caller is C then we could potentially use its argument area.
1855 // FIXME: for now we take the most conservative of these in both cases:
1856 // disallow all variadic memory operands.
1857 SmallVector<CCValAssign, 16> ArgLocs;
1858 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1859 getTargetMachine(), ArgLocs, *DAG.getContext());
1861 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1862 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
1863 if (!ArgLocs[i].isRegLoc())
1867 // If the calling conventions do not match, then we'd better make sure the
1868 // results are returned in the same way as what the caller expects.
1870 SmallVector<CCValAssign, 16> RVLocs1;
1871 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1872 getTargetMachine(), RVLocs1, *DAG.getContext());
1873 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC));
1875 SmallVector<CCValAssign, 16> RVLocs2;
1876 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1877 getTargetMachine(), RVLocs2, *DAG.getContext());
1878 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC));
1880 if (RVLocs1.size() != RVLocs2.size())
1882 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1883 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1885 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1887 if (RVLocs1[i].isRegLoc()) {
1888 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1891 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1897 // Nothing more to check if the callee is taking no arguments
1901 SmallVector<CCValAssign, 16> ArgLocs;
1902 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1903 getTargetMachine(), ArgLocs, *DAG.getContext());
1905 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1907 const AArch64MachineFunctionInfo *FuncInfo
1908 = MF.getInfo<AArch64MachineFunctionInfo>();
1910 // If the stack arguments for this call would fit into our own save area then
1911 // the call can be made tail.
1912 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
1915 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
1916 bool TailCallOpt) const {
1917 return CallCC == CallingConv::Fast && TailCallOpt;
1920 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
1921 return CallCC == CallingConv::Fast;
1924 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
1926 MachineFrameInfo *MFI,
1927 int ClobberedFI) const {
1928 SmallVector<SDValue, 8> ArgChains;
1929 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
1930 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
1932 // Include the original chain at the beginning of the list. When this is
1933 // used by target LowerCall hooks, this helps legalize find the
1934 // CALLSEQ_BEGIN node.
1935 ArgChains.push_back(Chain);
1937 // Add a chain value for each stack argument corresponding
1938 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
1939 UE = DAG.getEntryNode().getNode()->use_end(); U != UE; ++U)
1940 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
1941 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
1942 if (FI->getIndex() < 0) {
1943 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
1944 int64_t InLastByte = InFirstByte;
1945 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
1947 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1948 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1949 ArgChains.push_back(SDValue(L, 1));
1952 // Build a tokenfactor for all the chains.
1953 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other,
1954 &ArgChains[0], ArgChains.size());
1957 static A64CC::CondCodes IntCCToA64CC(ISD::CondCode CC) {
1959 case ISD::SETEQ: return A64CC::EQ;
1960 case ISD::SETGT: return A64CC::GT;
1961 case ISD::SETGE: return A64CC::GE;
1962 case ISD::SETLT: return A64CC::LT;
1963 case ISD::SETLE: return A64CC::LE;
1964 case ISD::SETNE: return A64CC::NE;
1965 case ISD::SETUGT: return A64CC::HI;
1966 case ISD::SETUGE: return A64CC::HS;
1967 case ISD::SETULT: return A64CC::LO;
1968 case ISD::SETULE: return A64CC::LS;
1969 default: llvm_unreachable("Unexpected condition code");
1973 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Val) const {
1974 // icmp is implemented using adds/subs immediate, which take an unsigned
1975 // 12-bit immediate, optionally shifted left by 12 bits.
1977 // Symmetric by using adds/subs
1981 return (Val & ~0xfff) == 0 || (Val & ~0xfff000) == 0;
1984 SDValue AArch64TargetLowering::getSelectableIntSetCC(SDValue LHS, SDValue RHS,
1985 ISD::CondCode CC, SDValue &A64cc,
1986 SelectionDAG &DAG, SDLoc &dl) const {
1987 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1989 EVT VT = RHSC->getValueType(0);
1990 bool knownInvalid = false;
1992 // I'm not convinced the rest of LLVM handles these edge cases properly, but
1993 // we can at least get it right.
1994 if (isSignedIntSetCC(CC)) {
1995 C = RHSC->getSExtValue();
1996 } else if (RHSC->getZExtValue() > INT64_MAX) {
1997 // A 64-bit constant not representable by a signed 64-bit integer is far
1998 // too big to fit into a SUBS immediate anyway.
1999 knownInvalid = true;
2001 C = RHSC->getZExtValue();
2004 if (!knownInvalid && !isLegalICmpImmediate(C)) {
2005 // Constant does not fit, try adjusting it by one?
2010 if (isLegalICmpImmediate(C-1)) {
2011 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2012 RHS = DAG.getConstant(C-1, VT);
2017 if (isLegalICmpImmediate(C-1)) {
2018 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2019 RHS = DAG.getConstant(C-1, VT);
2024 if (isLegalICmpImmediate(C+1)) {
2025 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2026 RHS = DAG.getConstant(C+1, VT);
2031 if (isLegalICmpImmediate(C+1)) {
2032 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2033 RHS = DAG.getConstant(C+1, VT);
2040 A64CC::CondCodes CondCode = IntCCToA64CC(CC);
2041 A64cc = DAG.getConstant(CondCode, MVT::i32);
2042 return DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2043 DAG.getCondCode(CC));
2046 static A64CC::CondCodes FPCCToA64CC(ISD::CondCode CC,
2047 A64CC::CondCodes &Alternative) {
2048 A64CC::CondCodes CondCode = A64CC::Invalid;
2049 Alternative = A64CC::Invalid;
2052 default: llvm_unreachable("Unknown FP condition!");
2054 case ISD::SETOEQ: CondCode = A64CC::EQ; break;
2056 case ISD::SETOGT: CondCode = A64CC::GT; break;
2058 case ISD::SETOGE: CondCode = A64CC::GE; break;
2059 case ISD::SETOLT: CondCode = A64CC::MI; break;
2060 case ISD::SETOLE: CondCode = A64CC::LS; break;
2061 case ISD::SETONE: CondCode = A64CC::MI; Alternative = A64CC::GT; break;
2062 case ISD::SETO: CondCode = A64CC::VC; break;
2063 case ISD::SETUO: CondCode = A64CC::VS; break;
2064 case ISD::SETUEQ: CondCode = A64CC::EQ; Alternative = A64CC::VS; break;
2065 case ISD::SETUGT: CondCode = A64CC::HI; break;
2066 case ISD::SETUGE: CondCode = A64CC::PL; break;
2068 case ISD::SETULT: CondCode = A64CC::LT; break;
2070 case ISD::SETULE: CondCode = A64CC::LE; break;
2072 case ISD::SETUNE: CondCode = A64CC::NE; break;
2078 AArch64TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
2080 EVT PtrVT = getPointerTy();
2081 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2083 switch(getTargetMachine().getCodeModel()) {
2084 case CodeModel::Small:
2085 // The most efficient code is PC-relative anyway for the small memory model,
2086 // so we don't need to worry about relocation model.
2087 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2088 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2089 AArch64II::MO_NO_FLAG),
2090 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2091 AArch64II::MO_LO12),
2092 DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
2093 case CodeModel::Large:
2095 AArch64ISD::WrapperLarge, DL, PtrVT,
2096 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G3),
2097 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2098 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2099 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2101 llvm_unreachable("Only small and large code models supported now");
2106 // (BRCOND chain, val, dest)
2108 AArch64TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
2110 SDValue Chain = Op.getOperand(0);
2111 SDValue TheBit = Op.getOperand(1);
2112 SDValue DestBB = Op.getOperand(2);
2114 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2115 // that as the consumer we are responsible for ignoring rubbish in higher
2117 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2118 DAG.getConstant(1, MVT::i32));
2120 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2121 DAG.getConstant(0, TheBit.getValueType()),
2122 DAG.getCondCode(ISD::SETNE));
2124 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other, Chain,
2125 A64CMP, DAG.getConstant(A64CC::NE, MVT::i32),
2129 // (BR_CC chain, condcode, lhs, rhs, dest)
2131 AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2133 SDValue Chain = Op.getOperand(0);
2134 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2135 SDValue LHS = Op.getOperand(2);
2136 SDValue RHS = Op.getOperand(3);
2137 SDValue DestBB = Op.getOperand(4);
2139 if (LHS.getValueType() == MVT::f128) {
2140 // f128 comparisons are lowered to runtime calls by a routine which sets
2141 // LHS, RHS and CC appropriately for the rest of this function to continue.
2142 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2144 // If softenSetCCOperands returned a scalar, we need to compare the result
2145 // against zero to select between true and false values.
2146 if (RHS.getNode() == 0) {
2147 RHS = DAG.getConstant(0, LHS.getValueType());
2152 if (LHS.getValueType().isInteger()) {
2155 // Integers are handled in a separate function because the combinations of
2156 // immediates and tests can get hairy and we may want to fiddle things.
2157 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2159 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2160 Chain, CmpOp, A64cc, DestBB);
2163 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2164 // conditional branch, hence FPCCToA64CC can set a second test, where either
2165 // passing is sufficient.
2166 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2167 CondCode = FPCCToA64CC(CC, Alternative);
2168 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2169 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2170 DAG.getCondCode(CC));
2171 SDValue A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2172 Chain, SetCC, A64cc, DestBB);
2174 if (Alternative != A64CC::Invalid) {
2175 A64cc = DAG.getConstant(Alternative, MVT::i32);
2176 A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2177 A64BR_CC, SetCC, A64cc, DestBB);
2185 AArch64TargetLowering::LowerF128ToCall(SDValue Op, SelectionDAG &DAG,
2186 RTLIB::Libcall Call) const {
2189 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
2190 EVT ArgVT = Op.getOperand(i).getValueType();
2191 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2192 Entry.Node = Op.getOperand(i); Entry.Ty = ArgTy;
2193 Entry.isSExt = false;
2194 Entry.isZExt = false;
2195 Args.push_back(Entry);
2197 SDValue Callee = DAG.getExternalSymbol(getLibcallName(Call), getPointerTy());
2199 Type *RetTy = Op.getValueType().getTypeForEVT(*DAG.getContext());
2201 // By default, the input chain to this libcall is the entry node of the
2202 // function. If the libcall is going to be emitted as a tail call then
2203 // isUsedByReturnOnly will change it to the right chain if the return
2204 // node which is being folded has a non-entry input chain.
2205 SDValue InChain = DAG.getEntryNode();
2207 // isTailCall may be true since the callee does not reference caller stack
2208 // frame. Check if it's in the right position.
2209 SDValue TCChain = InChain;
2210 bool isTailCall = isInTailCallPosition(DAG, Op.getNode(), TCChain);
2215 CallLoweringInfo CLI(InChain, RetTy, false, false, false, false,
2216 0, getLibcallCallingConv(Call), isTailCall,
2217 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2218 Callee, Args, DAG, SDLoc(Op));
2219 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
2221 if (!CallInfo.second.getNode())
2222 // It's a tailcall, return the chain (which is the DAG root).
2223 return DAG.getRoot();
2225 return CallInfo.first;
2229 AArch64TargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
2230 if (Op.getOperand(0).getValueType() != MVT::f128) {
2231 // It's legal except when f128 is involved
2236 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
2238 SDValue SrcVal = Op.getOperand(0);
2239 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
2240 /*isSigned*/ false, SDLoc(Op)).first;
2244 AArch64TargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
2245 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
2248 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
2250 return LowerF128ToCall(Op, DAG, LC);
2253 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2256 EVT VT = Op.getValueType();
2257 SDValue Vec = Op.getOperand(0);
2258 EVT OpVT = Vec.getValueType();
2259 unsigned Opc = IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
2261 if (VT.getVectorNumElements() == 1) {
2262 assert(OpVT == MVT::v1f64 && "Unexpected vector type!");
2263 if (VT.getSizeInBits() == OpVT.getSizeInBits())
2265 return DAG.UnrollVectorOp(Op.getNode());
2268 if (VT.getSizeInBits() > OpVT.getSizeInBits()) {
2269 assert(Vec.getValueType() == MVT::v2f32 && VT == MVT::v2i64 &&
2270 "Unexpected vector type!");
2271 Vec = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Vec);
2272 return DAG.getNode(Opc, dl, VT, Vec);
2273 } else if (VT.getSizeInBits() < OpVT.getSizeInBits()) {
2274 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2275 OpVT.getVectorElementType().getSizeInBits());
2277 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2278 Vec = DAG.getNode(Opc, dl, CastVT, Vec);
2279 return DAG.getNode(ISD::TRUNCATE, dl, VT, Vec);
2281 return DAG.getNode(Opc, dl, VT, Vec);
2284 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2285 // We custom lower concat_vectors with 4, 8, or 16 operands that are all the
2286 // same operand and of type v1* using the DUP instruction.
2287 unsigned NumOps = Op->getNumOperands();
2288 if (NumOps != 4 && NumOps != 8 && NumOps != 16)
2291 // Must be a single value for VDUP.
2292 bool isConstant = true;
2293 SDValue Op0 = Op.getOperand(0);
2294 for (unsigned i = 1; i < NumOps; ++i) {
2295 SDValue OpN = Op.getOperand(i);
2299 if (!isa<ConstantSDNode>(OpN->getOperand(0)))
2303 // Verify the value type.
2304 EVT EltVT = Op0.getValueType();
2306 default: llvm_unreachable("Unexpected number of operands");
2308 if (EltVT != MVT::v1i16 && EltVT != MVT::v1i32)
2312 if (EltVT != MVT::v1i8 && EltVT != MVT::v1i16)
2316 if (EltVT != MVT::v1i8)
2322 EVT VT = Op.getValueType();
2323 // VDUP produces better code for constants.
2325 return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Op0->getOperand(0));
2326 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, Op0,
2327 DAG.getConstant(0, MVT::i64));
2331 AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2332 bool IsSigned) const {
2333 if (Op.getValueType().isVector())
2334 return LowerVectorFP_TO_INT(Op, DAG, IsSigned);
2335 if (Op.getOperand(0).getValueType() != MVT::f128) {
2336 // It's legal except when f128 is involved
2342 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
2344 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
2346 return LowerF128ToCall(Op, DAG, LC);
2349 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2350 MachineFunction &MF = DAG.getMachineFunction();
2351 MachineFrameInfo *MFI = MF.getFrameInfo();
2352 MFI->setReturnAddressIsTaken(true);
2354 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2357 EVT VT = Op.getValueType();
2359 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2361 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2362 SDValue Offset = DAG.getConstant(8, MVT::i64);
2363 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2364 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2365 MachinePointerInfo(), false, false, false, 0);
2368 // Return X30, which contains the return address. Mark it an implicit live-in.
2369 unsigned Reg = MF.addLiveIn(AArch64::X30, getRegClassFor(MVT::i64));
2370 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, MVT::i64);
2374 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG)
2376 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2377 MFI->setFrameAddressIsTaken(true);
2379 EVT VT = Op.getValueType();
2381 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2382 unsigned FrameReg = AArch64::X29;
2383 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2385 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2386 MachinePointerInfo(),
2387 false, false, false, 0);
2392 AArch64TargetLowering::LowerGlobalAddressELFLarge(SDValue Op,
2393 SelectionDAG &DAG) const {
2394 assert(getTargetMachine().getCodeModel() == CodeModel::Large);
2395 assert(getTargetMachine().getRelocationModel() == Reloc::Static);
2397 EVT PtrVT = getPointerTy();
2399 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2400 const GlobalValue *GV = GN->getGlobal();
2402 SDValue GlobalAddr = DAG.getNode(
2403 AArch64ISD::WrapperLarge, dl, PtrVT,
2404 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G3),
2405 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2406 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2407 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2409 if (GN->getOffset() != 0)
2410 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2411 DAG.getConstant(GN->getOffset(), PtrVT));
2417 AArch64TargetLowering::LowerGlobalAddressELFSmall(SDValue Op,
2418 SelectionDAG &DAG) const {
2419 assert(getTargetMachine().getCodeModel() == CodeModel::Small);
2421 EVT PtrVT = getPointerTy();
2423 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2424 const GlobalValue *GV = GN->getGlobal();
2425 unsigned Alignment = GV->getAlignment();
2426 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2427 if (GV->isWeakForLinker() && GV->isDeclaration() && RelocM == Reloc::Static) {
2428 // Weak undefined symbols can't use ADRP/ADD pair since they should evaluate
2429 // to zero when they remain undefined. In PIC mode the GOT can take care of
2430 // this, but in absolute mode we use a constant pool load.
2432 PoolAddr = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2433 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2434 AArch64II::MO_NO_FLAG),
2435 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2436 AArch64II::MO_LO12),
2437 DAG.getConstant(8, MVT::i32));
2438 SDValue GlobalAddr = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), PoolAddr,
2439 MachinePointerInfo::getConstantPool(),
2440 /*isVolatile=*/ false,
2441 /*isNonTemporal=*/ true,
2442 /*isInvariant=*/ true, 8);
2443 if (GN->getOffset() != 0)
2444 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2445 DAG.getConstant(GN->getOffset(), PtrVT));
2450 if (Alignment == 0) {
2451 const PointerType *GVPtrTy = cast<PointerType>(GV->getType());
2452 if (GVPtrTy->getElementType()->isSized()) {
2454 = getDataLayout()->getABITypeAlignment(GVPtrTy->getElementType());
2456 // Be conservative if we can't guess, not that it really matters:
2457 // functions and labels aren't valid for loads, and the methods used to
2458 // actually calculate an address work with any alignment.
2463 unsigned char HiFixup, LoFixup;
2464 bool UseGOT = getSubtarget()->GVIsIndirectSymbol(GV, RelocM);
2467 HiFixup = AArch64II::MO_GOT;
2468 LoFixup = AArch64II::MO_GOT_LO12;
2471 HiFixup = AArch64II::MO_NO_FLAG;
2472 LoFixup = AArch64II::MO_LO12;
2475 // AArch64's small model demands the following sequence:
2476 // ADRP x0, somewhere
2477 // ADD x0, x0, #:lo12:somewhere ; (or LDR directly).
2478 SDValue GlobalRef = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2479 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2481 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2483 DAG.getConstant(Alignment, MVT::i32));
2486 GlobalRef = DAG.getNode(AArch64ISD::GOTLoad, dl, PtrVT, DAG.getEntryNode(),
2490 if (GN->getOffset() != 0)
2491 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalRef,
2492 DAG.getConstant(GN->getOffset(), PtrVT));
2498 AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
2499 SelectionDAG &DAG) const {
2500 // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
2501 // we make those distinctions here.
2503 switch (getTargetMachine().getCodeModel()) {
2504 case CodeModel::Small:
2505 return LowerGlobalAddressELFSmall(Op, DAG);
2506 case CodeModel::Large:
2507 return LowerGlobalAddressELFLarge(Op, DAG);
2509 llvm_unreachable("Only small and large code models supported now");
2514 AArch64TargetLowering::LowerConstantPool(SDValue Op,
2515 SelectionDAG &DAG) const {
2517 EVT PtrVT = getPointerTy();
2518 ConstantPoolSDNode *CN = cast<ConstantPoolSDNode>(Op);
2519 const Constant *C = CN->getConstVal();
2521 switch(getTargetMachine().getCodeModel()) {
2522 case CodeModel::Small:
2523 // The most efficient code is PC-relative anyway for the small memory model,
2524 // so we don't need to worry about relocation model.
2525 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2526 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2527 AArch64II::MO_NO_FLAG),
2528 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2529 AArch64II::MO_LO12),
2530 DAG.getConstant(CN->getAlignment(), MVT::i32));
2531 case CodeModel::Large:
2533 AArch64ISD::WrapperLarge, DL, PtrVT,
2534 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G3),
2535 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G2_NC),
2536 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G1_NC),
2537 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G0_NC));
2539 llvm_unreachable("Only small and large code models supported now");
2543 SDValue AArch64TargetLowering::LowerTLSDescCall(SDValue SymAddr,
2546 SelectionDAG &DAG) const {
2547 EVT PtrVT = getPointerTy();
2549 // The function we need to call is simply the first entry in the GOT for this
2550 // descriptor, load it in preparation.
2551 SDValue Func, Chain;
2552 Func = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2555 // The function takes only one argument: the address of the descriptor itself
2558 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
2559 Glue = Chain.getValue(1);
2561 // Finally, there's a special calling-convention which means that the lookup
2562 // must preserve all registers (except X0, obviously).
2563 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2564 const AArch64RegisterInfo *A64RI
2565 = static_cast<const AArch64RegisterInfo *>(TRI);
2566 const uint32_t *Mask = A64RI->getTLSDescCallPreservedMask();
2568 // We're now ready to populate the argument list, as with a normal call:
2569 std::vector<SDValue> Ops;
2570 Ops.push_back(Chain);
2571 Ops.push_back(Func);
2572 Ops.push_back(SymAddr);
2573 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
2574 Ops.push_back(DAG.getRegisterMask(Mask));
2575 Ops.push_back(Glue);
2577 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2578 Chain = DAG.getNode(AArch64ISD::TLSDESCCALL, DL, NodeTys, &Ops[0],
2580 Glue = Chain.getValue(1);
2582 // After the call, the offset from TPIDR_EL0 is in X0, copy it out and pass it
2583 // back to the generic handling code.
2584 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
2588 AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2589 SelectionDAG &DAG) const {
2590 assert(getSubtarget()->isTargetELF() &&
2591 "TLS not implemented for non-ELF targets");
2592 assert(getTargetMachine().getCodeModel() == CodeModel::Small
2593 && "TLS only supported in small memory model");
2594 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2596 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2599 EVT PtrVT = getPointerTy();
2601 const GlobalValue *GV = GA->getGlobal();
2603 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
2605 if (Model == TLSModel::InitialExec) {
2606 TPOff = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2607 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2608 AArch64II::MO_GOTTPREL),
2609 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2610 AArch64II::MO_GOTTPREL_LO12),
2611 DAG.getConstant(8, MVT::i32));
2612 TPOff = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2614 } else if (Model == TLSModel::LocalExec) {
2615 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2616 AArch64II::MO_TPREL_G1);
2617 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2618 AArch64II::MO_TPREL_G0_NC);
2620 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2621 DAG.getTargetConstant(1, MVT::i32)), 0);
2622 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2624 DAG.getTargetConstant(0, MVT::i32)), 0);
2625 } else if (Model == TLSModel::GeneralDynamic) {
2626 // Accesses used in this sequence go via the TLS descriptor which lives in
2627 // the GOT. Prepare an address we can use to handle this.
2628 SDValue HiDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2629 AArch64II::MO_TLSDESC);
2630 SDValue LoDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2631 AArch64II::MO_TLSDESC_LO12);
2632 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2634 DAG.getConstant(8, MVT::i32));
2635 SDValue SymAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0);
2637 TPOff = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2638 } else if (Model == TLSModel::LocalDynamic) {
2639 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2640 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2641 // the beginning of the module's TLS region, followed by a DTPREL offset
2644 // These accesses will need deduplicating if there's more than one.
2645 AArch64MachineFunctionInfo* MFI = DAG.getMachineFunction()
2646 .getInfo<AArch64MachineFunctionInfo>();
2647 MFI->incNumLocalDynamicTLSAccesses();
2650 // Get the location of _TLS_MODULE_BASE_:
2651 SDValue HiDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2652 AArch64II::MO_TLSDESC);
2653 SDValue LoDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2654 AArch64II::MO_TLSDESC_LO12);
2655 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2657 DAG.getConstant(8, MVT::i32));
2658 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT);
2660 ThreadBase = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2662 // Get the variable's offset from _TLS_MODULE_BASE_
2663 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2664 AArch64II::MO_DTPREL_G1);
2665 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2666 AArch64II::MO_DTPREL_G0_NC);
2668 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2669 DAG.getTargetConstant(0, MVT::i32)), 0);
2670 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2672 DAG.getTargetConstant(0, MVT::i32)), 0);
2674 llvm_unreachable("Unsupported TLS access model");
2677 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2680 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2683 EVT VT = Op.getValueType();
2684 SDValue Vec = Op.getOperand(0);
2685 unsigned Opc = IsSigned ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
2687 if (VT.getVectorNumElements() == 1) {
2688 assert(VT == MVT::v1f64 && "Unexpected vector type!");
2689 if (VT.getSizeInBits() == Vec.getValueSizeInBits())
2691 return DAG.UnrollVectorOp(Op.getNode());
2694 if (VT.getSizeInBits() < Vec.getValueSizeInBits()) {
2695 assert(Vec.getValueType() == MVT::v2i64 && VT == MVT::v2f32 &&
2696 "Unexpected vector type!");
2697 Vec = DAG.getNode(Opc, dl, MVT::v2f64, Vec);
2698 return DAG.getNode(ISD::FP_ROUND, dl, VT, Vec, DAG.getIntPtrConstant(0));
2699 } else if (VT.getSizeInBits() > Vec.getValueSizeInBits()) {
2700 unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2701 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2702 VT.getVectorElementType().getSizeInBits());
2704 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2705 Vec = DAG.getNode(CastOpc, dl, CastVT, Vec);
2708 return DAG.getNode(Opc, dl, VT, Vec);
2712 AArch64TargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2713 bool IsSigned) const {
2714 if (Op.getValueType().isVector())
2715 return LowerVectorINT_TO_FP(Op, DAG, IsSigned);
2716 if (Op.getValueType() != MVT::f128) {
2717 // Legal for everything except f128.
2723 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2725 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2727 return LowerF128ToCall(Op, DAG, LC);
2732 AArch64TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
2733 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2735 EVT PtrVT = getPointerTy();
2737 // When compiling PIC, jump tables get put in the code section so a static
2738 // relocation-style is acceptable for both cases.
2739 switch (getTargetMachine().getCodeModel()) {
2740 case CodeModel::Small:
2741 return DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2742 DAG.getTargetJumpTable(JT->getIndex(), PtrVT),
2743 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
2744 AArch64II::MO_LO12),
2745 DAG.getConstant(1, MVT::i32));
2746 case CodeModel::Large:
2748 AArch64ISD::WrapperLarge, dl, PtrVT,
2749 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G3),
2750 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G2_NC),
2751 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G1_NC),
2752 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G0_NC));
2754 llvm_unreachable("Only small and large code models supported now");
2758 // (SELECT testbit, iftrue, iffalse)
2760 AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2762 SDValue TheBit = Op.getOperand(0);
2763 SDValue IfTrue = Op.getOperand(1);
2764 SDValue IfFalse = Op.getOperand(2);
2766 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2767 // that as the consumer we are responsible for ignoring rubbish in higher
2769 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2770 DAG.getConstant(1, MVT::i32));
2771 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2772 DAG.getConstant(0, TheBit.getValueType()),
2773 DAG.getCondCode(ISD::SETNE));
2775 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2776 A64CMP, IfTrue, IfFalse,
2777 DAG.getConstant(A64CC::NE, MVT::i32));
2780 static SDValue LowerVectorSETCC(SDValue Op, SelectionDAG &DAG) {
2782 SDValue LHS = Op.getOperand(0);
2783 SDValue RHS = Op.getOperand(1);
2784 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2785 EVT VT = Op.getValueType();
2786 bool Invert = false;
2790 if (LHS.getValueType().isInteger()) {
2792 // Attempt to use Vector Integer Compare Mask Test instruction.
2793 // TST = icmp ne (and (op0, op1), zero).
2794 if (CC == ISD::SETNE) {
2795 if (((LHS.getOpcode() == ISD::AND) &&
2796 ISD::isBuildVectorAllZeros(RHS.getNode())) ||
2797 ((RHS.getOpcode() == ISD::AND) &&
2798 ISD::isBuildVectorAllZeros(LHS.getNode()))) {
2800 SDValue AndOp = (LHS.getOpcode() == ISD::AND) ? LHS : RHS;
2801 SDValue NewLHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(0));
2802 SDValue NewRHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(1));
2803 return DAG.getNode(AArch64ISD::NEON_TST, DL, VT, NewLHS, NewRHS);
2807 // Attempt to use Vector Integer Compare Mask against Zero instr (Signed).
2808 // Note: Compare against Zero does not support unsigned predicates.
2809 if ((ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2810 ISD::isBuildVectorAllZeros(LHS.getNode())) &&
2811 !isUnsignedIntSetCC(CC)) {
2813 // If LHS is the zero value, swap operands and CondCode.
2814 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2815 CC = getSetCCSwappedOperands(CC);
2820 // Ensure valid CondCode for Compare Mask against Zero instruction:
2821 // EQ, GE, GT, LE, LT.
2822 if (ISD::SETNE == CC) {
2827 // Using constant type to differentiate integer and FP compares with zero.
2828 Op1 = DAG.getConstant(0, MVT::i32);
2829 Opcode = AArch64ISD::NEON_CMPZ;
2832 // Attempt to use Vector Integer Compare Mask instr (Signed/Unsigned).
2833 // Ensure valid CondCode for Compare Mask instr: EQ, GE, GT, UGE, UGT.
2837 llvm_unreachable("Illegal integer comparison.");
2853 CC = getSetCCSwappedOperands(CC);
2857 std::swap(LHS, RHS);
2859 Opcode = AArch64ISD::NEON_CMP;
2864 // Generate Compare Mask instr or Compare Mask against Zero instr.
2866 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
2869 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
2874 // Now handle Floating Point cases.
2875 // Attempt to use Vector Floating Point Compare Mask against Zero instruction.
2876 if (ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2877 ISD::isBuildVectorAllZeros(LHS.getNode())) {
2879 // If LHS is the zero value, swap operands and CondCode.
2880 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2881 CC = getSetCCSwappedOperands(CC);
2886 // Using constant type to differentiate integer and FP compares with zero.
2887 Op1 = DAG.getConstantFP(0, MVT::f32);
2888 Opcode = AArch64ISD::NEON_CMPZ;
2890 // Attempt to use Vector Floating Point Compare Mask instruction.
2893 Opcode = AArch64ISD::NEON_CMP;
2897 // Some register compares have to be implemented with swapped CC and operands,
2898 // e.g.: OLT implemented as OGT with swapped operands.
2899 bool SwapIfRegArgs = false;
2901 // Ensure valid CondCode for FP Compare Mask against Zero instruction:
2902 // EQ, GE, GT, LE, LT.
2903 // And ensure valid CondCode for FP Compare Mask instruction: EQ, GE, GT.
2906 llvm_unreachable("Illegal FP comparison");
2909 Invert = true; // Fallthrough
2917 SwapIfRegArgs = true;
2926 SwapIfRegArgs = true;
2935 SwapIfRegArgs = true;
2944 SwapIfRegArgs = true;
2951 Invert = true; // Fallthrough
2953 // Expand this to (OGT |OLT).
2955 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGT));
2957 SwapIfRegArgs = true;
2960 Invert = true; // Fallthrough
2962 // Expand this to (OGE | OLT).
2964 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGE));
2966 SwapIfRegArgs = true;
2970 if (Opcode == AArch64ISD::NEON_CMP && SwapIfRegArgs) {
2971 CC = getSetCCSwappedOperands(CC);
2972 std::swap(Op0, Op1);
2975 // Generate FP Compare Mask instr or FP Compare Mask against Zero instr
2976 SDValue NeonCmp = DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
2978 if (NeonCmpAlt.getNode())
2979 NeonCmp = DAG.getNode(ISD::OR, DL, VT, NeonCmp, NeonCmpAlt);
2982 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
2987 // (SETCC lhs, rhs, condcode)
2989 AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2991 SDValue LHS = Op.getOperand(0);
2992 SDValue RHS = Op.getOperand(1);
2993 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2994 EVT VT = Op.getValueType();
2997 return LowerVectorSETCC(Op, DAG);
2999 if (LHS.getValueType() == MVT::f128) {
3000 // f128 comparisons will be lowered to libcalls giving a valid LHS and RHS
3001 // for the rest of the function (some i32 or i64 values).
3002 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3004 // If softenSetCCOperands returned a scalar, use it.
3005 if (RHS.getNode() == 0) {
3006 assert(LHS.getValueType() == Op.getValueType() &&
3007 "Unexpected setcc expansion!");
3012 if (LHS.getValueType().isInteger()) {
3015 // Integers are handled in a separate function because the combinations of
3016 // immediates and tests can get hairy and we may want to fiddle things.
3017 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
3019 return DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
3020 CmpOp, DAG.getConstant(1, VT), DAG.getConstant(0, VT),
3024 // Note that some LLVM floating-point CondCodes can't be lowered to a single
3025 // conditional branch, hence FPCCToA64CC can set a second test, where either
3026 // passing is sufficient.
3027 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3028 CondCode = FPCCToA64CC(CC, Alternative);
3029 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3030 SDValue CmpOp = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3031 DAG.getCondCode(CC));
3032 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
3033 CmpOp, DAG.getConstant(1, VT),
3034 DAG.getConstant(0, VT), A64cc);
3036 if (Alternative != A64CC::Invalid) {
3037 A64cc = DAG.getConstant(Alternative, MVT::i32);
3038 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
3039 DAG.getConstant(1, VT), A64SELECT_CC, A64cc);
3042 return A64SELECT_CC;
3045 static SDValue LowerVectorSELECT_CC(SDValue Op, SelectionDAG &DAG) {
3047 SDValue LHS = Op.getOperand(0);
3048 SDValue RHS = Op.getOperand(1);
3049 SDValue IfTrue = Op.getOperand(2);
3050 SDValue IfFalse = Op.getOperand(3);
3051 EVT IfTrueVT = IfTrue.getValueType();
3052 EVT CondVT = IfTrueVT.changeVectorElementTypeToInteger();
3053 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3055 // If LHS & RHS are floating point and IfTrue & IfFalse are vectors, we will
3056 // use NEON compare.
3057 if ((LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64)) {
3058 EVT EltVT = LHS.getValueType();
3059 unsigned EltNum = 128 / EltVT.getSizeInBits();
3060 EVT VT = EVT::getVectorVT(*DAG.getContext(), EltVT, EltNum);
3061 unsigned SubConstant =
3062 (LHS.getValueType() == MVT::f32) ? AArch64::sub_32 :AArch64::sub_64;
3063 EVT CEltT = (LHS.getValueType() == MVT::f32) ? MVT::i32 : MVT::i64;
3064 EVT CVT = EVT::getVectorVT(*DAG.getContext(), CEltT, EltNum);
3067 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
3068 VT, DAG.getTargetConstant(0, MVT::i32), LHS,
3069 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3071 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
3072 VT, DAG.getTargetConstant(0, MVT::i32), RHS,
3073 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3075 SDValue VSetCC = DAG.getSetCC(dl, CVT, LHS, RHS, CC);
3076 SDValue ResCC = LowerVectorSETCC(VSetCC, DAG);
3077 if (CEltT.getSizeInBits() < IfTrueVT.getSizeInBits()) {
3079 EVT::getVectorVT(*DAG.getContext(), CEltT,
3080 IfTrueVT.getSizeInBits() / CEltT.getSizeInBits());
3081 ResCC = DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, DUPVT, ResCC,
3082 DAG.getConstant(0, MVT::i64, false));
3084 ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
3086 // FIXME: If IfTrue & IfFalse hold v1i8, v1i16 or v1i32, this function
3087 // can't handle them and will hit this assert.
3088 assert(CEltT.getSizeInBits() == IfTrueVT.getSizeInBits() &&
3089 "Vector of IfTrue & IfFalse is too small.");
3092 EltNum * IfTrueVT.getSizeInBits() / ResCC.getValueSizeInBits();
3093 EVT ExVT = EVT::getVectorVT(*DAG.getContext(), CEltT, ExEltNum);
3094 ResCC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ExVT, ResCC,
3095 DAG.getConstant(0, MVT::i64, false));
3096 ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
3098 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3099 ResCC, IfTrue, IfFalse);
3103 // Here we handle the case that LHS & RHS are integer and IfTrue & IfFalse are
3105 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3106 CondCode = FPCCToA64CC(CC, Alternative);
3107 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3108 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3109 DAG.getCondCode(CC));
3110 EVT SEVT = MVT::i32;
3111 if (IfTrue.getValueType().getVectorElementType().getSizeInBits() > 32)
3113 SDValue AllOne = DAG.getConstant(-1, SEVT);
3114 SDValue AllZero = DAG.getConstant(0, SEVT);
3115 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, SEVT, SetCC,
3116 AllOne, AllZero, A64cc);
3118 if (Alternative != A64CC::Invalid) {
3119 A64cc = DAG.getConstant(Alternative, MVT::i32);
3120 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3121 SetCC, AllOne, A64SELECT_CC, A64cc);
3124 if (IfTrue.getValueType().getVectorNumElements() == 1)
3125 VDup = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, CondVT, A64SELECT_CC);
3127 VDup = DAG.getNode(AArch64ISD::NEON_VDUP, dl, CondVT, A64SELECT_CC);
3128 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3129 VDup, IfTrue, IfFalse);
3133 // (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
3135 AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3137 SDValue LHS = Op.getOperand(0);
3138 SDValue RHS = Op.getOperand(1);
3139 SDValue IfTrue = Op.getOperand(2);
3140 SDValue IfFalse = Op.getOperand(3);
3141 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3143 if (IfTrue.getValueType().isVector())
3144 return LowerVectorSELECT_CC(Op, DAG);
3146 if (LHS.getValueType() == MVT::f128) {
3147 // f128 comparisons are lowered to libcalls, but slot in nicely here
3149 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3151 // If softenSetCCOperands returned a scalar, we need to compare the result
3152 // against zero to select between true and false values.
3153 if (RHS.getNode() == 0) {
3154 RHS = DAG.getConstant(0, LHS.getValueType());
3159 if (LHS.getValueType().isInteger()) {
3162 // Integers are handled in a separate function because the combinations of
3163 // immediates and tests can get hairy and we may want to fiddle things.
3164 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
3166 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(), CmpOp,
3167 IfTrue, IfFalse, A64cc);
3170 // Note that some LLVM floating-point CondCodes can't be lowered to a single
3171 // conditional branch, hence FPCCToA64CC can set a second test, where either
3172 // passing is sufficient.
3173 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3174 CondCode = FPCCToA64CC(CC, Alternative);
3175 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3176 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3177 DAG.getCondCode(CC));
3178 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
3180 SetCC, IfTrue, IfFalse, A64cc);
3182 if (Alternative != A64CC::Invalid) {
3183 A64cc = DAG.getConstant(Alternative, MVT::i32);
3184 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3185 SetCC, IfTrue, A64SELECT_CC, A64cc);
3189 return A64SELECT_CC;
3193 AArch64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3194 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3195 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3197 // We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
3198 // rather than just 8.
3199 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op),
3200 Op.getOperand(1), Op.getOperand(2),
3201 DAG.getConstant(32, MVT::i32), 8, false, false,
3202 MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV));
3206 AArch64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3207 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3208 // Standard, section B.3.
3209 MachineFunction &MF = DAG.getMachineFunction();
3210 AArch64MachineFunctionInfo *FuncInfo
3211 = MF.getInfo<AArch64MachineFunctionInfo>();
3214 SDValue Chain = Op.getOperand(0);
3215 SDValue VAList = Op.getOperand(1);
3216 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3217 SmallVector<SDValue, 4> MemOps;
3219 // void *__stack at offset 0
3220 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVariadicStackIdx(),
3222 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3223 MachinePointerInfo(SV), false, false, 0));
3225 // void *__gr_top at offset 8
3226 int GPRSize = FuncInfo->getVariadicGPRSize();
3228 SDValue GRTop, GRTopAddr;
3230 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3231 DAG.getConstant(8, getPointerTy()));
3233 GRTop = DAG.getFrameIndex(FuncInfo->getVariadicGPRIdx(), getPointerTy());
3234 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3235 DAG.getConstant(GPRSize, getPointerTy()));
3237 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3238 MachinePointerInfo(SV, 8),
3242 // void *__vr_top at offset 16
3243 int FPRSize = FuncInfo->getVariadicFPRSize();
3245 SDValue VRTop, VRTopAddr;
3246 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3247 DAG.getConstant(16, getPointerTy()));
3249 VRTop = DAG.getFrameIndex(FuncInfo->getVariadicFPRIdx(), getPointerTy());
3250 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3251 DAG.getConstant(FPRSize, getPointerTy()));
3253 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3254 MachinePointerInfo(SV, 16),
3258 // int __gr_offs at offset 24
3259 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3260 DAG.getConstant(24, getPointerTy()));
3261 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3262 GROffsAddr, MachinePointerInfo(SV, 24),
3265 // int __vr_offs at offset 28
3266 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3267 DAG.getConstant(28, getPointerTy()));
3268 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3269 VROffsAddr, MachinePointerInfo(SV, 28),
3272 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3277 AArch64TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3278 switch (Op.getOpcode()) {
3279 default: llvm_unreachable("Don't know how to custom lower this!");
3280 case ISD::FADD: return LowerF128ToCall(Op, DAG, RTLIB::ADD_F128);
3281 case ISD::FSUB: return LowerF128ToCall(Op, DAG, RTLIB::SUB_F128);
3282 case ISD::FMUL: return LowerF128ToCall(Op, DAG, RTLIB::MUL_F128);
3283 case ISD::FDIV: return LowerF128ToCall(Op, DAG, RTLIB::DIV_F128);
3284 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, true);
3285 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG, false);
3286 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG, true);
3287 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG, false);
3288 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
3289 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
3290 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3291 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3293 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3294 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3295 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3296 case ISD::GlobalAddress: return LowerGlobalAddressELF(Op, DAG);
3297 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3298 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3299 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3300 case ISD::SELECT: return LowerSELECT(Op, DAG);
3301 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3302 case ISD::SETCC: return LowerSETCC(Op, DAG);
3303 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
3304 case ISD::VASTART: return LowerVASTART(Op, DAG);
3305 case ISD::BUILD_VECTOR:
3306 return LowerBUILD_VECTOR(Op, DAG, getSubtarget());
3307 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3308 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3314 /// Check if the specified splat value corresponds to a valid vector constant
3315 /// for a Neon instruction with a "modified immediate" operand (e.g., MOVI). If
3316 /// so, return the encoded 8-bit immediate and the OpCmode instruction fields
3318 static bool isNeonModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3319 unsigned SplatBitSize, SelectionDAG &DAG,
3320 bool is128Bits, NeonModImmType type, EVT &VT,
3321 unsigned &Imm, unsigned &OpCmode) {
3322 switch (SplatBitSize) {
3324 llvm_unreachable("unexpected size for isNeonModifiedImm");
3326 if (type != Neon_Mov_Imm)
3328 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3329 // Neon movi per byte: Op=0, Cmode=1110.
3332 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3336 // Neon move inst per halfword
3337 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3338 if ((SplatBits & ~0xff) == 0) {
3339 // Value = 0x00nn is 0x00nn LSL 0
3340 // movi: Op=0, Cmode=1000; mvni: Op=1, Cmode=1000
3341 // bic: Op=1, Cmode=1001; orr: Op=0, Cmode=1001
3347 if ((SplatBits & ~0xff00) == 0) {
3348 // Value = 0xnn00 is 0x00nn LSL 8
3349 // movi: Op=0, Cmode=1010; mvni: Op=1, Cmode=1010
3350 // bic: Op=1, Cmode=1011; orr: Op=0, Cmode=1011
3352 Imm = SplatBits >> 8;
3356 // can't handle any other
3361 // First the LSL variants (MSL is unusable by some interested instructions).
3363 // Neon move instr per word, shift zeros
3364 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3365 if ((SplatBits & ~0xff) == 0) {
3366 // Value = 0x000000nn is 0x000000nn LSL 0
3367 // movi: Op=0, Cmode= 0000; mvni: Op=1, Cmode= 0000
3368 // bic: Op=1, Cmode= 0001; orr: Op=0, Cmode= 0001
3374 if ((SplatBits & ~0xff00) == 0) {
3375 // Value = 0x0000nn00 is 0x000000nn LSL 8
3376 // movi: Op=0, Cmode= 0010; mvni: Op=1, Cmode= 0010
3377 // bic: Op=1, Cmode= 0011; orr : Op=0, Cmode= 0011
3379 Imm = SplatBits >> 8;
3383 if ((SplatBits & ~0xff0000) == 0) {
3384 // Value = 0x00nn0000 is 0x000000nn LSL 16
3385 // movi: Op=0, Cmode= 0100; mvni: Op=1, Cmode= 0100
3386 // bic: Op=1, Cmode= 0101; orr: Op=0, Cmode= 0101
3388 Imm = SplatBits >> 16;
3392 if ((SplatBits & ~0xff000000) == 0) {
3393 // Value = 0xnn000000 is 0x000000nn LSL 24
3394 // movi: Op=0, Cmode= 0110; mvni: Op=1, Cmode= 0110
3395 // bic: Op=1, Cmode= 0111; orr: Op=0, Cmode= 0111
3397 Imm = SplatBits >> 24;
3402 // Now the MSL immediates.
3404 // Neon move instr per word, shift ones
3405 if ((SplatBits & ~0xffff) == 0 &&
3406 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3407 // Value = 0x0000nnff is 0x000000nn MSL 8
3408 // movi: Op=0, Cmode= 1100; mvni: Op=1, Cmode= 1100
3410 Imm = SplatBits >> 8;
3414 if ((SplatBits & ~0xffffff) == 0 &&
3415 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3416 // Value = 0x00nnffff is 0x000000nn MSL 16
3417 // movi: Op=1, Cmode= 1101; mvni: Op=1, Cmode= 1101
3419 Imm = SplatBits >> 16;
3423 // can't handle any other
3428 if (type != Neon_Mov_Imm)
3430 // Neon move instr bytemask, where each byte is either 0x00 or 0xff.
3431 // movi Op=1, Cmode=1110.
3433 uint64_t BitMask = 0xff;
3435 unsigned ImmMask = 1;
3437 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3438 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3441 } else if ((SplatBits & BitMask) != 0) {
3448 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3456 static SDValue PerformANDCombine(SDNode *N,
3457 TargetLowering::DAGCombinerInfo &DCI) {
3459 SelectionDAG &DAG = DCI.DAG;
3461 EVT VT = N->getValueType(0);
3463 // We're looking for an SRA/SHL pair which form an SBFX.
3465 if (VT != MVT::i32 && VT != MVT::i64)
3468 if (!isa<ConstantSDNode>(N->getOperand(1)))
3471 uint64_t TruncMask = N->getConstantOperandVal(1);
3472 if (!isMask_64(TruncMask))
3475 uint64_t Width = CountPopulation_64(TruncMask);
3476 SDValue Shift = N->getOperand(0);
3478 if (Shift.getOpcode() != ISD::SRL)
3481 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3483 uint64_t LSB = Shift->getConstantOperandVal(1);
3485 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3488 return DAG.getNode(AArch64ISD::UBFX, DL, VT, Shift.getOperand(0),
3489 DAG.getConstant(LSB, MVT::i64),
3490 DAG.getConstant(LSB + Width - 1, MVT::i64));
3493 /// For a true bitfield insert, the bits getting into that contiguous mask
3494 /// should come from the low part of an existing value: they must be formed from
3495 /// a compatible SHL operation (unless they're already low). This function
3496 /// checks that condition and returns the least-significant bit that's
3497 /// intended. If the operation not a field preparation, -1 is returned.
3498 static int32_t getLSBForBFI(SelectionDAG &DAG, SDLoc DL, EVT VT,
3499 SDValue &MaskedVal, uint64_t Mask) {
3500 if (!isShiftedMask_64(Mask))
3503 // Now we need to alter MaskedVal so that it is an appropriate input for a BFI
3504 // instruction. BFI will do a left-shift by LSB before applying the mask we've
3505 // spotted, so in general we should pre-emptively "undo" that by making sure
3506 // the incoming bits have had a right-shift applied to them.
3508 // This right shift, however, will combine with existing left/right shifts. In
3509 // the simplest case of a completely straight bitfield operation, it will be
3510 // expected to completely cancel out with an existing SHL. More complicated
3511 // cases (e.g. bitfield to bitfield copy) may still need a real shift before
3514 uint64_t LSB = countTrailingZeros(Mask);
3515 int64_t ShiftRightRequired = LSB;
3516 if (MaskedVal.getOpcode() == ISD::SHL &&
3517 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3518 ShiftRightRequired -= MaskedVal.getConstantOperandVal(1);
3519 MaskedVal = MaskedVal.getOperand(0);
3520 } else if (MaskedVal.getOpcode() == ISD::SRL &&
3521 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3522 ShiftRightRequired += MaskedVal.getConstantOperandVal(1);
3523 MaskedVal = MaskedVal.getOperand(0);
3526 if (ShiftRightRequired > 0)
3527 MaskedVal = DAG.getNode(ISD::SRL, DL, VT, MaskedVal,
3528 DAG.getConstant(ShiftRightRequired, MVT::i64));
3529 else if (ShiftRightRequired < 0) {
3530 // We could actually end up with a residual left shift, for example with
3531 // "struc.bitfield = val << 1".
3532 MaskedVal = DAG.getNode(ISD::SHL, DL, VT, MaskedVal,
3533 DAG.getConstant(-ShiftRightRequired, MVT::i64));
3539 /// Searches from N for an existing AArch64ISD::BFI node, possibly surrounded by
3540 /// a mask and an extension. Returns true if a BFI was found and provides
3541 /// information on its surroundings.
3542 static bool findMaskedBFI(SDValue N, SDValue &BFI, uint64_t &Mask,
3545 if (N.getOpcode() == ISD::ZERO_EXTEND) {
3547 N = N.getOperand(0);
3550 if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
3551 Mask = N->getConstantOperandVal(1);
3552 N = N.getOperand(0);
3554 // Mask is the whole width.
3555 Mask = -1ULL >> (64 - N.getValueType().getSizeInBits());
3558 if (N.getOpcode() == AArch64ISD::BFI) {
3566 /// Try to combine a subtree (rooted at an OR) into a "masked BFI" node, which
3567 /// is roughly equivalent to (and (BFI ...), mask). This form is used because it
3568 /// can often be further combined with a larger mask. Ultimately, we want mask
3569 /// to be 2^32-1 or 2^64-1 so the AND can be skipped.
3570 static SDValue tryCombineToBFI(SDNode *N,
3571 TargetLowering::DAGCombinerInfo &DCI,
3572 const AArch64Subtarget *Subtarget) {
3573 SelectionDAG &DAG = DCI.DAG;
3575 EVT VT = N->getValueType(0);
3577 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3579 // We need the LHS to be (and SOMETHING, MASK). Find out what that mask is or
3580 // abandon the effort.
3581 SDValue LHS = N->getOperand(0);
3582 if (LHS.getOpcode() != ISD::AND)
3586 if (isa<ConstantSDNode>(LHS.getOperand(1)))
3587 LHSMask = LHS->getConstantOperandVal(1);
3591 // We also need the RHS to be (and SOMETHING, MASK). Find out what that mask
3592 // is or abandon the effort.
3593 SDValue RHS = N->getOperand(1);
3594 if (RHS.getOpcode() != ISD::AND)
3598 if (isa<ConstantSDNode>(RHS.getOperand(1)))
3599 RHSMask = RHS->getConstantOperandVal(1);
3603 // Can't do anything if the masks are incompatible.
3604 if (LHSMask & RHSMask)
3607 // Now we need one of the masks to be a contiguous field. Without loss of
3608 // generality that should be the RHS one.
3609 SDValue Bitfield = LHS.getOperand(0);
3610 if (getLSBForBFI(DAG, DL, VT, Bitfield, LHSMask) != -1) {
3611 // We know that LHS is a candidate new value, and RHS isn't already a better
3613 std::swap(LHS, RHS);
3614 std::swap(LHSMask, RHSMask);
3617 // We've done our best to put the right operands in the right places, all we
3618 // can do now is check whether a BFI exists.
3619 Bitfield = RHS.getOperand(0);
3620 int32_t LSB = getLSBForBFI(DAG, DL, VT, Bitfield, RHSMask);
3624 uint32_t Width = CountPopulation_64(RHSMask);
3625 assert(Width && "Expected non-zero bitfield width");
3627 SDValue BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3628 LHS.getOperand(0), Bitfield,
3629 DAG.getConstant(LSB, MVT::i64),
3630 DAG.getConstant(Width, MVT::i64));
3633 if ((LHSMask | RHSMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3636 return DAG.getNode(ISD::AND, DL, VT, BFI,
3637 DAG.getConstant(LHSMask | RHSMask, VT));
3640 /// Search for the bitwise combining (with careful masks) of a MaskedBFI and its
3641 /// original input. This is surprisingly common because SROA splits things up
3642 /// into i8 chunks, so the originally detected MaskedBFI may actually only act
3643 /// on the low (say) byte of a word. This is then orred into the rest of the
3644 /// word afterwards.
3646 /// Basic input: (or (and OLDFIELD, MASK1), (MaskedBFI MASK2, OLDFIELD, ...)).
3648 /// If MASK1 and MASK2 are compatible, we can fold the whole thing into the
3649 /// MaskedBFI. We can also deal with a certain amount of extend/truncate being
3651 static SDValue tryCombineToLargerBFI(SDNode *N,
3652 TargetLowering::DAGCombinerInfo &DCI,
3653 const AArch64Subtarget *Subtarget) {
3654 SelectionDAG &DAG = DCI.DAG;
3656 EVT VT = N->getValueType(0);
3658 // First job is to hunt for a MaskedBFI on either the left or right. Swap
3659 // operands if it's actually on the right.
3661 SDValue PossExtraMask;
3662 uint64_t ExistingMask = 0;
3663 bool Extended = false;
3664 if (findMaskedBFI(N->getOperand(0), BFI, ExistingMask, Extended))
3665 PossExtraMask = N->getOperand(1);
3666 else if (findMaskedBFI(N->getOperand(1), BFI, ExistingMask, Extended))
3667 PossExtraMask = N->getOperand(0);
3671 // We can only combine a BFI with another compatible mask.
3672 if (PossExtraMask.getOpcode() != ISD::AND ||
3673 !isa<ConstantSDNode>(PossExtraMask.getOperand(1)))
3676 uint64_t ExtraMask = PossExtraMask->getConstantOperandVal(1);
3678 // Masks must be compatible.
3679 if (ExtraMask & ExistingMask)
3682 SDValue OldBFIVal = BFI.getOperand(0);
3683 SDValue NewBFIVal = BFI.getOperand(1);
3685 // We skipped a ZERO_EXTEND above, so the input to the MaskedBFIs should be
3686 // 32-bit and we'll be forming a 64-bit MaskedBFI. The MaskedBFI arguments
3687 // need to be made compatible.
3688 assert(VT == MVT::i64 && BFI.getValueType() == MVT::i32
3689 && "Invalid types for BFI");
3690 OldBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, OldBFIVal);
3691 NewBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, NewBFIVal);
3694 // We need the MaskedBFI to be combined with a mask of the *same* value.
3695 if (PossExtraMask.getOperand(0) != OldBFIVal)
3698 BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3699 OldBFIVal, NewBFIVal,
3700 BFI.getOperand(2), BFI.getOperand(3));
3702 // If the masking is trivial, we don't need to create it.
3703 if ((ExtraMask | ExistingMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3706 return DAG.getNode(ISD::AND, DL, VT, BFI,
3707 DAG.getConstant(ExtraMask | ExistingMask, VT));
3710 /// An EXTR instruction is made up of two shifts, ORed together. This helper
3711 /// searches for and classifies those shifts.
3712 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
3714 if (N.getOpcode() == ISD::SHL)
3716 else if (N.getOpcode() == ISD::SRL)
3721 if (!isa<ConstantSDNode>(N.getOperand(1)))
3724 ShiftAmount = N->getConstantOperandVal(1);
3725 Src = N->getOperand(0);
3729 /// EXTR instruction extracts a contiguous chunk of bits from two existing
3730 /// registers viewed as a high/low pair. This function looks for the pattern:
3731 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
3732 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
3734 static SDValue tryCombineToEXTR(SDNode *N,
3735 TargetLowering::DAGCombinerInfo &DCI) {
3736 SelectionDAG &DAG = DCI.DAG;
3738 EVT VT = N->getValueType(0);
3740 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3742 if (VT != MVT::i32 && VT != MVT::i64)
3746 uint32_t ShiftLHS = 0;
3748 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
3752 uint32_t ShiftRHS = 0;
3754 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
3757 // If they're both trying to come from the high part of the register, they're
3758 // not really an EXTR.
3759 if (LHSFromHi == RHSFromHi)
3762 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
3766 std::swap(LHS, RHS);
3767 std::swap(ShiftLHS, ShiftRHS);
3770 return DAG.getNode(AArch64ISD::EXTR, DL, VT,
3772 DAG.getConstant(ShiftRHS, MVT::i64));
3775 /// Target-specific dag combine xforms for ISD::OR
3776 static SDValue PerformORCombine(SDNode *N,
3777 TargetLowering::DAGCombinerInfo &DCI,
3778 const AArch64Subtarget *Subtarget) {
3780 SelectionDAG &DAG = DCI.DAG;
3782 EVT VT = N->getValueType(0);
3784 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
3787 // Attempt to recognise bitfield-insert operations.
3788 SDValue Res = tryCombineToBFI(N, DCI, Subtarget);
3792 // Attempt to combine an existing MaskedBFI operation into one with a larger
3794 Res = tryCombineToLargerBFI(N, DCI, Subtarget);
3798 Res = tryCombineToEXTR(N, DCI);
3802 if (!Subtarget->hasNEON())
3805 // Attempt to use vector immediate-form BSL
3806 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
3808 SDValue N0 = N->getOperand(0);
3809 if (N0.getOpcode() != ISD::AND)
3812 SDValue N1 = N->getOperand(1);
3813 if (N1.getOpcode() != ISD::AND)
3816 if (VT.isVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
3818 unsigned SplatBitSize;
3820 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
3822 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
3825 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
3827 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
3828 HasAnyUndefs) && !HasAnyUndefs &&
3829 SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
3830 SplatBits0 == ~SplatBits1) {
3832 return DAG.getNode(ISD::VSELECT, DL, VT, N0->getOperand(1),
3833 N0->getOperand(0), N1->getOperand(0));
3841 /// Target-specific dag combine xforms for ISD::SRA
3842 static SDValue PerformSRACombine(SDNode *N,
3843 TargetLowering::DAGCombinerInfo &DCI) {
3845 SelectionDAG &DAG = DCI.DAG;
3847 EVT VT = N->getValueType(0);
3849 // We're looking for an SRA/SHL pair which form an SBFX.
3851 if (VT != MVT::i32 && VT != MVT::i64)
3854 if (!isa<ConstantSDNode>(N->getOperand(1)))
3857 uint64_t ExtraSignBits = N->getConstantOperandVal(1);
3858 SDValue Shift = N->getOperand(0);
3860 if (Shift.getOpcode() != ISD::SHL)
3863 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3866 uint64_t BitsOnLeft = Shift->getConstantOperandVal(1);
3867 uint64_t Width = VT.getSizeInBits() - ExtraSignBits;
3868 uint64_t LSB = VT.getSizeInBits() - Width - BitsOnLeft;
3870 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3873 return DAG.getNode(AArch64ISD::SBFX, DL, VT, Shift.getOperand(0),
3874 DAG.getConstant(LSB, MVT::i64),
3875 DAG.getConstant(LSB + Width - 1, MVT::i64));
3878 /// Check if this is a valid build_vector for the immediate operand of
3879 /// a vector shift operation, where all the elements of the build_vector
3880 /// must have the same constant integer value.
3881 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3882 // Ignore bit_converts.
3883 while (Op.getOpcode() == ISD::BITCAST)
3884 Op = Op.getOperand(0);
3885 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3886 APInt SplatBits, SplatUndef;
3887 unsigned SplatBitSize;
3889 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3890 HasAnyUndefs, ElementBits) ||
3891 SplatBitSize > ElementBits)
3893 Cnt = SplatBits.getSExtValue();
3897 /// Check if this is a valid build_vector for the immediate operand of
3898 /// a vector shift left operation. That value must be in the range:
3899 /// 0 <= Value < ElementBits
3900 static bool isVShiftLImm(SDValue Op, EVT VT, int64_t &Cnt) {
3901 assert(VT.isVector() && "vector shift count is not a vector type");
3902 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3903 if (!getVShiftImm(Op, ElementBits, Cnt))
3905 return (Cnt >= 0 && Cnt < ElementBits);
3908 /// Check if this is a valid build_vector for the immediate operand of a
3909 /// vector shift right operation. The value must be in the range:
3910 /// 1 <= Value <= ElementBits
3911 static bool isVShiftRImm(SDValue Op, EVT VT, int64_t &Cnt) {
3912 assert(VT.isVector() && "vector shift count is not a vector type");
3913 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3914 if (!getVShiftImm(Op, ElementBits, Cnt))
3916 return (Cnt >= 1 && Cnt <= ElementBits);
3919 static SDValue GenForSextInreg(SDNode *N,
3920 TargetLowering::DAGCombinerInfo &DCI,
3921 EVT SrcVT, EVT DestVT, EVT SubRegVT,
3922 const int *Mask, SDValue Src) {
3923 SelectionDAG &DAG = DCI.DAG;
3925 = DAG.getNode(ISD::BITCAST, SDLoc(N), SrcVT, Src);
3927 = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), DestVT, Bitcast);
3929 = DAG.getVectorShuffle(DestVT, SDLoc(N), Sext, DAG.getUNDEF(DestVT), Mask);
3930 SDValue ExtractSubreg
3931 = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, SDLoc(N),
3932 SubRegVT, ShuffleVec,
3933 DAG.getTargetConstant(AArch64::sub_64, MVT::i32)), 0);
3934 return ExtractSubreg;
3937 /// Checks for vector shifts and lowers them.
3938 static SDValue PerformShiftCombine(SDNode *N,
3939 TargetLowering::DAGCombinerInfo &DCI,
3940 const AArch64Subtarget *ST) {
3941 SelectionDAG &DAG = DCI.DAG;
3942 EVT VT = N->getValueType(0);
3943 if (N->getOpcode() == ISD::SRA && (VT == MVT::i32 || VT == MVT::i64))
3944 return PerformSRACombine(N, DCI);
3946 // We're looking for an SRA/SHL pair to help generating instruction
3947 // sshll v0.8h, v0.8b, #0
3948 // The instruction STXL is also the alias of this instruction.
3950 // For example, for DAG like below,
3951 // v2i32 = sra (v2i32 (shl v2i32, 16)), 16
3952 // we can transform it into
3953 // v2i32 = EXTRACT_SUBREG
3954 // (v4i32 (suffle_vector
3955 // (v4i32 (sext (v4i16 (bitcast v2i32))),
3956 // undef, (0, 2, u, u)),
3959 // With this transformation we expect to generate "SSHLL + UZIP1"
3960 // Sometimes UZIP1 can be optimized away by combining with other context.
3961 int64_t ShrCnt, ShlCnt;
3962 if (N->getOpcode() == ISD::SRA
3963 && (VT == MVT::v2i32 || VT == MVT::v4i16)
3964 && isVShiftRImm(N->getOperand(1), VT, ShrCnt)
3965 && N->getOperand(0).getOpcode() == ISD::SHL
3966 && isVShiftRImm(N->getOperand(0).getOperand(1), VT, ShlCnt)) {
3967 SDValue Src = N->getOperand(0).getOperand(0);
3968 if (VT == MVT::v2i32 && ShrCnt == 16 && ShlCnt == 16) {
3969 // sext_inreg(v2i32, v2i16)
3970 // We essentially only care the Mask {0, 2, u, u}
3971 int Mask[4] = {0, 2, 4, 6};
3972 return GenForSextInreg(N, DCI, MVT::v4i16, MVT::v4i32, MVT::v2i32,
3975 else if (VT == MVT::v2i32 && ShrCnt == 24 && ShlCnt == 24) {
3976 // sext_inreg(v2i16, v2i8)
3977 // We essentially only care the Mask {0, u, 4, u, u, u, u, u, u, u, u, u}
3978 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
3979 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v2i32,
3982 else if (VT == MVT::v4i16 && ShrCnt == 8 && ShlCnt == 8) {
3983 // sext_inreg(v4i16, v4i8)
3984 // We essentially only care the Mask {0, 2, 4, 6, u, u, u, u, u, u, u, u}
3985 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
3986 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v4i16,
3991 // Nothing to be done for scalar shifts.
3992 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3993 if (!VT.isVector() || !TLI.isTypeLegal(VT))
3996 assert(ST->hasNEON() && "unexpected vector shift");
3999 switch (N->getOpcode()) {
4001 llvm_unreachable("unexpected shift opcode");
4004 if (isVShiftLImm(N->getOperand(1), VT, Cnt)) {
4006 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
4007 DAG.getConstant(Cnt, MVT::i32));
4008 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0), RHS);
4014 if (isVShiftRImm(N->getOperand(1), VT, Cnt)) {
4016 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
4017 DAG.getConstant(Cnt, MVT::i32));
4018 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N->getOperand(0), RHS);
4026 /// ARM-specific DAG combining for intrinsics.
4027 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4028 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4032 // Don't do anything for most intrinsics.
4035 case Intrinsic::arm_neon_vqshifts:
4036 case Intrinsic::arm_neon_vqshiftu:
4037 EVT VT = N->getOperand(1).getValueType();
4039 if (!isVShiftLImm(N->getOperand(2), VT, Cnt))
4041 unsigned VShiftOpc = (IntNo == Intrinsic::arm_neon_vqshifts)
4042 ? AArch64ISD::NEON_QSHLs
4043 : AArch64ISD::NEON_QSHLu;
4044 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
4045 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4051 /// Target-specific DAG combine function for NEON load/store intrinsics
4052 /// to merge base address updates.
4053 static SDValue CombineBaseUpdate(SDNode *N,
4054 TargetLowering::DAGCombinerInfo &DCI) {
4055 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4058 SelectionDAG &DAG = DCI.DAG;
4059 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
4060 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
4061 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
4062 SDValue Addr = N->getOperand(AddrOpIdx);
4064 // Search for a use of the address operand that is an increment.
4065 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
4066 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
4068 if (User->getOpcode() != ISD::ADD ||
4069 UI.getUse().getResNo() != Addr.getResNo())
4072 // Check that the add is independent of the load/store. Otherwise, folding
4073 // it would create a cycle.
4074 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
4077 // Find the new opcode for the updating load/store.
4079 bool isLaneOp = false;
4080 unsigned NewOpc = 0;
4081 unsigned NumVecs = 0;
4083 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
4085 default: llvm_unreachable("unexpected intrinsic for Neon base update");
4086 case Intrinsic::arm_neon_vld1: NewOpc = AArch64ISD::NEON_LD1_UPD;
4088 case Intrinsic::arm_neon_vld2: NewOpc = AArch64ISD::NEON_LD2_UPD;
4090 case Intrinsic::arm_neon_vld3: NewOpc = AArch64ISD::NEON_LD3_UPD;
4092 case Intrinsic::arm_neon_vld4: NewOpc = AArch64ISD::NEON_LD4_UPD;
4094 case Intrinsic::arm_neon_vst1: NewOpc = AArch64ISD::NEON_ST1_UPD;
4095 NumVecs = 1; isLoad = false; break;
4096 case Intrinsic::arm_neon_vst2: NewOpc = AArch64ISD::NEON_ST2_UPD;
4097 NumVecs = 2; isLoad = false; break;
4098 case Intrinsic::arm_neon_vst3: NewOpc = AArch64ISD::NEON_ST3_UPD;
4099 NumVecs = 3; isLoad = false; break;
4100 case Intrinsic::arm_neon_vst4: NewOpc = AArch64ISD::NEON_ST4_UPD;
4101 NumVecs = 4; isLoad = false; break;
4102 case Intrinsic::aarch64_neon_vld1x2: NewOpc = AArch64ISD::NEON_LD1x2_UPD;
4104 case Intrinsic::aarch64_neon_vld1x3: NewOpc = AArch64ISD::NEON_LD1x3_UPD;
4106 case Intrinsic::aarch64_neon_vld1x4: NewOpc = AArch64ISD::NEON_LD1x4_UPD;
4108 case Intrinsic::aarch64_neon_vst1x2: NewOpc = AArch64ISD::NEON_ST1x2_UPD;
4109 NumVecs = 2; isLoad = false; break;
4110 case Intrinsic::aarch64_neon_vst1x3: NewOpc = AArch64ISD::NEON_ST1x3_UPD;
4111 NumVecs = 3; isLoad = false; break;
4112 case Intrinsic::aarch64_neon_vst1x4: NewOpc = AArch64ISD::NEON_ST1x4_UPD;
4113 NumVecs = 4; isLoad = false; break;
4114 case Intrinsic::arm_neon_vld2lane: NewOpc = AArch64ISD::NEON_LD2LN_UPD;
4115 NumVecs = 2; isLaneOp = true; break;
4116 case Intrinsic::arm_neon_vld3lane: NewOpc = AArch64ISD::NEON_LD3LN_UPD;
4117 NumVecs = 3; isLaneOp = true; break;
4118 case Intrinsic::arm_neon_vld4lane: NewOpc = AArch64ISD::NEON_LD4LN_UPD;
4119 NumVecs = 4; isLaneOp = true; break;
4120 case Intrinsic::arm_neon_vst2lane: NewOpc = AArch64ISD::NEON_ST2LN_UPD;
4121 NumVecs = 2; isLoad = false; isLaneOp = true; break;
4122 case Intrinsic::arm_neon_vst3lane: NewOpc = AArch64ISD::NEON_ST3LN_UPD;
4123 NumVecs = 3; isLoad = false; isLaneOp = true; break;
4124 case Intrinsic::arm_neon_vst4lane: NewOpc = AArch64ISD::NEON_ST4LN_UPD;
4125 NumVecs = 4; isLoad = false; isLaneOp = true; break;
4129 switch (N->getOpcode()) {
4130 default: llvm_unreachable("unexpected opcode for Neon base update");
4131 case AArch64ISD::NEON_LD2DUP: NewOpc = AArch64ISD::NEON_LD2DUP_UPD;
4133 case AArch64ISD::NEON_LD3DUP: NewOpc = AArch64ISD::NEON_LD3DUP_UPD;
4135 case AArch64ISD::NEON_LD4DUP: NewOpc = AArch64ISD::NEON_LD4DUP_UPD;
4140 // Find the size of memory referenced by the load/store.
4143 VecTy = N->getValueType(0);
4145 VecTy = N->getOperand(AddrOpIdx + 1).getValueType();
4146 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
4148 NumBytes /= VecTy.getVectorNumElements();
4150 // If the increment is a constant, it must match the memory ref size.
4151 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
4152 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
4153 uint32_t IncVal = CInc->getZExtValue();
4154 if (IncVal != NumBytes)
4156 Inc = DAG.getTargetConstant(IncVal, MVT::i32);
4159 // Create the new updating load/store node.
4161 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
4163 for (n = 0; n < NumResultVecs; ++n)
4165 Tys[n++] = MVT::i64;
4166 Tys[n] = MVT::Other;
4167 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs + 2);
4168 SmallVector<SDValue, 8> Ops;
4169 Ops.push_back(N->getOperand(0)); // incoming chain
4170 Ops.push_back(N->getOperand(AddrOpIdx));
4172 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
4173 Ops.push_back(N->getOperand(i));
4175 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
4176 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
4177 Ops.data(), Ops.size(),
4178 MemInt->getMemoryVT(),
4179 MemInt->getMemOperand());
4182 std::vector<SDValue> NewResults;
4183 for (unsigned i = 0; i < NumResultVecs; ++i) {
4184 NewResults.push_back(SDValue(UpdN.getNode(), i));
4186 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1)); // chain
4187 DCI.CombineTo(N, NewResults);
4188 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
4195 /// For a VDUPLANE node N, check if its source operand is a vldN-lane (N > 1)
4196 /// intrinsic, and if all the other uses of that intrinsic are also VDUPLANEs.
4197 /// If so, combine them to a vldN-dup operation and return true.
4198 static SDValue CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
4199 SelectionDAG &DAG = DCI.DAG;
4200 EVT VT = N->getValueType(0);
4202 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
4203 SDNode *VLD = N->getOperand(0).getNode();
4204 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
4206 unsigned NumVecs = 0;
4207 unsigned NewOpc = 0;
4208 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
4209 if (IntNo == Intrinsic::arm_neon_vld2lane) {
4211 NewOpc = AArch64ISD::NEON_LD2DUP;
4212 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
4214 NewOpc = AArch64ISD::NEON_LD3DUP;
4215 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
4217 NewOpc = AArch64ISD::NEON_LD4DUP;
4222 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
4223 // numbers match the load.
4224 unsigned VLDLaneNo =
4225 cast<ConstantSDNode>(VLD->getOperand(NumVecs + 3))->getZExtValue();
4226 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4228 // Ignore uses of the chain result.
4229 if (UI.getUse().getResNo() == NumVecs)
4232 if (User->getOpcode() != AArch64ISD::NEON_VDUPLANE ||
4233 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
4237 // Create the vldN-dup node.
4240 for (n = 0; n < NumVecs; ++n)
4242 Tys[n] = MVT::Other;
4243 SDVTList SDTys = DAG.getVTList(Tys, NumVecs + 1);
4244 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
4245 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
4246 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, Ops, 2,
4247 VLDMemInt->getMemoryVT(),
4248 VLDMemInt->getMemOperand());
4251 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4253 unsigned ResNo = UI.getUse().getResNo();
4254 // Ignore uses of the chain result.
4255 if (ResNo == NumVecs)
4258 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
4261 // Now the vldN-lane intrinsic is dead except for its chain result.
4262 // Update uses of the chain.
4263 std::vector<SDValue> VLDDupResults;
4264 for (unsigned n = 0; n < NumVecs; ++n)
4265 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
4266 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
4267 DCI.CombineTo(VLD, VLDDupResults);
4269 return SDValue(N, 0);
4273 // v1i1 (bitcast (i1 setcc (extract_vector_elt, extract_vector_elt))
4274 // FIXME: Currently the type legalizer can't handle SETCC having v1i1 as result.
4275 // If it can legalize "v1i1 SETCC" correctly, no need to combine such SETCC.
4276 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
4277 EVT ResVT = N->getValueType(0);
4279 if (!ResVT.isVector() || ResVT.getVectorNumElements() != 1 ||
4280 ResVT.getVectorElementType() != MVT::i1)
4283 SDValue LHS = N->getOperand(0);
4284 SDValue RHS = N->getOperand(1);
4285 EVT CmpVT = LHS.getValueType();
4286 LHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
4287 CmpVT.getVectorElementType(), LHS,
4288 DAG.getConstant(0, MVT::i64));
4289 RHS = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N),
4290 CmpVT.getVectorElementType(), RHS,
4291 DAG.getConstant(0, MVT::i64));
4293 DAG.getSetCC(SDLoc(N), MVT::i1, LHS, RHS,
4294 cast<CondCodeSDNode>(N->getOperand(2))->get());
4295 return DAG.getNode(ISD::BITCAST, SDLoc(N), ResVT, SetCC);
4298 // vselect (v1i1 setcc) ->
4299 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
4300 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
4301 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
4303 static SDValue PerformVSelectCombine(SDNode *N, SelectionDAG &DAG) {
4304 SDValue N0 = N->getOperand(0);
4305 EVT CCVT = N0.getValueType();
4307 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
4308 CCVT.getVectorElementType() != MVT::i1)
4311 EVT ResVT = N->getValueType(0);
4312 EVT CmpVT = N0.getOperand(0).getValueType();
4313 // Only combine when the result type is of the same size as the compared
4315 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
4318 SDValue IfTrue = N->getOperand(1);
4319 SDValue IfFalse = N->getOperand(2);
4321 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
4322 N0.getOperand(0), N0.getOperand(1),
4323 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4324 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
4328 // sign_extend (extract_vector_elt (v1i1 setcc)) ->
4329 // extract_vector_elt (v1iXX setcc)
4330 // (XX is the size of the compared operand type)
4331 static SDValue PerformSignExtendCombine(SDNode *N, SelectionDAG &DAG) {
4332 SDValue N0 = N->getOperand(0);
4333 SDValue Vec = N0.getOperand(0);
4335 if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4336 Vec.getOpcode() != ISD::SETCC)
4339 EVT ResVT = N->getValueType(0);
4340 EVT CmpVT = Vec.getOperand(0).getValueType();
4341 // Only optimize when the result type is of the same size as the element
4342 // type of the compared operand.
4343 if (ResVT.getSizeInBits() != CmpVT.getVectorElementType().getSizeInBits())
4346 SDValue Lane = N0.getOperand(1);
4348 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
4349 Vec.getOperand(0), Vec.getOperand(1),
4350 cast<CondCodeSDNode>(Vec.getOperand(2))->get());
4351 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), ResVT,
4356 AArch64TargetLowering::PerformDAGCombine(SDNode *N,
4357 DAGCombinerInfo &DCI) const {
4358 switch (N->getOpcode()) {
4360 case ISD::AND: return PerformANDCombine(N, DCI);
4361 case ISD::OR: return PerformORCombine(N, DCI, getSubtarget());
4365 return PerformShiftCombine(N, DCI, getSubtarget());
4366 case ISD::SETCC: return PerformSETCCCombine(N, DCI.DAG);
4367 case ISD::VSELECT: return PerformVSelectCombine(N, DCI.DAG);
4368 case ISD::SIGN_EXTEND: return PerformSignExtendCombine(N, DCI.DAG);
4369 case ISD::INTRINSIC_WO_CHAIN:
4370 return PerformIntrinsicCombine(N, DCI.DAG);
4371 case AArch64ISD::NEON_VDUPLANE:
4372 return CombineVLDDUP(N, DCI);
4373 case AArch64ISD::NEON_LD2DUP:
4374 case AArch64ISD::NEON_LD3DUP:
4375 case AArch64ISD::NEON_LD4DUP:
4376 return CombineBaseUpdate(N, DCI);
4377 case ISD::INTRINSIC_VOID:
4378 case ISD::INTRINSIC_W_CHAIN:
4379 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
4380 case Intrinsic::arm_neon_vld1:
4381 case Intrinsic::arm_neon_vld2:
4382 case Intrinsic::arm_neon_vld3:
4383 case Intrinsic::arm_neon_vld4:
4384 case Intrinsic::arm_neon_vst1:
4385 case Intrinsic::arm_neon_vst2:
4386 case Intrinsic::arm_neon_vst3:
4387 case Intrinsic::arm_neon_vst4:
4388 case Intrinsic::arm_neon_vld2lane:
4389 case Intrinsic::arm_neon_vld3lane:
4390 case Intrinsic::arm_neon_vld4lane:
4391 case Intrinsic::aarch64_neon_vld1x2:
4392 case Intrinsic::aarch64_neon_vld1x3:
4393 case Intrinsic::aarch64_neon_vld1x4:
4394 case Intrinsic::aarch64_neon_vst1x2:
4395 case Intrinsic::aarch64_neon_vst1x3:
4396 case Intrinsic::aarch64_neon_vst1x4:
4397 case Intrinsic::arm_neon_vst2lane:
4398 case Intrinsic::arm_neon_vst3lane:
4399 case Intrinsic::arm_neon_vst4lane:
4400 return CombineBaseUpdate(N, DCI);
4409 AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
4410 VT = VT.getScalarType();
4415 switch (VT.getSimpleVT().SimpleTy) {
4428 // Check whether a shuffle_vector could be presented as concat_vector.
4429 bool AArch64TargetLowering::isConcatVector(SDValue Op, SelectionDAG &DAG,
4430 SDValue V0, SDValue V1,
4432 SDValue &Res) const {
4434 EVT VT = Op.getValueType();
4435 if (VT.getSizeInBits() != 128)
4437 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4438 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4441 unsigned NumElts = VT.getVectorNumElements();
4442 bool isContactVector = true;
4443 bool splitV0 = false;
4444 if (V0.getValueType().getSizeInBits() == 128)
4447 for (int I = 0, E = NumElts / 2; I != E; I++) {
4449 isContactVector = false;
4454 if (isContactVector) {
4455 int offset = NumElts / 2;
4456 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4457 if (Mask[I] != I + splitV0 * offset) {
4458 isContactVector = false;
4464 if (isContactVector) {
4465 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4468 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4469 DAG.getConstant(0, MVT::i64));
4471 if (V1.getValueType().getSizeInBits() == 128) {
4472 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4473 DAG.getConstant(0, MVT::i64));
4475 Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4481 // Check whether a Build Vector could be presented as Shuffle Vector.
4482 // This Shuffle Vector maybe not legalized, so the length of its operand and
4483 // the length of result may not equal.
4484 bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
4485 SDValue &V0, SDValue &V1,
4488 EVT VT = Op.getValueType();
4489 unsigned NumElts = VT.getVectorNumElements();
4490 unsigned V0NumElts = 0;
4492 // Check if all elements are extracted from less than 3 vectors.
4493 for (unsigned i = 0; i < NumElts; ++i) {
4494 SDValue Elt = Op.getOperand(i);
4495 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4496 Elt.getOperand(0).getValueType().getVectorElementType() !=
4497 VT.getVectorElementType())
4500 if (V0.getNode() == 0) {
4501 V0 = Elt.getOperand(0);
4502 V0NumElts = V0.getValueType().getVectorNumElements();
4504 if (Elt.getOperand(0) == V0) {
4505 Mask[i] = (cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue());
4507 } else if (V1.getNode() == 0) {
4508 V1 = Elt.getOperand(0);
4510 if (Elt.getOperand(0) == V1) {
4511 unsigned Lane = cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue();
4512 Mask[i] = (Lane + V0NumElts);
4521 // If this is a case we can't handle, return null and let the default
4522 // expansion code take care of it.
4524 AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4525 const AArch64Subtarget *ST) const {
4527 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4529 EVT VT = Op.getValueType();
4531 APInt SplatBits, SplatUndef;
4532 unsigned SplatBitSize;
4535 unsigned UseNeonMov = VT.getSizeInBits() >= 64;
4537 // Note we favor lowering MOVI over MVNI.
4538 // This has implications on the definition of patterns in TableGen to select
4539 // BIC immediate instructions but not ORR immediate instructions.
4540 // If this lowering order is changed, TableGen patterns for BIC immediate and
4541 // ORR immediate instructions have to be updated.
4543 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4544 if (SplatBitSize <= 64) {
4545 // First attempt to use vector immediate-form MOVI
4548 unsigned OpCmode = 0;
4550 if (isNeonModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
4551 SplatBitSize, DAG, VT.is128BitVector(),
4552 Neon_Mov_Imm, NeonMovVT, Imm, OpCmode)) {
4553 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4554 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4556 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4557 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MOVIMM, DL, NeonMovVT,
4558 ImmVal, OpCmodeVal);
4559 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4563 // Then attempt to use vector immediate-form MVNI
4564 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4565 if (isNeonModifiedImm(NegatedImm, SplatUndef.getZExtValue(), SplatBitSize,
4566 DAG, VT.is128BitVector(), Neon_Mvn_Imm, NeonMovVT,
4568 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4569 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4570 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4571 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MVNIMM, DL, NeonMovVT,
4572 ImmVal, OpCmodeVal);
4573 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4577 // Attempt to use vector immediate-form FMOV
4578 if (((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) ||
4579 (VT == MVT::v2f64 && SplatBitSize == 64)) {
4581 SplatBitSize == 32 ? APFloat::IEEEsingle : APFloat::IEEEdouble,
4584 if (A64Imms::isFPImm(RealVal, ImmVal)) {
4585 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4586 return DAG.getNode(AArch64ISD::NEON_FMOVIMM, DL, VT, Val);
4592 unsigned NumElts = VT.getVectorNumElements();
4593 bool isOnlyLowElement = true;
4594 bool usesOnlyOneValue = true;
4595 bool hasDominantValue = false;
4596 bool isConstant = true;
4598 // Map of the number of times a particular SDValue appears in the
4600 DenseMap<SDValue, unsigned> ValueCounts;
4602 for (unsigned i = 0; i < NumElts; ++i) {
4603 SDValue V = Op.getOperand(i);
4604 if (V.getOpcode() == ISD::UNDEF)
4607 isOnlyLowElement = false;
4608 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4611 ValueCounts.insert(std::make_pair(V, 0));
4612 unsigned &Count = ValueCounts[V];
4614 // Is this value dominant? (takes up more than half of the lanes)
4615 if (++Count > (NumElts / 2)) {
4616 hasDominantValue = true;
4620 if (ValueCounts.size() != 1)
4621 usesOnlyOneValue = false;
4622 if (!Value.getNode() && ValueCounts.size() > 0)
4623 Value = ValueCounts.begin()->first;
4625 if (ValueCounts.size() == 0)
4626 return DAG.getUNDEF(VT);
4628 if (isOnlyLowElement)
4629 return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
4631 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4632 if (hasDominantValue && EltSize <= 64) {
4633 // Use VDUP for non-constant splats.
4637 // If we are DUPing a value that comes directly from a vector, we could
4638 // just use DUPLANE. We can only do this if the lane being extracted
4639 // is at a constant index, as the DUP from lane instructions only have
4640 // constant-index forms.
4642 // If there is a TRUNCATE between EXTRACT_VECTOR_ELT and DUP, we can
4643 // remove TRUNCATE for DUPLANE by apdating the source vector to
4644 // appropriate vector type and lane index.
4646 // FIXME: for now we have v1i8, v1i16, v1i32 legal vector types, if they
4647 // are not legal any more, no need to check the type size in bits should
4648 // be large than 64.
4650 if (Value->getOpcode() == ISD::TRUNCATE)
4651 V = Value->getOperand(0);
4652 if (V->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4653 isa<ConstantSDNode>(V->getOperand(1)) &&
4654 V->getOperand(0).getValueType().getSizeInBits() >= 64) {
4656 // If the element size of source vector is larger than DUPLANE
4657 // element size, we can do transformation by,
4658 // 1) bitcasting source register to smaller element vector
4659 // 2) mutiplying the lane index by SrcEltSize/ResEltSize
4660 // For example, we can lower
4661 // "v8i16 vdup_lane(v4i32, 1)"
4663 // "v8i16 vdup_lane(v8i16 bitcast(v4i32), 2)".
4664 SDValue SrcVec = V->getOperand(0);
4665 unsigned SrcEltSize =
4666 SrcVec.getValueType().getVectorElementType().getSizeInBits();
4667 unsigned ResEltSize = VT.getVectorElementType().getSizeInBits();
4668 if (SrcEltSize > ResEltSize) {
4669 assert((SrcEltSize % ResEltSize == 0) && "Invalid element size");
4671 unsigned SrcSize = SrcVec.getValueType().getSizeInBits();
4672 unsigned ResSize = VT.getSizeInBits();
4674 if (SrcSize > ResSize) {
4675 assert((SrcSize % ResSize == 0) && "Invalid vector size");
4677 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4678 SrcSize / ResEltSize);
4679 BitCast = DAG.getNode(ISD::BITCAST, DL, CastVT, SrcVec);
4681 assert((SrcSize == ResSize) && "Invalid vector size of source vec");
4682 BitCast = DAG.getNode(ISD::BITCAST, DL, VT, SrcVec);
4685 unsigned LaneIdx = V->getConstantOperandVal(1);
4687 DAG.getConstant((SrcEltSize / ResEltSize) * LaneIdx, MVT::i64);
4688 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, BitCast, Lane);
4690 assert((SrcEltSize == ResEltSize) &&
4691 "Invalid element size of source vec");
4692 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, V->getOperand(0),
4696 N = DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4698 if (!usesOnlyOneValue) {
4699 // The dominant value was splatted as 'N', but we now have to insert
4700 // all differing elements.
4701 for (unsigned I = 0; I < NumElts; ++I) {
4702 if (Op.getOperand(I) == Value)
4704 SmallVector<SDValue, 3> Ops;
4706 Ops.push_back(Op.getOperand(I));
4707 Ops.push_back(DAG.getConstant(I, MVT::i64));
4708 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, &Ops[0], 3);
4713 if (usesOnlyOneValue && isConstant) {
4714 return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4717 // If all elements are constants and the case above didn't get hit, fall back
4718 // to the default expansion, which will generate a load from the constant
4723 // Try to lower this in lowering ShuffleVector way.
4726 if (isKnownShuffleVector(Op, DAG, V0, V1, Mask)) {
4727 unsigned V0NumElts = V0.getValueType().getVectorNumElements();
4728 if (!V1.getNode() && V0NumElts == NumElts * 2) {
4729 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4730 DAG.getConstant(NumElts, MVT::i64));
4731 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4732 DAG.getConstant(0, MVT::i64));
4733 V0NumElts = V0.getValueType().getVectorNumElements();
4736 if (V1.getNode() && NumElts == V0NumElts &&
4737 V0NumElts == V1.getValueType().getVectorNumElements()) {
4738 SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
4739 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
4742 return LowerVECTOR_SHUFFLE(Shuffle, DAG);
4745 if (isConcatVector(Op, DAG, V0, V1, Mask, Res))
4750 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4751 // know the default expansion would otherwise fall back on something even
4752 // worse. For a vector with one or two non-undef values, that's
4753 // scalar_to_vector for the elements followed by a shuffle (provided the
4754 // shuffle is valid for the target) and materialization element by element
4755 // on the stack followed by a load for everything else.
4756 if (!isConstant && !usesOnlyOneValue) {
4757 SDValue Vec = DAG.getUNDEF(VT);
4758 for (unsigned i = 0 ; i < NumElts; ++i) {
4759 SDValue V = Op.getOperand(i);
4760 if (V.getOpcode() == ISD::UNDEF)
4762 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
4763 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V, LaneIdx);
4770 /// isREVMask - Check if a vector shuffle corresponds to a REV
4771 /// instruction with the specified blocksize. (The order of the elements
4772 /// within each block of the vector is reversed.)
4773 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4774 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4775 "Only possible block sizes for REV are: 16, 32, 64");
4777 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4781 unsigned NumElts = VT.getVectorNumElements();
4782 unsigned BlockElts = M[0] + 1;
4783 // If the first shuffle index is UNDEF, be optimistic.
4785 BlockElts = BlockSize / EltSz;
4787 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4790 for (unsigned i = 0; i < NumElts; ++i) {
4792 continue; // ignore UNDEF indices
4793 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4800 // isPermuteMask - Check whether the vector shuffle matches to UZP, ZIP and
4802 static unsigned isPermuteMask(ArrayRef<int> M, EVT VT, bool isV2undef) {
4803 unsigned NumElts = VT.getVectorNumElements();
4807 bool ismatch = true;
4810 for (unsigned i = 0; i < NumElts; ++i) {
4811 unsigned answer = i * 2;
4812 if (isV2undef && answer >= NumElts)
4814 if (M[i] != -1 && (unsigned)M[i] != answer) {
4820 return AArch64ISD::NEON_UZP1;
4824 for (unsigned i = 0; i < NumElts; ++i) {
4825 unsigned answer = i * 2 + 1;
4826 if (isV2undef && answer >= NumElts)
4828 if (M[i] != -1 && (unsigned)M[i] != answer) {
4834 return AArch64ISD::NEON_UZP2;
4838 for (unsigned i = 0; i < NumElts; ++i) {
4839 unsigned answer = i / 2 + NumElts * (i % 2);
4840 if (isV2undef && answer >= NumElts)
4842 if (M[i] != -1 && (unsigned)M[i] != answer) {
4848 return AArch64ISD::NEON_ZIP1;
4852 for (unsigned i = 0; i < NumElts; ++i) {
4853 unsigned answer = (NumElts + i) / 2 + NumElts * (i % 2);
4854 if (isV2undef && answer >= NumElts)
4856 if (M[i] != -1 && (unsigned)M[i] != answer) {
4862 return AArch64ISD::NEON_ZIP2;
4866 for (unsigned i = 0; i < NumElts; ++i) {
4867 unsigned answer = i + (NumElts - 1) * (i % 2);
4868 if (isV2undef && answer >= NumElts)
4870 if (M[i] != -1 && (unsigned)M[i] != answer) {
4876 return AArch64ISD::NEON_TRN1;
4880 for (unsigned i = 0; i < NumElts; ++i) {
4881 unsigned answer = 1 + i + (NumElts - 1) * (i % 2);
4882 if (isV2undef && answer >= NumElts)
4884 if (M[i] != -1 && (unsigned)M[i] != answer) {
4890 return AArch64ISD::NEON_TRN2;
4896 AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4897 SelectionDAG &DAG) const {
4898 SDValue V1 = Op.getOperand(0);
4899 SDValue V2 = Op.getOperand(1);
4901 EVT VT = Op.getValueType();
4902 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4904 // Convert shuffles that are directly supported on NEON to target-specific
4905 // DAG nodes, instead of keeping them as shuffles and matching them again
4906 // during code selection. This is more efficient and avoids the possibility
4907 // of inconsistencies between legalization and selection.
4908 ArrayRef<int> ShuffleMask = SVN->getMask();
4910 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4914 if (isREVMask(ShuffleMask, VT, 64))
4915 return DAG.getNode(AArch64ISD::NEON_REV64, dl, VT, V1);
4916 if (isREVMask(ShuffleMask, VT, 32))
4917 return DAG.getNode(AArch64ISD::NEON_REV32, dl, VT, V1);
4918 if (isREVMask(ShuffleMask, VT, 16))
4919 return DAG.getNode(AArch64ISD::NEON_REV16, dl, VT, V1);
4922 if (V2.getOpcode() == ISD::UNDEF)
4923 ISDNo = isPermuteMask(ShuffleMask, VT, true);
4925 ISDNo = isPermuteMask(ShuffleMask, VT, false);
4928 if (V2.getOpcode() == ISD::UNDEF)
4929 return DAG.getNode(ISDNo, dl, VT, V1, V1);
4931 return DAG.getNode(ISDNo, dl, VT, V1, V2);
4935 if (isConcatVector(Op, DAG, V1, V2, &ShuffleMask[0], Res))
4938 // If the element of shuffle mask are all the same constant, we can
4939 // transform it into either NEON_VDUP or NEON_VDUPLANE
4940 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4941 int Lane = SVN->getSplatIndex();
4942 // If this is undef splat, generate it via "just" vdup, if possible.
4943 if (Lane == -1) Lane = 0;
4945 // Test if V1 is a SCALAR_TO_VECTOR.
4946 if (V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4947 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT, V1.getOperand(0));
4949 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR.
4950 if (V1.getOpcode() == ISD::BUILD_VECTOR) {
4951 bool IsScalarToVector = true;
4952 for (unsigned i = 0, e = V1.getNumOperands(); i != e; ++i)
4953 if (V1.getOperand(i).getOpcode() != ISD::UNDEF &&
4954 i != (unsigned)Lane) {
4955 IsScalarToVector = false;
4958 if (IsScalarToVector)
4959 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT,
4960 V1.getOperand(Lane));
4963 // Test if V1 is a EXTRACT_SUBVECTOR.
4964 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4965 int ExtLane = cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4966 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1.getOperand(0),
4967 DAG.getConstant(Lane + ExtLane, MVT::i64));
4969 // Test if V1 is a CONCAT_VECTORS.
4970 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
4971 V1.getOperand(1).getOpcode() == ISD::UNDEF) {
4972 SDValue Op0 = V1.getOperand(0);
4973 assert((unsigned)Lane < Op0.getValueType().getVectorNumElements() &&
4974 "Invalid vector lane access");
4975 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, Op0,
4976 DAG.getConstant(Lane, MVT::i64));
4979 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1,
4980 DAG.getConstant(Lane, MVT::i64));
4983 int Length = ShuffleMask.size();
4984 int V1EltNum = V1.getValueType().getVectorNumElements();
4986 // If the number of v1 elements is the same as the number of shuffle mask
4987 // element and the shuffle masks are sequential values, we can transform
4988 // it into NEON_VEXTRACT.
4989 if (V1EltNum == Length) {
4990 // Check if the shuffle mask is sequential.
4992 while (ShuffleMask[SkipUndef] == -1) {
4995 int CurMask = ShuffleMask[SkipUndef];
4996 if (CurMask >= SkipUndef) {
4997 bool IsSequential = true;
4998 for (int I = SkipUndef; I < Length; ++I) {
4999 if (ShuffleMask[I] != -1 && ShuffleMask[I] != CurMask) {
5000 IsSequential = false;
5006 assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
5007 unsigned VecSize = EltSize * V1EltNum;
5008 unsigned Index = (EltSize / 8) * (ShuffleMask[SkipUndef] - SkipUndef);
5009 if (VecSize == 64 || VecSize == 128)
5010 return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
5011 DAG.getConstant(Index, MVT::i64));
5016 // For shuffle mask like "0, 1, 2, 3, 4, 5, 13, 7", try to generate insert
5017 // by element from V2 to V1 .
5018 // If shuffle mask is like "0, 1, 10, 11, 12, 13, 14, 15", V2 would be a
5019 // better choice to be inserted than V1 as less insert needed, so we count
5020 // element to be inserted for both V1 and V2, and select less one as insert
5023 // Collect elements need to be inserted and their index.
5024 SmallVector<int, 8> NV1Elt;
5025 SmallVector<int, 8> N1Index;
5026 SmallVector<int, 8> NV2Elt;
5027 SmallVector<int, 8> N2Index;
5028 for (int I = 0; I != Length; ++I) {
5029 if (ShuffleMask[I] != I) {
5030 NV1Elt.push_back(ShuffleMask[I]);
5031 N1Index.push_back(I);
5034 for (int I = 0; I != Length; ++I) {
5035 if (ShuffleMask[I] != (I + V1EltNum)) {
5036 NV2Elt.push_back(ShuffleMask[I]);
5037 N2Index.push_back(I);
5041 // Decide which to be inserted. If all lanes mismatch, neither V1 nor V2
5042 // will be inserted.
5044 SmallVector<int, 8> InsMasks = NV1Elt;
5045 SmallVector<int, 8> InsIndex = N1Index;
5046 if ((int)NV1Elt.size() != Length || (int)NV2Elt.size() != Length) {
5047 if (NV1Elt.size() > NV2Elt.size()) {
5053 InsV = DAG.getNode(ISD::UNDEF, dl, VT);
5056 for (int I = 0, E = InsMasks.size(); I != E; ++I) {
5058 int Mask = InsMasks[I];
5059 if (Mask >= V1EltNum) {
5063 // Any value type smaller than i32 is illegal in AArch64, and this lower
5064 // function is called after legalize pass, so we need to legalize
5067 if (VT.getVectorElementType().isFloatingPoint())
5068 EltVT = (EltSize == 64) ? MVT::f64 : MVT::f32;
5070 EltVT = (EltSize == 64) ? MVT::i64 : MVT::i32;
5073 ExtV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ExtV,
5074 DAG.getConstant(Mask, MVT::i64));
5075 InsV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, InsV, ExtV,
5076 DAG.getConstant(InsIndex[I], MVT::i64));
5082 AArch64TargetLowering::ConstraintType
5083 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
5084 if (Constraint.size() == 1) {
5085 switch (Constraint[0]) {
5087 case 'w': // An FP/SIMD vector register
5088 return C_RegisterClass;
5089 case 'I': // Constant that can be used with an ADD instruction
5090 case 'J': // Constant that can be used with a SUB instruction
5091 case 'K': // Constant that can be used with a 32-bit logical instruction
5092 case 'L': // Constant that can be used with a 64-bit logical instruction
5093 case 'M': // Constant that can be used as a 32-bit MOV immediate
5094 case 'N': // Constant that can be used as a 64-bit MOV immediate
5095 case 'Y': // Floating point constant zero
5096 case 'Z': // Integer constant zero
5098 case 'Q': // A memory reference with base register and no offset
5100 case 'S': // A symbolic address
5105 // FIXME: Ump, Utf, Usa, Ush
5106 // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes,
5107 // whatever they may be
5108 // Utf: A memory address suitable for ldp/stp in TF mode, whatever it may be
5109 // Usa: An absolute symbolic address
5110 // Ush: The high part (bits 32:12) of a pc-relative symbolic address
5111 assert(Constraint != "Ump" && Constraint != "Utf" && Constraint != "Usa"
5112 && Constraint != "Ush" && "Unimplemented constraints");
5114 return TargetLowering::getConstraintType(Constraint);
5117 TargetLowering::ConstraintWeight
5118 AArch64TargetLowering::getSingleConstraintMatchWeight(AsmOperandInfo &Info,
5119 const char *Constraint) const {
5121 llvm_unreachable("Constraint weight unimplemented");
5125 AArch64TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5126 std::string &Constraint,
5127 std::vector<SDValue> &Ops,
5128 SelectionDAG &DAG) const {
5129 SDValue Result(0, 0);
5131 // Only length 1 constraints are C_Other.
5132 if (Constraint.size() != 1) return;
5134 // Only C_Other constraints get lowered like this. That means constants for us
5135 // so return early if there's no hope the constraint can be lowered.
5137 switch(Constraint[0]) {
5139 case 'I': case 'J': case 'K': case 'L':
5140 case 'M': case 'N': case 'Z': {
5141 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5145 uint64_t CVal = C->getZExtValue();
5148 switch (Constraint[0]) {
5150 // FIXME: 'M' and 'N' are MOV pseudo-insts -- unsupported in assembly. 'J'
5151 // is a peculiarly useless SUB constraint.
5152 llvm_unreachable("Unimplemented C_Other constraint");
5158 if (A64Imms::isLogicalImm(32, CVal, Bits))
5162 if (A64Imms::isLogicalImm(64, CVal, Bits))
5171 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5175 // An absolute symbolic address or label reference.
5176 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5177 Result = DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
5178 GA->getValueType(0));
5179 } else if (const BlockAddressSDNode *BA
5180 = dyn_cast<BlockAddressSDNode>(Op)) {
5181 Result = DAG.getTargetBlockAddress(BA->getBlockAddress(),
5182 BA->getValueType(0));
5183 } else if (const ExternalSymbolSDNode *ES
5184 = dyn_cast<ExternalSymbolSDNode>(Op)) {
5185 Result = DAG.getTargetExternalSymbol(ES->getSymbol(),
5186 ES->getValueType(0));
5192 if (const ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) {
5193 if (CFP->isExactlyValue(0.0)) {
5194 Result = DAG.getTargetConstantFP(0.0, CFP->getValueType(0));
5201 if (Result.getNode()) {
5202 Ops.push_back(Result);
5206 // It's an unknown constraint for us. Let generic code have a go.
5207 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5210 std::pair<unsigned, const TargetRegisterClass*>
5211 AArch64TargetLowering::getRegForInlineAsmConstraint(
5212 const std::string &Constraint,
5214 if (Constraint.size() == 1) {
5215 switch (Constraint[0]) {
5217 if (VT.getSizeInBits() <= 32)
5218 return std::make_pair(0U, &AArch64::GPR32RegClass);
5219 else if (VT == MVT::i64)
5220 return std::make_pair(0U, &AArch64::GPR64RegClass);
5224 return std::make_pair(0U, &AArch64::FPR16RegClass);
5225 else if (VT == MVT::f32)
5226 return std::make_pair(0U, &AArch64::FPR32RegClass);
5227 else if (VT.getSizeInBits() == 64)
5228 return std::make_pair(0U, &AArch64::FPR64RegClass);
5229 else if (VT.getSizeInBits() == 128)
5230 return std::make_pair(0U, &AArch64::FPR128RegClass);
5235 // Use the default implementation in TargetLowering to convert the register
5236 // constraint into a member of a register class.
5237 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5240 /// Represent NEON load and store intrinsics as MemIntrinsicNodes.
5241 /// The associated MachineMemOperands record the alignment specified
5242 /// in the intrinsic calls.
5243 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5245 unsigned Intrinsic) const {
5246 switch (Intrinsic) {
5247 case Intrinsic::arm_neon_vld1:
5248 case Intrinsic::arm_neon_vld2:
5249 case Intrinsic::arm_neon_vld3:
5250 case Intrinsic::arm_neon_vld4:
5251 case Intrinsic::aarch64_neon_vld1x2:
5252 case Intrinsic::aarch64_neon_vld1x3:
5253 case Intrinsic::aarch64_neon_vld1x4:
5254 case Intrinsic::arm_neon_vld2lane:
5255 case Intrinsic::arm_neon_vld3lane:
5256 case Intrinsic::arm_neon_vld4lane: {
5257 Info.opc = ISD::INTRINSIC_W_CHAIN;
5258 // Conservatively set memVT to the entire set of vectors loaded.
5259 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5260 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5261 Info.ptrVal = I.getArgOperand(0);
5263 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5264 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5265 Info.vol = false; // volatile loads with NEON intrinsics not supported
5266 Info.readMem = true;
5267 Info.writeMem = false;
5270 case Intrinsic::arm_neon_vst1:
5271 case Intrinsic::arm_neon_vst2:
5272 case Intrinsic::arm_neon_vst3:
5273 case Intrinsic::arm_neon_vst4:
5274 case Intrinsic::aarch64_neon_vst1x2:
5275 case Intrinsic::aarch64_neon_vst1x3:
5276 case Intrinsic::aarch64_neon_vst1x4:
5277 case Intrinsic::arm_neon_vst2lane:
5278 case Intrinsic::arm_neon_vst3lane:
5279 case Intrinsic::arm_neon_vst4lane: {
5280 Info.opc = ISD::INTRINSIC_VOID;
5281 // Conservatively set memVT to the entire set of vectors stored.
5282 unsigned NumElts = 0;
5283 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5284 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5285 if (!ArgTy->isVectorTy())
5287 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5289 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5290 Info.ptrVal = I.getArgOperand(0);
5292 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5293 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5294 Info.vol = false; // volatile stores with NEON intrinsics not supported
5295 Info.readMem = false;
5296 Info.writeMem = true;