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"
32 static TargetLoweringObjectFile *createTLOF(AArch64TargetMachine &TM) {
33 assert (TM.getSubtarget<AArch64Subtarget>().isTargetELF() &&
34 "unknown subtarget type");
35 return new AArch64ElfTargetObjectFile();
38 AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
39 : TargetLowering(TM, createTLOF(TM)), Itins(TM.getInstrItineraryData()) {
41 const AArch64Subtarget *Subtarget = &TM.getSubtarget<AArch64Subtarget>();
43 // SIMD compares set the entire lane's bits to 1
44 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
46 // Scalar register <-> type mapping
47 addRegisterClass(MVT::i32, &AArch64::GPR32RegClass);
48 addRegisterClass(MVT::i64, &AArch64::GPR64RegClass);
50 if (Subtarget->hasFPARMv8()) {
51 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
52 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
53 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
54 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
57 if (Subtarget->hasNEON()) {
59 addRegisterClass(MVT::v1i8, &AArch64::FPR8RegClass);
60 addRegisterClass(MVT::v1i16, &AArch64::FPR16RegClass);
61 addRegisterClass(MVT::v1i32, &AArch64::FPR32RegClass);
62 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
63 addRegisterClass(MVT::v1f64, &AArch64::FPR64RegClass);
64 addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
65 addRegisterClass(MVT::v4i16, &AArch64::FPR64RegClass);
66 addRegisterClass(MVT::v2i32, &AArch64::FPR64RegClass);
67 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
68 addRegisterClass(MVT::v2f32, &AArch64::FPR64RegClass);
69 addRegisterClass(MVT::v16i8, &AArch64::FPR128RegClass);
70 addRegisterClass(MVT::v8i16, &AArch64::FPR128RegClass);
71 addRegisterClass(MVT::v4i32, &AArch64::FPR128RegClass);
72 addRegisterClass(MVT::v2i64, &AArch64::FPR128RegClass);
73 addRegisterClass(MVT::v4f32, &AArch64::FPR128RegClass);
74 addRegisterClass(MVT::v2f64, &AArch64::FPR128RegClass);
77 computeRegisterProperties();
79 // We combine OR nodes for bitfield and NEON BSL operations.
80 setTargetDAGCombine(ISD::OR);
82 setTargetDAGCombine(ISD::AND);
83 setTargetDAGCombine(ISD::SRA);
84 setTargetDAGCombine(ISD::SRL);
85 setTargetDAGCombine(ISD::SHL);
87 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
88 setTargetDAGCombine(ISD::INTRINSIC_VOID);
89 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
91 // AArch64 does not have i1 loads, or much of anything for i1 really.
92 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
93 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
94 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
96 setStackPointerRegisterToSaveRestore(AArch64::XSP);
97 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
98 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
99 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
101 // We'll lower globals to wrappers for selection.
102 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
103 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
105 // A64 instructions have the comparison predicate attached to the user of the
106 // result, but having a separate comparison is valuable for matching.
107 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
108 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
109 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
110 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
112 setOperationAction(ISD::SELECT, MVT::i32, Custom);
113 setOperationAction(ISD::SELECT, MVT::i64, Custom);
114 setOperationAction(ISD::SELECT, MVT::f32, Custom);
115 setOperationAction(ISD::SELECT, MVT::f64, Custom);
117 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
118 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
119 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
120 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
122 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
124 setOperationAction(ISD::SETCC, MVT::i32, Custom);
125 setOperationAction(ISD::SETCC, MVT::i64, Custom);
126 setOperationAction(ISD::SETCC, MVT::f32, Custom);
127 setOperationAction(ISD::SETCC, MVT::f64, Custom);
129 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
130 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
131 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
133 setOperationAction(ISD::VASTART, MVT::Other, Custom);
134 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
135 setOperationAction(ISD::VAEND, MVT::Other, Expand);
136 setOperationAction(ISD::VAARG, MVT::Other, Expand);
138 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
139 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
141 setOperationAction(ISD::ROTL, MVT::i32, Expand);
142 setOperationAction(ISD::ROTL, MVT::i64, Expand);
144 setOperationAction(ISD::UREM, MVT::i32, Expand);
145 setOperationAction(ISD::UREM, MVT::i64, Expand);
146 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
147 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
149 setOperationAction(ISD::SREM, MVT::i32, Expand);
150 setOperationAction(ISD::SREM, MVT::i64, Expand);
151 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
152 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
154 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
155 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
156 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
157 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
159 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
160 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
162 // Legal floating-point operations.
163 setOperationAction(ISD::FABS, MVT::f32, Legal);
164 setOperationAction(ISD::FABS, MVT::f64, Legal);
166 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
167 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
169 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
170 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
172 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
173 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
175 setOperationAction(ISD::FNEG, MVT::f32, Legal);
176 setOperationAction(ISD::FNEG, MVT::f64, Legal);
178 setOperationAction(ISD::FRINT, MVT::f32, Legal);
179 setOperationAction(ISD::FRINT, MVT::f64, Legal);
181 setOperationAction(ISD::FSQRT, MVT::f32, Legal);
182 setOperationAction(ISD::FSQRT, MVT::f64, Legal);
184 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
185 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
187 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
188 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
189 setOperationAction(ISD::ConstantFP, MVT::f128, Legal);
191 // Illegal floating-point operations.
192 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
193 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
195 setOperationAction(ISD::FCOS, MVT::f32, Expand);
196 setOperationAction(ISD::FCOS, MVT::f64, Expand);
198 setOperationAction(ISD::FEXP, MVT::f32, Expand);
199 setOperationAction(ISD::FEXP, MVT::f64, Expand);
201 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
202 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
204 setOperationAction(ISD::FLOG, MVT::f32, Expand);
205 setOperationAction(ISD::FLOG, MVT::f64, Expand);
207 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
208 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
210 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
211 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
213 setOperationAction(ISD::FPOW, MVT::f32, Expand);
214 setOperationAction(ISD::FPOW, MVT::f64, Expand);
216 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
217 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
219 setOperationAction(ISD::FREM, MVT::f32, Expand);
220 setOperationAction(ISD::FREM, MVT::f64, Expand);
222 setOperationAction(ISD::FSIN, MVT::f32, Expand);
223 setOperationAction(ISD::FSIN, MVT::f64, Expand);
225 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
226 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
228 // Virtually no operation on f128 is legal, but LLVM can't expand them when
229 // there's a valid register class, so we need custom operations in most cases.
230 setOperationAction(ISD::FABS, MVT::f128, Expand);
231 setOperationAction(ISD::FADD, MVT::f128, Custom);
232 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
233 setOperationAction(ISD::FCOS, MVT::f128, Expand);
234 setOperationAction(ISD::FDIV, MVT::f128, Custom);
235 setOperationAction(ISD::FMA, MVT::f128, Expand);
236 setOperationAction(ISD::FMUL, MVT::f128, Custom);
237 setOperationAction(ISD::FNEG, MVT::f128, Expand);
238 setOperationAction(ISD::FP_EXTEND, MVT::f128, Expand);
239 setOperationAction(ISD::FP_ROUND, MVT::f128, Expand);
240 setOperationAction(ISD::FPOW, MVT::f128, Expand);
241 setOperationAction(ISD::FREM, MVT::f128, Expand);
242 setOperationAction(ISD::FRINT, MVT::f128, Expand);
243 setOperationAction(ISD::FSIN, MVT::f128, Expand);
244 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
245 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
246 setOperationAction(ISD::FSUB, MVT::f128, Custom);
247 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
248 setOperationAction(ISD::SETCC, MVT::f128, Custom);
249 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
250 setOperationAction(ISD::SELECT, MVT::f128, Expand);
251 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
252 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
254 // Lowering for many of the conversions is actually specified by the non-f128
255 // type. The LowerXXX function will be trivial when f128 isn't involved.
256 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
257 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
258 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
259 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
260 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
261 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
262 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
263 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
264 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
265 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
266 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
267 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
268 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
269 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
271 // This prevents LLVM trying to compress double constants into a floating
272 // constant-pool entry and trying to load from there. It's of doubtful benefit
273 // for A64: we'd need LDR followed by FCVT, I believe.
274 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
275 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
276 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
278 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
279 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
280 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
281 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
282 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
283 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
285 setExceptionPointerRegister(AArch64::X0);
286 setExceptionSelectorRegister(AArch64::X1);
288 if (Subtarget->hasNEON()) {
289 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i8, Expand);
290 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Expand);
291 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Expand);
292 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v1i64, Expand);
293 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v16i8, Expand);
294 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i16, Expand);
295 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Expand);
296 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Expand);
298 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i8, Custom);
299 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
300 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
301 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i16, Custom);
302 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
303 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
304 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i32, Custom);
305 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
306 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
307 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
308 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
309 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
310 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
311 setOperationAction(ISD::BUILD_VECTOR, MVT::v1f64, Custom);
312 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
314 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
315 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
316 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
317 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i16, Custom);
318 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
319 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i32, Custom);
320 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
321 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
322 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f32, Custom);
323 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
324 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1f64, Custom);
325 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
327 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i32, Legal);
328 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Legal);
329 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Legal);
330 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Legal);
331 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Legal);
332 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Legal);
333 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Legal);
335 setOperationAction(ISD::SETCC, MVT::v8i8, Custom);
336 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
337 setOperationAction(ISD::SETCC, MVT::v4i16, Custom);
338 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
339 setOperationAction(ISD::SETCC, MVT::v2i32, Custom);
340 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
341 setOperationAction(ISD::SETCC, MVT::v1i64, Custom);
342 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
343 setOperationAction(ISD::SETCC, MVT::v2f32, Custom);
344 setOperationAction(ISD::SETCC, MVT::v4f32, Custom);
345 setOperationAction(ISD::SETCC, MVT::v1f64, Custom);
346 setOperationAction(ISD::SETCC, MVT::v2f64, Custom);
348 setOperationAction(ISD::FFLOOR, MVT::v2f32, Legal);
349 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
350 setOperationAction(ISD::FFLOOR, MVT::v1f64, Legal);
351 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
353 setOperationAction(ISD::FCEIL, MVT::v2f32, Legal);
354 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
355 setOperationAction(ISD::FCEIL, MVT::v1f64, Legal);
356 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
358 setOperationAction(ISD::FTRUNC, MVT::v2f32, Legal);
359 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
360 setOperationAction(ISD::FTRUNC, MVT::v1f64, Legal);
361 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
363 setOperationAction(ISD::FRINT, MVT::v2f32, Legal);
364 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
365 setOperationAction(ISD::FRINT, MVT::v1f64, Legal);
366 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
368 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Legal);
369 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
370 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Legal);
371 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
373 setOperationAction(ISD::FROUND, MVT::v2f32, Legal);
374 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
375 setOperationAction(ISD::FROUND, MVT::v1f64, Legal);
376 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
378 setOperationAction(ISD::SINT_TO_FP, MVT::v1i8, Custom);
379 setOperationAction(ISD::SINT_TO_FP, MVT::v1i16, Custom);
380 setOperationAction(ISD::SINT_TO_FP, MVT::v1i32, Custom);
381 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
382 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
383 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
385 setOperationAction(ISD::UINT_TO_FP, MVT::v1i8, Custom);
386 setOperationAction(ISD::UINT_TO_FP, MVT::v1i16, Custom);
387 setOperationAction(ISD::UINT_TO_FP, MVT::v1i32, Custom);
388 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
389 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
390 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
392 setOperationAction(ISD::FP_TO_SINT, MVT::v1i8, Custom);
393 setOperationAction(ISD::FP_TO_SINT, MVT::v1i16, Custom);
394 setOperationAction(ISD::FP_TO_SINT, MVT::v1i32, Custom);
395 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
396 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
397 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Custom);
399 setOperationAction(ISD::FP_TO_UINT, MVT::v1i8, Custom);
400 setOperationAction(ISD::FP_TO_UINT, MVT::v1i16, Custom);
401 setOperationAction(ISD::FP_TO_UINT, MVT::v1i32, Custom);
402 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
403 setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom);
404 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Custom);
406 // Neon does not support vector divide/remainder operations except
407 // floating-point divide.
408 setOperationAction(ISD::SDIV, MVT::v1i8, Expand);
409 setOperationAction(ISD::SDIV, MVT::v8i8, Expand);
410 setOperationAction(ISD::SDIV, MVT::v16i8, Expand);
411 setOperationAction(ISD::SDIV, MVT::v1i16, Expand);
412 setOperationAction(ISD::SDIV, MVT::v4i16, Expand);
413 setOperationAction(ISD::SDIV, MVT::v8i16, Expand);
414 setOperationAction(ISD::SDIV, MVT::v1i32, Expand);
415 setOperationAction(ISD::SDIV, MVT::v2i32, Expand);
416 setOperationAction(ISD::SDIV, MVT::v4i32, Expand);
417 setOperationAction(ISD::SDIV, MVT::v1i64, Expand);
418 setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
420 setOperationAction(ISD::UDIV, MVT::v1i8, Expand);
421 setOperationAction(ISD::UDIV, MVT::v8i8, Expand);
422 setOperationAction(ISD::UDIV, MVT::v16i8, Expand);
423 setOperationAction(ISD::UDIV, MVT::v1i16, Expand);
424 setOperationAction(ISD::UDIV, MVT::v4i16, Expand);
425 setOperationAction(ISD::UDIV, MVT::v8i16, Expand);
426 setOperationAction(ISD::UDIV, MVT::v1i32, Expand);
427 setOperationAction(ISD::UDIV, MVT::v2i32, Expand);
428 setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
429 setOperationAction(ISD::UDIV, MVT::v1i64, Expand);
430 setOperationAction(ISD::UDIV, MVT::v2i64, Expand);
432 setOperationAction(ISD::SREM, MVT::v1i8, Expand);
433 setOperationAction(ISD::SREM, MVT::v8i8, Expand);
434 setOperationAction(ISD::SREM, MVT::v16i8, Expand);
435 setOperationAction(ISD::SREM, MVT::v1i16, Expand);
436 setOperationAction(ISD::SREM, MVT::v4i16, Expand);
437 setOperationAction(ISD::SREM, MVT::v8i16, Expand);
438 setOperationAction(ISD::SREM, MVT::v1i32, Expand);
439 setOperationAction(ISD::SREM, MVT::v2i32, Expand);
440 setOperationAction(ISD::SREM, MVT::v4i32, Expand);
441 setOperationAction(ISD::SREM, MVT::v1i64, Expand);
442 setOperationAction(ISD::SREM, MVT::v2i64, Expand);
444 setOperationAction(ISD::UREM, MVT::v1i8, Expand);
445 setOperationAction(ISD::UREM, MVT::v8i8, Expand);
446 setOperationAction(ISD::UREM, MVT::v16i8, Expand);
447 setOperationAction(ISD::UREM, MVT::v1i16, Expand);
448 setOperationAction(ISD::UREM, MVT::v4i16, Expand);
449 setOperationAction(ISD::UREM, MVT::v8i16, Expand);
450 setOperationAction(ISD::UREM, MVT::v1i32, Expand);
451 setOperationAction(ISD::UREM, MVT::v2i32, Expand);
452 setOperationAction(ISD::UREM, MVT::v4i32, Expand);
453 setOperationAction(ISD::UREM, MVT::v1i64, Expand);
454 setOperationAction(ISD::UREM, MVT::v2i64, Expand);
456 setOperationAction(ISD::FREM, MVT::v2f32, Expand);
457 setOperationAction(ISD::FREM, MVT::v4f32, Expand);
458 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
459 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
461 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
462 setOperationAction(ISD::SELECT, MVT::v16i8, Expand);
463 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
464 setOperationAction(ISD::SELECT, MVT::v8i16, Expand);
465 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
466 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
467 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
468 setOperationAction(ISD::SELECT, MVT::v2i64, Expand);
469 setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
470 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
471 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
472 setOperationAction(ISD::SELECT, MVT::v2f64, Expand);
474 setOperationAction(ISD::SELECT_CC, MVT::v8i8, Custom);
475 setOperationAction(ISD::SELECT_CC, MVT::v16i8, Custom);
476 setOperationAction(ISD::SELECT_CC, MVT::v4i16, Custom);
477 setOperationAction(ISD::SELECT_CC, MVT::v8i16, Custom);
478 setOperationAction(ISD::SELECT_CC, MVT::v2i32, Custom);
479 setOperationAction(ISD::SELECT_CC, MVT::v4i32, Custom);
480 setOperationAction(ISD::SELECT_CC, MVT::v1i64, Custom);
481 setOperationAction(ISD::SELECT_CC, MVT::v2i64, Custom);
482 setOperationAction(ISD::SELECT_CC, MVT::v2f32, Custom);
483 setOperationAction(ISD::SELECT_CC, MVT::v4f32, Custom);
484 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Custom);
485 setOperationAction(ISD::SELECT_CC, MVT::v2f64, Custom);
487 // Vector ExtLoad and TruncStore are expanded.
488 for (unsigned I = MVT::FIRST_VECTOR_VALUETYPE;
489 I <= MVT::LAST_VECTOR_VALUETYPE; ++I) {
490 MVT VT = (MVT::SimpleValueType) I;
491 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
492 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
493 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
494 for (unsigned II = MVT::FIRST_VECTOR_VALUETYPE;
495 II <= MVT::LAST_VECTOR_VALUETYPE; ++II) {
496 MVT VT1 = (MVT::SimpleValueType) II;
497 // A TruncStore has two vector types of the same number of elements
498 // and different element sizes.
499 if (VT.getVectorNumElements() == VT1.getVectorNumElements() &&
500 VT.getVectorElementType().getSizeInBits()
501 > VT1.getVectorElementType().getSizeInBits())
502 setTruncStoreAction(VT, VT1, Expand);
506 // There is no v1i64/v2i64 multiply, expand v1i64/v2i64 to GPR i64 multiply.
507 // FIXME: For a v2i64 multiply, we copy VPR to GPR and do 2 i64 multiplies,
508 // and then copy back to VPR. This solution may be optimized by Following 3
509 // NEON instructions:
510 // pmull v2.1q, v0.1d, v1.1d
511 // pmull2 v3.1q, v0.2d, v1.2d
512 // ins v2.d[1], v3.d[0]
513 // As currently we can't verify the correctness of such assumption, we can
514 // do such optimization in the future.
515 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
516 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
520 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
521 // It's reasonably important that this value matches the "natural" legal
522 // promotion from i1 for scalar types. Otherwise LegalizeTypes can get itself
523 // in a twist (e.g. inserting an any_extend which then becomes i64 -> i64).
524 if (!VT.isVector()) return MVT::i32;
525 return VT.changeVectorElementTypeToInteger();
528 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
531 static const unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
532 AArch64::LDXR_word, AArch64::LDXR_dword};
533 static const unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
534 AArch64::LDAXR_word, AArch64::LDAXR_dword};
535 static const unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
536 AArch64::STXR_word, AArch64::STXR_dword};
537 static const unsigned StoreRels[] = {AArch64::STLXR_byte,AArch64::STLXR_hword,
538 AArch64::STLXR_word, AArch64::STLXR_dword};
540 const unsigned *LoadOps, *StoreOps;
541 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
546 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
547 StoreOps = StoreRels;
549 StoreOps = StoreBares;
551 assert(isPowerOf2_32(Size) && Size <= 8 &&
552 "unsupported size for atomic binary op!");
554 LdrOpc = LoadOps[Log2_32(Size)];
555 StrOpc = StoreOps[Log2_32(Size)];
558 // FIXME: AArch64::DTripleRegClass and AArch64::QTripleRegClass don't really
559 // have value type mapped, and they are both being defined as MVT::untyped.
560 // Without knowing the MVT type, MachineLICM::getRegisterClassIDAndCost
561 // would fail to figure out the register pressure correctly.
562 std::pair<const TargetRegisterClass*, uint8_t>
563 AArch64TargetLowering::findRepresentativeClass(MVT VT) const{
564 const TargetRegisterClass *RRC = 0;
566 switch (VT.SimpleTy) {
568 return TargetLowering::findRepresentativeClass(VT);
570 RRC = &AArch64::QPairRegClass;
574 RRC = &AArch64::QQuadRegClass;
578 return std::make_pair(RRC, Cost);
582 AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
584 unsigned BinOpcode) const {
585 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
586 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
588 const BasicBlock *LLVM_BB = BB->getBasicBlock();
589 MachineFunction *MF = BB->getParent();
590 MachineFunction::iterator It = BB;
593 unsigned dest = MI->getOperand(0).getReg();
594 unsigned ptr = MI->getOperand(1).getReg();
595 unsigned incr = MI->getOperand(2).getReg();
596 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
597 DebugLoc dl = MI->getDebugLoc();
599 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
601 unsigned ldrOpc, strOpc;
602 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
604 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
605 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
606 MF->insert(It, loopMBB);
607 MF->insert(It, exitMBB);
609 // Transfer the remainder of BB and its successor edges to exitMBB.
610 exitMBB->splice(exitMBB->begin(), BB,
611 llvm::next(MachineBasicBlock::iterator(MI)),
613 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
615 const TargetRegisterClass *TRC
616 = Size == 8 ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
617 unsigned scratch = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
621 // fallthrough --> loopMBB
622 BB->addSuccessor(loopMBB);
626 // <binop> scratch, dest, incr
627 // stxr stxr_status, scratch, ptr
628 // cbnz stxr_status, loopMBB
629 // fallthrough --> exitMBB
631 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
633 // All arithmetic operations we'll be creating are designed to take an extra
634 // shift or extend operand, which we can conveniently set to zero.
636 // Operand order needs to go the other way for NAND.
637 if (BinOpcode == AArch64::BICwww_lsl || BinOpcode == AArch64::BICxxx_lsl)
638 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
639 .addReg(incr).addReg(dest).addImm(0);
641 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
642 .addReg(dest).addReg(incr).addImm(0);
645 // From the stxr, the register is GPR32; from the cmp it's GPR32wsp
646 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
647 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
649 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(scratch).addReg(ptr);
650 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
651 .addReg(stxr_status).addMBB(loopMBB);
653 BB->addSuccessor(loopMBB);
654 BB->addSuccessor(exitMBB);
660 MI->eraseFromParent(); // The instruction is gone now.
666 AArch64TargetLowering::emitAtomicBinaryMinMax(MachineInstr *MI,
667 MachineBasicBlock *BB,
670 A64CC::CondCodes Cond) const {
671 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
673 const BasicBlock *LLVM_BB = BB->getBasicBlock();
674 MachineFunction *MF = BB->getParent();
675 MachineFunction::iterator It = BB;
678 unsigned dest = MI->getOperand(0).getReg();
679 unsigned ptr = MI->getOperand(1).getReg();
680 unsigned incr = MI->getOperand(2).getReg();
681 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
683 unsigned oldval = dest;
684 DebugLoc dl = MI->getDebugLoc();
686 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
687 const TargetRegisterClass *TRC, *TRCsp;
689 TRC = &AArch64::GPR64RegClass;
690 TRCsp = &AArch64::GPR64xspRegClass;
692 TRC = &AArch64::GPR32RegClass;
693 TRCsp = &AArch64::GPR32wspRegClass;
696 unsigned ldrOpc, strOpc;
697 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
699 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
700 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
701 MF->insert(It, loopMBB);
702 MF->insert(It, exitMBB);
704 // Transfer the remainder of BB and its successor edges to exitMBB.
705 exitMBB->splice(exitMBB->begin(), BB,
706 llvm::next(MachineBasicBlock::iterator(MI)),
708 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
710 unsigned scratch = MRI.createVirtualRegister(TRC);
711 MRI.constrainRegClass(scratch, TRCsp);
715 // fallthrough --> loopMBB
716 BB->addSuccessor(loopMBB);
720 // cmp incr, dest (, sign extend if necessary)
721 // csel scratch, dest, incr, cond
722 // stxr stxr_status, scratch, ptr
723 // cbnz stxr_status, loopMBB
724 // fallthrough --> exitMBB
726 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
728 // Build compare and cmov instructions.
729 MRI.constrainRegClass(incr, TRCsp);
730 BuildMI(BB, dl, TII->get(CmpOp))
731 .addReg(incr).addReg(oldval).addImm(0);
733 BuildMI(BB, dl, TII->get(Size == 8 ? AArch64::CSELxxxc : AArch64::CSELwwwc),
735 .addReg(oldval).addReg(incr).addImm(Cond);
737 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
738 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
740 BuildMI(BB, dl, TII->get(strOpc), stxr_status)
741 .addReg(scratch).addReg(ptr);
742 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
743 .addReg(stxr_status).addMBB(loopMBB);
745 BB->addSuccessor(loopMBB);
746 BB->addSuccessor(exitMBB);
752 MI->eraseFromParent(); // The instruction is gone now.
758 AArch64TargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
759 MachineBasicBlock *BB,
760 unsigned Size) const {
761 unsigned dest = MI->getOperand(0).getReg();
762 unsigned ptr = MI->getOperand(1).getReg();
763 unsigned oldval = MI->getOperand(2).getReg();
764 unsigned newval = MI->getOperand(3).getReg();
765 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
766 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
767 DebugLoc dl = MI->getDebugLoc();
769 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
770 const TargetRegisterClass *TRCsp;
771 TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
773 unsigned ldrOpc, strOpc;
774 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
776 MachineFunction *MF = BB->getParent();
777 const BasicBlock *LLVM_BB = BB->getBasicBlock();
778 MachineFunction::iterator It = BB;
779 ++It; // insert the new blocks after the current block
781 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
782 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
783 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
784 MF->insert(It, loop1MBB);
785 MF->insert(It, loop2MBB);
786 MF->insert(It, exitMBB);
788 // Transfer the remainder of BB and its successor edges to exitMBB.
789 exitMBB->splice(exitMBB->begin(), BB,
790 llvm::next(MachineBasicBlock::iterator(MI)),
792 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
796 // fallthrough --> loop1MBB
797 BB->addSuccessor(loop1MBB);
804 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
806 unsigned CmpOp = Size == 8 ? AArch64::CMPxx_lsl : AArch64::CMPww_lsl;
807 MRI.constrainRegClass(dest, TRCsp);
808 BuildMI(BB, dl, TII->get(CmpOp))
809 .addReg(dest).addReg(oldval).addImm(0);
810 BuildMI(BB, dl, TII->get(AArch64::Bcc))
811 .addImm(A64CC::NE).addMBB(exitMBB);
812 BB->addSuccessor(loop2MBB);
813 BB->addSuccessor(exitMBB);
816 // strex stxr_status, newval, [ptr]
817 // cbnz stxr_status, loop1MBB
819 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
820 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
822 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(newval).addReg(ptr);
823 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
824 .addReg(stxr_status).addMBB(loop1MBB);
825 BB->addSuccessor(loop1MBB);
826 BB->addSuccessor(exitMBB);
832 MI->eraseFromParent(); // The instruction is gone now.
838 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
839 MachineBasicBlock *MBB) const {
840 // We materialise the F128CSEL pseudo-instruction using conditional branches
841 // and loads, giving an instruciton sequence like:
850 // Using virtual registers would probably not be beneficial since COPY
851 // instructions are expensive for f128 (there's no actual instruction to
854 // An alternative would be to do an integer-CSEL on some address. E.g.:
859 // csel x0, x0, x1, ne
862 // It's unclear which approach is actually optimal.
863 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
864 MachineFunction *MF = MBB->getParent();
865 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
866 DebugLoc DL = MI->getDebugLoc();
867 MachineFunction::iterator It = MBB;
870 unsigned DestReg = MI->getOperand(0).getReg();
871 unsigned IfTrueReg = MI->getOperand(1).getReg();
872 unsigned IfFalseReg = MI->getOperand(2).getReg();
873 unsigned CondCode = MI->getOperand(3).getImm();
874 bool NZCVKilled = MI->getOperand(4).isKill();
876 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
877 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
878 MF->insert(It, TrueBB);
879 MF->insert(It, EndBB);
881 // Transfer rest of current basic-block to EndBB
882 EndBB->splice(EndBB->begin(), MBB,
883 llvm::next(MachineBasicBlock::iterator(MI)),
885 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
887 // We need somewhere to store the f128 value needed.
888 int ScratchFI = MF->getFrameInfo()->CreateSpillStackObject(16, 16);
890 // [... start of incoming MBB ...]
891 // str qIFFALSE, [sp]
894 BuildMI(MBB, DL, TII->get(AArch64::LSFP128_STR))
896 .addFrameIndex(ScratchFI)
898 BuildMI(MBB, DL, TII->get(AArch64::Bcc))
901 BuildMI(MBB, DL, TII->get(AArch64::Bimm))
903 MBB->addSuccessor(TrueBB);
904 MBB->addSuccessor(EndBB);
907 // NZCV is live-through TrueBB.
908 TrueBB->addLiveIn(AArch64::NZCV);
909 EndBB->addLiveIn(AArch64::NZCV);
914 BuildMI(TrueBB, DL, TII->get(AArch64::LSFP128_STR))
916 .addFrameIndex(ScratchFI)
919 // Note: fallthrough. We can rely on LLVM adding a branch if it reorders the
921 TrueBB->addSuccessor(EndBB);
925 // [... rest of incoming MBB ...]
926 MachineInstr *StartOfEnd = EndBB->begin();
927 BuildMI(*EndBB, StartOfEnd, DL, TII->get(AArch64::LSFP128_LDR), DestReg)
928 .addFrameIndex(ScratchFI)
931 MI->eraseFromParent();
936 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
937 MachineBasicBlock *MBB) const {
938 switch (MI->getOpcode()) {
939 default: llvm_unreachable("Unhandled instruction with custom inserter");
940 case AArch64::F128CSEL:
941 return EmitF128CSEL(MI, MBB);
942 case AArch64::ATOMIC_LOAD_ADD_I8:
943 return emitAtomicBinary(MI, MBB, 1, AArch64::ADDwww_lsl);
944 case AArch64::ATOMIC_LOAD_ADD_I16:
945 return emitAtomicBinary(MI, MBB, 2, AArch64::ADDwww_lsl);
946 case AArch64::ATOMIC_LOAD_ADD_I32:
947 return emitAtomicBinary(MI, MBB, 4, AArch64::ADDwww_lsl);
948 case AArch64::ATOMIC_LOAD_ADD_I64:
949 return emitAtomicBinary(MI, MBB, 8, AArch64::ADDxxx_lsl);
951 case AArch64::ATOMIC_LOAD_SUB_I8:
952 return emitAtomicBinary(MI, MBB, 1, AArch64::SUBwww_lsl);
953 case AArch64::ATOMIC_LOAD_SUB_I16:
954 return emitAtomicBinary(MI, MBB, 2, AArch64::SUBwww_lsl);
955 case AArch64::ATOMIC_LOAD_SUB_I32:
956 return emitAtomicBinary(MI, MBB, 4, AArch64::SUBwww_lsl);
957 case AArch64::ATOMIC_LOAD_SUB_I64:
958 return emitAtomicBinary(MI, MBB, 8, AArch64::SUBxxx_lsl);
960 case AArch64::ATOMIC_LOAD_AND_I8:
961 return emitAtomicBinary(MI, MBB, 1, AArch64::ANDwww_lsl);
962 case AArch64::ATOMIC_LOAD_AND_I16:
963 return emitAtomicBinary(MI, MBB, 2, AArch64::ANDwww_lsl);
964 case AArch64::ATOMIC_LOAD_AND_I32:
965 return emitAtomicBinary(MI, MBB, 4, AArch64::ANDwww_lsl);
966 case AArch64::ATOMIC_LOAD_AND_I64:
967 return emitAtomicBinary(MI, MBB, 8, AArch64::ANDxxx_lsl);
969 case AArch64::ATOMIC_LOAD_OR_I8:
970 return emitAtomicBinary(MI, MBB, 1, AArch64::ORRwww_lsl);
971 case AArch64::ATOMIC_LOAD_OR_I16:
972 return emitAtomicBinary(MI, MBB, 2, AArch64::ORRwww_lsl);
973 case AArch64::ATOMIC_LOAD_OR_I32:
974 return emitAtomicBinary(MI, MBB, 4, AArch64::ORRwww_lsl);
975 case AArch64::ATOMIC_LOAD_OR_I64:
976 return emitAtomicBinary(MI, MBB, 8, AArch64::ORRxxx_lsl);
978 case AArch64::ATOMIC_LOAD_XOR_I8:
979 return emitAtomicBinary(MI, MBB, 1, AArch64::EORwww_lsl);
980 case AArch64::ATOMIC_LOAD_XOR_I16:
981 return emitAtomicBinary(MI, MBB, 2, AArch64::EORwww_lsl);
982 case AArch64::ATOMIC_LOAD_XOR_I32:
983 return emitAtomicBinary(MI, MBB, 4, AArch64::EORwww_lsl);
984 case AArch64::ATOMIC_LOAD_XOR_I64:
985 return emitAtomicBinary(MI, MBB, 8, AArch64::EORxxx_lsl);
987 case AArch64::ATOMIC_LOAD_NAND_I8:
988 return emitAtomicBinary(MI, MBB, 1, AArch64::BICwww_lsl);
989 case AArch64::ATOMIC_LOAD_NAND_I16:
990 return emitAtomicBinary(MI, MBB, 2, AArch64::BICwww_lsl);
991 case AArch64::ATOMIC_LOAD_NAND_I32:
992 return emitAtomicBinary(MI, MBB, 4, AArch64::BICwww_lsl);
993 case AArch64::ATOMIC_LOAD_NAND_I64:
994 return emitAtomicBinary(MI, MBB, 8, AArch64::BICxxx_lsl);
996 case AArch64::ATOMIC_LOAD_MIN_I8:
997 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::GT);
998 case AArch64::ATOMIC_LOAD_MIN_I16:
999 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::GT);
1000 case AArch64::ATOMIC_LOAD_MIN_I32:
1001 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::GT);
1002 case AArch64::ATOMIC_LOAD_MIN_I64:
1003 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::GT);
1005 case AArch64::ATOMIC_LOAD_MAX_I8:
1006 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::LT);
1007 case AArch64::ATOMIC_LOAD_MAX_I16:
1008 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::LT);
1009 case AArch64::ATOMIC_LOAD_MAX_I32:
1010 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LT);
1011 case AArch64::ATOMIC_LOAD_MAX_I64:
1012 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LT);
1014 case AArch64::ATOMIC_LOAD_UMIN_I8:
1015 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::HI);
1016 case AArch64::ATOMIC_LOAD_UMIN_I16:
1017 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::HI);
1018 case AArch64::ATOMIC_LOAD_UMIN_I32:
1019 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::HI);
1020 case AArch64::ATOMIC_LOAD_UMIN_I64:
1021 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::HI);
1023 case AArch64::ATOMIC_LOAD_UMAX_I8:
1024 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::LO);
1025 case AArch64::ATOMIC_LOAD_UMAX_I16:
1026 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::LO);
1027 case AArch64::ATOMIC_LOAD_UMAX_I32:
1028 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LO);
1029 case AArch64::ATOMIC_LOAD_UMAX_I64:
1030 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LO);
1032 case AArch64::ATOMIC_SWAP_I8:
1033 return emitAtomicBinary(MI, MBB, 1, 0);
1034 case AArch64::ATOMIC_SWAP_I16:
1035 return emitAtomicBinary(MI, MBB, 2, 0);
1036 case AArch64::ATOMIC_SWAP_I32:
1037 return emitAtomicBinary(MI, MBB, 4, 0);
1038 case AArch64::ATOMIC_SWAP_I64:
1039 return emitAtomicBinary(MI, MBB, 8, 0);
1041 case AArch64::ATOMIC_CMP_SWAP_I8:
1042 return emitAtomicCmpSwap(MI, MBB, 1);
1043 case AArch64::ATOMIC_CMP_SWAP_I16:
1044 return emitAtomicCmpSwap(MI, MBB, 2);
1045 case AArch64::ATOMIC_CMP_SWAP_I32:
1046 return emitAtomicCmpSwap(MI, MBB, 4);
1047 case AArch64::ATOMIC_CMP_SWAP_I64:
1048 return emitAtomicCmpSwap(MI, MBB, 8);
1053 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
1055 case AArch64ISD::BR_CC: return "AArch64ISD::BR_CC";
1056 case AArch64ISD::Call: return "AArch64ISD::Call";
1057 case AArch64ISD::FPMOV: return "AArch64ISD::FPMOV";
1058 case AArch64ISD::GOTLoad: return "AArch64ISD::GOTLoad";
1059 case AArch64ISD::BFI: return "AArch64ISD::BFI";
1060 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
1061 case AArch64ISD::Ret: return "AArch64ISD::Ret";
1062 case AArch64ISD::SBFX: return "AArch64ISD::SBFX";
1063 case AArch64ISD::SELECT_CC: return "AArch64ISD::SELECT_CC";
1064 case AArch64ISD::SETCC: return "AArch64ISD::SETCC";
1065 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
1066 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
1067 case AArch64ISD::TLSDESCCALL: return "AArch64ISD::TLSDESCCALL";
1068 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
1069 case AArch64ISD::WrapperSmall: return "AArch64ISD::WrapperSmall";
1071 case AArch64ISD::NEON_MOVIMM:
1072 return "AArch64ISD::NEON_MOVIMM";
1073 case AArch64ISD::NEON_MVNIMM:
1074 return "AArch64ISD::NEON_MVNIMM";
1075 case AArch64ISD::NEON_FMOVIMM:
1076 return "AArch64ISD::NEON_FMOVIMM";
1077 case AArch64ISD::NEON_CMP:
1078 return "AArch64ISD::NEON_CMP";
1079 case AArch64ISD::NEON_CMPZ:
1080 return "AArch64ISD::NEON_CMPZ";
1081 case AArch64ISD::NEON_TST:
1082 return "AArch64ISD::NEON_TST";
1083 case AArch64ISD::NEON_QSHLs:
1084 return "AArch64ISD::NEON_QSHLs";
1085 case AArch64ISD::NEON_QSHLu:
1086 return "AArch64ISD::NEON_QSHLu";
1087 case AArch64ISD::NEON_VDUP:
1088 return "AArch64ISD::NEON_VDUP";
1089 case AArch64ISD::NEON_VDUPLANE:
1090 return "AArch64ISD::NEON_VDUPLANE";
1091 case AArch64ISD::NEON_REV16:
1092 return "AArch64ISD::NEON_REV16";
1093 case AArch64ISD::NEON_REV32:
1094 return "AArch64ISD::NEON_REV32";
1095 case AArch64ISD::NEON_REV64:
1096 return "AArch64ISD::NEON_REV64";
1097 case AArch64ISD::NEON_UZP1:
1098 return "AArch64ISD::NEON_UZP1";
1099 case AArch64ISD::NEON_UZP2:
1100 return "AArch64ISD::NEON_UZP2";
1101 case AArch64ISD::NEON_ZIP1:
1102 return "AArch64ISD::NEON_ZIP1";
1103 case AArch64ISD::NEON_ZIP2:
1104 return "AArch64ISD::NEON_ZIP2";
1105 case AArch64ISD::NEON_TRN1:
1106 return "AArch64ISD::NEON_TRN1";
1107 case AArch64ISD::NEON_TRN2:
1108 return "AArch64ISD::NEON_TRN2";
1109 case AArch64ISD::NEON_LD1_UPD:
1110 return "AArch64ISD::NEON_LD1_UPD";
1111 case AArch64ISD::NEON_LD2_UPD:
1112 return "AArch64ISD::NEON_LD2_UPD";
1113 case AArch64ISD::NEON_LD3_UPD:
1114 return "AArch64ISD::NEON_LD3_UPD";
1115 case AArch64ISD::NEON_LD4_UPD:
1116 return "AArch64ISD::NEON_LD4_UPD";
1117 case AArch64ISD::NEON_ST1_UPD:
1118 return "AArch64ISD::NEON_ST1_UPD";
1119 case AArch64ISD::NEON_ST2_UPD:
1120 return "AArch64ISD::NEON_ST2_UPD";
1121 case AArch64ISD::NEON_ST3_UPD:
1122 return "AArch64ISD::NEON_ST3_UPD";
1123 case AArch64ISD::NEON_ST4_UPD:
1124 return "AArch64ISD::NEON_ST4_UPD";
1125 case AArch64ISD::NEON_LD1x2_UPD:
1126 return "AArch64ISD::NEON_LD1x2_UPD";
1127 case AArch64ISD::NEON_LD1x3_UPD:
1128 return "AArch64ISD::NEON_LD1x3_UPD";
1129 case AArch64ISD::NEON_LD1x4_UPD:
1130 return "AArch64ISD::NEON_LD1x4_UPD";
1131 case AArch64ISD::NEON_ST1x2_UPD:
1132 return "AArch64ISD::NEON_ST1x2_UPD";
1133 case AArch64ISD::NEON_ST1x3_UPD:
1134 return "AArch64ISD::NEON_ST1x3_UPD";
1135 case AArch64ISD::NEON_ST1x4_UPD:
1136 return "AArch64ISD::NEON_ST1x4_UPD";
1137 case AArch64ISD::NEON_LD2DUP:
1138 return "AArch64ISD::NEON_LD2DUP";
1139 case AArch64ISD::NEON_LD3DUP:
1140 return "AArch64ISD::NEON_LD3DUP";
1141 case AArch64ISD::NEON_LD4DUP:
1142 return "AArch64ISD::NEON_LD4DUP";
1143 case AArch64ISD::NEON_LD2DUP_UPD:
1144 return "AArch64ISD::NEON_LD2DUP_UPD";
1145 case AArch64ISD::NEON_LD3DUP_UPD:
1146 return "AArch64ISD::NEON_LD3DUP_UPD";
1147 case AArch64ISD::NEON_LD4DUP_UPD:
1148 return "AArch64ISD::NEON_LD4DUP_UPD";
1149 case AArch64ISD::NEON_LD2LN_UPD:
1150 return "AArch64ISD::NEON_LD2LN_UPD";
1151 case AArch64ISD::NEON_LD3LN_UPD:
1152 return "AArch64ISD::NEON_LD3LN_UPD";
1153 case AArch64ISD::NEON_LD4LN_UPD:
1154 return "AArch64ISD::NEON_LD4LN_UPD";
1155 case AArch64ISD::NEON_ST2LN_UPD:
1156 return "AArch64ISD::NEON_ST2LN_UPD";
1157 case AArch64ISD::NEON_ST3LN_UPD:
1158 return "AArch64ISD::NEON_ST3LN_UPD";
1159 case AArch64ISD::NEON_ST4LN_UPD:
1160 return "AArch64ISD::NEON_ST4LN_UPD";
1161 case AArch64ISD::NEON_VEXTRACT:
1162 return "AArch64ISD::NEON_VEXTRACT";
1168 static const uint16_t AArch64FPRArgRegs[] = {
1169 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
1170 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7
1172 static const unsigned NumFPRArgRegs = llvm::array_lengthof(AArch64FPRArgRegs);
1174 static const uint16_t AArch64ArgRegs[] = {
1175 AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3,
1176 AArch64::X4, AArch64::X5, AArch64::X6, AArch64::X7
1178 static const unsigned NumArgRegs = llvm::array_lengthof(AArch64ArgRegs);
1180 static bool CC_AArch64NoMoreRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
1181 CCValAssign::LocInfo LocInfo,
1182 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1183 // Mark all remaining general purpose registers as allocated. We don't
1184 // backtrack: if (for example) an i128 gets put on the stack, no subsequent
1185 // i64 will go in registers (C.11).
1186 for (unsigned i = 0; i < NumArgRegs; ++i)
1187 State.AllocateReg(AArch64ArgRegs[i]);
1192 #include "AArch64GenCallingConv.inc"
1194 CCAssignFn *AArch64TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1197 default: llvm_unreachable("Unsupported calling convention");
1198 case CallingConv::Fast:
1199 case CallingConv::C:
1205 AArch64TargetLowering::SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG,
1206 SDLoc DL, SDValue &Chain) const {
1207 MachineFunction &MF = DAG.getMachineFunction();
1208 MachineFrameInfo *MFI = MF.getFrameInfo();
1209 AArch64MachineFunctionInfo *FuncInfo
1210 = MF.getInfo<AArch64MachineFunctionInfo>();
1212 SmallVector<SDValue, 8> MemOps;
1214 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(AArch64ArgRegs,
1216 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(AArch64FPRArgRegs,
1219 unsigned GPRSaveSize = 8 * (NumArgRegs - FirstVariadicGPR);
1221 if (GPRSaveSize != 0) {
1222 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1224 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1226 for (unsigned i = FirstVariadicGPR; i < NumArgRegs; ++i) {
1227 unsigned VReg = MF.addLiveIn(AArch64ArgRegs[i], &AArch64::GPR64RegClass);
1228 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1229 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1230 MachinePointerInfo::getStack(i * 8),
1232 MemOps.push_back(Store);
1233 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1234 DAG.getConstant(8, getPointerTy()));
1238 if (getSubtarget()->hasFPARMv8()) {
1239 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1241 // According to the AArch64 Procedure Call Standard, section B.1/B.3, we
1242 // can omit a register save area if we know we'll never use registers of
1244 if (FPRSaveSize != 0) {
1245 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1247 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1249 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1250 unsigned VReg = MF.addLiveIn(AArch64FPRArgRegs[i],
1251 &AArch64::FPR128RegClass);
1252 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1253 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1254 MachinePointerInfo::getStack(i * 16),
1256 MemOps.push_back(Store);
1257 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1258 DAG.getConstant(16, getPointerTy()));
1261 FuncInfo->setVariadicFPRIdx(FPRIdx);
1262 FuncInfo->setVariadicFPRSize(FPRSaveSize);
1265 int StackIdx = MFI->CreateFixedObject(8, CCInfo.getNextStackOffset(), true);
1267 FuncInfo->setVariadicStackIdx(StackIdx);
1268 FuncInfo->setVariadicGPRIdx(GPRIdx);
1269 FuncInfo->setVariadicGPRSize(GPRSaveSize);
1271 if (!MemOps.empty()) {
1272 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
1279 AArch64TargetLowering::LowerFormalArguments(SDValue Chain,
1280 CallingConv::ID CallConv, bool isVarArg,
1281 const SmallVectorImpl<ISD::InputArg> &Ins,
1282 SDLoc dl, SelectionDAG &DAG,
1283 SmallVectorImpl<SDValue> &InVals) const {
1284 MachineFunction &MF = DAG.getMachineFunction();
1285 AArch64MachineFunctionInfo *FuncInfo
1286 = MF.getInfo<AArch64MachineFunctionInfo>();
1287 MachineFrameInfo *MFI = MF.getFrameInfo();
1288 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1290 SmallVector<CCValAssign, 16> ArgLocs;
1291 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1292 getTargetMachine(), ArgLocs, *DAG.getContext());
1293 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1295 SmallVector<SDValue, 16> ArgValues;
1298 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1299 CCValAssign &VA = ArgLocs[i];
1300 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1302 if (Flags.isByVal()) {
1303 // Byval is used for small structs and HFAs in the PCS, but the system
1304 // should work in a non-compliant manner for larger structs.
1305 EVT PtrTy = getPointerTy();
1306 int Size = Flags.getByValSize();
1307 unsigned NumRegs = (Size + 7) / 8;
1309 unsigned FrameIdx = MFI->CreateFixedObject(8 * NumRegs,
1310 VA.getLocMemOffset(),
1312 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1313 InVals.push_back(FrameIdxN);
1316 } else if (VA.isRegLoc()) {
1317 MVT RegVT = VA.getLocVT();
1318 const TargetRegisterClass *RC = getRegClassFor(RegVT);
1319 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1321 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1322 } else { // VA.isRegLoc()
1323 assert(VA.isMemLoc());
1325 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
1326 VA.getLocMemOffset(), true);
1328 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1329 ArgValue = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1330 MachinePointerInfo::getFixedStack(FI),
1331 false, false, false, 0);
1336 switch (VA.getLocInfo()) {
1337 default: llvm_unreachable("Unknown loc info!");
1338 case CCValAssign::Full: break;
1339 case CCValAssign::BCvt:
1340 ArgValue = DAG.getNode(ISD::BITCAST,dl, VA.getValVT(), ArgValue);
1342 case CCValAssign::SExt:
1343 case CCValAssign::ZExt:
1344 case CCValAssign::AExt:
1345 case CCValAssign::FPExt: {
1346 unsigned DestSize = VA.getValVT().getSizeInBits();
1347 unsigned DestSubReg;
1350 case 8: DestSubReg = AArch64::sub_8; break;
1351 case 16: DestSubReg = AArch64::sub_16; break;
1352 case 32: DestSubReg = AArch64::sub_32; break;
1353 case 64: DestSubReg = AArch64::sub_64; break;
1354 default: llvm_unreachable("Unexpected argument promotion");
1357 ArgValue = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
1358 VA.getValVT(), ArgValue,
1359 DAG.getTargetConstant(DestSubReg, MVT::i32)),
1365 InVals.push_back(ArgValue);
1369 SaveVarArgRegisters(CCInfo, DAG, dl, Chain);
1371 unsigned StackArgSize = CCInfo.getNextStackOffset();
1372 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
1373 // This is a non-standard ABI so by fiat I say we're allowed to make full
1374 // use of the stack area to be popped, which must be aligned to 16 bytes in
1376 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
1378 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
1379 // a multiple of 16.
1380 FuncInfo->setArgumentStackToRestore(StackArgSize);
1382 // This realignment carries over to the available bytes below. Our own
1383 // callers will guarantee the space is free by giving an aligned value to
1386 // Even if we're not expected to free up the space, it's useful to know how
1387 // much is there while considering tail calls (because we can reuse it).
1388 FuncInfo->setBytesInStackArgArea(StackArgSize);
1394 AArch64TargetLowering::LowerReturn(SDValue Chain,
1395 CallingConv::ID CallConv, bool isVarArg,
1396 const SmallVectorImpl<ISD::OutputArg> &Outs,
1397 const SmallVectorImpl<SDValue> &OutVals,
1398 SDLoc dl, SelectionDAG &DAG) const {
1399 // CCValAssign - represent the assignment of the return value to a location.
1400 SmallVector<CCValAssign, 16> RVLocs;
1402 // CCState - Info about the registers and stack slots.
1403 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1404 getTargetMachine(), RVLocs, *DAG.getContext());
1406 // Analyze outgoing return values.
1407 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv));
1410 SmallVector<SDValue, 4> RetOps(1, Chain);
1412 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1413 // PCS: "If the type, T, of the result of a function is such that
1414 // void func(T arg) would require that arg be passed as a value in a
1415 // register (or set of registers) according to the rules in 5.4, then the
1416 // result is returned in the same registers as would be used for such an
1419 // Otherwise, the caller shall reserve a block of memory of sufficient
1420 // size and alignment to hold the result. The address of the memory block
1421 // shall be passed as an additional argument to the function in x8."
1423 // This is implemented in two places. The register-return values are dealt
1424 // with here, more complex returns are passed as an sret parameter, which
1425 // means we don't have to worry about it during actual return.
1426 CCValAssign &VA = RVLocs[i];
1427 assert(VA.isRegLoc() && "Only register-returns should be created by PCS");
1430 SDValue Arg = OutVals[i];
1432 // There's no convenient note in the ABI about this as there is for normal
1433 // arguments, but it says return values are passed in the same registers as
1434 // an argument would be. I believe that includes the comments about
1435 // unspecified higher bits, putting the burden of widening on the *caller*
1436 // for return values.
1437 switch (VA.getLocInfo()) {
1438 default: llvm_unreachable("Unknown loc info");
1439 case CCValAssign::Full: break;
1440 case CCValAssign::SExt:
1441 case CCValAssign::ZExt:
1442 case CCValAssign::AExt:
1443 // Floating-point values should only be extended when they're going into
1444 // memory, which can't happen here so an integer extend is acceptable.
1445 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1447 case CCValAssign::BCvt:
1448 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1452 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1453 Flag = Chain.getValue(1);
1454 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1457 RetOps[0] = Chain; // Update chain.
1459 // Add the flag if we have it.
1461 RetOps.push_back(Flag);
1463 return DAG.getNode(AArch64ISD::Ret, dl, MVT::Other,
1464 &RetOps[0], RetOps.size());
1467 unsigned AArch64TargetLowering::getByValTypeAlignment(Type *Ty) const {
1468 // This is a new backend. For anything more precise than this a FE should
1469 // set an explicit alignment.
1474 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1475 SmallVectorImpl<SDValue> &InVals) const {
1476 SelectionDAG &DAG = CLI.DAG;
1478 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1479 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1480 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1481 SDValue Chain = CLI.Chain;
1482 SDValue Callee = CLI.Callee;
1483 bool &IsTailCall = CLI.IsTailCall;
1484 CallingConv::ID CallConv = CLI.CallConv;
1485 bool IsVarArg = CLI.IsVarArg;
1487 MachineFunction &MF = DAG.getMachineFunction();
1488 AArch64MachineFunctionInfo *FuncInfo
1489 = MF.getInfo<AArch64MachineFunctionInfo>();
1490 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1491 bool IsStructRet = !Outs.empty() && Outs[0].Flags.isSRet();
1492 bool IsSibCall = false;
1495 IsTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1496 IsVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1497 Outs, OutVals, Ins, DAG);
1499 // A sibling call is one where we're under the usual C ABI and not planning
1500 // to change that but can still do a tail call:
1501 if (!TailCallOpt && IsTailCall)
1505 SmallVector<CCValAssign, 16> ArgLocs;
1506 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1507 getTargetMachine(), ArgLocs, *DAG.getContext());
1508 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1510 // On AArch64 (and all other architectures I'm aware of) the most this has to
1511 // do is adjust the stack pointer.
1512 unsigned NumBytes = RoundUpToAlignment(CCInfo.getNextStackOffset(), 16);
1514 // Since we're not changing the ABI to make this a tail call, the memory
1515 // operands are already available in the caller's incoming argument space.
1519 // FPDiff is the byte offset of the call's argument area from the callee's.
1520 // Stores to callee stack arguments will be placed in FixedStackSlots offset
1521 // by this amount for a tail call. In a sibling call it must be 0 because the
1522 // caller will deallocate the entire stack and the callee still expects its
1523 // arguments to begin at SP+0. Completely unused for non-tail calls.
1526 if (IsTailCall && !IsSibCall) {
1527 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
1529 // FPDiff will be negative if this tail call requires more space than we
1530 // would automatically have in our incoming argument space. Positive if we
1531 // can actually shrink the stack.
1532 FPDiff = NumReusableBytes - NumBytes;
1534 // The stack pointer must be 16-byte aligned at all times it's used for a
1535 // memory operation, which in practice means at *all* times and in
1536 // particular across call boundaries. Therefore our own arguments started at
1537 // a 16-byte aligned SP and the delta applied for the tail call should
1538 // satisfy the same constraint.
1539 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
1543 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1546 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, AArch64::XSP,
1549 SmallVector<SDValue, 8> MemOpChains;
1550 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1552 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1553 CCValAssign &VA = ArgLocs[i];
1554 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1555 SDValue Arg = OutVals[i];
1557 // Callee does the actual widening, so all extensions just use an implicit
1558 // definition of the rest of the Loc. Aesthetically, this would be nicer as
1559 // an ANY_EXTEND, but that isn't valid for floating-point types and this
1560 // alternative works on integer types too.
1561 switch (VA.getLocInfo()) {
1562 default: llvm_unreachable("Unknown loc info!");
1563 case CCValAssign::Full: break;
1564 case CCValAssign::SExt:
1565 case CCValAssign::ZExt:
1566 case CCValAssign::AExt:
1567 case CCValAssign::FPExt: {
1568 unsigned SrcSize = VA.getValVT().getSizeInBits();
1572 case 8: SrcSubReg = AArch64::sub_8; break;
1573 case 16: SrcSubReg = AArch64::sub_16; break;
1574 case 32: SrcSubReg = AArch64::sub_32; break;
1575 case 64: SrcSubReg = AArch64::sub_64; break;
1576 default: llvm_unreachable("Unexpected argument promotion");
1579 Arg = SDValue(DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1581 DAG.getUNDEF(VA.getLocVT()),
1583 DAG.getTargetConstant(SrcSubReg, MVT::i32)),
1588 case CCValAssign::BCvt:
1589 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1593 if (VA.isRegLoc()) {
1594 // A normal register (sub-) argument. For now we just note it down because
1595 // we want to copy things into registers as late as possible to avoid
1596 // register-pressure (and possibly worse).
1597 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1601 assert(VA.isMemLoc() && "unexpected argument location");
1604 MachinePointerInfo DstInfo;
1606 uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize() :
1607 VA.getLocVT().getSizeInBits();
1608 OpSize = (OpSize + 7) / 8;
1609 int32_t Offset = VA.getLocMemOffset() + FPDiff;
1610 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
1612 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
1613 DstInfo = MachinePointerInfo::getFixedStack(FI);
1615 // Make sure any stack arguments overlapping with where we're storing are
1616 // loaded before this eventual operation. Otherwise they'll be clobbered.
1617 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
1619 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
1621 DstAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1622 DstInfo = MachinePointerInfo::getStack(VA.getLocMemOffset());
1625 if (Flags.isByVal()) {
1626 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i64);
1627 SDValue Cpy = DAG.getMemcpy(Chain, dl, DstAddr, Arg, SizeNode,
1628 Flags.getByValAlign(),
1629 /*isVolatile = */ false,
1630 /*alwaysInline = */ false,
1631 DstInfo, MachinePointerInfo(0));
1632 MemOpChains.push_back(Cpy);
1634 // Normal stack argument, put it where it's needed.
1635 SDValue Store = DAG.getStore(Chain, dl, Arg, DstAddr, DstInfo,
1637 MemOpChains.push_back(Store);
1641 // The loads and stores generated above shouldn't clash with each
1642 // other. Combining them with this TokenFactor notes that fact for the rest of
1644 if (!MemOpChains.empty())
1645 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1646 &MemOpChains[0], MemOpChains.size());
1648 // Most of the rest of the instructions need to be glued together; we don't
1649 // want assignments to actual registers used by a call to be rearranged by a
1650 // well-meaning scheduler.
1653 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1654 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1655 RegsToPass[i].second, InFlag);
1656 InFlag = Chain.getValue(1);
1659 // The linker is responsible for inserting veneers when necessary to put a
1660 // function call destination in range, so we don't need to bother with a
1662 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1663 const GlobalValue *GV = G->getGlobal();
1664 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1665 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1666 const char *Sym = S->getSymbol();
1667 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1670 // We don't usually want to end the call-sequence here because we would tidy
1671 // the frame up *after* the call, however in the ABI-changing tail-call case
1672 // we've carefully laid out the parameters so that when sp is reset they'll be
1673 // in the correct location.
1674 if (IsTailCall && !IsSibCall) {
1675 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1676 DAG.getIntPtrConstant(0, true), InFlag, dl);
1677 InFlag = Chain.getValue(1);
1680 // We produce the following DAG scheme for the actual call instruction:
1681 // (AArch64Call Chain, Callee, reg1, ..., regn, preserveMask, inflag?
1683 // Most arguments aren't going to be used and just keep the values live as
1684 // far as LLVM is concerned. It's expected to be selected as simply "bl
1685 // callee" (for a direct, non-tail call).
1686 std::vector<SDValue> Ops;
1687 Ops.push_back(Chain);
1688 Ops.push_back(Callee);
1691 // Each tail call may have to adjust the stack by a different amount, so
1692 // this information must travel along with the operation for eventual
1693 // consumption by emitEpilogue.
1694 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
1697 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1698 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1699 RegsToPass[i].second.getValueType()));
1702 // Add a register mask operand representing the call-preserved registers. This
1703 // is used later in codegen to constrain register-allocation.
1704 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1705 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1706 assert(Mask && "Missing call preserved mask for calling convention");
1707 Ops.push_back(DAG.getRegisterMask(Mask));
1709 // If we needed glue, put it in as the last argument.
1710 if (InFlag.getNode())
1711 Ops.push_back(InFlag);
1713 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1716 return DAG.getNode(AArch64ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1719 Chain = DAG.getNode(AArch64ISD::Call, dl, NodeTys, &Ops[0], Ops.size());
1720 InFlag = Chain.getValue(1);
1722 // Now we can reclaim the stack, just as well do it before working out where
1723 // our return value is.
1725 uint64_t CalleePopBytes
1726 = DoesCalleeRestoreStack(CallConv, TailCallOpt) ? NumBytes : 0;
1728 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1729 DAG.getIntPtrConstant(CalleePopBytes, true),
1731 InFlag = Chain.getValue(1);
1734 return LowerCallResult(Chain, InFlag, CallConv,
1735 IsVarArg, Ins, dl, DAG, InVals);
1739 AArch64TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1740 CallingConv::ID CallConv, bool IsVarArg,
1741 const SmallVectorImpl<ISD::InputArg> &Ins,
1742 SDLoc dl, SelectionDAG &DAG,
1743 SmallVectorImpl<SDValue> &InVals) const {
1744 // Assign locations to each value returned by this call.
1745 SmallVector<CCValAssign, 16> RVLocs;
1746 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1747 getTargetMachine(), RVLocs, *DAG.getContext());
1748 CCInfo.AnalyzeCallResult(Ins, CCAssignFnForNode(CallConv));
1750 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1751 CCValAssign VA = RVLocs[i];
1753 // Return values that are too big to fit into registers should use an sret
1754 // pointer, so this can be a lot simpler than the main argument code.
1755 assert(VA.isRegLoc() && "Memory locations not expected for call return");
1757 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1759 Chain = Val.getValue(1);
1760 InFlag = Val.getValue(2);
1762 switch (VA.getLocInfo()) {
1763 default: llvm_unreachable("Unknown loc info!");
1764 case CCValAssign::Full: break;
1765 case CCValAssign::BCvt:
1766 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1768 case CCValAssign::ZExt:
1769 case CCValAssign::SExt:
1770 case CCValAssign::AExt:
1771 // Floating-point arguments only get extended/truncated if they're going
1772 // in memory, so using the integer operation is acceptable here.
1773 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
1777 InVals.push_back(Val);
1784 AArch64TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1785 CallingConv::ID CalleeCC,
1787 bool IsCalleeStructRet,
1788 bool IsCallerStructRet,
1789 const SmallVectorImpl<ISD::OutputArg> &Outs,
1790 const SmallVectorImpl<SDValue> &OutVals,
1791 const SmallVectorImpl<ISD::InputArg> &Ins,
1792 SelectionDAG& DAG) const {
1794 // For CallingConv::C this function knows whether the ABI needs
1795 // changing. That's not true for other conventions so they will have to opt in
1797 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1800 const MachineFunction &MF = DAG.getMachineFunction();
1801 const Function *CallerF = MF.getFunction();
1802 CallingConv::ID CallerCC = CallerF->getCallingConv();
1803 bool CCMatch = CallerCC == CalleeCC;
1805 // Byval parameters hand the function a pointer directly into the stack area
1806 // we want to reuse during a tail call. Working around this *is* possible (see
1807 // X86) but less efficient and uglier in LowerCall.
1808 for (Function::const_arg_iterator i = CallerF->arg_begin(),
1809 e = CallerF->arg_end(); i != e; ++i)
1810 if (i->hasByValAttr())
1813 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
1814 if (IsTailCallConvention(CalleeCC) && CCMatch)
1819 // Now we search for cases where we can use a tail call without changing the
1820 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
1823 // I want anyone implementing a new calling convention to think long and hard
1824 // about this assert.
1825 assert((!IsVarArg || CalleeCC == CallingConv::C)
1826 && "Unexpected variadic calling convention");
1828 if (IsVarArg && !Outs.empty()) {
1829 // At least two cases here: if caller is fastcc then we can't have any
1830 // memory arguments (we'd be expected to clean up the stack afterwards). If
1831 // caller is C then we could potentially use its argument area.
1833 // FIXME: for now we take the most conservative of these in both cases:
1834 // disallow all variadic memory operands.
1835 SmallVector<CCValAssign, 16> ArgLocs;
1836 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1837 getTargetMachine(), ArgLocs, *DAG.getContext());
1839 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1840 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
1841 if (!ArgLocs[i].isRegLoc())
1845 // If the calling conventions do not match, then we'd better make sure the
1846 // results are returned in the same way as what the caller expects.
1848 SmallVector<CCValAssign, 16> RVLocs1;
1849 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1850 getTargetMachine(), RVLocs1, *DAG.getContext());
1851 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC));
1853 SmallVector<CCValAssign, 16> RVLocs2;
1854 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1855 getTargetMachine(), RVLocs2, *DAG.getContext());
1856 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC));
1858 if (RVLocs1.size() != RVLocs2.size())
1860 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1861 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1863 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1865 if (RVLocs1[i].isRegLoc()) {
1866 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1869 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1875 // Nothing more to check if the callee is taking no arguments
1879 SmallVector<CCValAssign, 16> ArgLocs;
1880 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1881 getTargetMachine(), ArgLocs, *DAG.getContext());
1883 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1885 const AArch64MachineFunctionInfo *FuncInfo
1886 = MF.getInfo<AArch64MachineFunctionInfo>();
1888 // If the stack arguments for this call would fit into our own save area then
1889 // the call can be made tail.
1890 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
1893 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
1894 bool TailCallOpt) const {
1895 return CallCC == CallingConv::Fast && TailCallOpt;
1898 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
1899 return CallCC == CallingConv::Fast;
1902 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
1904 MachineFrameInfo *MFI,
1905 int ClobberedFI) const {
1906 SmallVector<SDValue, 8> ArgChains;
1907 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
1908 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
1910 // Include the original chain at the beginning of the list. When this is
1911 // used by target LowerCall hooks, this helps legalize find the
1912 // CALLSEQ_BEGIN node.
1913 ArgChains.push_back(Chain);
1915 // Add a chain value for each stack argument corresponding
1916 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
1917 UE = DAG.getEntryNode().getNode()->use_end(); U != UE; ++U)
1918 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
1919 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
1920 if (FI->getIndex() < 0) {
1921 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
1922 int64_t InLastByte = InFirstByte;
1923 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
1925 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1926 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1927 ArgChains.push_back(SDValue(L, 1));
1930 // Build a tokenfactor for all the chains.
1931 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other,
1932 &ArgChains[0], ArgChains.size());
1935 static A64CC::CondCodes IntCCToA64CC(ISD::CondCode CC) {
1937 case ISD::SETEQ: return A64CC::EQ;
1938 case ISD::SETGT: return A64CC::GT;
1939 case ISD::SETGE: return A64CC::GE;
1940 case ISD::SETLT: return A64CC::LT;
1941 case ISD::SETLE: return A64CC::LE;
1942 case ISD::SETNE: return A64CC::NE;
1943 case ISD::SETUGT: return A64CC::HI;
1944 case ISD::SETUGE: return A64CC::HS;
1945 case ISD::SETULT: return A64CC::LO;
1946 case ISD::SETULE: return A64CC::LS;
1947 default: llvm_unreachable("Unexpected condition code");
1951 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Val) const {
1952 // icmp is implemented using adds/subs immediate, which take an unsigned
1953 // 12-bit immediate, optionally shifted left by 12 bits.
1955 // Symmetric by using adds/subs
1959 return (Val & ~0xfff) == 0 || (Val & ~0xfff000) == 0;
1962 SDValue AArch64TargetLowering::getSelectableIntSetCC(SDValue LHS, SDValue RHS,
1963 ISD::CondCode CC, SDValue &A64cc,
1964 SelectionDAG &DAG, SDLoc &dl) const {
1965 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1967 EVT VT = RHSC->getValueType(0);
1968 bool knownInvalid = false;
1970 // I'm not convinced the rest of LLVM handles these edge cases properly, but
1971 // we can at least get it right.
1972 if (isSignedIntSetCC(CC)) {
1973 C = RHSC->getSExtValue();
1974 } else if (RHSC->getZExtValue() > INT64_MAX) {
1975 // A 64-bit constant not representable by a signed 64-bit integer is far
1976 // too big to fit into a SUBS immediate anyway.
1977 knownInvalid = true;
1979 C = RHSC->getZExtValue();
1982 if (!knownInvalid && !isLegalICmpImmediate(C)) {
1983 // Constant does not fit, try adjusting it by one?
1988 if (isLegalICmpImmediate(C-1)) {
1989 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1990 RHS = DAG.getConstant(C-1, VT);
1995 if (isLegalICmpImmediate(C-1)) {
1996 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1997 RHS = DAG.getConstant(C-1, VT);
2002 if (isLegalICmpImmediate(C+1)) {
2003 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2004 RHS = DAG.getConstant(C+1, VT);
2009 if (isLegalICmpImmediate(C+1)) {
2010 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2011 RHS = DAG.getConstant(C+1, VT);
2018 A64CC::CondCodes CondCode = IntCCToA64CC(CC);
2019 A64cc = DAG.getConstant(CondCode, MVT::i32);
2020 return DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2021 DAG.getCondCode(CC));
2024 static A64CC::CondCodes FPCCToA64CC(ISD::CondCode CC,
2025 A64CC::CondCodes &Alternative) {
2026 A64CC::CondCodes CondCode = A64CC::Invalid;
2027 Alternative = A64CC::Invalid;
2030 default: llvm_unreachable("Unknown FP condition!");
2032 case ISD::SETOEQ: CondCode = A64CC::EQ; break;
2034 case ISD::SETOGT: CondCode = A64CC::GT; break;
2036 case ISD::SETOGE: CondCode = A64CC::GE; break;
2037 case ISD::SETOLT: CondCode = A64CC::MI; break;
2038 case ISD::SETOLE: CondCode = A64CC::LS; break;
2039 case ISD::SETONE: CondCode = A64CC::MI; Alternative = A64CC::GT; break;
2040 case ISD::SETO: CondCode = A64CC::VC; break;
2041 case ISD::SETUO: CondCode = A64CC::VS; break;
2042 case ISD::SETUEQ: CondCode = A64CC::EQ; Alternative = A64CC::VS; break;
2043 case ISD::SETUGT: CondCode = A64CC::HI; break;
2044 case ISD::SETUGE: CondCode = A64CC::PL; break;
2046 case ISD::SETULT: CondCode = A64CC::LT; break;
2048 case ISD::SETULE: CondCode = A64CC::LE; break;
2050 case ISD::SETUNE: CondCode = A64CC::NE; break;
2056 AArch64TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
2058 EVT PtrVT = getPointerTy();
2059 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2061 switch(getTargetMachine().getCodeModel()) {
2062 case CodeModel::Small:
2063 // The most efficient code is PC-relative anyway for the small memory model,
2064 // so we don't need to worry about relocation model.
2065 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2066 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2067 AArch64II::MO_NO_FLAG),
2068 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2069 AArch64II::MO_LO12),
2070 DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
2071 case CodeModel::Large:
2073 AArch64ISD::WrapperLarge, DL, PtrVT,
2074 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G3),
2075 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2076 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2077 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2079 llvm_unreachable("Only small and large code models supported now");
2084 // (BRCOND chain, val, dest)
2086 AArch64TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
2088 SDValue Chain = Op.getOperand(0);
2089 SDValue TheBit = Op.getOperand(1);
2090 SDValue DestBB = Op.getOperand(2);
2092 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2093 // that as the consumer we are responsible for ignoring rubbish in higher
2095 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2096 DAG.getConstant(1, MVT::i32));
2098 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2099 DAG.getConstant(0, TheBit.getValueType()),
2100 DAG.getCondCode(ISD::SETNE));
2102 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other, Chain,
2103 A64CMP, DAG.getConstant(A64CC::NE, MVT::i32),
2107 // (BR_CC chain, condcode, lhs, rhs, dest)
2109 AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2111 SDValue Chain = Op.getOperand(0);
2112 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2113 SDValue LHS = Op.getOperand(2);
2114 SDValue RHS = Op.getOperand(3);
2115 SDValue DestBB = Op.getOperand(4);
2117 if (LHS.getValueType() == MVT::f128) {
2118 // f128 comparisons are lowered to runtime calls by a routine which sets
2119 // LHS, RHS and CC appropriately for the rest of this function to continue.
2120 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2122 // If softenSetCCOperands returned a scalar, we need to compare the result
2123 // against zero to select between true and false values.
2124 if (RHS.getNode() == 0) {
2125 RHS = DAG.getConstant(0, LHS.getValueType());
2130 if (LHS.getValueType().isInteger()) {
2133 // Integers are handled in a separate function because the combinations of
2134 // immediates and tests can get hairy and we may want to fiddle things.
2135 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2137 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2138 Chain, CmpOp, A64cc, DestBB);
2141 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2142 // conditional branch, hence FPCCToA64CC can set a second test, where either
2143 // passing is sufficient.
2144 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2145 CondCode = FPCCToA64CC(CC, Alternative);
2146 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2147 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2148 DAG.getCondCode(CC));
2149 SDValue A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2150 Chain, SetCC, A64cc, DestBB);
2152 if (Alternative != A64CC::Invalid) {
2153 A64cc = DAG.getConstant(Alternative, MVT::i32);
2154 A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2155 A64BR_CC, SetCC, A64cc, DestBB);
2163 AArch64TargetLowering::LowerF128ToCall(SDValue Op, SelectionDAG &DAG,
2164 RTLIB::Libcall Call) const {
2167 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
2168 EVT ArgVT = Op.getOperand(i).getValueType();
2169 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2170 Entry.Node = Op.getOperand(i); Entry.Ty = ArgTy;
2171 Entry.isSExt = false;
2172 Entry.isZExt = false;
2173 Args.push_back(Entry);
2175 SDValue Callee = DAG.getExternalSymbol(getLibcallName(Call), getPointerTy());
2177 Type *RetTy = Op.getValueType().getTypeForEVT(*DAG.getContext());
2179 // By default, the input chain to this libcall is the entry node of the
2180 // function. If the libcall is going to be emitted as a tail call then
2181 // isUsedByReturnOnly will change it to the right chain if the return
2182 // node which is being folded has a non-entry input chain.
2183 SDValue InChain = DAG.getEntryNode();
2185 // isTailCall may be true since the callee does not reference caller stack
2186 // frame. Check if it's in the right position.
2187 SDValue TCChain = InChain;
2188 bool isTailCall = isInTailCallPosition(DAG, Op.getNode(), TCChain);
2193 CallLoweringInfo CLI(InChain, RetTy, false, false, false, false,
2194 0, getLibcallCallingConv(Call), isTailCall,
2195 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
2196 Callee, Args, DAG, SDLoc(Op));
2197 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
2199 if (!CallInfo.second.getNode())
2200 // It's a tailcall, return the chain (which is the DAG root).
2201 return DAG.getRoot();
2203 return CallInfo.first;
2207 AArch64TargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
2208 if (Op.getOperand(0).getValueType() != MVT::f128) {
2209 // It's legal except when f128 is involved
2214 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
2216 SDValue SrcVal = Op.getOperand(0);
2217 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
2218 /*isSigned*/ false, SDLoc(Op)).first;
2222 AArch64TargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
2223 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
2226 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
2228 return LowerF128ToCall(Op, DAG, LC);
2231 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2234 EVT VT = Op.getValueType();
2235 SDValue Vec = Op.getOperand(0);
2236 EVT OpVT = Vec.getValueType();
2237 unsigned Opc = IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
2239 if (VT.getVectorNumElements() == 1) {
2240 assert(OpVT == MVT::v1f64 && "Unexpected vector type!");
2241 if (VT.getSizeInBits() == OpVT.getSizeInBits())
2243 return DAG.UnrollVectorOp(Op.getNode());
2246 if (VT.getSizeInBits() > OpVT.getSizeInBits()) {
2247 assert(Vec.getValueType() == MVT::v2f32 && VT == MVT::v2i64 &&
2248 "Unexpected vector type!");
2249 Vec = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Vec);
2250 return DAG.getNode(Opc, dl, VT, Vec);
2251 } else if (VT.getSizeInBits() < OpVT.getSizeInBits()) {
2252 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2253 OpVT.getVectorElementType().getSizeInBits());
2255 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2256 Vec = DAG.getNode(Opc, dl, CastVT, Vec);
2257 return DAG.getNode(ISD::TRUNCATE, dl, VT, Vec);
2259 return DAG.getNode(Opc, dl, VT, Vec);
2263 AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2264 bool IsSigned) const {
2265 if (Op.getValueType().isVector())
2266 return LowerVectorFP_TO_INT(Op, DAG, IsSigned);
2267 if (Op.getOperand(0).getValueType() != MVT::f128) {
2268 // It's legal except when f128 is involved
2274 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
2276 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
2278 return LowerF128ToCall(Op, DAG, LC);
2281 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2282 MachineFunction &MF = DAG.getMachineFunction();
2283 MachineFrameInfo *MFI = MF.getFrameInfo();
2284 MFI->setReturnAddressIsTaken(true);
2286 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2289 EVT VT = Op.getValueType();
2291 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2293 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2294 SDValue Offset = DAG.getConstant(8, MVT::i64);
2295 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2296 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2297 MachinePointerInfo(), false, false, false, 0);
2300 // Return X30, which contains the return address. Mark it an implicit live-in.
2301 unsigned Reg = MF.addLiveIn(AArch64::X30, getRegClassFor(MVT::i64));
2302 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, MVT::i64);
2306 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG)
2308 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2309 MFI->setFrameAddressIsTaken(true);
2311 EVT VT = Op.getValueType();
2313 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2314 unsigned FrameReg = AArch64::X29;
2315 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2317 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2318 MachinePointerInfo(),
2319 false, false, false, 0);
2324 AArch64TargetLowering::LowerGlobalAddressELFLarge(SDValue Op,
2325 SelectionDAG &DAG) const {
2326 assert(getTargetMachine().getCodeModel() == CodeModel::Large);
2327 assert(getTargetMachine().getRelocationModel() == Reloc::Static);
2329 EVT PtrVT = getPointerTy();
2331 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2332 const GlobalValue *GV = GN->getGlobal();
2334 SDValue GlobalAddr = DAG.getNode(
2335 AArch64ISD::WrapperLarge, dl, PtrVT,
2336 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G3),
2337 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2338 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2339 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2341 if (GN->getOffset() != 0)
2342 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2343 DAG.getConstant(GN->getOffset(), PtrVT));
2349 AArch64TargetLowering::LowerGlobalAddressELFSmall(SDValue Op,
2350 SelectionDAG &DAG) const {
2351 assert(getTargetMachine().getCodeModel() == CodeModel::Small);
2353 EVT PtrVT = getPointerTy();
2355 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2356 const GlobalValue *GV = GN->getGlobal();
2357 unsigned Alignment = GV->getAlignment();
2358 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2359 if (GV->isWeakForLinker() && GV->isDeclaration() && RelocM == Reloc::Static) {
2360 // Weak undefined symbols can't use ADRP/ADD pair since they should evaluate
2361 // to zero when they remain undefined. In PIC mode the GOT can take care of
2362 // this, but in absolute mode we use a constant pool load.
2364 PoolAddr = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2365 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2366 AArch64II::MO_NO_FLAG),
2367 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2368 AArch64II::MO_LO12),
2369 DAG.getConstant(8, MVT::i32));
2370 SDValue GlobalAddr = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), PoolAddr,
2371 MachinePointerInfo::getConstantPool(),
2372 /*isVolatile=*/ false,
2373 /*isNonTemporal=*/ true,
2374 /*isInvariant=*/ true, 8);
2375 if (GN->getOffset() != 0)
2376 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2377 DAG.getConstant(GN->getOffset(), PtrVT));
2382 if (Alignment == 0) {
2383 const PointerType *GVPtrTy = cast<PointerType>(GV->getType());
2384 if (GVPtrTy->getElementType()->isSized()) {
2386 = getDataLayout()->getABITypeAlignment(GVPtrTy->getElementType());
2388 // Be conservative if we can't guess, not that it really matters:
2389 // functions and labels aren't valid for loads, and the methods used to
2390 // actually calculate an address work with any alignment.
2395 unsigned char HiFixup, LoFixup;
2396 bool UseGOT = getSubtarget()->GVIsIndirectSymbol(GV, RelocM);
2399 HiFixup = AArch64II::MO_GOT;
2400 LoFixup = AArch64II::MO_GOT_LO12;
2403 HiFixup = AArch64II::MO_NO_FLAG;
2404 LoFixup = AArch64II::MO_LO12;
2407 // AArch64's small model demands the following sequence:
2408 // ADRP x0, somewhere
2409 // ADD x0, x0, #:lo12:somewhere ; (or LDR directly).
2410 SDValue GlobalRef = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2411 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2413 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2415 DAG.getConstant(Alignment, MVT::i32));
2418 GlobalRef = DAG.getNode(AArch64ISD::GOTLoad, dl, PtrVT, DAG.getEntryNode(),
2422 if (GN->getOffset() != 0)
2423 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalRef,
2424 DAG.getConstant(GN->getOffset(), PtrVT));
2430 AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
2431 SelectionDAG &DAG) const {
2432 // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
2433 // we make those distinctions here.
2435 switch (getTargetMachine().getCodeModel()) {
2436 case CodeModel::Small:
2437 return LowerGlobalAddressELFSmall(Op, DAG);
2438 case CodeModel::Large:
2439 return LowerGlobalAddressELFLarge(Op, DAG);
2441 llvm_unreachable("Only small and large code models supported now");
2446 AArch64TargetLowering::LowerConstantPool(SDValue Op,
2447 SelectionDAG &DAG) const {
2449 EVT PtrVT = getPointerTy();
2450 ConstantPoolSDNode *CN = cast<ConstantPoolSDNode>(Op);
2451 const Constant *C = CN->getConstVal();
2453 switch(getTargetMachine().getCodeModel()) {
2454 case CodeModel::Small:
2455 // The most efficient code is PC-relative anyway for the small memory model,
2456 // so we don't need to worry about relocation model.
2457 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2458 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2459 AArch64II::MO_NO_FLAG),
2460 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2461 AArch64II::MO_LO12),
2462 DAG.getConstant(CN->getAlignment(), MVT::i32));
2463 case CodeModel::Large:
2465 AArch64ISD::WrapperLarge, DL, PtrVT,
2466 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G3),
2467 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G2_NC),
2468 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G1_NC),
2469 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G0_NC));
2471 llvm_unreachable("Only small and large code models supported now");
2475 SDValue AArch64TargetLowering::LowerTLSDescCall(SDValue SymAddr,
2478 SelectionDAG &DAG) const {
2479 EVT PtrVT = getPointerTy();
2481 // The function we need to call is simply the first entry in the GOT for this
2482 // descriptor, load it in preparation.
2483 SDValue Func, Chain;
2484 Func = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2487 // The function takes only one argument: the address of the descriptor itself
2490 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
2491 Glue = Chain.getValue(1);
2493 // Finally, there's a special calling-convention which means that the lookup
2494 // must preserve all registers (except X0, obviously).
2495 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2496 const AArch64RegisterInfo *A64RI
2497 = static_cast<const AArch64RegisterInfo *>(TRI);
2498 const uint32_t *Mask = A64RI->getTLSDescCallPreservedMask();
2500 // We're now ready to populate the argument list, as with a normal call:
2501 std::vector<SDValue> Ops;
2502 Ops.push_back(Chain);
2503 Ops.push_back(Func);
2504 Ops.push_back(SymAddr);
2505 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
2506 Ops.push_back(DAG.getRegisterMask(Mask));
2507 Ops.push_back(Glue);
2509 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2510 Chain = DAG.getNode(AArch64ISD::TLSDESCCALL, DL, NodeTys, &Ops[0],
2512 Glue = Chain.getValue(1);
2514 // After the call, the offset from TPIDR_EL0 is in X0, copy it out and pass it
2515 // back to the generic handling code.
2516 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
2520 AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2521 SelectionDAG &DAG) const {
2522 assert(getSubtarget()->isTargetELF() &&
2523 "TLS not implemented for non-ELF targets");
2524 assert(getTargetMachine().getCodeModel() == CodeModel::Small
2525 && "TLS only supported in small memory model");
2526 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2528 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2531 EVT PtrVT = getPointerTy();
2533 const GlobalValue *GV = GA->getGlobal();
2535 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
2537 if (Model == TLSModel::InitialExec) {
2538 TPOff = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2539 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2540 AArch64II::MO_GOTTPREL),
2541 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2542 AArch64II::MO_GOTTPREL_LO12),
2543 DAG.getConstant(8, MVT::i32));
2544 TPOff = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2546 } else if (Model == TLSModel::LocalExec) {
2547 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2548 AArch64II::MO_TPREL_G1);
2549 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2550 AArch64II::MO_TPREL_G0_NC);
2552 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2553 DAG.getTargetConstant(1, MVT::i32)), 0);
2554 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2556 DAG.getTargetConstant(0, MVT::i32)), 0);
2557 } else if (Model == TLSModel::GeneralDynamic) {
2558 // Accesses used in this sequence go via the TLS descriptor which lives in
2559 // the GOT. Prepare an address we can use to handle this.
2560 SDValue HiDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2561 AArch64II::MO_TLSDESC);
2562 SDValue LoDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2563 AArch64II::MO_TLSDESC_LO12);
2564 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2566 DAG.getConstant(8, MVT::i32));
2567 SDValue SymAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0);
2569 TPOff = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2570 } else if (Model == TLSModel::LocalDynamic) {
2571 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2572 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2573 // the beginning of the module's TLS region, followed by a DTPREL offset
2576 // These accesses will need deduplicating if there's more than one.
2577 AArch64MachineFunctionInfo* MFI = DAG.getMachineFunction()
2578 .getInfo<AArch64MachineFunctionInfo>();
2579 MFI->incNumLocalDynamicTLSAccesses();
2582 // Get the location of _TLS_MODULE_BASE_:
2583 SDValue HiDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2584 AArch64II::MO_TLSDESC);
2585 SDValue LoDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2586 AArch64II::MO_TLSDESC_LO12);
2587 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2589 DAG.getConstant(8, MVT::i32));
2590 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT);
2592 ThreadBase = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2594 // Get the variable's offset from _TLS_MODULE_BASE_
2595 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2596 AArch64II::MO_DTPREL_G1);
2597 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2598 AArch64II::MO_DTPREL_G0_NC);
2600 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2601 DAG.getTargetConstant(0, MVT::i32)), 0);
2602 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2604 DAG.getTargetConstant(0, MVT::i32)), 0);
2606 llvm_unreachable("Unsupported TLS access model");
2609 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2612 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2615 EVT VT = Op.getValueType();
2616 SDValue Vec = Op.getOperand(0);
2617 unsigned Opc = IsSigned ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
2619 if (VT.getVectorNumElements() == 1) {
2620 assert(VT == MVT::v1f64 && "Unexpected vector type!");
2621 if (VT.getSizeInBits() == Vec.getValueSizeInBits())
2623 return DAG.UnrollVectorOp(Op.getNode());
2626 if (VT.getSizeInBits() < Vec.getValueSizeInBits()) {
2627 assert(Vec.getValueType() == MVT::v2i64 && VT == MVT::v2f32 &&
2628 "Unexpected vector type!");
2629 Vec = DAG.getNode(Opc, dl, MVT::v2f64, Vec);
2630 return DAG.getNode(ISD::FP_ROUND, dl, VT, Vec, DAG.getIntPtrConstant(0));
2631 } else if (VT.getSizeInBits() > Vec.getValueSizeInBits()) {
2632 unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2633 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2634 VT.getVectorElementType().getSizeInBits());
2636 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2637 Vec = DAG.getNode(CastOpc, dl, CastVT, Vec);
2640 return DAG.getNode(Opc, dl, VT, Vec);
2644 AArch64TargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2645 bool IsSigned) const {
2646 if (Op.getValueType().isVector())
2647 return LowerVectorINT_TO_FP(Op, DAG, IsSigned);
2648 if (Op.getValueType() != MVT::f128) {
2649 // Legal for everything except f128.
2655 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2657 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2659 return LowerF128ToCall(Op, DAG, LC);
2664 AArch64TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
2665 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2667 EVT PtrVT = getPointerTy();
2669 // When compiling PIC, jump tables get put in the code section so a static
2670 // relocation-style is acceptable for both cases.
2671 switch (getTargetMachine().getCodeModel()) {
2672 case CodeModel::Small:
2673 return DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2674 DAG.getTargetJumpTable(JT->getIndex(), PtrVT),
2675 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
2676 AArch64II::MO_LO12),
2677 DAG.getConstant(1, MVT::i32));
2678 case CodeModel::Large:
2680 AArch64ISD::WrapperLarge, dl, PtrVT,
2681 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G3),
2682 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G2_NC),
2683 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G1_NC),
2684 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G0_NC));
2686 llvm_unreachable("Only small and large code models supported now");
2690 // (SELECT testbit, iftrue, iffalse)
2692 AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2694 SDValue TheBit = Op.getOperand(0);
2695 SDValue IfTrue = Op.getOperand(1);
2696 SDValue IfFalse = Op.getOperand(2);
2698 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2699 // that as the consumer we are responsible for ignoring rubbish in higher
2701 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2702 DAG.getConstant(1, MVT::i32));
2703 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2704 DAG.getConstant(0, TheBit.getValueType()),
2705 DAG.getCondCode(ISD::SETNE));
2707 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2708 A64CMP, IfTrue, IfFalse,
2709 DAG.getConstant(A64CC::NE, MVT::i32));
2712 static SDValue LowerVectorSETCC(SDValue Op, SelectionDAG &DAG) {
2714 SDValue LHS = Op.getOperand(0);
2715 SDValue RHS = Op.getOperand(1);
2716 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2717 EVT VT = Op.getValueType();
2718 bool Invert = false;
2722 if (LHS.getValueType().isInteger()) {
2724 // Attempt to use Vector Integer Compare Mask Test instruction.
2725 // TST = icmp ne (and (op0, op1), zero).
2726 if (CC == ISD::SETNE) {
2727 if (((LHS.getOpcode() == ISD::AND) &&
2728 ISD::isBuildVectorAllZeros(RHS.getNode())) ||
2729 ((RHS.getOpcode() == ISD::AND) &&
2730 ISD::isBuildVectorAllZeros(LHS.getNode()))) {
2732 SDValue AndOp = (LHS.getOpcode() == ISD::AND) ? LHS : RHS;
2733 SDValue NewLHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(0));
2734 SDValue NewRHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(1));
2735 return DAG.getNode(AArch64ISD::NEON_TST, DL, VT, NewLHS, NewRHS);
2739 // Attempt to use Vector Integer Compare Mask against Zero instr (Signed).
2740 // Note: Compare against Zero does not support unsigned predicates.
2741 if ((ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2742 ISD::isBuildVectorAllZeros(LHS.getNode())) &&
2743 !isUnsignedIntSetCC(CC)) {
2745 // If LHS is the zero value, swap operands and CondCode.
2746 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2747 CC = getSetCCSwappedOperands(CC);
2752 // Ensure valid CondCode for Compare Mask against Zero instruction:
2753 // EQ, GE, GT, LE, LT.
2754 if (ISD::SETNE == CC) {
2759 // Using constant type to differentiate integer and FP compares with zero.
2760 Op1 = DAG.getConstant(0, MVT::i32);
2761 Opcode = AArch64ISD::NEON_CMPZ;
2764 // Attempt to use Vector Integer Compare Mask instr (Signed/Unsigned).
2765 // Ensure valid CondCode for Compare Mask instr: EQ, GE, GT, UGE, UGT.
2769 llvm_unreachable("Illegal integer comparison.");
2785 CC = getSetCCSwappedOperands(CC);
2789 std::swap(LHS, RHS);
2791 Opcode = AArch64ISD::NEON_CMP;
2796 // Generate Compare Mask instr or Compare Mask against Zero instr.
2798 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
2801 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
2806 // Now handle Floating Point cases.
2807 // Attempt to use Vector Floating Point Compare Mask against Zero instruction.
2808 if (ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2809 ISD::isBuildVectorAllZeros(LHS.getNode())) {
2811 // If LHS is the zero value, swap operands and CondCode.
2812 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2813 CC = getSetCCSwappedOperands(CC);
2818 // Using constant type to differentiate integer and FP compares with zero.
2819 Op1 = DAG.getConstantFP(0, MVT::f32);
2820 Opcode = AArch64ISD::NEON_CMPZ;
2822 // Attempt to use Vector Floating Point Compare Mask instruction.
2825 Opcode = AArch64ISD::NEON_CMP;
2829 // Some register compares have to be implemented with swapped CC and operands,
2830 // e.g.: OLT implemented as OGT with swapped operands.
2831 bool SwapIfRegArgs = false;
2833 // Ensure valid CondCode for FP Compare Mask against Zero instruction:
2834 // EQ, GE, GT, LE, LT.
2835 // And ensure valid CondCode for FP Compare Mask instruction: EQ, GE, GT.
2838 llvm_unreachable("Illegal FP comparison");
2841 Invert = true; // Fallthrough
2849 SwapIfRegArgs = true;
2858 SwapIfRegArgs = true;
2867 SwapIfRegArgs = true;
2876 SwapIfRegArgs = true;
2883 Invert = true; // Fallthrough
2885 // Expand this to (OGT |OLT).
2887 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGT));
2889 SwapIfRegArgs = true;
2892 Invert = true; // Fallthrough
2894 // Expand this to (OGE | OLT).
2896 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGE));
2898 SwapIfRegArgs = true;
2902 if (Opcode == AArch64ISD::NEON_CMP && SwapIfRegArgs) {
2903 CC = getSetCCSwappedOperands(CC);
2904 std::swap(Op0, Op1);
2907 // Generate FP Compare Mask instr or FP Compare Mask against Zero instr
2908 SDValue NeonCmp = DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
2910 if (NeonCmpAlt.getNode())
2911 NeonCmp = DAG.getNode(ISD::OR, DL, VT, NeonCmp, NeonCmpAlt);
2914 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
2919 // (SETCC lhs, rhs, condcode)
2921 AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2923 SDValue LHS = Op.getOperand(0);
2924 SDValue RHS = Op.getOperand(1);
2925 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2926 EVT VT = Op.getValueType();
2929 return LowerVectorSETCC(Op, DAG);
2931 if (LHS.getValueType() == MVT::f128) {
2932 // f128 comparisons will be lowered to libcalls giving a valid LHS and RHS
2933 // for the rest of the function (some i32 or i64 values).
2934 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2936 // If softenSetCCOperands returned a scalar, use it.
2937 if (RHS.getNode() == 0) {
2938 assert(LHS.getValueType() == Op.getValueType() &&
2939 "Unexpected setcc expansion!");
2944 if (LHS.getValueType().isInteger()) {
2947 // Integers are handled in a separate function because the combinations of
2948 // immediates and tests can get hairy and we may want to fiddle things.
2949 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2951 return DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
2952 CmpOp, DAG.getConstant(1, VT), DAG.getConstant(0, VT),
2956 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2957 // conditional branch, hence FPCCToA64CC can set a second test, where either
2958 // passing is sufficient.
2959 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2960 CondCode = FPCCToA64CC(CC, Alternative);
2961 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2962 SDValue CmpOp = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2963 DAG.getCondCode(CC));
2964 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
2965 CmpOp, DAG.getConstant(1, VT),
2966 DAG.getConstant(0, VT), A64cc);
2968 if (Alternative != A64CC::Invalid) {
2969 A64cc = DAG.getConstant(Alternative, MVT::i32);
2970 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
2971 DAG.getConstant(1, VT), A64SELECT_CC, A64cc);
2974 return A64SELECT_CC;
2977 static SDValue LowerVectorSELECT_CC(SDValue Op, SelectionDAG &DAG) {
2979 SDValue LHS = Op.getOperand(0);
2980 SDValue RHS = Op.getOperand(1);
2981 SDValue IfTrue = Op.getOperand(2);
2982 SDValue IfFalse = Op.getOperand(3);
2983 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2985 // If LHS & RHS are floating point and IfTrue & IfFalse are vectors, we will
2986 // use NEON compare.
2987 if ((LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64)) {
2988 EVT EltVT = LHS.getValueType();
2989 unsigned EltNum = 128 / EltVT.getSizeInBits();
2990 EVT VT = EVT::getVectorVT(*DAG.getContext(), EltVT, EltNum);
2991 unsigned SubConstant =
2992 (LHS.getValueType() == MVT::f32) ? AArch64::sub_32 :AArch64::sub_64;
2993 EVT CEltT = (LHS.getValueType() == MVT::f32) ? MVT::i32 : MVT::i64;
2994 EVT CVT = EVT::getVectorVT(*DAG.getContext(), CEltT, EltNum);
2997 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
2998 VT, DAG.getTargetConstant(0, MVT::i32), LHS,
2999 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3001 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
3002 VT, DAG.getTargetConstant(0, MVT::i32), RHS,
3003 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3005 SDValue VSetCC = DAG.getSetCC(dl, CVT, LHS, RHS, CC);
3006 SDValue ResCC = LowerVectorSETCC(VSetCC, DAG);
3007 EVT IfTrueVT = IfTrue.getValueType();
3009 MVT::getIntegerVT(IfTrueVT.getVectorElementType().getSizeInBits());
3010 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), CastEltT,
3011 IfTrueVT.getVectorNumElements());
3012 if (CEltT.getSizeInBits() < IfTrueVT.getSizeInBits()) {
3014 EVT::getVectorVT(*DAG.getContext(), CEltT,
3015 IfTrueVT.getSizeInBits() / CEltT.getSizeInBits());
3016 ResCC = DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, DUPVT, ResCC,
3017 DAG.getConstant(0, MVT::i64, false));
3019 ResCC = DAG.getNode(ISD::BITCAST, dl, CastVT, ResCC);
3021 // FIXME: If IfTrue & IfFalse hold v1i8, v1i16 or v1i32, this function
3022 // can't handle them and will hit this assert.
3023 assert(CEltT.getSizeInBits() == IfTrueVT.getSizeInBits() &&
3024 "Vector of IfTrue & IfFalse is too small.");
3027 EltNum * IfTrueVT.getSizeInBits() / ResCC.getValueSizeInBits();
3028 EVT ExVT = EVT::getVectorVT(*DAG.getContext(), CEltT, ExEltNum);
3029 ResCC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ExVT, ResCC,
3030 DAG.getConstant(0, MVT::i64, false));
3031 ResCC = DAG.getNode(ISD::BITCAST, dl, CastVT, ResCC);
3033 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3034 ResCC, IfTrue, IfFalse);
3038 // Here we handle the case that LHS & RHS are integer and IfTrue & IfFalse are
3040 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3041 CondCode = FPCCToA64CC(CC, Alternative);
3042 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3043 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3044 DAG.getCondCode(CC));
3045 EVT SEVT = MVT::i32;
3046 if (IfTrue.getValueType().getVectorElementType().getSizeInBits() > 32)
3048 SDValue AllOne = DAG.getConstant(-1, SEVT);
3049 SDValue AllZero = DAG.getConstant(0, SEVT);
3050 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, SEVT, SetCC,
3051 AllOne, AllZero, A64cc);
3053 if (Alternative != A64CC::Invalid) {
3054 A64cc = DAG.getConstant(Alternative, MVT::i32);
3055 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3056 SetCC, AllOne, A64SELECT_CC, A64cc);
3059 if (IfTrue.getValueType().getVectorNumElements() == 1)
3060 VDup = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, IfTrue.getValueType(),
3063 VDup = DAG.getNode(AArch64ISD::NEON_VDUP, dl, IfTrue.getValueType(),
3065 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3066 VDup, IfTrue, IfFalse);
3070 // (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
3072 AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3074 SDValue LHS = Op.getOperand(0);
3075 SDValue RHS = Op.getOperand(1);
3076 SDValue IfTrue = Op.getOperand(2);
3077 SDValue IfFalse = Op.getOperand(3);
3078 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3080 if (IfTrue.getValueType().isVector())
3081 return LowerVectorSELECT_CC(Op, DAG);
3083 if (LHS.getValueType() == MVT::f128) {
3084 // f128 comparisons are lowered to libcalls, but slot in nicely here
3086 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3088 // If softenSetCCOperands returned a scalar, we need to compare the result
3089 // against zero to select between true and false values.
3090 if (RHS.getNode() == 0) {
3091 RHS = DAG.getConstant(0, LHS.getValueType());
3096 if (LHS.getValueType().isInteger()) {
3099 // Integers are handled in a separate function because the combinations of
3100 // immediates and tests can get hairy and we may want to fiddle things.
3101 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
3103 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(), CmpOp,
3104 IfTrue, IfFalse, A64cc);
3107 // Note that some LLVM floating-point CondCodes can't be lowered to a single
3108 // conditional branch, hence FPCCToA64CC can set a second test, where either
3109 // passing is sufficient.
3110 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3111 CondCode = FPCCToA64CC(CC, Alternative);
3112 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3113 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3114 DAG.getCondCode(CC));
3115 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
3117 SetCC, IfTrue, IfFalse, A64cc);
3119 if (Alternative != A64CC::Invalid) {
3120 A64cc = DAG.getConstant(Alternative, MVT::i32);
3121 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3122 SetCC, IfTrue, A64SELECT_CC, A64cc);
3126 return A64SELECT_CC;
3130 AArch64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3131 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3132 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3134 // We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
3135 // rather than just 8.
3136 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op),
3137 Op.getOperand(1), Op.getOperand(2),
3138 DAG.getConstant(32, MVT::i32), 8, false, false,
3139 MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV));
3143 AArch64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3144 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3145 // Standard, section B.3.
3146 MachineFunction &MF = DAG.getMachineFunction();
3147 AArch64MachineFunctionInfo *FuncInfo
3148 = MF.getInfo<AArch64MachineFunctionInfo>();
3151 SDValue Chain = Op.getOperand(0);
3152 SDValue VAList = Op.getOperand(1);
3153 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3154 SmallVector<SDValue, 4> MemOps;
3156 // void *__stack at offset 0
3157 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVariadicStackIdx(),
3159 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3160 MachinePointerInfo(SV), false, false, 0));
3162 // void *__gr_top at offset 8
3163 int GPRSize = FuncInfo->getVariadicGPRSize();
3165 SDValue GRTop, GRTopAddr;
3167 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3168 DAG.getConstant(8, getPointerTy()));
3170 GRTop = DAG.getFrameIndex(FuncInfo->getVariadicGPRIdx(), getPointerTy());
3171 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3172 DAG.getConstant(GPRSize, getPointerTy()));
3174 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3175 MachinePointerInfo(SV, 8),
3179 // void *__vr_top at offset 16
3180 int FPRSize = FuncInfo->getVariadicFPRSize();
3182 SDValue VRTop, VRTopAddr;
3183 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3184 DAG.getConstant(16, getPointerTy()));
3186 VRTop = DAG.getFrameIndex(FuncInfo->getVariadicFPRIdx(), getPointerTy());
3187 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3188 DAG.getConstant(FPRSize, getPointerTy()));
3190 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3191 MachinePointerInfo(SV, 16),
3195 // int __gr_offs at offset 24
3196 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3197 DAG.getConstant(24, getPointerTy()));
3198 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3199 GROffsAddr, MachinePointerInfo(SV, 24),
3202 // int __vr_offs at offset 28
3203 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3204 DAG.getConstant(28, getPointerTy()));
3205 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3206 VROffsAddr, MachinePointerInfo(SV, 28),
3209 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
3214 AArch64TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3215 switch (Op.getOpcode()) {
3216 default: llvm_unreachable("Don't know how to custom lower this!");
3217 case ISD::FADD: return LowerF128ToCall(Op, DAG, RTLIB::ADD_F128);
3218 case ISD::FSUB: return LowerF128ToCall(Op, DAG, RTLIB::SUB_F128);
3219 case ISD::FMUL: return LowerF128ToCall(Op, DAG, RTLIB::MUL_F128);
3220 case ISD::FDIV: return LowerF128ToCall(Op, DAG, RTLIB::DIV_F128);
3221 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, true);
3222 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG, false);
3223 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG, true);
3224 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG, false);
3225 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
3226 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
3227 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3228 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3230 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3231 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3232 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3233 case ISD::GlobalAddress: return LowerGlobalAddressELF(Op, DAG);
3234 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3235 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3236 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3237 case ISD::SELECT: return LowerSELECT(Op, DAG);
3238 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3239 case ISD::SETCC: return LowerSETCC(Op, DAG);
3240 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
3241 case ISD::VASTART: return LowerVASTART(Op, DAG);
3242 case ISD::BUILD_VECTOR:
3243 return LowerBUILD_VECTOR(Op, DAG, getSubtarget());
3244 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3250 /// Check if the specified splat value corresponds to a valid vector constant
3251 /// for a Neon instruction with a "modified immediate" operand (e.g., MOVI). If
3252 /// so, return the encoded 8-bit immediate and the OpCmode instruction fields
3254 static bool isNeonModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3255 unsigned SplatBitSize, SelectionDAG &DAG,
3256 bool is128Bits, NeonModImmType type, EVT &VT,
3257 unsigned &Imm, unsigned &OpCmode) {
3258 switch (SplatBitSize) {
3260 llvm_unreachable("unexpected size for isNeonModifiedImm");
3262 if (type != Neon_Mov_Imm)
3264 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3265 // Neon movi per byte: Op=0, Cmode=1110.
3268 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3272 // Neon move inst per halfword
3273 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3274 if ((SplatBits & ~0xff) == 0) {
3275 // Value = 0x00nn is 0x00nn LSL 0
3276 // movi: Op=0, Cmode=1000; mvni: Op=1, Cmode=1000
3277 // bic: Op=1, Cmode=1001; orr: Op=0, Cmode=1001
3283 if ((SplatBits & ~0xff00) == 0) {
3284 // Value = 0xnn00 is 0x00nn LSL 8
3285 // movi: Op=0, Cmode=1010; mvni: Op=1, Cmode=1010
3286 // bic: Op=1, Cmode=1011; orr: Op=0, Cmode=1011
3288 Imm = SplatBits >> 8;
3292 // can't handle any other
3297 // First the LSL variants (MSL is unusable by some interested instructions).
3299 // Neon move instr per word, shift zeros
3300 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3301 if ((SplatBits & ~0xff) == 0) {
3302 // Value = 0x000000nn is 0x000000nn LSL 0
3303 // movi: Op=0, Cmode= 0000; mvni: Op=1, Cmode= 0000
3304 // bic: Op=1, Cmode= 0001; orr: Op=0, Cmode= 0001
3310 if ((SplatBits & ~0xff00) == 0) {
3311 // Value = 0x0000nn00 is 0x000000nn LSL 8
3312 // movi: Op=0, Cmode= 0010; mvni: Op=1, Cmode= 0010
3313 // bic: Op=1, Cmode= 0011; orr : Op=0, Cmode= 0011
3315 Imm = SplatBits >> 8;
3319 if ((SplatBits & ~0xff0000) == 0) {
3320 // Value = 0x00nn0000 is 0x000000nn LSL 16
3321 // movi: Op=0, Cmode= 0100; mvni: Op=1, Cmode= 0100
3322 // bic: Op=1, Cmode= 0101; orr: Op=0, Cmode= 0101
3324 Imm = SplatBits >> 16;
3328 if ((SplatBits & ~0xff000000) == 0) {
3329 // Value = 0xnn000000 is 0x000000nn LSL 24
3330 // movi: Op=0, Cmode= 0110; mvni: Op=1, Cmode= 0110
3331 // bic: Op=1, Cmode= 0111; orr: Op=0, Cmode= 0111
3333 Imm = SplatBits >> 24;
3338 // Now the MSL immediates.
3340 // Neon move instr per word, shift ones
3341 if ((SplatBits & ~0xffff) == 0 &&
3342 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3343 // Value = 0x0000nnff is 0x000000nn MSL 8
3344 // movi: Op=0, Cmode= 1100; mvni: Op=1, Cmode= 1100
3346 Imm = SplatBits >> 8;
3350 if ((SplatBits & ~0xffffff) == 0 &&
3351 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3352 // Value = 0x00nnffff is 0x000000nn MSL 16
3353 // movi: Op=1, Cmode= 1101; mvni: Op=1, Cmode= 1101
3355 Imm = SplatBits >> 16;
3359 // can't handle any other
3364 if (type != Neon_Mov_Imm)
3366 // Neon move instr bytemask, where each byte is either 0x00 or 0xff.
3367 // movi Op=1, Cmode=1110.
3369 uint64_t BitMask = 0xff;
3371 unsigned ImmMask = 1;
3373 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3374 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3377 } else if ((SplatBits & BitMask) != 0) {
3384 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3392 static SDValue PerformANDCombine(SDNode *N,
3393 TargetLowering::DAGCombinerInfo &DCI) {
3395 SelectionDAG &DAG = DCI.DAG;
3397 EVT VT = N->getValueType(0);
3399 // We're looking for an SRA/SHL pair which form an SBFX.
3401 if (VT != MVT::i32 && VT != MVT::i64)
3404 if (!isa<ConstantSDNode>(N->getOperand(1)))
3407 uint64_t TruncMask = N->getConstantOperandVal(1);
3408 if (!isMask_64(TruncMask))
3411 uint64_t Width = CountPopulation_64(TruncMask);
3412 SDValue Shift = N->getOperand(0);
3414 if (Shift.getOpcode() != ISD::SRL)
3417 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3419 uint64_t LSB = Shift->getConstantOperandVal(1);
3421 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3424 return DAG.getNode(AArch64ISD::UBFX, DL, VT, Shift.getOperand(0),
3425 DAG.getConstant(LSB, MVT::i64),
3426 DAG.getConstant(LSB + Width - 1, MVT::i64));
3429 /// For a true bitfield insert, the bits getting into that contiguous mask
3430 /// should come from the low part of an existing value: they must be formed from
3431 /// a compatible SHL operation (unless they're already low). This function
3432 /// checks that condition and returns the least-significant bit that's
3433 /// intended. If the operation not a field preparation, -1 is returned.
3434 static int32_t getLSBForBFI(SelectionDAG &DAG, SDLoc DL, EVT VT,
3435 SDValue &MaskedVal, uint64_t Mask) {
3436 if (!isShiftedMask_64(Mask))
3439 // Now we need to alter MaskedVal so that it is an appropriate input for a BFI
3440 // instruction. BFI will do a left-shift by LSB before applying the mask we've
3441 // spotted, so in general we should pre-emptively "undo" that by making sure
3442 // the incoming bits have had a right-shift applied to them.
3444 // This right shift, however, will combine with existing left/right shifts. In
3445 // the simplest case of a completely straight bitfield operation, it will be
3446 // expected to completely cancel out with an existing SHL. More complicated
3447 // cases (e.g. bitfield to bitfield copy) may still need a real shift before
3450 uint64_t LSB = countTrailingZeros(Mask);
3451 int64_t ShiftRightRequired = LSB;
3452 if (MaskedVal.getOpcode() == ISD::SHL &&
3453 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3454 ShiftRightRequired -= MaskedVal.getConstantOperandVal(1);
3455 MaskedVal = MaskedVal.getOperand(0);
3456 } else if (MaskedVal.getOpcode() == ISD::SRL &&
3457 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3458 ShiftRightRequired += MaskedVal.getConstantOperandVal(1);
3459 MaskedVal = MaskedVal.getOperand(0);
3462 if (ShiftRightRequired > 0)
3463 MaskedVal = DAG.getNode(ISD::SRL, DL, VT, MaskedVal,
3464 DAG.getConstant(ShiftRightRequired, MVT::i64));
3465 else if (ShiftRightRequired < 0) {
3466 // We could actually end up with a residual left shift, for example with
3467 // "struc.bitfield = val << 1".
3468 MaskedVal = DAG.getNode(ISD::SHL, DL, VT, MaskedVal,
3469 DAG.getConstant(-ShiftRightRequired, MVT::i64));
3475 /// Searches from N for an existing AArch64ISD::BFI node, possibly surrounded by
3476 /// a mask and an extension. Returns true if a BFI was found and provides
3477 /// information on its surroundings.
3478 static bool findMaskedBFI(SDValue N, SDValue &BFI, uint64_t &Mask,
3481 if (N.getOpcode() == ISD::ZERO_EXTEND) {
3483 N = N.getOperand(0);
3486 if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
3487 Mask = N->getConstantOperandVal(1);
3488 N = N.getOperand(0);
3490 // Mask is the whole width.
3491 Mask = -1ULL >> (64 - N.getValueType().getSizeInBits());
3494 if (N.getOpcode() == AArch64ISD::BFI) {
3502 /// Try to combine a subtree (rooted at an OR) into a "masked BFI" node, which
3503 /// is roughly equivalent to (and (BFI ...), mask). This form is used because it
3504 /// can often be further combined with a larger mask. Ultimately, we want mask
3505 /// to be 2^32-1 or 2^64-1 so the AND can be skipped.
3506 static SDValue tryCombineToBFI(SDNode *N,
3507 TargetLowering::DAGCombinerInfo &DCI,
3508 const AArch64Subtarget *Subtarget) {
3509 SelectionDAG &DAG = DCI.DAG;
3511 EVT VT = N->getValueType(0);
3513 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3515 // We need the LHS to be (and SOMETHING, MASK). Find out what that mask is or
3516 // abandon the effort.
3517 SDValue LHS = N->getOperand(0);
3518 if (LHS.getOpcode() != ISD::AND)
3522 if (isa<ConstantSDNode>(LHS.getOperand(1)))
3523 LHSMask = LHS->getConstantOperandVal(1);
3527 // We also need the RHS to be (and SOMETHING, MASK). Find out what that mask
3528 // is or abandon the effort.
3529 SDValue RHS = N->getOperand(1);
3530 if (RHS.getOpcode() != ISD::AND)
3534 if (isa<ConstantSDNode>(RHS.getOperand(1)))
3535 RHSMask = RHS->getConstantOperandVal(1);
3539 // Can't do anything if the masks are incompatible.
3540 if (LHSMask & RHSMask)
3543 // Now we need one of the masks to be a contiguous field. Without loss of
3544 // generality that should be the RHS one.
3545 SDValue Bitfield = LHS.getOperand(0);
3546 if (getLSBForBFI(DAG, DL, VT, Bitfield, LHSMask) != -1) {
3547 // We know that LHS is a candidate new value, and RHS isn't already a better
3549 std::swap(LHS, RHS);
3550 std::swap(LHSMask, RHSMask);
3553 // We've done our best to put the right operands in the right places, all we
3554 // can do now is check whether a BFI exists.
3555 Bitfield = RHS.getOperand(0);
3556 int32_t LSB = getLSBForBFI(DAG, DL, VT, Bitfield, RHSMask);
3560 uint32_t Width = CountPopulation_64(RHSMask);
3561 assert(Width && "Expected non-zero bitfield width");
3563 SDValue BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3564 LHS.getOperand(0), Bitfield,
3565 DAG.getConstant(LSB, MVT::i64),
3566 DAG.getConstant(Width, MVT::i64));
3569 if ((LHSMask | RHSMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3572 return DAG.getNode(ISD::AND, DL, VT, BFI,
3573 DAG.getConstant(LHSMask | RHSMask, VT));
3576 /// Search for the bitwise combining (with careful masks) of a MaskedBFI and its
3577 /// original input. This is surprisingly common because SROA splits things up
3578 /// into i8 chunks, so the originally detected MaskedBFI may actually only act
3579 /// on the low (say) byte of a word. This is then orred into the rest of the
3580 /// word afterwards.
3582 /// Basic input: (or (and OLDFIELD, MASK1), (MaskedBFI MASK2, OLDFIELD, ...)).
3584 /// If MASK1 and MASK2 are compatible, we can fold the whole thing into the
3585 /// MaskedBFI. We can also deal with a certain amount of extend/truncate being
3587 static SDValue tryCombineToLargerBFI(SDNode *N,
3588 TargetLowering::DAGCombinerInfo &DCI,
3589 const AArch64Subtarget *Subtarget) {
3590 SelectionDAG &DAG = DCI.DAG;
3592 EVT VT = N->getValueType(0);
3594 // First job is to hunt for a MaskedBFI on either the left or right. Swap
3595 // operands if it's actually on the right.
3597 SDValue PossExtraMask;
3598 uint64_t ExistingMask = 0;
3599 bool Extended = false;
3600 if (findMaskedBFI(N->getOperand(0), BFI, ExistingMask, Extended))
3601 PossExtraMask = N->getOperand(1);
3602 else if (findMaskedBFI(N->getOperand(1), BFI, ExistingMask, Extended))
3603 PossExtraMask = N->getOperand(0);
3607 // We can only combine a BFI with another compatible mask.
3608 if (PossExtraMask.getOpcode() != ISD::AND ||
3609 !isa<ConstantSDNode>(PossExtraMask.getOperand(1)))
3612 uint64_t ExtraMask = PossExtraMask->getConstantOperandVal(1);
3614 // Masks must be compatible.
3615 if (ExtraMask & ExistingMask)
3618 SDValue OldBFIVal = BFI.getOperand(0);
3619 SDValue NewBFIVal = BFI.getOperand(1);
3621 // We skipped a ZERO_EXTEND above, so the input to the MaskedBFIs should be
3622 // 32-bit and we'll be forming a 64-bit MaskedBFI. The MaskedBFI arguments
3623 // need to be made compatible.
3624 assert(VT == MVT::i64 && BFI.getValueType() == MVT::i32
3625 && "Invalid types for BFI");
3626 OldBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, OldBFIVal);
3627 NewBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, NewBFIVal);
3630 // We need the MaskedBFI to be combined with a mask of the *same* value.
3631 if (PossExtraMask.getOperand(0) != OldBFIVal)
3634 BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3635 OldBFIVal, NewBFIVal,
3636 BFI.getOperand(2), BFI.getOperand(3));
3638 // If the masking is trivial, we don't need to create it.
3639 if ((ExtraMask | ExistingMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3642 return DAG.getNode(ISD::AND, DL, VT, BFI,
3643 DAG.getConstant(ExtraMask | ExistingMask, VT));
3646 /// An EXTR instruction is made up of two shifts, ORed together. This helper
3647 /// searches for and classifies those shifts.
3648 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
3650 if (N.getOpcode() == ISD::SHL)
3652 else if (N.getOpcode() == ISD::SRL)
3657 if (!isa<ConstantSDNode>(N.getOperand(1)))
3660 ShiftAmount = N->getConstantOperandVal(1);
3661 Src = N->getOperand(0);
3665 /// EXTR instruction extracts a contiguous chunk of bits from two existing
3666 /// registers viewed as a high/low pair. This function looks for the pattern:
3667 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
3668 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
3670 static SDValue tryCombineToEXTR(SDNode *N,
3671 TargetLowering::DAGCombinerInfo &DCI) {
3672 SelectionDAG &DAG = DCI.DAG;
3674 EVT VT = N->getValueType(0);
3676 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3678 if (VT != MVT::i32 && VT != MVT::i64)
3682 uint32_t ShiftLHS = 0;
3684 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
3688 uint32_t ShiftRHS = 0;
3690 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
3693 // If they're both trying to come from the high part of the register, they're
3694 // not really an EXTR.
3695 if (LHSFromHi == RHSFromHi)
3698 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
3702 std::swap(LHS, RHS);
3703 std::swap(ShiftLHS, ShiftRHS);
3706 return DAG.getNode(AArch64ISD::EXTR, DL, VT,
3708 DAG.getConstant(ShiftRHS, MVT::i64));
3711 /// Target-specific dag combine xforms for ISD::OR
3712 static SDValue PerformORCombine(SDNode *N,
3713 TargetLowering::DAGCombinerInfo &DCI,
3714 const AArch64Subtarget *Subtarget) {
3716 SelectionDAG &DAG = DCI.DAG;
3718 EVT VT = N->getValueType(0);
3720 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
3723 // Attempt to recognise bitfield-insert operations.
3724 SDValue Res = tryCombineToBFI(N, DCI, Subtarget);
3728 // Attempt to combine an existing MaskedBFI operation into one with a larger
3730 Res = tryCombineToLargerBFI(N, DCI, Subtarget);
3734 Res = tryCombineToEXTR(N, DCI);
3738 if (!Subtarget->hasNEON())
3741 // Attempt to use vector immediate-form BSL
3742 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
3744 SDValue N0 = N->getOperand(0);
3745 if (N0.getOpcode() != ISD::AND)
3748 SDValue N1 = N->getOperand(1);
3749 if (N1.getOpcode() != ISD::AND)
3752 if (VT.isVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
3754 unsigned SplatBitSize;
3756 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
3758 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
3761 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
3763 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
3764 HasAnyUndefs) && !HasAnyUndefs &&
3765 SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
3766 SplatBits0 == ~SplatBits1) {
3768 return DAG.getNode(ISD::VSELECT, DL, VT, N0->getOperand(1),
3769 N0->getOperand(0), N1->getOperand(0));
3777 /// Target-specific dag combine xforms for ISD::SRA
3778 static SDValue PerformSRACombine(SDNode *N,
3779 TargetLowering::DAGCombinerInfo &DCI) {
3781 SelectionDAG &DAG = DCI.DAG;
3783 EVT VT = N->getValueType(0);
3785 // We're looking for an SRA/SHL pair which form an SBFX.
3787 if (VT != MVT::i32 && VT != MVT::i64)
3790 if (!isa<ConstantSDNode>(N->getOperand(1)))
3793 uint64_t ExtraSignBits = N->getConstantOperandVal(1);
3794 SDValue Shift = N->getOperand(0);
3796 if (Shift.getOpcode() != ISD::SHL)
3799 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3802 uint64_t BitsOnLeft = Shift->getConstantOperandVal(1);
3803 uint64_t Width = VT.getSizeInBits() - ExtraSignBits;
3804 uint64_t LSB = VT.getSizeInBits() - Width - BitsOnLeft;
3806 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3809 return DAG.getNode(AArch64ISD::SBFX, DL, VT, Shift.getOperand(0),
3810 DAG.getConstant(LSB, MVT::i64),
3811 DAG.getConstant(LSB + Width - 1, MVT::i64));
3814 /// Check if this is a valid build_vector for the immediate operand of
3815 /// a vector shift operation, where all the elements of the build_vector
3816 /// must have the same constant integer value.
3817 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3818 // Ignore bit_converts.
3819 while (Op.getOpcode() == ISD::BITCAST)
3820 Op = Op.getOperand(0);
3821 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3822 APInt SplatBits, SplatUndef;
3823 unsigned SplatBitSize;
3825 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3826 HasAnyUndefs, ElementBits) ||
3827 SplatBitSize > ElementBits)
3829 Cnt = SplatBits.getSExtValue();
3833 /// Check if this is a valid build_vector for the immediate operand of
3834 /// a vector shift left operation. That value must be in the range:
3835 /// 0 <= Value < ElementBits
3836 static bool isVShiftLImm(SDValue Op, EVT VT, int64_t &Cnt) {
3837 assert(VT.isVector() && "vector shift count is not a vector type");
3838 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3839 if (!getVShiftImm(Op, ElementBits, Cnt))
3841 return (Cnt >= 0 && Cnt < ElementBits);
3844 /// Check if this is a valid build_vector for the immediate operand of a
3845 /// vector shift right operation. The value must be in the range:
3846 /// 1 <= Value <= ElementBits
3847 static bool isVShiftRImm(SDValue Op, EVT VT, int64_t &Cnt) {
3848 assert(VT.isVector() && "vector shift count is not a vector type");
3849 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3850 if (!getVShiftImm(Op, ElementBits, Cnt))
3852 return (Cnt >= 1 && Cnt <= ElementBits);
3855 static SDValue GenForSextInreg(SDNode *N,
3856 TargetLowering::DAGCombinerInfo &DCI,
3857 EVT SrcVT, EVT DestVT, EVT SubRegVT,
3858 const int *Mask, SDValue Src) {
3859 SelectionDAG &DAG = DCI.DAG;
3861 = DAG.getNode(ISD::BITCAST, SDLoc(N), SrcVT, Src);
3863 = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), DestVT, Bitcast);
3865 = DAG.getVectorShuffle(DestVT, SDLoc(N), Sext, DAG.getUNDEF(DestVT), Mask);
3866 SDValue ExtractSubreg
3867 = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, SDLoc(N),
3868 SubRegVT, ShuffleVec,
3869 DAG.getTargetConstant(AArch64::sub_64, MVT::i32)), 0);
3870 return ExtractSubreg;
3873 /// Checks for vector shifts and lowers them.
3874 static SDValue PerformShiftCombine(SDNode *N,
3875 TargetLowering::DAGCombinerInfo &DCI,
3876 const AArch64Subtarget *ST) {
3877 SelectionDAG &DAG = DCI.DAG;
3878 EVT VT = N->getValueType(0);
3879 if (N->getOpcode() == ISD::SRA && (VT == MVT::i32 || VT == MVT::i64))
3880 return PerformSRACombine(N, DCI);
3882 // We're looking for an SRA/SHL pair to help generating instruction
3883 // sshll v0.8h, v0.8b, #0
3884 // The instruction STXL is also the alias of this instruction.
3886 // For example, for DAG like below,
3887 // v2i32 = sra (v2i32 (shl v2i32, 16)), 16
3888 // we can transform it into
3889 // v2i32 = EXTRACT_SUBREG
3890 // (v4i32 (suffle_vector
3891 // (v4i32 (sext (v4i16 (bitcast v2i32))),
3892 // undef, (0, 2, u, u)),
3895 // With this transformation we expect to generate "SSHLL + UZIP1"
3896 // Sometimes UZIP1 can be optimized away by combining with other context.
3897 int64_t ShrCnt, ShlCnt;
3898 if (N->getOpcode() == ISD::SRA
3899 && (VT == MVT::v2i32 || VT == MVT::v4i16)
3900 && isVShiftRImm(N->getOperand(1), VT, ShrCnt)
3901 && N->getOperand(0).getOpcode() == ISD::SHL
3902 && isVShiftRImm(N->getOperand(0).getOperand(1), VT, ShlCnt)) {
3903 SDValue Src = N->getOperand(0).getOperand(0);
3904 if (VT == MVT::v2i32 && ShrCnt == 16 && ShlCnt == 16) {
3905 // sext_inreg(v2i32, v2i16)
3906 // We essentially only care the Mask {0, 2, u, u}
3907 int Mask[4] = {0, 2, 4, 6};
3908 return GenForSextInreg(N, DCI, MVT::v4i16, MVT::v4i32, MVT::v2i32,
3911 else if (VT == MVT::v2i32 && ShrCnt == 24 && ShlCnt == 24) {
3912 // sext_inreg(v2i16, v2i8)
3913 // We essentially only care the Mask {0, u, 4, u, u, u, u, u, u, u, u, u}
3914 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
3915 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v2i32,
3918 else if (VT == MVT::v4i16 && ShrCnt == 8 && ShlCnt == 8) {
3919 // sext_inreg(v4i16, v4i8)
3920 // We essentially only care the Mask {0, 2, 4, 6, u, u, u, u, u, u, u, u}
3921 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
3922 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v4i16,
3927 // Nothing to be done for scalar shifts.
3928 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
3929 if (!VT.isVector() || !TLI.isTypeLegal(VT))
3932 assert(ST->hasNEON() && "unexpected vector shift");
3935 switch (N->getOpcode()) {
3937 llvm_unreachable("unexpected shift opcode");
3940 if (isVShiftLImm(N->getOperand(1), VT, Cnt)) {
3942 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
3943 DAG.getConstant(Cnt, MVT::i32));
3944 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0), RHS);
3950 if (isVShiftRImm(N->getOperand(1), VT, Cnt)) {
3952 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
3953 DAG.getConstant(Cnt, MVT::i32));
3954 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N->getOperand(0), RHS);
3962 /// ARM-specific DAG combining for intrinsics.
3963 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
3964 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
3968 // Don't do anything for most intrinsics.
3971 case Intrinsic::arm_neon_vqshifts:
3972 case Intrinsic::arm_neon_vqshiftu:
3973 EVT VT = N->getOperand(1).getValueType();
3975 if (!isVShiftLImm(N->getOperand(2), VT, Cnt))
3977 unsigned VShiftOpc = (IntNo == Intrinsic::arm_neon_vqshifts)
3978 ? AArch64ISD::NEON_QSHLs
3979 : AArch64ISD::NEON_QSHLu;
3980 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
3981 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
3987 /// Target-specific DAG combine function for NEON load/store intrinsics
3988 /// to merge base address updates.
3989 static SDValue CombineBaseUpdate(SDNode *N,
3990 TargetLowering::DAGCombinerInfo &DCI) {
3991 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
3994 SelectionDAG &DAG = DCI.DAG;
3995 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
3996 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
3997 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
3998 SDValue Addr = N->getOperand(AddrOpIdx);
4000 // Search for a use of the address operand that is an increment.
4001 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
4002 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
4004 if (User->getOpcode() != ISD::ADD ||
4005 UI.getUse().getResNo() != Addr.getResNo())
4008 // Check that the add is independent of the load/store. Otherwise, folding
4009 // it would create a cycle.
4010 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
4013 // Find the new opcode for the updating load/store.
4015 bool isLaneOp = false;
4016 unsigned NewOpc = 0;
4017 unsigned NumVecs = 0;
4019 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
4021 default: llvm_unreachable("unexpected intrinsic for Neon base update");
4022 case Intrinsic::arm_neon_vld1: NewOpc = AArch64ISD::NEON_LD1_UPD;
4024 case Intrinsic::arm_neon_vld2: NewOpc = AArch64ISD::NEON_LD2_UPD;
4026 case Intrinsic::arm_neon_vld3: NewOpc = AArch64ISD::NEON_LD3_UPD;
4028 case Intrinsic::arm_neon_vld4: NewOpc = AArch64ISD::NEON_LD4_UPD;
4030 case Intrinsic::arm_neon_vst1: NewOpc = AArch64ISD::NEON_ST1_UPD;
4031 NumVecs = 1; isLoad = false; break;
4032 case Intrinsic::arm_neon_vst2: NewOpc = AArch64ISD::NEON_ST2_UPD;
4033 NumVecs = 2; isLoad = false; break;
4034 case Intrinsic::arm_neon_vst3: NewOpc = AArch64ISD::NEON_ST3_UPD;
4035 NumVecs = 3; isLoad = false; break;
4036 case Intrinsic::arm_neon_vst4: NewOpc = AArch64ISD::NEON_ST4_UPD;
4037 NumVecs = 4; isLoad = false; break;
4038 case Intrinsic::aarch64_neon_vld1x2: NewOpc = AArch64ISD::NEON_LD1x2_UPD;
4040 case Intrinsic::aarch64_neon_vld1x3: NewOpc = AArch64ISD::NEON_LD1x3_UPD;
4042 case Intrinsic::aarch64_neon_vld1x4: NewOpc = AArch64ISD::NEON_LD1x4_UPD;
4044 case Intrinsic::aarch64_neon_vst1x2: NewOpc = AArch64ISD::NEON_ST1x2_UPD;
4045 NumVecs = 2; isLoad = false; break;
4046 case Intrinsic::aarch64_neon_vst1x3: NewOpc = AArch64ISD::NEON_ST1x3_UPD;
4047 NumVecs = 3; isLoad = false; break;
4048 case Intrinsic::aarch64_neon_vst1x4: NewOpc = AArch64ISD::NEON_ST1x4_UPD;
4049 NumVecs = 4; isLoad = false; break;
4050 case Intrinsic::arm_neon_vld2lane: NewOpc = AArch64ISD::NEON_LD2LN_UPD;
4051 NumVecs = 2; isLaneOp = true; break;
4052 case Intrinsic::arm_neon_vld3lane: NewOpc = AArch64ISD::NEON_LD3LN_UPD;
4053 NumVecs = 3; isLaneOp = true; break;
4054 case Intrinsic::arm_neon_vld4lane: NewOpc = AArch64ISD::NEON_LD4LN_UPD;
4055 NumVecs = 4; isLaneOp = true; break;
4056 case Intrinsic::arm_neon_vst2lane: NewOpc = AArch64ISD::NEON_ST2LN_UPD;
4057 NumVecs = 2; isLoad = false; isLaneOp = true; break;
4058 case Intrinsic::arm_neon_vst3lane: NewOpc = AArch64ISD::NEON_ST3LN_UPD;
4059 NumVecs = 3; isLoad = false; isLaneOp = true; break;
4060 case Intrinsic::arm_neon_vst4lane: NewOpc = AArch64ISD::NEON_ST4LN_UPD;
4061 NumVecs = 4; isLoad = false; isLaneOp = true; break;
4065 switch (N->getOpcode()) {
4066 default: llvm_unreachable("unexpected opcode for Neon base update");
4067 case AArch64ISD::NEON_LD2DUP: NewOpc = AArch64ISD::NEON_LD2DUP_UPD;
4069 case AArch64ISD::NEON_LD3DUP: NewOpc = AArch64ISD::NEON_LD3DUP_UPD;
4071 case AArch64ISD::NEON_LD4DUP: NewOpc = AArch64ISD::NEON_LD4DUP_UPD;
4076 // Find the size of memory referenced by the load/store.
4079 VecTy = N->getValueType(0);
4081 VecTy = N->getOperand(AddrOpIdx + 1).getValueType();
4082 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
4084 NumBytes /= VecTy.getVectorNumElements();
4086 // If the increment is a constant, it must match the memory ref size.
4087 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
4088 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
4089 uint32_t IncVal = CInc->getZExtValue();
4090 if (IncVal != NumBytes)
4092 Inc = DAG.getTargetConstant(IncVal, MVT::i32);
4095 // Create the new updating load/store node.
4097 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
4099 for (n = 0; n < NumResultVecs; ++n)
4101 Tys[n++] = MVT::i64;
4102 Tys[n] = MVT::Other;
4103 SDVTList SDTys = DAG.getVTList(Tys, NumResultVecs + 2);
4104 SmallVector<SDValue, 8> Ops;
4105 Ops.push_back(N->getOperand(0)); // incoming chain
4106 Ops.push_back(N->getOperand(AddrOpIdx));
4108 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
4109 Ops.push_back(N->getOperand(i));
4111 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
4112 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
4113 Ops.data(), Ops.size(),
4114 MemInt->getMemoryVT(),
4115 MemInt->getMemOperand());
4118 std::vector<SDValue> NewResults;
4119 for (unsigned i = 0; i < NumResultVecs; ++i) {
4120 NewResults.push_back(SDValue(UpdN.getNode(), i));
4122 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1)); // chain
4123 DCI.CombineTo(N, NewResults);
4124 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
4131 /// For a VDUPLANE node N, check if its source operand is a vldN-lane (N > 1)
4132 /// intrinsic, and if all the other uses of that intrinsic are also VDUPLANEs.
4133 /// If so, combine them to a vldN-dup operation and return true.
4134 static SDValue CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
4135 SelectionDAG &DAG = DCI.DAG;
4136 EVT VT = N->getValueType(0);
4138 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
4139 SDNode *VLD = N->getOperand(0).getNode();
4140 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
4142 unsigned NumVecs = 0;
4143 unsigned NewOpc = 0;
4144 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
4145 if (IntNo == Intrinsic::arm_neon_vld2lane) {
4147 NewOpc = AArch64ISD::NEON_LD2DUP;
4148 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
4150 NewOpc = AArch64ISD::NEON_LD3DUP;
4151 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
4153 NewOpc = AArch64ISD::NEON_LD4DUP;
4158 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
4159 // numbers match the load.
4160 unsigned VLDLaneNo =
4161 cast<ConstantSDNode>(VLD->getOperand(NumVecs + 3))->getZExtValue();
4162 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4164 // Ignore uses of the chain result.
4165 if (UI.getUse().getResNo() == NumVecs)
4168 if (User->getOpcode() != AArch64ISD::NEON_VDUPLANE ||
4169 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
4173 // Create the vldN-dup node.
4176 for (n = 0; n < NumVecs; ++n)
4178 Tys[n] = MVT::Other;
4179 SDVTList SDTys = DAG.getVTList(Tys, NumVecs + 1);
4180 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
4181 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
4182 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, Ops, 2,
4183 VLDMemInt->getMemoryVT(),
4184 VLDMemInt->getMemOperand());
4187 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4189 unsigned ResNo = UI.getUse().getResNo();
4190 // Ignore uses of the chain result.
4191 if (ResNo == NumVecs)
4194 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
4197 // Now the vldN-lane intrinsic is dead except for its chain result.
4198 // Update uses of the chain.
4199 std::vector<SDValue> VLDDupResults;
4200 for (unsigned n = 0; n < NumVecs; ++n)
4201 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
4202 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
4203 DCI.CombineTo(VLD, VLDDupResults);
4205 return SDValue(N, 0);
4209 AArch64TargetLowering::PerformDAGCombine(SDNode *N,
4210 DAGCombinerInfo &DCI) const {
4211 switch (N->getOpcode()) {
4213 case ISD::AND: return PerformANDCombine(N, DCI);
4214 case ISD::OR: return PerformORCombine(N, DCI, getSubtarget());
4218 return PerformShiftCombine(N, DCI, getSubtarget());
4219 case ISD::INTRINSIC_WO_CHAIN:
4220 return PerformIntrinsicCombine(N, DCI.DAG);
4221 case AArch64ISD::NEON_VDUPLANE:
4222 return CombineVLDDUP(N, DCI);
4223 case AArch64ISD::NEON_LD2DUP:
4224 case AArch64ISD::NEON_LD3DUP:
4225 case AArch64ISD::NEON_LD4DUP:
4226 return CombineBaseUpdate(N, DCI);
4227 case ISD::INTRINSIC_VOID:
4228 case ISD::INTRINSIC_W_CHAIN:
4229 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
4230 case Intrinsic::arm_neon_vld1:
4231 case Intrinsic::arm_neon_vld2:
4232 case Intrinsic::arm_neon_vld3:
4233 case Intrinsic::arm_neon_vld4:
4234 case Intrinsic::arm_neon_vst1:
4235 case Intrinsic::arm_neon_vst2:
4236 case Intrinsic::arm_neon_vst3:
4237 case Intrinsic::arm_neon_vst4:
4238 case Intrinsic::arm_neon_vld2lane:
4239 case Intrinsic::arm_neon_vld3lane:
4240 case Intrinsic::arm_neon_vld4lane:
4241 case Intrinsic::aarch64_neon_vld1x2:
4242 case Intrinsic::aarch64_neon_vld1x3:
4243 case Intrinsic::aarch64_neon_vld1x4:
4244 case Intrinsic::aarch64_neon_vst1x2:
4245 case Intrinsic::aarch64_neon_vst1x3:
4246 case Intrinsic::aarch64_neon_vst1x4:
4247 case Intrinsic::arm_neon_vst2lane:
4248 case Intrinsic::arm_neon_vst3lane:
4249 case Intrinsic::arm_neon_vst4lane:
4250 return CombineBaseUpdate(N, DCI);
4259 AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
4260 VT = VT.getScalarType();
4265 switch (VT.getSimpleVT().SimpleTy) {
4278 // Check whether a shuffle_vector could be presented as concat_vector.
4279 bool AArch64TargetLowering::isConcatVector(SDValue Op, SelectionDAG &DAG,
4280 SDValue V0, SDValue V1,
4282 SDValue &Res) const {
4284 EVT VT = Op.getValueType();
4285 if (VT.getSizeInBits() != 128)
4287 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4288 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4291 unsigned NumElts = VT.getVectorNumElements();
4292 bool isContactVector = true;
4293 bool splitV0 = false;
4294 if (V0.getValueType().getSizeInBits() == 128)
4297 for (int I = 0, E = NumElts / 2; I != E; I++) {
4299 isContactVector = false;
4304 if (isContactVector) {
4305 int offset = NumElts / 2;
4306 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4307 if (Mask[I] != I + splitV0 * offset) {
4308 isContactVector = false;
4314 if (isContactVector) {
4315 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4318 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4319 DAG.getConstant(0, MVT::i64));
4321 if (V1.getValueType().getSizeInBits() == 128) {
4322 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4323 DAG.getConstant(0, MVT::i64));
4325 Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4331 // Check whether a Build Vector could be presented as Shuffle Vector.
4332 // This Shuffle Vector maybe not legalized, so the length of its operand and
4333 // the length of result may not equal.
4334 bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
4335 SDValue &V0, SDValue &V1,
4338 EVT VT = Op.getValueType();
4339 unsigned NumElts = VT.getVectorNumElements();
4340 unsigned V0NumElts = 0;
4342 // Check if all elements are extracted from less than 3 vectors.
4343 for (unsigned i = 0; i < NumElts; ++i) {
4344 SDValue Elt = Op.getOperand(i);
4345 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4346 Elt.getOperand(0).getValueType().getVectorElementType() !=
4347 VT.getVectorElementType())
4350 if (V0.getNode() == 0) {
4351 V0 = Elt.getOperand(0);
4352 V0NumElts = V0.getValueType().getVectorNumElements();
4354 if (Elt.getOperand(0) == V0) {
4355 Mask[i] = (cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue());
4357 } else if (V1.getNode() == 0) {
4358 V1 = Elt.getOperand(0);
4360 if (Elt.getOperand(0) == V1) {
4361 unsigned Lane = cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue();
4362 Mask[i] = (Lane + V0NumElts);
4371 // If this is a case we can't handle, return null and let the default
4372 // expansion code take care of it.
4374 AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4375 const AArch64Subtarget *ST) const {
4377 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4379 EVT VT = Op.getValueType();
4381 APInt SplatBits, SplatUndef;
4382 unsigned SplatBitSize;
4385 unsigned UseNeonMov = VT.getSizeInBits() >= 64;
4387 // Note we favor lowering MOVI over MVNI.
4388 // This has implications on the definition of patterns in TableGen to select
4389 // BIC immediate instructions but not ORR immediate instructions.
4390 // If this lowering order is changed, TableGen patterns for BIC immediate and
4391 // ORR immediate instructions have to be updated.
4393 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4394 if (SplatBitSize <= 64) {
4395 // First attempt to use vector immediate-form MOVI
4398 unsigned OpCmode = 0;
4400 if (isNeonModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
4401 SplatBitSize, DAG, VT.is128BitVector(),
4402 Neon_Mov_Imm, NeonMovVT, Imm, OpCmode)) {
4403 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4404 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4406 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4407 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MOVIMM, DL, NeonMovVT,
4408 ImmVal, OpCmodeVal);
4409 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4413 // Then attempt to use vector immediate-form MVNI
4414 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4415 if (isNeonModifiedImm(NegatedImm, SplatUndef.getZExtValue(), SplatBitSize,
4416 DAG, VT.is128BitVector(), Neon_Mvn_Imm, NeonMovVT,
4418 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4419 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4420 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4421 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MVNIMM, DL, NeonMovVT,
4422 ImmVal, OpCmodeVal);
4423 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4427 // Attempt to use vector immediate-form FMOV
4428 if (((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) ||
4429 (VT == MVT::v2f64 && SplatBitSize == 64)) {
4431 SplatBitSize == 32 ? APFloat::IEEEsingle : APFloat::IEEEdouble,
4434 if (A64Imms::isFPImm(RealVal, ImmVal)) {
4435 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4436 return DAG.getNode(AArch64ISD::NEON_FMOVIMM, DL, VT, Val);
4442 unsigned NumElts = VT.getVectorNumElements();
4443 bool isOnlyLowElement = true;
4444 bool usesOnlyOneValue = true;
4445 bool hasDominantValue = false;
4446 bool isConstant = true;
4448 // Map of the number of times a particular SDValue appears in the
4450 DenseMap<SDValue, unsigned> ValueCounts;
4452 for (unsigned i = 0; i < NumElts; ++i) {
4453 SDValue V = Op.getOperand(i);
4454 if (V.getOpcode() == ISD::UNDEF)
4457 isOnlyLowElement = false;
4458 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4461 ValueCounts.insert(std::make_pair(V, 0));
4462 unsigned &Count = ValueCounts[V];
4464 // Is this value dominant? (takes up more than half of the lanes)
4465 if (++Count > (NumElts / 2)) {
4466 hasDominantValue = true;
4470 if (ValueCounts.size() != 1)
4471 usesOnlyOneValue = false;
4472 if (!Value.getNode() && ValueCounts.size() > 0)
4473 Value = ValueCounts.begin()->first;
4475 if (ValueCounts.size() == 0)
4476 return DAG.getUNDEF(VT);
4478 if (isOnlyLowElement)
4479 return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
4481 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4482 if (hasDominantValue && EltSize <= 64) {
4483 // Use VDUP for non-constant splats.
4487 // If we are DUPing a value that comes directly from a vector, we could
4488 // just use DUPLANE. We can only do this if the lane being extracted
4489 // is at a constant index, as the DUP from lane instructions only have
4490 // constant-index forms.
4492 // If there is a TRUNCATE between EXTRACT_VECTOR_ELT and DUP, we can
4493 // remove TRUNCATE for DUPLANE by apdating the source vector to
4494 // appropriate vector type and lane index.
4496 // FIXME: for now we have v1i8, v1i16, v1i32 legal vector types, if they
4497 // are not legal any more, no need to check the type size in bits should
4498 // be large than 64.
4500 if (Value->getOpcode() == ISD::TRUNCATE)
4501 V = Value->getOperand(0);
4502 if (V->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4503 isa<ConstantSDNode>(V->getOperand(1)) &&
4504 V->getOperand(0).getValueType().getSizeInBits() >= 64) {
4506 // If the element size of source vector is larger than DUPLANE
4507 // element size, we can do transformation by,
4508 // 1) bitcasting source register to smaller element vector
4509 // 2) mutiplying the lane index by SrcEltSize/ResEltSize
4510 // For example, we can lower
4511 // "v8i16 vdup_lane(v4i32, 1)"
4513 // "v8i16 vdup_lane(v8i16 bitcast(v4i32), 2)".
4514 SDValue SrcVec = V->getOperand(0);
4515 unsigned SrcEltSize =
4516 SrcVec.getValueType().getVectorElementType().getSizeInBits();
4517 unsigned ResEltSize = VT.getVectorElementType().getSizeInBits();
4518 if (SrcEltSize > ResEltSize) {
4519 assert((SrcEltSize % ResEltSize == 0) && "Invalid element size");
4521 unsigned SrcSize = SrcVec.getValueType().getSizeInBits();
4522 unsigned ResSize = VT.getSizeInBits();
4524 if (SrcSize > ResSize) {
4525 assert((SrcSize % ResSize == 0) && "Invalid vector size");
4527 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4528 SrcSize / ResEltSize);
4529 BitCast = DAG.getNode(ISD::BITCAST, DL, CastVT, SrcVec);
4531 assert((SrcSize == ResSize) && "Invalid vector size of source vec");
4532 BitCast = DAG.getNode(ISD::BITCAST, DL, VT, SrcVec);
4535 unsigned LaneIdx = V->getConstantOperandVal(1);
4537 DAG.getConstant((SrcEltSize / ResEltSize) * LaneIdx, MVT::i64);
4538 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, BitCast, Lane);
4540 assert((SrcEltSize == ResEltSize) &&
4541 "Invalid element size of source vec");
4542 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, V->getOperand(0),
4546 N = DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4548 if (!usesOnlyOneValue) {
4549 // The dominant value was splatted as 'N', but we now have to insert
4550 // all differing elements.
4551 for (unsigned I = 0; I < NumElts; ++I) {
4552 if (Op.getOperand(I) == Value)
4554 SmallVector<SDValue, 3> Ops;
4556 Ops.push_back(Op.getOperand(I));
4557 Ops.push_back(DAG.getConstant(I, MVT::i64));
4558 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, &Ops[0], 3);
4563 if (usesOnlyOneValue && isConstant) {
4564 return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4567 // If all elements are constants and the case above didn't get hit, fall back
4568 // to the default expansion, which will generate a load from the constant
4573 // Try to lower this in lowering ShuffleVector way.
4576 if (isKnownShuffleVector(Op, DAG, V0, V1, Mask)) {
4577 unsigned V0NumElts = V0.getValueType().getVectorNumElements();
4578 if (!V1.getNode() && V0NumElts == NumElts * 2) {
4579 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4580 DAG.getConstant(NumElts, MVT::i64));
4581 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4582 DAG.getConstant(0, MVT::i64));
4583 V0NumElts = V0.getValueType().getVectorNumElements();
4586 if (V1.getNode() && NumElts == V0NumElts &&
4587 V0NumElts == V1.getValueType().getVectorNumElements()) {
4588 SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
4589 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
4592 return LowerVECTOR_SHUFFLE(Shuffle, DAG);
4595 if (isConcatVector(Op, DAG, V0, V1, Mask, Res))
4600 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4601 // know the default expansion would otherwise fall back on something even
4602 // worse. For a vector with one or two non-undef values, that's
4603 // scalar_to_vector for the elements followed by a shuffle (provided the
4604 // shuffle is valid for the target) and materialization element by element
4605 // on the stack followed by a load for everything else.
4606 if (!isConstant && !usesOnlyOneValue) {
4607 SDValue Vec = DAG.getUNDEF(VT);
4608 for (unsigned i = 0 ; i < NumElts; ++i) {
4609 SDValue V = Op.getOperand(i);
4610 if (V.getOpcode() == ISD::UNDEF)
4612 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
4613 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V, LaneIdx);
4620 /// isREVMask - Check if a vector shuffle corresponds to a REV
4621 /// instruction with the specified blocksize. (The order of the elements
4622 /// within each block of the vector is reversed.)
4623 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4624 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4625 "Only possible block sizes for REV are: 16, 32, 64");
4627 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4631 unsigned NumElts = VT.getVectorNumElements();
4632 unsigned BlockElts = M[0] + 1;
4633 // If the first shuffle index is UNDEF, be optimistic.
4635 BlockElts = BlockSize / EltSz;
4637 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4640 for (unsigned i = 0; i < NumElts; ++i) {
4642 continue; // ignore UNDEF indices
4643 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4650 // isPermuteMask - Check whether the vector shuffle matches to UZP, ZIP and
4652 static unsigned isPermuteMask(ArrayRef<int> M, EVT VT, bool isV2undef) {
4653 unsigned NumElts = VT.getVectorNumElements();
4657 bool ismatch = true;
4660 for (unsigned i = 0; i < NumElts; ++i) {
4661 unsigned answer = i * 2;
4662 if (isV2undef && answer >= NumElts)
4664 if (M[i] != -1 && (unsigned)M[i] != answer) {
4670 return AArch64ISD::NEON_UZP1;
4674 for (unsigned i = 0; i < NumElts; ++i) {
4675 unsigned answer = i * 2 + 1;
4676 if (isV2undef && answer >= NumElts)
4678 if (M[i] != -1 && (unsigned)M[i] != answer) {
4684 return AArch64ISD::NEON_UZP2;
4688 for (unsigned i = 0; i < NumElts; ++i) {
4689 unsigned answer = i / 2 + NumElts * (i % 2);
4690 if (isV2undef && answer >= NumElts)
4692 if (M[i] != -1 && (unsigned)M[i] != answer) {
4698 return AArch64ISD::NEON_ZIP1;
4702 for (unsigned i = 0; i < NumElts; ++i) {
4703 unsigned answer = (NumElts + i) / 2 + NumElts * (i % 2);
4704 if (isV2undef && answer >= NumElts)
4706 if (M[i] != -1 && (unsigned)M[i] != answer) {
4712 return AArch64ISD::NEON_ZIP2;
4716 for (unsigned i = 0; i < NumElts; ++i) {
4717 unsigned answer = i + (NumElts - 1) * (i % 2);
4718 if (isV2undef && answer >= NumElts)
4720 if (M[i] != -1 && (unsigned)M[i] != answer) {
4726 return AArch64ISD::NEON_TRN1;
4730 for (unsigned i = 0; i < NumElts; ++i) {
4731 unsigned answer = 1 + i + (NumElts - 1) * (i % 2);
4732 if (isV2undef && answer >= NumElts)
4734 if (M[i] != -1 && (unsigned)M[i] != answer) {
4740 return AArch64ISD::NEON_TRN2;
4746 AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
4747 SelectionDAG &DAG) const {
4748 SDValue V1 = Op.getOperand(0);
4749 SDValue V2 = Op.getOperand(1);
4751 EVT VT = Op.getValueType();
4752 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
4754 // Convert shuffles that are directly supported on NEON to target-specific
4755 // DAG nodes, instead of keeping them as shuffles and matching them again
4756 // during code selection. This is more efficient and avoids the possibility
4757 // of inconsistencies between legalization and selection.
4758 ArrayRef<int> ShuffleMask = SVN->getMask();
4760 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4764 if (isREVMask(ShuffleMask, VT, 64))
4765 return DAG.getNode(AArch64ISD::NEON_REV64, dl, VT, V1);
4766 if (isREVMask(ShuffleMask, VT, 32))
4767 return DAG.getNode(AArch64ISD::NEON_REV32, dl, VT, V1);
4768 if (isREVMask(ShuffleMask, VT, 16))
4769 return DAG.getNode(AArch64ISD::NEON_REV16, dl, VT, V1);
4772 if (V2.getOpcode() == ISD::UNDEF)
4773 ISDNo = isPermuteMask(ShuffleMask, VT, true);
4775 ISDNo = isPermuteMask(ShuffleMask, VT, false);
4778 if (V2.getOpcode() == ISD::UNDEF)
4779 return DAG.getNode(ISDNo, dl, VT, V1, V1);
4781 return DAG.getNode(ISDNo, dl, VT, V1, V2);
4785 if (isConcatVector(Op, DAG, V1, V2, &ShuffleMask[0], Res))
4788 // If the element of shuffle mask are all the same constant, we can
4789 // transform it into either NEON_VDUP or NEON_VDUPLANE
4790 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
4791 int Lane = SVN->getSplatIndex();
4792 // If this is undef splat, generate it via "just" vdup, if possible.
4793 if (Lane == -1) Lane = 0;
4795 // Test if V1 is a SCALAR_TO_VECTOR.
4796 if (V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
4797 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT, V1.getOperand(0));
4799 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR.
4800 if (V1.getOpcode() == ISD::BUILD_VECTOR) {
4801 bool IsScalarToVector = true;
4802 for (unsigned i = 0, e = V1.getNumOperands(); i != e; ++i)
4803 if (V1.getOperand(i).getOpcode() != ISD::UNDEF &&
4804 i != (unsigned)Lane) {
4805 IsScalarToVector = false;
4808 if (IsScalarToVector)
4809 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT,
4810 V1.getOperand(Lane));
4813 // Test if V1 is a EXTRACT_SUBVECTOR.
4814 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
4815 int ExtLane = cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
4816 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1.getOperand(0),
4817 DAG.getConstant(Lane + ExtLane, MVT::i64));
4819 // Test if V1 is a CONCAT_VECTORS.
4820 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
4821 V1.getOperand(1).getOpcode() == ISD::UNDEF) {
4822 SDValue Op0 = V1.getOperand(0);
4823 assert((unsigned)Lane < Op0.getValueType().getVectorNumElements() &&
4824 "Invalid vector lane access");
4825 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, Op0,
4826 DAG.getConstant(Lane, MVT::i64));
4829 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1,
4830 DAG.getConstant(Lane, MVT::i64));
4833 int Length = ShuffleMask.size();
4834 int V1EltNum = V1.getValueType().getVectorNumElements();
4836 // If the number of v1 elements is the same as the number of shuffle mask
4837 // element and the shuffle masks are sequential values, we can transform
4838 // it into NEON_VEXTRACT.
4839 if (V1EltNum == Length) {
4840 // Check if the shuffle mask is sequential.
4842 while (ShuffleMask[SkipUndef] == -1) {
4845 int CurMask = ShuffleMask[SkipUndef];
4846 if (CurMask >= SkipUndef) {
4847 bool IsSequential = true;
4848 for (int I = SkipUndef; I < Length; ++I) {
4849 if (ShuffleMask[I] != -1 && ShuffleMask[I] != CurMask) {
4850 IsSequential = false;
4856 assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
4857 unsigned VecSize = EltSize * V1EltNum;
4858 unsigned Index = (EltSize / 8) * (ShuffleMask[SkipUndef] - SkipUndef);
4859 if (VecSize == 64 || VecSize == 128)
4860 return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
4861 DAG.getConstant(Index, MVT::i64));
4866 // For shuffle mask like "0, 1, 2, 3, 4, 5, 13, 7", try to generate insert
4867 // by element from V2 to V1 .
4868 // If shuffle mask is like "0, 1, 10, 11, 12, 13, 14, 15", V2 would be a
4869 // better choice to be inserted than V1 as less insert needed, so we count
4870 // element to be inserted for both V1 and V2, and select less one as insert
4873 // Collect elements need to be inserted and their index.
4874 SmallVector<int, 8> NV1Elt;
4875 SmallVector<int, 8> N1Index;
4876 SmallVector<int, 8> NV2Elt;
4877 SmallVector<int, 8> N2Index;
4878 for (int I = 0; I != Length; ++I) {
4879 if (ShuffleMask[I] != I) {
4880 NV1Elt.push_back(ShuffleMask[I]);
4881 N1Index.push_back(I);
4884 for (int I = 0; I != Length; ++I) {
4885 if (ShuffleMask[I] != (I + V1EltNum)) {
4886 NV2Elt.push_back(ShuffleMask[I]);
4887 N2Index.push_back(I);
4891 // Decide which to be inserted. If all lanes mismatch, neither V1 nor V2
4892 // will be inserted.
4894 SmallVector<int, 8> InsMasks = NV1Elt;
4895 SmallVector<int, 8> InsIndex = N1Index;
4896 if ((int)NV1Elt.size() != Length || (int)NV2Elt.size() != Length) {
4897 if (NV1Elt.size() > NV2Elt.size()) {
4903 InsV = DAG.getNode(ISD::UNDEF, dl, VT);
4906 for (int I = 0, E = InsMasks.size(); I != E; ++I) {
4908 int Mask = InsMasks[I];
4909 if (Mask >= V1EltNum) {
4913 // Any value type smaller than i32 is illegal in AArch64, and this lower
4914 // function is called after legalize pass, so we need to legalize
4917 if (VT.getVectorElementType().isFloatingPoint())
4918 EltVT = (EltSize == 64) ? MVT::f64 : MVT::f32;
4920 EltVT = (EltSize == 64) ? MVT::i64 : MVT::i32;
4923 ExtV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ExtV,
4924 DAG.getConstant(Mask, MVT::i64));
4925 InsV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, InsV, ExtV,
4926 DAG.getConstant(InsIndex[I], MVT::i64));
4932 AArch64TargetLowering::ConstraintType
4933 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
4934 if (Constraint.size() == 1) {
4935 switch (Constraint[0]) {
4937 case 'w': // An FP/SIMD vector register
4938 return C_RegisterClass;
4939 case 'I': // Constant that can be used with an ADD instruction
4940 case 'J': // Constant that can be used with a SUB instruction
4941 case 'K': // Constant that can be used with a 32-bit logical instruction
4942 case 'L': // Constant that can be used with a 64-bit logical instruction
4943 case 'M': // Constant that can be used as a 32-bit MOV immediate
4944 case 'N': // Constant that can be used as a 64-bit MOV immediate
4945 case 'Y': // Floating point constant zero
4946 case 'Z': // Integer constant zero
4948 case 'Q': // A memory reference with base register and no offset
4950 case 'S': // A symbolic address
4955 // FIXME: Ump, Utf, Usa, Ush
4956 // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes,
4957 // whatever they may be
4958 // Utf: A memory address suitable for ldp/stp in TF mode, whatever it may be
4959 // Usa: An absolute symbolic address
4960 // Ush: The high part (bits 32:12) of a pc-relative symbolic address
4961 assert(Constraint != "Ump" && Constraint != "Utf" && Constraint != "Usa"
4962 && Constraint != "Ush" && "Unimplemented constraints");
4964 return TargetLowering::getConstraintType(Constraint);
4967 TargetLowering::ConstraintWeight
4968 AArch64TargetLowering::getSingleConstraintMatchWeight(AsmOperandInfo &Info,
4969 const char *Constraint) const {
4971 llvm_unreachable("Constraint weight unimplemented");
4975 AArch64TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
4976 std::string &Constraint,
4977 std::vector<SDValue> &Ops,
4978 SelectionDAG &DAG) const {
4979 SDValue Result(0, 0);
4981 // Only length 1 constraints are C_Other.
4982 if (Constraint.size() != 1) return;
4984 // Only C_Other constraints get lowered like this. That means constants for us
4985 // so return early if there's no hope the constraint can be lowered.
4987 switch(Constraint[0]) {
4989 case 'I': case 'J': case 'K': case 'L':
4990 case 'M': case 'N': case 'Z': {
4991 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
4995 uint64_t CVal = C->getZExtValue();
4998 switch (Constraint[0]) {
5000 // FIXME: 'M' and 'N' are MOV pseudo-insts -- unsupported in assembly. 'J'
5001 // is a peculiarly useless SUB constraint.
5002 llvm_unreachable("Unimplemented C_Other constraint");
5008 if (A64Imms::isLogicalImm(32, CVal, Bits))
5012 if (A64Imms::isLogicalImm(64, CVal, Bits))
5021 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5025 // An absolute symbolic address or label reference.
5026 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5027 Result = DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
5028 GA->getValueType(0));
5029 } else if (const BlockAddressSDNode *BA
5030 = dyn_cast<BlockAddressSDNode>(Op)) {
5031 Result = DAG.getTargetBlockAddress(BA->getBlockAddress(),
5032 BA->getValueType(0));
5033 } else if (const ExternalSymbolSDNode *ES
5034 = dyn_cast<ExternalSymbolSDNode>(Op)) {
5035 Result = DAG.getTargetExternalSymbol(ES->getSymbol(),
5036 ES->getValueType(0));
5042 if (const ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) {
5043 if (CFP->isExactlyValue(0.0)) {
5044 Result = DAG.getTargetConstantFP(0.0, CFP->getValueType(0));
5051 if (Result.getNode()) {
5052 Ops.push_back(Result);
5056 // It's an unknown constraint for us. Let generic code have a go.
5057 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5060 std::pair<unsigned, const TargetRegisterClass*>
5061 AArch64TargetLowering::getRegForInlineAsmConstraint(
5062 const std::string &Constraint,
5064 if (Constraint.size() == 1) {
5065 switch (Constraint[0]) {
5067 if (VT.getSizeInBits() <= 32)
5068 return std::make_pair(0U, &AArch64::GPR32RegClass);
5069 else if (VT == MVT::i64)
5070 return std::make_pair(0U, &AArch64::GPR64RegClass);
5074 return std::make_pair(0U, &AArch64::FPR16RegClass);
5075 else if (VT == MVT::f32)
5076 return std::make_pair(0U, &AArch64::FPR32RegClass);
5077 else if (VT.getSizeInBits() == 64)
5078 return std::make_pair(0U, &AArch64::FPR64RegClass);
5079 else if (VT.getSizeInBits() == 128)
5080 return std::make_pair(0U, &AArch64::FPR128RegClass);
5085 // Use the default implementation in TargetLowering to convert the register
5086 // constraint into a member of a register class.
5087 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5090 /// Represent NEON load and store intrinsics as MemIntrinsicNodes.
5091 /// The associated MachineMemOperands record the alignment specified
5092 /// in the intrinsic calls.
5093 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5095 unsigned Intrinsic) const {
5096 switch (Intrinsic) {
5097 case Intrinsic::arm_neon_vld1:
5098 case Intrinsic::arm_neon_vld2:
5099 case Intrinsic::arm_neon_vld3:
5100 case Intrinsic::arm_neon_vld4:
5101 case Intrinsic::aarch64_neon_vld1x2:
5102 case Intrinsic::aarch64_neon_vld1x3:
5103 case Intrinsic::aarch64_neon_vld1x4:
5104 case Intrinsic::arm_neon_vld2lane:
5105 case Intrinsic::arm_neon_vld3lane:
5106 case Intrinsic::arm_neon_vld4lane: {
5107 Info.opc = ISD::INTRINSIC_W_CHAIN;
5108 // Conservatively set memVT to the entire set of vectors loaded.
5109 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5110 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5111 Info.ptrVal = I.getArgOperand(0);
5113 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5114 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5115 Info.vol = false; // volatile loads with NEON intrinsics not supported
5116 Info.readMem = true;
5117 Info.writeMem = false;
5120 case Intrinsic::arm_neon_vst1:
5121 case Intrinsic::arm_neon_vst2:
5122 case Intrinsic::arm_neon_vst3:
5123 case Intrinsic::arm_neon_vst4:
5124 case Intrinsic::aarch64_neon_vst1x2:
5125 case Intrinsic::aarch64_neon_vst1x3:
5126 case Intrinsic::aarch64_neon_vst1x4:
5127 case Intrinsic::arm_neon_vst2lane:
5128 case Intrinsic::arm_neon_vst3lane:
5129 case Intrinsic::arm_neon_vst4lane: {
5130 Info.opc = ISD::INTRINSIC_VOID;
5131 // Conservatively set memVT to the entire set of vectors stored.
5132 unsigned NumElts = 0;
5133 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5134 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5135 if (!ArgTy->isVectorTy())
5137 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5139 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5140 Info.ptrVal = I.getArgOperand(0);
5142 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5143 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5144 Info.vol = false; // volatile stores with NEON intrinsics not supported
5145 Info.readMem = false;
5146 Info.writeMem = true;