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 //===----------------------------------------------------------------------===//
16 #include "AArch64ISelLowering.h"
17 #include "AArch64MachineFunctionInfo.h"
18 #include "AArch64Subtarget.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "Utils/AArch64BaseInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/IR/CallingConv.h"
29 #include "llvm/Support/MathExtras.h"
33 #define DEBUG_TYPE "aarch64-isel"
35 static TargetLoweringObjectFile *createTLOF(AArch64TargetMachine &TM) {
36 assert (TM.getSubtarget<AArch64Subtarget>().isTargetELF() &&
37 "unknown subtarget type");
38 return new AArch64ElfTargetObjectFile();
41 AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
42 : TargetLowering(TM, createTLOF(TM)), Itins(TM.getInstrItineraryData()) {
44 const AArch64Subtarget *Subtarget = &TM.getSubtarget<AArch64Subtarget>();
46 // SIMD compares set the entire lane's bits to 1
47 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
49 // Scalar register <-> type mapping
50 addRegisterClass(MVT::i32, &AArch64::GPR32RegClass);
51 addRegisterClass(MVT::i64, &AArch64::GPR64RegClass);
53 if (Subtarget->hasFPARMv8()) {
54 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
55 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
56 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
57 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
60 if (Subtarget->hasNEON()) {
62 addRegisterClass(MVT::v1i8, &AArch64::FPR8RegClass);
63 addRegisterClass(MVT::v1i16, &AArch64::FPR16RegClass);
64 addRegisterClass(MVT::v1i32, &AArch64::FPR32RegClass);
65 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
66 addRegisterClass(MVT::v1f64, &AArch64::FPR64RegClass);
67 addRegisterClass(MVT::v8i8, &AArch64::FPR64RegClass);
68 addRegisterClass(MVT::v4i16, &AArch64::FPR64RegClass);
69 addRegisterClass(MVT::v2i32, &AArch64::FPR64RegClass);
70 addRegisterClass(MVT::v1i64, &AArch64::FPR64RegClass);
71 addRegisterClass(MVT::v2f32, &AArch64::FPR64RegClass);
72 addRegisterClass(MVT::v16i8, &AArch64::FPR128RegClass);
73 addRegisterClass(MVT::v8i16, &AArch64::FPR128RegClass);
74 addRegisterClass(MVT::v4i32, &AArch64::FPR128RegClass);
75 addRegisterClass(MVT::v2i64, &AArch64::FPR128RegClass);
76 addRegisterClass(MVT::v4f32, &AArch64::FPR128RegClass);
77 addRegisterClass(MVT::v2f64, &AArch64::FPR128RegClass);
80 computeRegisterProperties();
82 // We combine OR nodes for bitfield and NEON BSL operations.
83 setTargetDAGCombine(ISD::OR);
85 setTargetDAGCombine(ISD::AND);
86 setTargetDAGCombine(ISD::SRA);
87 setTargetDAGCombine(ISD::SRL);
88 setTargetDAGCombine(ISD::SHL);
90 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
91 setTargetDAGCombine(ISD::INTRINSIC_VOID);
92 setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
94 // AArch64 does not have i1 loads, or much of anything for i1 really.
95 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
96 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
97 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
99 setStackPointerRegisterToSaveRestore(AArch64::XSP);
100 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
101 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
102 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
104 // We'll lower globals to wrappers for selection.
105 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
106 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
108 // A64 instructions have the comparison predicate attached to the user of the
109 // result, but having a separate comparison is valuable for matching.
110 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
111 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
112 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
113 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
115 setOperationAction(ISD::SELECT, MVT::i32, Custom);
116 setOperationAction(ISD::SELECT, MVT::i64, Custom);
117 setOperationAction(ISD::SELECT, MVT::f32, Custom);
118 setOperationAction(ISD::SELECT, MVT::f64, Custom);
120 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
121 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
122 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
123 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
125 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
127 setOperationAction(ISD::SETCC, MVT::i32, Custom);
128 setOperationAction(ISD::SETCC, MVT::i64, Custom);
129 setOperationAction(ISD::SETCC, MVT::f32, Custom);
130 setOperationAction(ISD::SETCC, MVT::f64, Custom);
132 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
133 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
134 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
136 setOperationAction(ISD::VASTART, MVT::Other, Custom);
137 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
138 setOperationAction(ISD::VAEND, MVT::Other, Expand);
139 setOperationAction(ISD::VAARG, MVT::Other, Expand);
141 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
142 setOperationAction(ISD::ConstantPool, MVT::i64, Custom);
144 setOperationAction(ISD::ROTL, MVT::i32, Expand);
145 setOperationAction(ISD::ROTL, MVT::i64, Expand);
147 setOperationAction(ISD::UREM, MVT::i32, Expand);
148 setOperationAction(ISD::UREM, MVT::i64, Expand);
149 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
150 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
152 setOperationAction(ISD::SREM, MVT::i32, Expand);
153 setOperationAction(ISD::SREM, MVT::i64, Expand);
154 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
155 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
157 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
158 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
159 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
160 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
162 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
163 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
165 // Legal floating-point operations.
166 setOperationAction(ISD::FABS, MVT::f32, Legal);
167 setOperationAction(ISD::FABS, MVT::f64, Legal);
169 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
170 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
172 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
173 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
175 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
176 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
178 setOperationAction(ISD::FNEG, MVT::f32, Legal);
179 setOperationAction(ISD::FNEG, MVT::f64, Legal);
181 setOperationAction(ISD::FRINT, MVT::f32, Legal);
182 setOperationAction(ISD::FRINT, MVT::f64, Legal);
184 setOperationAction(ISD::FSQRT, MVT::f32, Legal);
185 setOperationAction(ISD::FSQRT, MVT::f64, Legal);
187 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
188 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
190 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
191 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
192 setOperationAction(ISD::ConstantFP, MVT::f128, Legal);
194 // Illegal floating-point operations.
195 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
196 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
198 setOperationAction(ISD::FCOS, MVT::f32, Expand);
199 setOperationAction(ISD::FCOS, MVT::f64, Expand);
201 setOperationAction(ISD::FEXP, MVT::f32, Expand);
202 setOperationAction(ISD::FEXP, MVT::f64, Expand);
204 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
205 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
207 setOperationAction(ISD::FLOG, MVT::f32, Expand);
208 setOperationAction(ISD::FLOG, MVT::f64, Expand);
210 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
211 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
213 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
214 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
216 setOperationAction(ISD::FPOW, MVT::f32, Expand);
217 setOperationAction(ISD::FPOW, MVT::f64, Expand);
219 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
220 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
222 setOperationAction(ISD::FREM, MVT::f32, Expand);
223 setOperationAction(ISD::FREM, MVT::f64, Expand);
225 setOperationAction(ISD::FSIN, MVT::f32, Expand);
226 setOperationAction(ISD::FSIN, MVT::f64, Expand);
228 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
229 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
231 // Virtually no operation on f128 is legal, but LLVM can't expand them when
232 // there's a valid register class, so we need custom operations in most cases.
233 setOperationAction(ISD::FABS, MVT::f128, Expand);
234 setOperationAction(ISD::FADD, MVT::f128, Custom);
235 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
236 setOperationAction(ISD::FCOS, MVT::f128, Expand);
237 setOperationAction(ISD::FDIV, MVT::f128, Custom);
238 setOperationAction(ISD::FMA, MVT::f128, Expand);
239 setOperationAction(ISD::FMUL, MVT::f128, Custom);
240 setOperationAction(ISD::FNEG, MVT::f128, Expand);
241 setOperationAction(ISD::FP_EXTEND, MVT::f128, Expand);
242 setOperationAction(ISD::FP_ROUND, MVT::f128, Expand);
243 setOperationAction(ISD::FPOW, MVT::f128, Expand);
244 setOperationAction(ISD::FREM, MVT::f128, Expand);
245 setOperationAction(ISD::FRINT, MVT::f128, Expand);
246 setOperationAction(ISD::FSIN, MVT::f128, Expand);
247 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
248 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
249 setOperationAction(ISD::FSUB, MVT::f128, Custom);
250 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
251 setOperationAction(ISD::SETCC, MVT::f128, Custom);
252 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
253 setOperationAction(ISD::SELECT, MVT::f128, Expand);
254 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
255 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
257 // Lowering for many of the conversions is actually specified by the non-f128
258 // type. The LowerXXX function will be trivial when f128 isn't involved.
259 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
260 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
261 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
262 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
263 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
264 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
265 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
266 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
267 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
268 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
269 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
270 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
271 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
272 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
274 // i128 shift operation support
275 setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
276 setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
277 setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);
279 // This prevents LLVM trying to compress double constants into a floating
280 // constant-pool entry and trying to load from there. It's of doubtful benefit
281 // for A64: we'd need LDR followed by FCVT, I believe.
282 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
283 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
284 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
286 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
287 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
288 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
289 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
290 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
291 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
293 setExceptionPointerRegister(AArch64::X0);
294 setExceptionSelectorRegister(AArch64::X1);
296 if (Subtarget->hasNEON()) {
297 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i8, Expand);
298 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Expand);
299 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i32, Expand);
300 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v1i64, Expand);
301 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v16i8, Expand);
302 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v8i16, Expand);
303 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i32, Expand);
304 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i64, Expand);
306 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i8, Custom);
307 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
308 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
309 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i16, Custom);
310 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
311 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
312 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i32, Custom);
313 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
314 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
315 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
316 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
317 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
318 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
319 setOperationAction(ISD::BUILD_VECTOR, MVT::v1f64, Custom);
320 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
322 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
323 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i8, Custom);
324 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
325 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i16, Custom);
326 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
327 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i32, Custom);
328 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
329 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
330 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f32, Custom);
331 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
332 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1f64, Custom);
333 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
335 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i32, Legal);
336 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Legal);
337 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Legal);
338 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Legal);
339 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Legal);
340 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Legal);
341 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Legal);
343 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i8, Custom);
344 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i16, Custom);
345 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
346 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
347 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
349 setOperationAction(ISD::SETCC, MVT::v8i8, Custom);
350 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
351 setOperationAction(ISD::SETCC, MVT::v4i16, Custom);
352 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
353 setOperationAction(ISD::SETCC, MVT::v2i32, Custom);
354 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
355 setOperationAction(ISD::SETCC, MVT::v1i64, Custom);
356 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
357 setOperationAction(ISD::SETCC, MVT::v2f32, Custom);
358 setOperationAction(ISD::SETCC, MVT::v4f32, Custom);
359 setOperationAction(ISD::SETCC, MVT::v1f64, Custom);
360 setOperationAction(ISD::SETCC, MVT::v2f64, Custom);
362 setOperationAction(ISD::FFLOOR, MVT::v2f32, Legal);
363 setOperationAction(ISD::FFLOOR, MVT::v4f32, Legal);
364 setOperationAction(ISD::FFLOOR, MVT::v1f64, Legal);
365 setOperationAction(ISD::FFLOOR, MVT::v2f64, Legal);
367 setOperationAction(ISD::FCEIL, MVT::v2f32, Legal);
368 setOperationAction(ISD::FCEIL, MVT::v4f32, Legal);
369 setOperationAction(ISD::FCEIL, MVT::v1f64, Legal);
370 setOperationAction(ISD::FCEIL, MVT::v2f64, Legal);
372 setOperationAction(ISD::FTRUNC, MVT::v2f32, Legal);
373 setOperationAction(ISD::FTRUNC, MVT::v4f32, Legal);
374 setOperationAction(ISD::FTRUNC, MVT::v1f64, Legal);
375 setOperationAction(ISD::FTRUNC, MVT::v2f64, Legal);
377 setOperationAction(ISD::FRINT, MVT::v2f32, Legal);
378 setOperationAction(ISD::FRINT, MVT::v4f32, Legal);
379 setOperationAction(ISD::FRINT, MVT::v1f64, Legal);
380 setOperationAction(ISD::FRINT, MVT::v2f64, Legal);
382 setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Legal);
383 setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Legal);
384 setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Legal);
385 setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Legal);
387 setOperationAction(ISD::FROUND, MVT::v2f32, Legal);
388 setOperationAction(ISD::FROUND, MVT::v4f32, Legal);
389 setOperationAction(ISD::FROUND, MVT::v1f64, Legal);
390 setOperationAction(ISD::FROUND, MVT::v2f64, Legal);
392 setOperationAction(ISD::SINT_TO_FP, MVT::v1i8, Custom);
393 setOperationAction(ISD::SINT_TO_FP, MVT::v1i16, Custom);
394 setOperationAction(ISD::SINT_TO_FP, MVT::v1i32, Custom);
395 setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
396 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
397 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
399 setOperationAction(ISD::UINT_TO_FP, MVT::v1i8, Custom);
400 setOperationAction(ISD::UINT_TO_FP, MVT::v1i16, Custom);
401 setOperationAction(ISD::UINT_TO_FP, MVT::v1i32, Custom);
402 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
403 setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
404 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
406 setOperationAction(ISD::FP_TO_SINT, MVT::v1i8, Custom);
407 setOperationAction(ISD::FP_TO_SINT, MVT::v1i16, Custom);
408 setOperationAction(ISD::FP_TO_SINT, MVT::v1i32, Custom);
409 setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);
410 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
411 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Custom);
413 setOperationAction(ISD::FP_TO_UINT, MVT::v1i8, Custom);
414 setOperationAction(ISD::FP_TO_UINT, MVT::v1i16, Custom);
415 setOperationAction(ISD::FP_TO_UINT, MVT::v1i32, Custom);
416 setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
417 setOperationAction(ISD::FP_TO_UINT, MVT::v2i32, Custom);
418 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Custom);
420 // Neon does not support vector divide/remainder operations except
421 // floating-point divide.
422 setOperationAction(ISD::SDIV, MVT::v1i8, Expand);
423 setOperationAction(ISD::SDIV, MVT::v8i8, Expand);
424 setOperationAction(ISD::SDIV, MVT::v16i8, Expand);
425 setOperationAction(ISD::SDIV, MVT::v1i16, Expand);
426 setOperationAction(ISD::SDIV, MVT::v4i16, Expand);
427 setOperationAction(ISD::SDIV, MVT::v8i16, Expand);
428 setOperationAction(ISD::SDIV, MVT::v1i32, Expand);
429 setOperationAction(ISD::SDIV, MVT::v2i32, Expand);
430 setOperationAction(ISD::SDIV, MVT::v4i32, Expand);
431 setOperationAction(ISD::SDIV, MVT::v1i64, Expand);
432 setOperationAction(ISD::SDIV, MVT::v2i64, Expand);
434 setOperationAction(ISD::UDIV, MVT::v1i8, Expand);
435 setOperationAction(ISD::UDIV, MVT::v8i8, Expand);
436 setOperationAction(ISD::UDIV, MVT::v16i8, Expand);
437 setOperationAction(ISD::UDIV, MVT::v1i16, Expand);
438 setOperationAction(ISD::UDIV, MVT::v4i16, Expand);
439 setOperationAction(ISD::UDIV, MVT::v8i16, Expand);
440 setOperationAction(ISD::UDIV, MVT::v1i32, Expand);
441 setOperationAction(ISD::UDIV, MVT::v2i32, Expand);
442 setOperationAction(ISD::UDIV, MVT::v4i32, Expand);
443 setOperationAction(ISD::UDIV, MVT::v1i64, Expand);
444 setOperationAction(ISD::UDIV, MVT::v2i64, Expand);
446 setOperationAction(ISD::SREM, MVT::v1i8, Expand);
447 setOperationAction(ISD::SREM, MVT::v8i8, Expand);
448 setOperationAction(ISD::SREM, MVT::v16i8, Expand);
449 setOperationAction(ISD::SREM, MVT::v1i16, Expand);
450 setOperationAction(ISD::SREM, MVT::v4i16, Expand);
451 setOperationAction(ISD::SREM, MVT::v8i16, Expand);
452 setOperationAction(ISD::SREM, MVT::v1i32, Expand);
453 setOperationAction(ISD::SREM, MVT::v2i32, Expand);
454 setOperationAction(ISD::SREM, MVT::v4i32, Expand);
455 setOperationAction(ISD::SREM, MVT::v1i64, Expand);
456 setOperationAction(ISD::SREM, MVT::v2i64, Expand);
458 setOperationAction(ISD::UREM, MVT::v1i8, Expand);
459 setOperationAction(ISD::UREM, MVT::v8i8, Expand);
460 setOperationAction(ISD::UREM, MVT::v16i8, Expand);
461 setOperationAction(ISD::UREM, MVT::v1i16, Expand);
462 setOperationAction(ISD::UREM, MVT::v4i16, Expand);
463 setOperationAction(ISD::UREM, MVT::v8i16, Expand);
464 setOperationAction(ISD::UREM, MVT::v1i32, Expand);
465 setOperationAction(ISD::UREM, MVT::v2i32, Expand);
466 setOperationAction(ISD::UREM, MVT::v4i32, Expand);
467 setOperationAction(ISD::UREM, MVT::v1i64, Expand);
468 setOperationAction(ISD::UREM, MVT::v2i64, Expand);
470 setOperationAction(ISD::FREM, MVT::v2f32, Expand);
471 setOperationAction(ISD::FREM, MVT::v4f32, Expand);
472 setOperationAction(ISD::FREM, MVT::v1f64, Expand);
473 setOperationAction(ISD::FREM, MVT::v2f64, Expand);
475 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
476 setOperationAction(ISD::SELECT, MVT::v16i8, Expand);
477 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
478 setOperationAction(ISD::SELECT, MVT::v8i16, Expand);
479 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
480 setOperationAction(ISD::SELECT, MVT::v4i32, Expand);
481 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
482 setOperationAction(ISD::SELECT, MVT::v2i64, Expand);
483 setOperationAction(ISD::SELECT, MVT::v2f32, Expand);
484 setOperationAction(ISD::SELECT, MVT::v4f32, Expand);
485 setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
486 setOperationAction(ISD::SELECT, MVT::v2f64, Expand);
488 setOperationAction(ISD::SELECT_CC, MVT::v8i8, Custom);
489 setOperationAction(ISD::SELECT_CC, MVT::v16i8, Custom);
490 setOperationAction(ISD::SELECT_CC, MVT::v4i16, Custom);
491 setOperationAction(ISD::SELECT_CC, MVT::v8i16, Custom);
492 setOperationAction(ISD::SELECT_CC, MVT::v2i32, Custom);
493 setOperationAction(ISD::SELECT_CC, MVT::v4i32, Custom);
494 setOperationAction(ISD::SELECT_CC, MVT::v1i64, Custom);
495 setOperationAction(ISD::SELECT_CC, MVT::v2i64, Custom);
496 setOperationAction(ISD::SELECT_CC, MVT::v2f32, Custom);
497 setOperationAction(ISD::SELECT_CC, MVT::v4f32, Custom);
498 setOperationAction(ISD::SELECT_CC, MVT::v1f64, Custom);
499 setOperationAction(ISD::SELECT_CC, MVT::v2f64, Custom);
501 // Vector ExtLoad and TruncStore are expanded.
502 for (unsigned I = MVT::FIRST_VECTOR_VALUETYPE;
503 I <= MVT::LAST_VECTOR_VALUETYPE; ++I) {
504 MVT VT = (MVT::SimpleValueType) I;
505 setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
506 setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
507 setLoadExtAction(ISD::EXTLOAD, VT, Expand);
508 for (unsigned II = MVT::FIRST_VECTOR_VALUETYPE;
509 II <= MVT::LAST_VECTOR_VALUETYPE; ++II) {
510 MVT VT1 = (MVT::SimpleValueType) II;
511 // A TruncStore has two vector types of the same number of elements
512 // and different element sizes.
513 if (VT.getVectorNumElements() == VT1.getVectorNumElements() &&
514 VT.getVectorElementType().getSizeInBits()
515 > VT1.getVectorElementType().getSizeInBits())
516 setTruncStoreAction(VT, VT1, Expand);
519 setOperationAction(ISD::MULHS, VT, Expand);
520 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
521 setOperationAction(ISD::MULHU, VT, Expand);
522 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
524 setOperationAction(ISD::BSWAP, VT, Expand);
527 // There is no v1i64/v2i64 multiply, expand v1i64/v2i64 to GPR i64 multiply.
528 // FIXME: For a v2i64 multiply, we copy VPR to GPR and do 2 i64 multiplies,
529 // and then copy back to VPR. This solution may be optimized by Following 3
530 // NEON instructions:
531 // pmull v2.1q, v0.1d, v1.1d
532 // pmull2 v3.1q, v0.2d, v1.2d
533 // ins v2.d[1], v3.d[0]
534 // As currently we can't verify the correctness of such assumption, we can
535 // do such optimization in the future.
536 setOperationAction(ISD::MUL, MVT::v1i64, Expand);
537 setOperationAction(ISD::MUL, MVT::v2i64, Expand);
539 setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
540 setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
541 setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
542 setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
543 setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
544 setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
545 setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
546 setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
547 setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
550 setTargetDAGCombine(ISD::SIGN_EXTEND);
551 setTargetDAGCombine(ISD::VSELECT);
553 MaskAndBranchFoldingIsLegal = true;
556 EVT AArch64TargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
557 // It's reasonably important that this value matches the "natural" legal
558 // promotion from i1 for scalar types. Otherwise LegalizeTypes can get itself
559 // in a twist (e.g. inserting an any_extend which then becomes i64 -> i64).
560 if (!VT.isVector()) return MVT::i32;
561 return VT.changeVectorElementTypeToInteger();
564 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
567 static const unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
568 AArch64::LDXR_word, AArch64::LDXR_dword};
569 static const unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
570 AArch64::LDAXR_word, AArch64::LDAXR_dword};
571 static const unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
572 AArch64::STXR_word, AArch64::STXR_dword};
573 static const unsigned StoreRels[] = {AArch64::STLXR_byte,AArch64::STLXR_hword,
574 AArch64::STLXR_word, AArch64::STLXR_dword};
576 const unsigned *LoadOps, *StoreOps;
577 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
582 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
583 StoreOps = StoreRels;
585 StoreOps = StoreBares;
587 assert(isPowerOf2_32(Size) && Size <= 8 &&
588 "unsupported size for atomic binary op!");
590 LdrOpc = LoadOps[Log2_32(Size)];
591 StrOpc = StoreOps[Log2_32(Size)];
594 // FIXME: AArch64::DTripleRegClass and AArch64::QTripleRegClass don't really
595 // have value type mapped, and they are both being defined as MVT::untyped.
596 // Without knowing the MVT type, MachineLICM::getRegisterClassIDAndCost
597 // would fail to figure out the register pressure correctly.
598 std::pair<const TargetRegisterClass*, uint8_t>
599 AArch64TargetLowering::findRepresentativeClass(MVT VT) const{
600 const TargetRegisterClass *RRC = nullptr;
602 switch (VT.SimpleTy) {
604 return TargetLowering::findRepresentativeClass(VT);
606 RRC = &AArch64::QPairRegClass;
610 RRC = &AArch64::QQuadRegClass;
614 return std::make_pair(RRC, Cost);
618 AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
620 unsigned BinOpcode) const {
621 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
622 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
624 const BasicBlock *LLVM_BB = BB->getBasicBlock();
625 MachineFunction *MF = BB->getParent();
626 MachineFunction::iterator It = BB;
629 unsigned dest = MI->getOperand(0).getReg();
630 unsigned ptr = MI->getOperand(1).getReg();
631 unsigned incr = MI->getOperand(2).getReg();
632 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
633 DebugLoc dl = MI->getDebugLoc();
635 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
637 unsigned ldrOpc, strOpc;
638 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
640 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
641 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
642 MF->insert(It, loopMBB);
643 MF->insert(It, exitMBB);
645 // Transfer the remainder of BB and its successor edges to exitMBB.
646 exitMBB->splice(exitMBB->begin(), BB,
647 std::next(MachineBasicBlock::iterator(MI)), BB->end());
648 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
650 const TargetRegisterClass *TRC
651 = Size == 8 ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
652 unsigned scratch = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
656 // fallthrough --> loopMBB
657 BB->addSuccessor(loopMBB);
661 // <binop> scratch, dest, incr
662 // stxr stxr_status, scratch, ptr
663 // cbnz stxr_status, loopMBB
664 // fallthrough --> exitMBB
666 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
668 // All arithmetic operations we'll be creating are designed to take an extra
669 // shift or extend operand, which we can conveniently set to zero.
671 // Operand order needs to go the other way for NAND.
672 if (BinOpcode == AArch64::BICwww_lsl || BinOpcode == AArch64::BICxxx_lsl)
673 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
674 .addReg(incr).addReg(dest).addImm(0);
676 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
677 .addReg(dest).addReg(incr).addImm(0);
680 // From the stxr, the register is GPR32; from the cmp it's GPR32wsp
681 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
682 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
684 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(scratch).addReg(ptr);
685 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
686 .addReg(stxr_status).addMBB(loopMBB);
688 BB->addSuccessor(loopMBB);
689 BB->addSuccessor(exitMBB);
695 MI->eraseFromParent(); // The instruction is gone now.
701 AArch64TargetLowering::emitAtomicBinaryMinMax(MachineInstr *MI,
702 MachineBasicBlock *BB,
705 A64CC::CondCodes Cond) const {
706 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
708 const BasicBlock *LLVM_BB = BB->getBasicBlock();
709 MachineFunction *MF = BB->getParent();
710 MachineFunction::iterator It = BB;
713 unsigned dest = MI->getOperand(0).getReg();
714 unsigned ptr = MI->getOperand(1).getReg();
715 unsigned incr = MI->getOperand(2).getReg();
716 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
718 unsigned oldval = dest;
719 DebugLoc dl = MI->getDebugLoc();
721 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
722 const TargetRegisterClass *TRC, *TRCsp;
724 TRC = &AArch64::GPR64RegClass;
725 TRCsp = &AArch64::GPR64xspRegClass;
727 TRC = &AArch64::GPR32RegClass;
728 TRCsp = &AArch64::GPR32wspRegClass;
731 unsigned ldrOpc, strOpc;
732 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
734 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
735 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
736 MF->insert(It, loopMBB);
737 MF->insert(It, exitMBB);
739 // Transfer the remainder of BB and its successor edges to exitMBB.
740 exitMBB->splice(exitMBB->begin(), BB,
741 std::next(MachineBasicBlock::iterator(MI)), BB->end());
742 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
744 unsigned scratch = MRI.createVirtualRegister(TRC);
745 MRI.constrainRegClass(scratch, TRCsp);
749 // fallthrough --> loopMBB
750 BB->addSuccessor(loopMBB);
754 // cmp incr, dest (, sign extend if necessary)
755 // csel scratch, dest, incr, cond
756 // stxr stxr_status, scratch, ptr
757 // cbnz stxr_status, loopMBB
758 // fallthrough --> exitMBB
760 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
762 // Build compare and cmov instructions.
763 MRI.constrainRegClass(incr, TRCsp);
764 BuildMI(BB, dl, TII->get(CmpOp))
765 .addReg(incr).addReg(oldval).addImm(0);
767 BuildMI(BB, dl, TII->get(Size == 8 ? AArch64::CSELxxxc : AArch64::CSELwwwc),
769 .addReg(oldval).addReg(incr).addImm(Cond);
771 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
772 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
774 BuildMI(BB, dl, TII->get(strOpc), stxr_status)
775 .addReg(scratch).addReg(ptr);
776 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
777 .addReg(stxr_status).addMBB(loopMBB);
779 BB->addSuccessor(loopMBB);
780 BB->addSuccessor(exitMBB);
786 MI->eraseFromParent(); // The instruction is gone now.
792 AArch64TargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
793 MachineBasicBlock *BB,
794 unsigned Size) const {
795 unsigned dest = MI->getOperand(0).getReg();
796 unsigned ptr = MI->getOperand(1).getReg();
797 unsigned oldval = MI->getOperand(2).getReg();
798 unsigned newval = MI->getOperand(3).getReg();
799 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
800 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
801 DebugLoc dl = MI->getDebugLoc();
803 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
804 const TargetRegisterClass *TRCsp;
805 TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
807 unsigned ldrOpc, strOpc;
808 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
810 MachineFunction *MF = BB->getParent();
811 const BasicBlock *LLVM_BB = BB->getBasicBlock();
812 MachineFunction::iterator It = BB;
813 ++It; // insert the new blocks after the current block
815 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
816 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
817 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
818 MF->insert(It, loop1MBB);
819 MF->insert(It, loop2MBB);
820 MF->insert(It, exitMBB);
822 // Transfer the remainder of BB and its successor edges to exitMBB.
823 exitMBB->splice(exitMBB->begin(), BB,
824 std::next(MachineBasicBlock::iterator(MI)), BB->end());
825 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
829 // fallthrough --> loop1MBB
830 BB->addSuccessor(loop1MBB);
837 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
839 unsigned CmpOp = Size == 8 ? AArch64::CMPxx_lsl : AArch64::CMPww_lsl;
840 MRI.constrainRegClass(dest, TRCsp);
841 BuildMI(BB, dl, TII->get(CmpOp))
842 .addReg(dest).addReg(oldval).addImm(0);
843 BuildMI(BB, dl, TII->get(AArch64::Bcc))
844 .addImm(A64CC::NE).addMBB(exitMBB);
845 BB->addSuccessor(loop2MBB);
846 BB->addSuccessor(exitMBB);
849 // strex stxr_status, newval, [ptr]
850 // cbnz stxr_status, loop1MBB
852 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
853 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
855 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(newval).addReg(ptr);
856 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
857 .addReg(stxr_status).addMBB(loop1MBB);
858 BB->addSuccessor(loop1MBB);
859 BB->addSuccessor(exitMBB);
865 MI->eraseFromParent(); // The instruction is gone now.
871 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
872 MachineBasicBlock *MBB) const {
873 // We materialise the F128CSEL pseudo-instruction using conditional branches
874 // and loads, giving an instruciton sequence like:
883 // Using virtual registers would probably not be beneficial since COPY
884 // instructions are expensive for f128 (there's no actual instruction to
887 // An alternative would be to do an integer-CSEL on some address. E.g.:
892 // csel x0, x0, x1, ne
895 // It's unclear which approach is actually optimal.
896 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
897 MachineFunction *MF = MBB->getParent();
898 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
899 DebugLoc DL = MI->getDebugLoc();
900 MachineFunction::iterator It = MBB;
903 unsigned DestReg = MI->getOperand(0).getReg();
904 unsigned IfTrueReg = MI->getOperand(1).getReg();
905 unsigned IfFalseReg = MI->getOperand(2).getReg();
906 unsigned CondCode = MI->getOperand(3).getImm();
907 bool NZCVKilled = MI->getOperand(4).isKill();
909 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
910 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
911 MF->insert(It, TrueBB);
912 MF->insert(It, EndBB);
914 // Transfer rest of current basic-block to EndBB
915 EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
917 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
919 // We need somewhere to store the f128 value needed.
920 int ScratchFI = MF->getFrameInfo()->CreateSpillStackObject(16, 16);
922 // [... start of incoming MBB ...]
923 // str qIFFALSE, [sp]
926 BuildMI(MBB, DL, TII->get(AArch64::LSFP128_STR))
928 .addFrameIndex(ScratchFI)
930 BuildMI(MBB, DL, TII->get(AArch64::Bcc))
933 BuildMI(MBB, DL, TII->get(AArch64::Bimm))
935 MBB->addSuccessor(TrueBB);
936 MBB->addSuccessor(EndBB);
939 // NZCV is live-through TrueBB.
940 TrueBB->addLiveIn(AArch64::NZCV);
941 EndBB->addLiveIn(AArch64::NZCV);
946 BuildMI(TrueBB, DL, TII->get(AArch64::LSFP128_STR))
948 .addFrameIndex(ScratchFI)
951 // Note: fallthrough. We can rely on LLVM adding a branch if it reorders the
953 TrueBB->addSuccessor(EndBB);
957 // [... rest of incoming MBB ...]
958 MachineInstr *StartOfEnd = EndBB->begin();
959 BuildMI(*EndBB, StartOfEnd, DL, TII->get(AArch64::LSFP128_LDR), DestReg)
960 .addFrameIndex(ScratchFI)
963 MI->eraseFromParent();
968 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
969 MachineBasicBlock *MBB) const {
970 switch (MI->getOpcode()) {
971 default: llvm_unreachable("Unhandled instruction with custom inserter");
972 case AArch64::F128CSEL:
973 return EmitF128CSEL(MI, MBB);
974 case AArch64::ATOMIC_LOAD_ADD_I8:
975 return emitAtomicBinary(MI, MBB, 1, AArch64::ADDwww_lsl);
976 case AArch64::ATOMIC_LOAD_ADD_I16:
977 return emitAtomicBinary(MI, MBB, 2, AArch64::ADDwww_lsl);
978 case AArch64::ATOMIC_LOAD_ADD_I32:
979 return emitAtomicBinary(MI, MBB, 4, AArch64::ADDwww_lsl);
980 case AArch64::ATOMIC_LOAD_ADD_I64:
981 return emitAtomicBinary(MI, MBB, 8, AArch64::ADDxxx_lsl);
983 case AArch64::ATOMIC_LOAD_SUB_I8:
984 return emitAtomicBinary(MI, MBB, 1, AArch64::SUBwww_lsl);
985 case AArch64::ATOMIC_LOAD_SUB_I16:
986 return emitAtomicBinary(MI, MBB, 2, AArch64::SUBwww_lsl);
987 case AArch64::ATOMIC_LOAD_SUB_I32:
988 return emitAtomicBinary(MI, MBB, 4, AArch64::SUBwww_lsl);
989 case AArch64::ATOMIC_LOAD_SUB_I64:
990 return emitAtomicBinary(MI, MBB, 8, AArch64::SUBxxx_lsl);
992 case AArch64::ATOMIC_LOAD_AND_I8:
993 return emitAtomicBinary(MI, MBB, 1, AArch64::ANDwww_lsl);
994 case AArch64::ATOMIC_LOAD_AND_I16:
995 return emitAtomicBinary(MI, MBB, 2, AArch64::ANDwww_lsl);
996 case AArch64::ATOMIC_LOAD_AND_I32:
997 return emitAtomicBinary(MI, MBB, 4, AArch64::ANDwww_lsl);
998 case AArch64::ATOMIC_LOAD_AND_I64:
999 return emitAtomicBinary(MI, MBB, 8, AArch64::ANDxxx_lsl);
1001 case AArch64::ATOMIC_LOAD_OR_I8:
1002 return emitAtomicBinary(MI, MBB, 1, AArch64::ORRwww_lsl);
1003 case AArch64::ATOMIC_LOAD_OR_I16:
1004 return emitAtomicBinary(MI, MBB, 2, AArch64::ORRwww_lsl);
1005 case AArch64::ATOMIC_LOAD_OR_I32:
1006 return emitAtomicBinary(MI, MBB, 4, AArch64::ORRwww_lsl);
1007 case AArch64::ATOMIC_LOAD_OR_I64:
1008 return emitAtomicBinary(MI, MBB, 8, AArch64::ORRxxx_lsl);
1010 case AArch64::ATOMIC_LOAD_XOR_I8:
1011 return emitAtomicBinary(MI, MBB, 1, AArch64::EORwww_lsl);
1012 case AArch64::ATOMIC_LOAD_XOR_I16:
1013 return emitAtomicBinary(MI, MBB, 2, AArch64::EORwww_lsl);
1014 case AArch64::ATOMIC_LOAD_XOR_I32:
1015 return emitAtomicBinary(MI, MBB, 4, AArch64::EORwww_lsl);
1016 case AArch64::ATOMIC_LOAD_XOR_I64:
1017 return emitAtomicBinary(MI, MBB, 8, AArch64::EORxxx_lsl);
1019 case AArch64::ATOMIC_LOAD_NAND_I8:
1020 return emitAtomicBinary(MI, MBB, 1, AArch64::BICwww_lsl);
1021 case AArch64::ATOMIC_LOAD_NAND_I16:
1022 return emitAtomicBinary(MI, MBB, 2, AArch64::BICwww_lsl);
1023 case AArch64::ATOMIC_LOAD_NAND_I32:
1024 return emitAtomicBinary(MI, MBB, 4, AArch64::BICwww_lsl);
1025 case AArch64::ATOMIC_LOAD_NAND_I64:
1026 return emitAtomicBinary(MI, MBB, 8, AArch64::BICxxx_lsl);
1028 case AArch64::ATOMIC_LOAD_MIN_I8:
1029 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::GT);
1030 case AArch64::ATOMIC_LOAD_MIN_I16:
1031 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::GT);
1032 case AArch64::ATOMIC_LOAD_MIN_I32:
1033 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::GT);
1034 case AArch64::ATOMIC_LOAD_MIN_I64:
1035 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::GT);
1037 case AArch64::ATOMIC_LOAD_MAX_I8:
1038 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::LT);
1039 case AArch64::ATOMIC_LOAD_MAX_I16:
1040 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::LT);
1041 case AArch64::ATOMIC_LOAD_MAX_I32:
1042 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LT);
1043 case AArch64::ATOMIC_LOAD_MAX_I64:
1044 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LT);
1046 case AArch64::ATOMIC_LOAD_UMIN_I8:
1047 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::HI);
1048 case AArch64::ATOMIC_LOAD_UMIN_I16:
1049 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::HI);
1050 case AArch64::ATOMIC_LOAD_UMIN_I32:
1051 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::HI);
1052 case AArch64::ATOMIC_LOAD_UMIN_I64:
1053 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::HI);
1055 case AArch64::ATOMIC_LOAD_UMAX_I8:
1056 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::LO);
1057 case AArch64::ATOMIC_LOAD_UMAX_I16:
1058 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::LO);
1059 case AArch64::ATOMIC_LOAD_UMAX_I32:
1060 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LO);
1061 case AArch64::ATOMIC_LOAD_UMAX_I64:
1062 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LO);
1064 case AArch64::ATOMIC_SWAP_I8:
1065 return emitAtomicBinary(MI, MBB, 1, 0);
1066 case AArch64::ATOMIC_SWAP_I16:
1067 return emitAtomicBinary(MI, MBB, 2, 0);
1068 case AArch64::ATOMIC_SWAP_I32:
1069 return emitAtomicBinary(MI, MBB, 4, 0);
1070 case AArch64::ATOMIC_SWAP_I64:
1071 return emitAtomicBinary(MI, MBB, 8, 0);
1073 case AArch64::ATOMIC_CMP_SWAP_I8:
1074 return emitAtomicCmpSwap(MI, MBB, 1);
1075 case AArch64::ATOMIC_CMP_SWAP_I16:
1076 return emitAtomicCmpSwap(MI, MBB, 2);
1077 case AArch64::ATOMIC_CMP_SWAP_I32:
1078 return emitAtomicCmpSwap(MI, MBB, 4);
1079 case AArch64::ATOMIC_CMP_SWAP_I64:
1080 return emitAtomicCmpSwap(MI, MBB, 8);
1085 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
1087 case AArch64ISD::BR_CC: return "AArch64ISD::BR_CC";
1088 case AArch64ISD::Call: return "AArch64ISD::Call";
1089 case AArch64ISD::FPMOV: return "AArch64ISD::FPMOV";
1090 case AArch64ISD::GOTLoad: return "AArch64ISD::GOTLoad";
1091 case AArch64ISD::BFI: return "AArch64ISD::BFI";
1092 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
1093 case AArch64ISD::Ret: return "AArch64ISD::Ret";
1094 case AArch64ISD::SBFX: return "AArch64ISD::SBFX";
1095 case AArch64ISD::SELECT_CC: return "AArch64ISD::SELECT_CC";
1096 case AArch64ISD::SETCC: return "AArch64ISD::SETCC";
1097 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
1098 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
1099 case AArch64ISD::TLSDESCCALL: return "AArch64ISD::TLSDESCCALL";
1100 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
1101 case AArch64ISD::WrapperSmall: return "AArch64ISD::WrapperSmall";
1103 case AArch64ISD::NEON_MOVIMM:
1104 return "AArch64ISD::NEON_MOVIMM";
1105 case AArch64ISD::NEON_MVNIMM:
1106 return "AArch64ISD::NEON_MVNIMM";
1107 case AArch64ISD::NEON_FMOVIMM:
1108 return "AArch64ISD::NEON_FMOVIMM";
1109 case AArch64ISD::NEON_CMP:
1110 return "AArch64ISD::NEON_CMP";
1111 case AArch64ISD::NEON_CMPZ:
1112 return "AArch64ISD::NEON_CMPZ";
1113 case AArch64ISD::NEON_TST:
1114 return "AArch64ISD::NEON_TST";
1115 case AArch64ISD::NEON_QSHLs:
1116 return "AArch64ISD::NEON_QSHLs";
1117 case AArch64ISD::NEON_QSHLu:
1118 return "AArch64ISD::NEON_QSHLu";
1119 case AArch64ISD::NEON_VDUP:
1120 return "AArch64ISD::NEON_VDUP";
1121 case AArch64ISD::NEON_VDUPLANE:
1122 return "AArch64ISD::NEON_VDUPLANE";
1123 case AArch64ISD::NEON_REV16:
1124 return "AArch64ISD::NEON_REV16";
1125 case AArch64ISD::NEON_REV32:
1126 return "AArch64ISD::NEON_REV32";
1127 case AArch64ISD::NEON_REV64:
1128 return "AArch64ISD::NEON_REV64";
1129 case AArch64ISD::NEON_UZP1:
1130 return "AArch64ISD::NEON_UZP1";
1131 case AArch64ISD::NEON_UZP2:
1132 return "AArch64ISD::NEON_UZP2";
1133 case AArch64ISD::NEON_ZIP1:
1134 return "AArch64ISD::NEON_ZIP1";
1135 case AArch64ISD::NEON_ZIP2:
1136 return "AArch64ISD::NEON_ZIP2";
1137 case AArch64ISD::NEON_TRN1:
1138 return "AArch64ISD::NEON_TRN1";
1139 case AArch64ISD::NEON_TRN2:
1140 return "AArch64ISD::NEON_TRN2";
1141 case AArch64ISD::NEON_LD1_UPD:
1142 return "AArch64ISD::NEON_LD1_UPD";
1143 case AArch64ISD::NEON_LD2_UPD:
1144 return "AArch64ISD::NEON_LD2_UPD";
1145 case AArch64ISD::NEON_LD3_UPD:
1146 return "AArch64ISD::NEON_LD3_UPD";
1147 case AArch64ISD::NEON_LD4_UPD:
1148 return "AArch64ISD::NEON_LD4_UPD";
1149 case AArch64ISD::NEON_ST1_UPD:
1150 return "AArch64ISD::NEON_ST1_UPD";
1151 case AArch64ISD::NEON_ST2_UPD:
1152 return "AArch64ISD::NEON_ST2_UPD";
1153 case AArch64ISD::NEON_ST3_UPD:
1154 return "AArch64ISD::NEON_ST3_UPD";
1155 case AArch64ISD::NEON_ST4_UPD:
1156 return "AArch64ISD::NEON_ST4_UPD";
1157 case AArch64ISD::NEON_LD1x2_UPD:
1158 return "AArch64ISD::NEON_LD1x2_UPD";
1159 case AArch64ISD::NEON_LD1x3_UPD:
1160 return "AArch64ISD::NEON_LD1x3_UPD";
1161 case AArch64ISD::NEON_LD1x4_UPD:
1162 return "AArch64ISD::NEON_LD1x4_UPD";
1163 case AArch64ISD::NEON_ST1x2_UPD:
1164 return "AArch64ISD::NEON_ST1x2_UPD";
1165 case AArch64ISD::NEON_ST1x3_UPD:
1166 return "AArch64ISD::NEON_ST1x3_UPD";
1167 case AArch64ISD::NEON_ST1x4_UPD:
1168 return "AArch64ISD::NEON_ST1x4_UPD";
1169 case AArch64ISD::NEON_LD2DUP:
1170 return "AArch64ISD::NEON_LD2DUP";
1171 case AArch64ISD::NEON_LD3DUP:
1172 return "AArch64ISD::NEON_LD3DUP";
1173 case AArch64ISD::NEON_LD4DUP:
1174 return "AArch64ISD::NEON_LD4DUP";
1175 case AArch64ISD::NEON_LD2DUP_UPD:
1176 return "AArch64ISD::NEON_LD2DUP_UPD";
1177 case AArch64ISD::NEON_LD3DUP_UPD:
1178 return "AArch64ISD::NEON_LD3DUP_UPD";
1179 case AArch64ISD::NEON_LD4DUP_UPD:
1180 return "AArch64ISD::NEON_LD4DUP_UPD";
1181 case AArch64ISD::NEON_LD2LN_UPD:
1182 return "AArch64ISD::NEON_LD2LN_UPD";
1183 case AArch64ISD::NEON_LD3LN_UPD:
1184 return "AArch64ISD::NEON_LD3LN_UPD";
1185 case AArch64ISD::NEON_LD4LN_UPD:
1186 return "AArch64ISD::NEON_LD4LN_UPD";
1187 case AArch64ISD::NEON_ST2LN_UPD:
1188 return "AArch64ISD::NEON_ST2LN_UPD";
1189 case AArch64ISD::NEON_ST3LN_UPD:
1190 return "AArch64ISD::NEON_ST3LN_UPD";
1191 case AArch64ISD::NEON_ST4LN_UPD:
1192 return "AArch64ISD::NEON_ST4LN_UPD";
1193 case AArch64ISD::NEON_VEXTRACT:
1194 return "AArch64ISD::NEON_VEXTRACT";
1200 static const MCPhysReg AArch64FPRArgRegs[] = {
1201 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
1202 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7
1204 static const unsigned NumFPRArgRegs = llvm::array_lengthof(AArch64FPRArgRegs);
1206 static const MCPhysReg AArch64ArgRegs[] = {
1207 AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3,
1208 AArch64::X4, AArch64::X5, AArch64::X6, AArch64::X7
1210 static const unsigned NumArgRegs = llvm::array_lengthof(AArch64ArgRegs);
1212 static bool CC_AArch64NoMoreRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
1213 CCValAssign::LocInfo LocInfo,
1214 ISD::ArgFlagsTy ArgFlags, CCState &State) {
1215 // Mark all remaining general purpose registers as allocated. We don't
1216 // backtrack: if (for example) an i128 gets put on the stack, no subsequent
1217 // i64 will go in registers (C.11).
1218 for (unsigned i = 0; i < NumArgRegs; ++i)
1219 State.AllocateReg(AArch64ArgRegs[i]);
1224 #include "AArch64GenCallingConv.inc"
1226 CCAssignFn *AArch64TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1229 default: llvm_unreachable("Unsupported calling convention");
1230 case CallingConv::Fast:
1231 case CallingConv::C:
1237 AArch64TargetLowering::SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG,
1238 SDLoc DL, SDValue &Chain) const {
1239 MachineFunction &MF = DAG.getMachineFunction();
1240 MachineFrameInfo *MFI = MF.getFrameInfo();
1241 AArch64MachineFunctionInfo *FuncInfo
1242 = MF.getInfo<AArch64MachineFunctionInfo>();
1244 SmallVector<SDValue, 8> MemOps;
1246 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(AArch64ArgRegs,
1248 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(AArch64FPRArgRegs,
1251 unsigned GPRSaveSize = 8 * (NumArgRegs - FirstVariadicGPR);
1253 if (GPRSaveSize != 0) {
1254 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
1256 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
1258 for (unsigned i = FirstVariadicGPR; i < NumArgRegs; ++i) {
1259 unsigned VReg = MF.addLiveIn(AArch64ArgRegs[i], &AArch64::GPR64RegClass);
1260 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
1261 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1262 MachinePointerInfo::getStack(i * 8),
1264 MemOps.push_back(Store);
1265 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1266 DAG.getConstant(8, getPointerTy()));
1270 if (getSubtarget()->hasFPARMv8()) {
1271 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
1273 // According to the AArch64 Procedure Call Standard, section B.1/B.3, we
1274 // can omit a register save area if we know we'll never use registers of
1276 if (FPRSaveSize != 0) {
1277 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
1279 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
1281 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
1282 unsigned VReg = MF.addLiveIn(AArch64FPRArgRegs[i],
1283 &AArch64::FPR128RegClass);
1284 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
1285 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
1286 MachinePointerInfo::getStack(i * 16),
1288 MemOps.push_back(Store);
1289 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
1290 DAG.getConstant(16, getPointerTy()));
1293 FuncInfo->setVariadicFPRIdx(FPRIdx);
1294 FuncInfo->setVariadicFPRSize(FPRSaveSize);
1297 unsigned StackOffset = RoundUpToAlignment(CCInfo.getNextStackOffset(), 8);
1298 int StackIdx = MFI->CreateFixedObject(8, StackOffset, true);
1300 FuncInfo->setVariadicStackIdx(StackIdx);
1301 FuncInfo->setVariadicGPRIdx(GPRIdx);
1302 FuncInfo->setVariadicGPRSize(GPRSaveSize);
1304 if (!MemOps.empty()) {
1305 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
1311 AArch64TargetLowering::LowerFormalArguments(SDValue Chain,
1312 CallingConv::ID CallConv, bool isVarArg,
1313 const SmallVectorImpl<ISD::InputArg> &Ins,
1314 SDLoc dl, SelectionDAG &DAG,
1315 SmallVectorImpl<SDValue> &InVals) const {
1316 MachineFunction &MF = DAG.getMachineFunction();
1317 AArch64MachineFunctionInfo *FuncInfo
1318 = MF.getInfo<AArch64MachineFunctionInfo>();
1319 MachineFrameInfo *MFI = MF.getFrameInfo();
1320 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1322 SmallVector<CCValAssign, 16> ArgLocs;
1323 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1324 getTargetMachine(), ArgLocs, *DAG.getContext());
1325 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1327 SmallVector<SDValue, 16> ArgValues;
1330 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1331 CCValAssign &VA = ArgLocs[i];
1332 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1334 if (Flags.isByVal()) {
1335 // Byval is used for small structs and HFAs in the PCS, but the system
1336 // should work in a non-compliant manner for larger structs.
1337 EVT PtrTy = getPointerTy();
1338 int Size = Flags.getByValSize();
1339 unsigned NumRegs = (Size + 7) / 8;
1341 uint32_t BEAlign = 0;
1342 if (Size < 8 && !getSubtarget()->isLittle())
1344 unsigned FrameIdx = MFI->CreateFixedObject(8 * NumRegs,
1345 VA.getLocMemOffset() + BEAlign,
1347 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
1348 InVals.push_back(FrameIdxN);
1351 } else if (VA.isRegLoc()) {
1352 MVT RegVT = VA.getLocVT();
1353 const TargetRegisterClass *RC = getRegClassFor(RegVT);
1354 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1356 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1357 } else { // VA.isRegLoc()
1358 assert(VA.isMemLoc());
1360 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
1361 VA.getLocMemOffset(), true);
1363 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1364 ArgValue = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
1365 MachinePointerInfo::getFixedStack(FI),
1366 false, false, false, 0);
1371 switch (VA.getLocInfo()) {
1372 default: llvm_unreachable("Unknown loc info!");
1373 case CCValAssign::Full: break;
1374 case CCValAssign::BCvt:
1375 ArgValue = DAG.getNode(ISD::BITCAST,dl, VA.getValVT(), ArgValue);
1377 case CCValAssign::SExt:
1378 case CCValAssign::ZExt:
1379 case CCValAssign::AExt:
1380 case CCValAssign::FPExt: {
1381 unsigned DestSize = VA.getValVT().getSizeInBits();
1382 unsigned DestSubReg;
1385 case 8: DestSubReg = AArch64::sub_8; break;
1386 case 16: DestSubReg = AArch64::sub_16; break;
1387 case 32: DestSubReg = AArch64::sub_32; break;
1388 case 64: DestSubReg = AArch64::sub_64; break;
1389 default: llvm_unreachable("Unexpected argument promotion");
1392 ArgValue = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
1393 VA.getValVT(), ArgValue,
1394 DAG.getTargetConstant(DestSubReg, MVT::i32)),
1400 InVals.push_back(ArgValue);
1404 SaveVarArgRegisters(CCInfo, DAG, dl, Chain);
1406 unsigned StackArgSize = CCInfo.getNextStackOffset();
1407 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
1408 // This is a non-standard ABI so by fiat I say we're allowed to make full
1409 // use of the stack area to be popped, which must be aligned to 16 bytes in
1411 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
1413 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
1414 // a multiple of 16.
1415 FuncInfo->setArgumentStackToRestore(StackArgSize);
1417 // This realignment carries over to the available bytes below. Our own
1418 // callers will guarantee the space is free by giving an aligned value to
1421 // Even if we're not expected to free up the space, it's useful to know how
1422 // much is there while considering tail calls (because we can reuse it).
1423 FuncInfo->setBytesInStackArgArea(StackArgSize);
1429 AArch64TargetLowering::LowerReturn(SDValue Chain,
1430 CallingConv::ID CallConv, bool isVarArg,
1431 const SmallVectorImpl<ISD::OutputArg> &Outs,
1432 const SmallVectorImpl<SDValue> &OutVals,
1433 SDLoc dl, SelectionDAG &DAG) const {
1434 // CCValAssign - represent the assignment of the return value to a location.
1435 SmallVector<CCValAssign, 16> RVLocs;
1437 // CCState - Info about the registers and stack slots.
1438 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1439 getTargetMachine(), RVLocs, *DAG.getContext());
1441 // Analyze outgoing return values.
1442 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv));
1445 SmallVector<SDValue, 4> RetOps(1, Chain);
1447 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1448 // PCS: "If the type, T, of the result of a function is such that
1449 // void func(T arg) would require that arg be passed as a value in a
1450 // register (or set of registers) according to the rules in 5.4, then the
1451 // result is returned in the same registers as would be used for such an
1454 // Otherwise, the caller shall reserve a block of memory of sufficient
1455 // size and alignment to hold the result. The address of the memory block
1456 // shall be passed as an additional argument to the function in x8."
1458 // This is implemented in two places. The register-return values are dealt
1459 // with here, more complex returns are passed as an sret parameter, which
1460 // means we don't have to worry about it during actual return.
1461 CCValAssign &VA = RVLocs[i];
1462 assert(VA.isRegLoc() && "Only register-returns should be created by PCS");
1465 SDValue Arg = OutVals[i];
1467 // There's no convenient note in the ABI about this as there is for normal
1468 // arguments, but it says return values are passed in the same registers as
1469 // an argument would be. I believe that includes the comments about
1470 // unspecified higher bits, putting the burden of widening on the *caller*
1471 // for return values.
1472 switch (VA.getLocInfo()) {
1473 default: llvm_unreachable("Unknown loc info");
1474 case CCValAssign::Full: break;
1475 case CCValAssign::SExt:
1476 case CCValAssign::ZExt:
1477 case CCValAssign::AExt:
1478 // Floating-point values should only be extended when they're going into
1479 // memory, which can't happen here so an integer extend is acceptable.
1480 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1482 case CCValAssign::BCvt:
1483 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1487 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1488 Flag = Chain.getValue(1);
1489 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1492 RetOps[0] = Chain; // Update chain.
1494 // Add the flag if we have it.
1496 RetOps.push_back(Flag);
1498 return DAG.getNode(AArch64ISD::Ret, dl, MVT::Other, RetOps);
1501 unsigned AArch64TargetLowering::getByValTypeAlignment(Type *Ty) const {
1502 // This is a new backend. For anything more precise than this a FE should
1503 // set an explicit alignment.
1508 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1509 SmallVectorImpl<SDValue> &InVals) const {
1510 SelectionDAG &DAG = CLI.DAG;
1512 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1513 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1514 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1515 SDValue Chain = CLI.Chain;
1516 SDValue Callee = CLI.Callee;
1517 bool &IsTailCall = CLI.IsTailCall;
1518 CallingConv::ID CallConv = CLI.CallConv;
1519 bool IsVarArg = CLI.IsVarArg;
1521 MachineFunction &MF = DAG.getMachineFunction();
1522 AArch64MachineFunctionInfo *FuncInfo
1523 = MF.getInfo<AArch64MachineFunctionInfo>();
1524 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1525 bool IsStructRet = !Outs.empty() && Outs[0].Flags.isSRet();
1526 bool IsSibCall = false;
1529 IsTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1530 IsVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1531 Outs, OutVals, Ins, DAG);
1533 if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
1534 report_fatal_error("failed to perform tail call elimination on a call "
1535 "site marked musttail");
1537 // A sibling call is one where we're under the usual C ABI and not planning
1538 // to change that but can still do a tail call:
1539 if (!TailCallOpt && IsTailCall)
1543 SmallVector<CCValAssign, 16> ArgLocs;
1544 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1545 getTargetMachine(), ArgLocs, *DAG.getContext());
1546 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1548 // On AArch64 (and all other architectures I'm aware of) the most this has to
1549 // do is adjust the stack pointer.
1550 unsigned NumBytes = RoundUpToAlignment(CCInfo.getNextStackOffset(), 16);
1552 // Since we're not changing the ABI to make this a tail call, the memory
1553 // operands are already available in the caller's incoming argument space.
1557 // FPDiff is the byte offset of the call's argument area from the callee's.
1558 // Stores to callee stack arguments will be placed in FixedStackSlots offset
1559 // by this amount for a tail call. In a sibling call it must be 0 because the
1560 // caller will deallocate the entire stack and the callee still expects its
1561 // arguments to begin at SP+0. Completely unused for non-tail calls.
1564 if (IsTailCall && !IsSibCall) {
1565 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
1567 // FPDiff will be negative if this tail call requires more space than we
1568 // would automatically have in our incoming argument space. Positive if we
1569 // can actually shrink the stack.
1570 FPDiff = NumReusableBytes - NumBytes;
1572 // The stack pointer must be 16-byte aligned at all times it's used for a
1573 // memory operation, which in practice means at *all* times and in
1574 // particular across call boundaries. Therefore our own arguments started at
1575 // a 16-byte aligned SP and the delta applied for the tail call should
1576 // satisfy the same constraint.
1577 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
1581 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
1584 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, AArch64::XSP,
1587 SmallVector<SDValue, 8> MemOpChains;
1588 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1590 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1591 CCValAssign &VA = ArgLocs[i];
1592 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1593 SDValue Arg = OutVals[i];
1595 // Callee does the actual widening, so all extensions just use an implicit
1596 // definition of the rest of the Loc. Aesthetically, this would be nicer as
1597 // an ANY_EXTEND, but that isn't valid for floating-point types and this
1598 // alternative works on integer types too.
1599 switch (VA.getLocInfo()) {
1600 default: llvm_unreachable("Unknown loc info!");
1601 case CCValAssign::Full: break;
1602 case CCValAssign::SExt:
1603 case CCValAssign::ZExt:
1604 case CCValAssign::AExt:
1605 case CCValAssign::FPExt: {
1606 unsigned SrcSize = VA.getValVT().getSizeInBits();
1610 case 8: SrcSubReg = AArch64::sub_8; break;
1611 case 16: SrcSubReg = AArch64::sub_16; break;
1612 case 32: SrcSubReg = AArch64::sub_32; break;
1613 case 64: SrcSubReg = AArch64::sub_64; break;
1614 default: llvm_unreachable("Unexpected argument promotion");
1617 Arg = SDValue(DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1619 DAG.getUNDEF(VA.getLocVT()),
1621 DAG.getTargetConstant(SrcSubReg, MVT::i32)),
1626 case CCValAssign::BCvt:
1627 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1631 if (VA.isRegLoc()) {
1632 // A normal register (sub-) argument. For now we just note it down because
1633 // we want to copy things into registers as late as possible to avoid
1634 // register-pressure (and possibly worse).
1635 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1639 assert(VA.isMemLoc() && "unexpected argument location");
1642 MachinePointerInfo DstInfo;
1644 uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize() :
1645 VA.getLocVT().getSizeInBits();
1646 OpSize = (OpSize + 7) / 8;
1647 int32_t Offset = VA.getLocMemOffset() + FPDiff;
1648 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
1650 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
1651 DstInfo = MachinePointerInfo::getFixedStack(FI);
1653 // Make sure any stack arguments overlapping with where we're storing are
1654 // loaded before this eventual operation. Otherwise they'll be clobbered.
1655 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
1657 uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize()*8 :
1658 VA.getLocVT().getSizeInBits();
1659 OpSize = (OpSize + 7) / 8;
1660 uint32_t BEAlign = 0;
1661 if (OpSize < 8 && !getSubtarget()->isLittle())
1663 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset() + BEAlign);
1665 DstAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1666 DstInfo = MachinePointerInfo::getStack(VA.getLocMemOffset());
1669 if (Flags.isByVal()) {
1670 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i64);
1671 SDValue Cpy = DAG.getMemcpy(Chain, dl, DstAddr, Arg, SizeNode,
1672 Flags.getByValAlign(),
1673 /*isVolatile = */ false,
1674 /*alwaysInline = */ false,
1675 DstInfo, MachinePointerInfo());
1676 MemOpChains.push_back(Cpy);
1678 // Normal stack argument, put it where it's needed.
1679 SDValue Store = DAG.getStore(Chain, dl, Arg, DstAddr, DstInfo,
1681 MemOpChains.push_back(Store);
1685 // The loads and stores generated above shouldn't clash with each
1686 // other. Combining them with this TokenFactor notes that fact for the rest of
1688 if (!MemOpChains.empty())
1689 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
1691 // Most of the rest of the instructions need to be glued together; we don't
1692 // want assignments to actual registers used by a call to be rearranged by a
1693 // well-meaning scheduler.
1696 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1697 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1698 RegsToPass[i].second, InFlag);
1699 InFlag = Chain.getValue(1);
1702 // The linker is responsible for inserting veneers when necessary to put a
1703 // function call destination in range, so we don't need to bother with a
1705 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1706 const GlobalValue *GV = G->getGlobal();
1707 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1708 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1709 const char *Sym = S->getSymbol();
1710 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1713 // We don't usually want to end the call-sequence here because we would tidy
1714 // the frame up *after* the call, however in the ABI-changing tail-call case
1715 // we've carefully laid out the parameters so that when sp is reset they'll be
1716 // in the correct location.
1717 if (IsTailCall && !IsSibCall) {
1718 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1719 DAG.getIntPtrConstant(0, true), InFlag, dl);
1720 InFlag = Chain.getValue(1);
1723 // We produce the following DAG scheme for the actual call instruction:
1724 // (AArch64Call Chain, Callee, reg1, ..., regn, preserveMask, inflag?
1726 // Most arguments aren't going to be used and just keep the values live as
1727 // far as LLVM is concerned. It's expected to be selected as simply "bl
1728 // callee" (for a direct, non-tail call).
1729 std::vector<SDValue> Ops;
1730 Ops.push_back(Chain);
1731 Ops.push_back(Callee);
1734 // Each tail call may have to adjust the stack by a different amount, so
1735 // this information must travel along with the operation for eventual
1736 // consumption by emitEpilogue.
1737 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
1740 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1741 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1742 RegsToPass[i].second.getValueType()));
1745 // Add a register mask operand representing the call-preserved registers. This
1746 // is used later in codegen to constrain register-allocation.
1747 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1748 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1749 assert(Mask && "Missing call preserved mask for calling convention");
1750 Ops.push_back(DAG.getRegisterMask(Mask));
1752 // If we needed glue, put it in as the last argument.
1753 if (InFlag.getNode())
1754 Ops.push_back(InFlag);
1756 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1759 return DAG.getNode(AArch64ISD::TC_RETURN, dl, NodeTys, Ops);
1762 Chain = DAG.getNode(AArch64ISD::Call, dl, NodeTys, Ops);
1763 InFlag = Chain.getValue(1);
1765 // Now we can reclaim the stack, just as well do it before working out where
1766 // our return value is.
1768 uint64_t CalleePopBytes
1769 = DoesCalleeRestoreStack(CallConv, TailCallOpt) ? NumBytes : 0;
1771 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1772 DAG.getIntPtrConstant(CalleePopBytes, true),
1774 InFlag = Chain.getValue(1);
1777 return LowerCallResult(Chain, InFlag, CallConv,
1778 IsVarArg, Ins, dl, DAG, InVals);
1782 AArch64TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1783 CallingConv::ID CallConv, bool IsVarArg,
1784 const SmallVectorImpl<ISD::InputArg> &Ins,
1785 SDLoc dl, SelectionDAG &DAG,
1786 SmallVectorImpl<SDValue> &InVals) const {
1787 // Assign locations to each value returned by this call.
1788 SmallVector<CCValAssign, 16> RVLocs;
1789 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1790 getTargetMachine(), RVLocs, *DAG.getContext());
1791 CCInfo.AnalyzeCallResult(Ins, CCAssignFnForNode(CallConv));
1793 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1794 CCValAssign VA = RVLocs[i];
1796 // Return values that are too big to fit into registers should use an sret
1797 // pointer, so this can be a lot simpler than the main argument code.
1798 assert(VA.isRegLoc() && "Memory locations not expected for call return");
1800 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1802 Chain = Val.getValue(1);
1803 InFlag = Val.getValue(2);
1805 switch (VA.getLocInfo()) {
1806 default: llvm_unreachable("Unknown loc info!");
1807 case CCValAssign::Full: break;
1808 case CCValAssign::BCvt:
1809 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1811 case CCValAssign::ZExt:
1812 case CCValAssign::SExt:
1813 case CCValAssign::AExt:
1814 // Floating-point arguments only get extended/truncated if they're going
1815 // in memory, so using the integer operation is acceptable here.
1816 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
1820 InVals.push_back(Val);
1827 AArch64TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1828 CallingConv::ID CalleeCC,
1830 bool IsCalleeStructRet,
1831 bool IsCallerStructRet,
1832 const SmallVectorImpl<ISD::OutputArg> &Outs,
1833 const SmallVectorImpl<SDValue> &OutVals,
1834 const SmallVectorImpl<ISD::InputArg> &Ins,
1835 SelectionDAG& DAG) const {
1837 // For CallingConv::C this function knows whether the ABI needs
1838 // changing. That's not true for other conventions so they will have to opt in
1840 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1843 const MachineFunction &MF = DAG.getMachineFunction();
1844 const Function *CallerF = MF.getFunction();
1845 CallingConv::ID CallerCC = CallerF->getCallingConv();
1846 bool CCMatch = CallerCC == CalleeCC;
1848 // Byval parameters hand the function a pointer directly into the stack area
1849 // we want to reuse during a tail call. Working around this *is* possible (see
1850 // X86) but less efficient and uglier in LowerCall.
1851 for (Function::const_arg_iterator i = CallerF->arg_begin(),
1852 e = CallerF->arg_end(); i != e; ++i)
1853 if (i->hasByValAttr())
1856 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
1857 if (IsTailCallConvention(CalleeCC) && CCMatch)
1862 // Now we search for cases where we can use a tail call without changing the
1863 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
1866 // I want anyone implementing a new calling convention to think long and hard
1867 // about this assert.
1868 assert((!IsVarArg || CalleeCC == CallingConv::C)
1869 && "Unexpected variadic calling convention");
1871 if (IsVarArg && !Outs.empty()) {
1872 // At least two cases here: if caller is fastcc then we can't have any
1873 // memory arguments (we'd be expected to clean up the stack afterwards). If
1874 // caller is C then we could potentially use its argument area.
1876 // FIXME: for now we take the most conservative of these in both cases:
1877 // disallow all variadic memory operands.
1878 SmallVector<CCValAssign, 16> ArgLocs;
1879 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1880 getTargetMachine(), ArgLocs, *DAG.getContext());
1882 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1883 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
1884 if (!ArgLocs[i].isRegLoc())
1888 // If the calling conventions do not match, then we'd better make sure the
1889 // results are returned in the same way as what the caller expects.
1891 SmallVector<CCValAssign, 16> RVLocs1;
1892 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1893 getTargetMachine(), RVLocs1, *DAG.getContext());
1894 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC));
1896 SmallVector<CCValAssign, 16> RVLocs2;
1897 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1898 getTargetMachine(), RVLocs2, *DAG.getContext());
1899 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC));
1901 if (RVLocs1.size() != RVLocs2.size())
1903 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1904 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1906 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1908 if (RVLocs1[i].isRegLoc()) {
1909 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1912 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1918 // Nothing more to check if the callee is taking no arguments
1922 SmallVector<CCValAssign, 16> ArgLocs;
1923 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1924 getTargetMachine(), ArgLocs, *DAG.getContext());
1926 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1928 const AArch64MachineFunctionInfo *FuncInfo
1929 = MF.getInfo<AArch64MachineFunctionInfo>();
1931 // If the stack arguments for this call would fit into our own save area then
1932 // the call can be made tail.
1933 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
1936 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
1937 bool TailCallOpt) const {
1938 return CallCC == CallingConv::Fast && TailCallOpt;
1941 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
1942 return CallCC == CallingConv::Fast;
1945 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
1947 MachineFrameInfo *MFI,
1948 int ClobberedFI) const {
1949 SmallVector<SDValue, 8> ArgChains;
1950 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
1951 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
1953 // Include the original chain at the beginning of the list. When this is
1954 // used by target LowerCall hooks, this helps legalize find the
1955 // CALLSEQ_BEGIN node.
1956 ArgChains.push_back(Chain);
1958 // Add a chain value for each stack argument corresponding
1959 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
1960 UE = DAG.getEntryNode().getNode()->use_end(); U != UE; ++U)
1961 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
1962 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
1963 if (FI->getIndex() < 0) {
1964 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
1965 int64_t InLastByte = InFirstByte;
1966 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
1968 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1969 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1970 ArgChains.push_back(SDValue(L, 1));
1973 // Build a tokenfactor for all the chains.
1974 return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
1977 static A64CC::CondCodes IntCCToA64CC(ISD::CondCode CC) {
1979 case ISD::SETEQ: return A64CC::EQ;
1980 case ISD::SETGT: return A64CC::GT;
1981 case ISD::SETGE: return A64CC::GE;
1982 case ISD::SETLT: return A64CC::LT;
1983 case ISD::SETLE: return A64CC::LE;
1984 case ISD::SETNE: return A64CC::NE;
1985 case ISD::SETUGT: return A64CC::HI;
1986 case ISD::SETUGE: return A64CC::HS;
1987 case ISD::SETULT: return A64CC::LO;
1988 case ISD::SETULE: return A64CC::LS;
1989 default: llvm_unreachable("Unexpected condition code");
1993 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Val) const {
1994 // icmp is implemented using adds/subs immediate, which take an unsigned
1995 // 12-bit immediate, optionally shifted left by 12 bits.
1997 // Symmetric by using adds/subs
2001 return (Val & ~0xfff) == 0 || (Val & ~0xfff000) == 0;
2004 SDValue AArch64TargetLowering::getSelectableIntSetCC(SDValue LHS, SDValue RHS,
2005 ISD::CondCode CC, SDValue &A64cc,
2006 SelectionDAG &DAG, SDLoc &dl) const {
2007 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
2009 EVT VT = RHSC->getValueType(0);
2010 bool knownInvalid = false;
2012 // I'm not convinced the rest of LLVM handles these edge cases properly, but
2013 // we can at least get it right.
2014 if (isSignedIntSetCC(CC)) {
2015 C = RHSC->getSExtValue();
2016 } else if (RHSC->getZExtValue() > INT64_MAX) {
2017 // A 64-bit constant not representable by a signed 64-bit integer is far
2018 // too big to fit into a SUBS immediate anyway.
2019 knownInvalid = true;
2021 C = RHSC->getZExtValue();
2024 if (!knownInvalid && !isLegalICmpImmediate(C)) {
2025 // Constant does not fit, try adjusting it by one?
2030 if (isLegalICmpImmediate(C-1)) {
2031 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
2032 RHS = DAG.getConstant(C-1, VT);
2037 if (isLegalICmpImmediate(C-1)) {
2038 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
2039 RHS = DAG.getConstant(C-1, VT);
2044 if (isLegalICmpImmediate(C+1)) {
2045 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
2046 RHS = DAG.getConstant(C+1, VT);
2051 if (isLegalICmpImmediate(C+1)) {
2052 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
2053 RHS = DAG.getConstant(C+1, VT);
2060 A64CC::CondCodes CondCode = IntCCToA64CC(CC);
2061 A64cc = DAG.getConstant(CondCode, MVT::i32);
2062 return DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2063 DAG.getCondCode(CC));
2066 static A64CC::CondCodes FPCCToA64CC(ISD::CondCode CC,
2067 A64CC::CondCodes &Alternative) {
2068 A64CC::CondCodes CondCode = A64CC::Invalid;
2069 Alternative = A64CC::Invalid;
2072 default: llvm_unreachable("Unknown FP condition!");
2074 case ISD::SETOEQ: CondCode = A64CC::EQ; break;
2076 case ISD::SETOGT: CondCode = A64CC::GT; break;
2078 case ISD::SETOGE: CondCode = A64CC::GE; break;
2079 case ISD::SETOLT: CondCode = A64CC::MI; break;
2080 case ISD::SETOLE: CondCode = A64CC::LS; break;
2081 case ISD::SETONE: CondCode = A64CC::MI; Alternative = A64CC::GT; break;
2082 case ISD::SETO: CondCode = A64CC::VC; break;
2083 case ISD::SETUO: CondCode = A64CC::VS; break;
2084 case ISD::SETUEQ: CondCode = A64CC::EQ; Alternative = A64CC::VS; break;
2085 case ISD::SETUGT: CondCode = A64CC::HI; break;
2086 case ISD::SETUGE: CondCode = A64CC::PL; break;
2088 case ISD::SETULT: CondCode = A64CC::LT; break;
2090 case ISD::SETULE: CondCode = A64CC::LE; break;
2092 case ISD::SETUNE: CondCode = A64CC::NE; break;
2098 AArch64TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
2100 EVT PtrVT = getPointerTy();
2101 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
2103 switch(getTargetMachine().getCodeModel()) {
2104 case CodeModel::Small:
2105 // The most efficient code is PC-relative anyway for the small memory model,
2106 // so we don't need to worry about relocation model.
2107 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2108 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2109 AArch64II::MO_NO_FLAG),
2110 DAG.getTargetBlockAddress(BA, PtrVT, 0,
2111 AArch64II::MO_LO12),
2112 DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
2113 case CodeModel::Large:
2115 AArch64ISD::WrapperLarge, DL, PtrVT,
2116 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G3),
2117 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2118 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2119 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2121 llvm_unreachable("Only small and large code models supported now");
2126 // (BRCOND chain, val, dest)
2128 AArch64TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
2130 SDValue Chain = Op.getOperand(0);
2131 SDValue TheBit = Op.getOperand(1);
2132 SDValue DestBB = Op.getOperand(2);
2134 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2135 // that as the consumer we are responsible for ignoring rubbish in higher
2137 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2138 DAG.getConstant(1, MVT::i32));
2140 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2141 DAG.getConstant(0, TheBit.getValueType()),
2142 DAG.getCondCode(ISD::SETNE));
2144 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other, Chain,
2145 A64CMP, DAG.getConstant(A64CC::NE, MVT::i32),
2149 // (BR_CC chain, condcode, lhs, rhs, dest)
2151 AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
2153 SDValue Chain = Op.getOperand(0);
2154 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
2155 SDValue LHS = Op.getOperand(2);
2156 SDValue RHS = Op.getOperand(3);
2157 SDValue DestBB = Op.getOperand(4);
2159 if (LHS.getValueType() == MVT::f128) {
2160 // f128 comparisons are lowered to runtime calls by a routine which sets
2161 // LHS, RHS and CC appropriately for the rest of this function to continue.
2162 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2164 // If softenSetCCOperands returned a scalar, we need to compare the result
2165 // against zero to select between true and false values.
2166 if (!RHS.getNode()) {
2167 RHS = DAG.getConstant(0, LHS.getValueType());
2172 if (LHS.getValueType().isInteger()) {
2175 // Integers are handled in a separate function because the combinations of
2176 // immediates and tests can get hairy and we may want to fiddle things.
2177 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2179 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2180 Chain, CmpOp, A64cc, DestBB);
2183 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2184 // conditional branch, hence FPCCToA64CC can set a second test, where either
2185 // passing is sufficient.
2186 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2187 CondCode = FPCCToA64CC(CC, Alternative);
2188 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2189 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2190 DAG.getCondCode(CC));
2191 SDValue A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2192 Chain, SetCC, A64cc, DestBB);
2194 if (Alternative != A64CC::Invalid) {
2195 A64cc = DAG.getConstant(Alternative, MVT::i32);
2196 A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
2197 A64BR_CC, SetCC, A64cc, DestBB);
2205 AArch64TargetLowering::LowerF128ToCall(SDValue Op, SelectionDAG &DAG,
2206 RTLIB::Libcall Call) const {
2209 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
2210 EVT ArgVT = Op.getOperand(i).getValueType();
2211 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
2212 Entry.Node = Op.getOperand(i); Entry.Ty = ArgTy;
2213 Entry.isSExt = false;
2214 Entry.isZExt = false;
2215 Args.push_back(Entry);
2217 SDValue Callee = DAG.getExternalSymbol(getLibcallName(Call), getPointerTy());
2219 Type *RetTy = Op.getValueType().getTypeForEVT(*DAG.getContext());
2221 // By default, the input chain to this libcall is the entry node of the
2222 // function. If the libcall is going to be emitted as a tail call then
2223 // isUsedByReturnOnly will change it to the right chain if the return
2224 // node which is being folded has a non-entry input chain.
2225 SDValue InChain = DAG.getEntryNode();
2227 // isTailCall may be true since the callee does not reference caller stack
2228 // frame. Check if it's in the right position.
2229 SDValue TCChain = InChain;
2230 bool isTailCall = isInTailCallPosition(DAG, Op.getNode(), TCChain);
2234 TargetLowering::CallLoweringInfo CLI(DAG);
2235 CLI.setDebugLoc(SDLoc(Op)).setChain(InChain)
2236 .setCallee(getLibcallCallingConv(Call), RetTy, Callee, &Args, 0)
2237 .setTailCall(isTailCall);
2239 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
2241 if (!CallInfo.second.getNode())
2242 // It's a tailcall, return the chain (which is the DAG root).
2243 return DAG.getRoot();
2245 return CallInfo.first;
2249 AArch64TargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
2250 if (Op.getOperand(0).getValueType() != MVT::f128) {
2251 // It's legal except when f128 is involved
2256 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
2258 SDValue SrcVal = Op.getOperand(0);
2259 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
2260 /*isSigned*/ false, SDLoc(Op)).first;
2264 AArch64TargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
2265 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
2268 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
2270 return LowerF128ToCall(Op, DAG, LC);
2273 static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2276 EVT VT = Op.getValueType();
2277 SDValue Vec = Op.getOperand(0);
2278 EVT OpVT = Vec.getValueType();
2279 unsigned Opc = IsSigned ? ISD::FP_TO_SINT : ISD::FP_TO_UINT;
2281 if (VT.getVectorNumElements() == 1) {
2282 assert(OpVT == MVT::v1f64 && "Unexpected vector type!");
2283 if (VT.getSizeInBits() == OpVT.getSizeInBits())
2285 return DAG.UnrollVectorOp(Op.getNode());
2288 if (VT.getSizeInBits() > OpVT.getSizeInBits()) {
2289 assert(Vec.getValueType() == MVT::v2f32 && VT == MVT::v2i64 &&
2290 "Unexpected vector type!");
2291 Vec = DAG.getNode(ISD::FP_EXTEND, dl, MVT::v2f64, Vec);
2292 return DAG.getNode(Opc, dl, VT, Vec);
2293 } else if (VT.getSizeInBits() < OpVT.getSizeInBits()) {
2294 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2295 OpVT.getVectorElementType().getSizeInBits());
2297 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2298 Vec = DAG.getNode(Opc, dl, CastVT, Vec);
2299 return DAG.getNode(ISD::TRUNCATE, dl, VT, Vec);
2301 return DAG.getNode(Opc, dl, VT, Vec);
2304 static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
2305 // We custom lower concat_vectors with 4, 8, or 16 operands that are all the
2306 // same operand and of type v1* using the DUP instruction.
2307 unsigned NumOps = Op->getNumOperands();
2309 assert(Op.getValueType().getSizeInBits() == 128 && "unexpected concat");
2313 if (NumOps != 4 && NumOps != 8 && NumOps != 16)
2316 // Must be a single value for VDUP.
2317 SDValue Op0 = Op.getOperand(0);
2318 for (unsigned i = 1; i < NumOps; ++i) {
2319 SDValue OpN = Op.getOperand(i);
2324 // Verify the value type.
2325 EVT EltVT = Op0.getValueType();
2327 default: llvm_unreachable("Unexpected number of operands");
2329 if (EltVT != MVT::v1i16 && EltVT != MVT::v1i32)
2333 if (EltVT != MVT::v1i8 && EltVT != MVT::v1i16)
2337 if (EltVT != MVT::v1i8)
2343 EVT VT = Op.getValueType();
2344 // VDUP produces better code for constants.
2345 if (Op0->getOpcode() == ISD::BUILD_VECTOR)
2346 return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Op0->getOperand(0));
2347 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, Op0,
2348 DAG.getConstant(0, MVT::i64));
2352 AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
2353 bool IsSigned) const {
2354 if (Op.getValueType().isVector())
2355 return LowerVectorFP_TO_INT(Op, DAG, IsSigned);
2356 if (Op.getOperand(0).getValueType() != MVT::f128) {
2357 // It's legal except when f128 is involved
2363 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
2365 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
2367 return LowerF128ToCall(Op, DAG, LC);
2370 SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
2371 MachineFunction &MF = DAG.getMachineFunction();
2372 MachineFrameInfo *MFI = MF.getFrameInfo();
2373 MFI->setReturnAddressIsTaken(true);
2375 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
2378 EVT VT = Op.getValueType();
2380 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2382 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
2383 SDValue Offset = DAG.getConstant(8, MVT::i64);
2384 return DAG.getLoad(VT, dl, DAG.getEntryNode(),
2385 DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
2386 MachinePointerInfo(), false, false, false, 0);
2389 // Return X30, which contains the return address. Mark it an implicit live-in.
2390 unsigned Reg = MF.addLiveIn(AArch64::X30, getRegClassFor(MVT::i64));
2391 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, MVT::i64);
2395 SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG)
2397 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
2398 MFI->setFrameAddressIsTaken(true);
2400 EVT VT = Op.getValueType();
2402 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
2403 unsigned FrameReg = AArch64::X29;
2404 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
2406 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
2407 MachinePointerInfo(),
2408 false, false, false, 0);
2412 // FIXME? Maybe this could be a TableGen attribute on some registers and
2413 // this table could be generated automatically from RegInfo.
2414 unsigned AArch64TargetLowering::getRegisterByName(const char* RegName,
2416 unsigned Reg = StringSwitch<unsigned>(RegName)
2417 .Case("sp", AArch64::XSP)
2421 report_fatal_error("Invalid register name global variable");
2425 AArch64TargetLowering::LowerGlobalAddressELFLarge(SDValue Op,
2426 SelectionDAG &DAG) const {
2427 assert(getTargetMachine().getCodeModel() == CodeModel::Large);
2428 assert(getTargetMachine().getRelocationModel() == Reloc::Static);
2430 EVT PtrVT = getPointerTy();
2432 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2433 const GlobalValue *GV = GN->getGlobal();
2435 SDValue GlobalAddr = DAG.getNode(
2436 AArch64ISD::WrapperLarge, dl, PtrVT,
2437 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G3),
2438 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
2439 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
2440 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
2442 if (GN->getOffset() != 0)
2443 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2444 DAG.getConstant(GN->getOffset(), PtrVT));
2450 AArch64TargetLowering::LowerGlobalAddressELFSmall(SDValue Op,
2451 SelectionDAG &DAG) const {
2452 assert(getTargetMachine().getCodeModel() == CodeModel::Small);
2454 EVT PtrVT = getPointerTy();
2456 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
2457 const GlobalValue *GV = GN->getGlobal();
2458 unsigned Alignment = GV->getAlignment();
2459 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
2460 if (GV->isWeakForLinker() && GV->isDeclaration() && RelocM == Reloc::Static) {
2461 // Weak undefined symbols can't use ADRP/ADD pair since they should evaluate
2462 // to zero when they remain undefined. In PIC mode the GOT can take care of
2463 // this, but in absolute mode we use a constant pool load.
2465 PoolAddr = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2466 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2467 AArch64II::MO_NO_FLAG),
2468 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
2469 AArch64II::MO_LO12),
2470 DAG.getConstant(8, MVT::i32));
2471 SDValue GlobalAddr = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), PoolAddr,
2472 MachinePointerInfo::getConstantPool(),
2473 /*isVolatile=*/ false,
2474 /*isNonTemporal=*/ true,
2475 /*isInvariant=*/ true, 8);
2476 if (GN->getOffset() != 0)
2477 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
2478 DAG.getConstant(GN->getOffset(), PtrVT));
2483 if (Alignment == 0) {
2484 const PointerType *GVPtrTy = cast<PointerType>(GV->getType());
2485 if (GVPtrTy->getElementType()->isSized()) {
2487 = getDataLayout()->getABITypeAlignment(GVPtrTy->getElementType());
2489 // Be conservative if we can't guess, not that it really matters:
2490 // functions and labels aren't valid for loads, and the methods used to
2491 // actually calculate an address work with any alignment.
2496 unsigned char HiFixup, LoFixup;
2497 bool UseGOT = getSubtarget()->GVIsIndirectSymbol(GV, RelocM);
2500 HiFixup = AArch64II::MO_GOT;
2501 LoFixup = AArch64II::MO_GOT_LO12;
2504 HiFixup = AArch64II::MO_NO_FLAG;
2505 LoFixup = AArch64II::MO_LO12;
2508 // AArch64's small model demands the following sequence:
2509 // ADRP x0, somewhere
2510 // ADD x0, x0, #:lo12:somewhere ; (or LDR directly).
2511 SDValue GlobalRef = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2512 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2514 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
2516 DAG.getConstant(Alignment, MVT::i32));
2519 GlobalRef = DAG.getNode(AArch64ISD::GOTLoad, dl, PtrVT, DAG.getEntryNode(),
2523 if (GN->getOffset() != 0)
2524 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalRef,
2525 DAG.getConstant(GN->getOffset(), PtrVT));
2531 AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
2532 SelectionDAG &DAG) const {
2533 // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
2534 // we make those distinctions here.
2536 switch (getTargetMachine().getCodeModel()) {
2537 case CodeModel::Small:
2538 return LowerGlobalAddressELFSmall(Op, DAG);
2539 case CodeModel::Large:
2540 return LowerGlobalAddressELFLarge(Op, DAG);
2542 llvm_unreachable("Only small and large code models supported now");
2547 AArch64TargetLowering::LowerConstantPool(SDValue Op,
2548 SelectionDAG &DAG) const {
2550 EVT PtrVT = getPointerTy();
2551 ConstantPoolSDNode *CN = cast<ConstantPoolSDNode>(Op);
2552 const Constant *C = CN->getConstVal();
2554 switch(getTargetMachine().getCodeModel()) {
2555 case CodeModel::Small:
2556 // The most efficient code is PC-relative anyway for the small memory model,
2557 // so we don't need to worry about relocation model.
2558 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2559 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2560 AArch64II::MO_NO_FLAG),
2561 DAG.getTargetConstantPool(C, PtrVT, 0, 0,
2562 AArch64II::MO_LO12),
2563 DAG.getConstant(CN->getAlignment(), MVT::i32));
2564 case CodeModel::Large:
2566 AArch64ISD::WrapperLarge, DL, PtrVT,
2567 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G3),
2568 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G2_NC),
2569 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G1_NC),
2570 DAG.getTargetConstantPool(C, PtrVT, 0, 0, AArch64II::MO_ABS_G0_NC));
2572 llvm_unreachable("Only small and large code models supported now");
2576 SDValue AArch64TargetLowering::LowerTLSDescCall(SDValue SymAddr,
2579 SelectionDAG &DAG) const {
2580 EVT PtrVT = getPointerTy();
2582 // The function we need to call is simply the first entry in the GOT for this
2583 // descriptor, load it in preparation.
2584 SDValue Func, Chain;
2585 Func = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2588 // The function takes only one argument: the address of the descriptor itself
2591 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
2592 Glue = Chain.getValue(1);
2594 // Finally, there's a special calling-convention which means that the lookup
2595 // must preserve all registers (except X0, obviously).
2596 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2597 const AArch64RegisterInfo *A64RI
2598 = static_cast<const AArch64RegisterInfo *>(TRI);
2599 const uint32_t *Mask = A64RI->getTLSDescCallPreservedMask();
2601 // We're now ready to populate the argument list, as with a normal call:
2602 std::vector<SDValue> Ops;
2603 Ops.push_back(Chain);
2604 Ops.push_back(Func);
2605 Ops.push_back(SymAddr);
2606 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
2607 Ops.push_back(DAG.getRegisterMask(Mask));
2608 Ops.push_back(Glue);
2610 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2611 Chain = DAG.getNode(AArch64ISD::TLSDESCCALL, DL, NodeTys, Ops);
2612 Glue = Chain.getValue(1);
2614 // After the call, the offset from TPIDR_EL0 is in X0, copy it out and pass it
2615 // back to the generic handling code.
2616 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
2620 AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2621 SelectionDAG &DAG) const {
2622 assert(getSubtarget()->isTargetELF() &&
2623 "TLS not implemented for non-ELF targets");
2624 assert(getTargetMachine().getCodeModel() == CodeModel::Small
2625 && "TLS only supported in small memory model");
2626 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2628 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2631 EVT PtrVT = getPointerTy();
2633 const GlobalValue *GV = GA->getGlobal();
2635 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
2637 if (Model == TLSModel::InitialExec) {
2638 TPOff = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2639 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2640 AArch64II::MO_GOTTPREL),
2641 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2642 AArch64II::MO_GOTTPREL_LO12),
2643 DAG.getConstant(8, MVT::i32));
2644 TPOff = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2646 } else if (Model == TLSModel::LocalExec) {
2647 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2648 AArch64II::MO_TPREL_G1);
2649 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2650 AArch64II::MO_TPREL_G0_NC);
2652 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2653 DAG.getTargetConstant(1, MVT::i32)), 0);
2654 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2656 DAG.getTargetConstant(0, MVT::i32)), 0);
2657 } else if (Model == TLSModel::GeneralDynamic) {
2658 // Accesses used in this sequence go via the TLS descriptor which lives in
2659 // the GOT. Prepare an address we can use to handle this.
2660 SDValue HiDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2661 AArch64II::MO_TLSDESC);
2662 SDValue LoDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2663 AArch64II::MO_TLSDESC_LO12);
2664 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2666 DAG.getConstant(8, MVT::i32));
2667 SDValue SymAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0);
2669 TPOff = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2670 } else if (Model == TLSModel::LocalDynamic) {
2671 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2672 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2673 // the beginning of the module's TLS region, followed by a DTPREL offset
2676 // These accesses will need deduplicating if there's more than one.
2677 AArch64MachineFunctionInfo* MFI = DAG.getMachineFunction()
2678 .getInfo<AArch64MachineFunctionInfo>();
2679 MFI->incNumLocalDynamicTLSAccesses();
2682 // Get the location of _TLS_MODULE_BASE_:
2683 SDValue HiDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2684 AArch64II::MO_TLSDESC);
2685 SDValue LoDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2686 AArch64II::MO_TLSDESC_LO12);
2687 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2689 DAG.getConstant(8, MVT::i32));
2690 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT);
2692 ThreadBase = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2694 // Get the variable's offset from _TLS_MODULE_BASE_
2695 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2696 AArch64II::MO_DTPREL_G1);
2697 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2698 AArch64II::MO_DTPREL_G0_NC);
2700 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2701 DAG.getTargetConstant(0, MVT::i32)), 0);
2702 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2704 DAG.getTargetConstant(0, MVT::i32)), 0);
2706 llvm_unreachable("Unsupported TLS access model");
2709 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2712 static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2715 EVT VT = Op.getValueType();
2716 SDValue Vec = Op.getOperand(0);
2717 unsigned Opc = IsSigned ? ISD::SINT_TO_FP : ISD::UINT_TO_FP;
2719 if (VT.getVectorNumElements() == 1) {
2720 assert(VT == MVT::v1f64 && "Unexpected vector type!");
2721 if (VT.getSizeInBits() == Vec.getValueSizeInBits())
2723 return DAG.UnrollVectorOp(Op.getNode());
2726 if (VT.getSizeInBits() < Vec.getValueSizeInBits()) {
2727 assert(Vec.getValueType() == MVT::v2i64 && VT == MVT::v2f32 &&
2728 "Unexpected vector type!");
2729 Vec = DAG.getNode(Opc, dl, MVT::v2f64, Vec);
2730 return DAG.getNode(ISD::FP_ROUND, dl, VT, Vec, DAG.getIntPtrConstant(0));
2731 } else if (VT.getSizeInBits() > Vec.getValueSizeInBits()) {
2732 unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
2733 EVT CastVT = EVT::getIntegerVT(*DAG.getContext(),
2734 VT.getVectorElementType().getSizeInBits());
2736 EVT::getVectorVT(*DAG.getContext(), CastVT, VT.getVectorNumElements());
2737 Vec = DAG.getNode(CastOpc, dl, CastVT, Vec);
2740 return DAG.getNode(Opc, dl, VT, Vec);
2744 AArch64TargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2745 bool IsSigned) const {
2746 if (Op.getValueType().isVector())
2747 return LowerVectorINT_TO_FP(Op, DAG, IsSigned);
2748 if (Op.getValueType() != MVT::f128) {
2749 // Legal for everything except f128.
2755 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2757 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2759 return LowerF128ToCall(Op, DAG, LC);
2764 AArch64TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
2765 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2767 EVT PtrVT = getPointerTy();
2769 // When compiling PIC, jump tables get put in the code section so a static
2770 // relocation-style is acceptable for both cases.
2771 switch (getTargetMachine().getCodeModel()) {
2772 case CodeModel::Small:
2773 return DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2774 DAG.getTargetJumpTable(JT->getIndex(), PtrVT),
2775 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
2776 AArch64II::MO_LO12),
2777 DAG.getConstant(1, MVT::i32));
2778 case CodeModel::Large:
2780 AArch64ISD::WrapperLarge, dl, PtrVT,
2781 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G3),
2782 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G2_NC),
2783 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G1_NC),
2784 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G0_NC));
2786 llvm_unreachable("Only small and large code models supported now");
2790 // (SELECT testbit, iftrue, iffalse)
2792 AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2794 SDValue TheBit = Op.getOperand(0);
2795 SDValue IfTrue = Op.getOperand(1);
2796 SDValue IfFalse = Op.getOperand(2);
2798 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2799 // that as the consumer we are responsible for ignoring rubbish in higher
2801 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2802 DAG.getConstant(1, MVT::i32));
2803 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2804 DAG.getConstant(0, TheBit.getValueType()),
2805 DAG.getCondCode(ISD::SETNE));
2807 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2808 A64CMP, IfTrue, IfFalse,
2809 DAG.getConstant(A64CC::NE, MVT::i32));
2812 static SDValue LowerVectorSETCC(SDValue Op, SelectionDAG &DAG) {
2814 SDValue LHS = Op.getOperand(0);
2815 SDValue RHS = Op.getOperand(1);
2816 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2817 EVT VT = Op.getValueType();
2818 bool Invert = false;
2822 if (LHS.getValueType().isInteger()) {
2824 // Attempt to use Vector Integer Compare Mask Test instruction.
2825 // TST = icmp ne (and (op0, op1), zero).
2826 if (CC == ISD::SETNE) {
2827 if (((LHS.getOpcode() == ISD::AND) &&
2828 ISD::isBuildVectorAllZeros(RHS.getNode())) ||
2829 ((RHS.getOpcode() == ISD::AND) &&
2830 ISD::isBuildVectorAllZeros(LHS.getNode()))) {
2832 SDValue AndOp = (LHS.getOpcode() == ISD::AND) ? LHS : RHS;
2833 SDValue NewLHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(0));
2834 SDValue NewRHS = DAG.getNode(ISD::BITCAST, DL, VT, AndOp.getOperand(1));
2835 return DAG.getNode(AArch64ISD::NEON_TST, DL, VT, NewLHS, NewRHS);
2839 // Attempt to use Vector Integer Compare Mask against Zero instr (Signed).
2840 // Note: Compare against Zero does not support unsigned predicates.
2841 if ((ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2842 ISD::isBuildVectorAllZeros(LHS.getNode())) &&
2843 !isUnsignedIntSetCC(CC)) {
2845 // If LHS is the zero value, swap operands and CondCode.
2846 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2847 CC = getSetCCSwappedOperands(CC);
2852 // Ensure valid CondCode for Compare Mask against Zero instruction:
2853 // EQ, GE, GT, LE, LT.
2854 if (ISD::SETNE == CC) {
2859 // Using constant type to differentiate integer and FP compares with zero.
2860 Op1 = DAG.getConstant(0, MVT::i32);
2861 Opcode = AArch64ISD::NEON_CMPZ;
2864 // Attempt to use Vector Integer Compare Mask instr (Signed/Unsigned).
2865 // Ensure valid CondCode for Compare Mask instr: EQ, GE, GT, UGE, UGT.
2869 llvm_unreachable("Illegal integer comparison.");
2885 CC = getSetCCSwappedOperands(CC);
2889 std::swap(LHS, RHS);
2891 Opcode = AArch64ISD::NEON_CMP;
2896 // Generate Compare Mask instr or Compare Mask against Zero instr.
2898 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
2901 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
2906 // Now handle Floating Point cases.
2907 // Attempt to use Vector Floating Point Compare Mask against Zero instruction.
2908 if (ISD::isBuildVectorAllZeros(RHS.getNode()) ||
2909 ISD::isBuildVectorAllZeros(LHS.getNode())) {
2911 // If LHS is the zero value, swap operands and CondCode.
2912 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
2913 CC = getSetCCSwappedOperands(CC);
2918 // Using constant type to differentiate integer and FP compares with zero.
2919 Op1 = DAG.getConstantFP(0, MVT::f32);
2920 Opcode = AArch64ISD::NEON_CMPZ;
2922 // Attempt to use Vector Floating Point Compare Mask instruction.
2925 Opcode = AArch64ISD::NEON_CMP;
2929 // Some register compares have to be implemented with swapped CC and operands,
2930 // e.g.: OLT implemented as OGT with swapped operands.
2931 bool SwapIfRegArgs = false;
2933 // Ensure valid CondCode for FP Compare Mask against Zero instruction:
2934 // EQ, GE, GT, LE, LT.
2935 // And ensure valid CondCode for FP Compare Mask instruction: EQ, GE, GT.
2938 llvm_unreachable("Illegal FP comparison");
2941 Invert = true; // Fallthrough
2949 SwapIfRegArgs = true;
2958 SwapIfRegArgs = true;
2967 SwapIfRegArgs = true;
2976 SwapIfRegArgs = true;
2983 Invert = true; // Fallthrough
2985 // Expand this to (OGT |OLT).
2987 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGT));
2989 SwapIfRegArgs = true;
2992 Invert = true; // Fallthrough
2994 // Expand this to (OGE | OLT).
2996 DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(ISD::SETGE));
2998 SwapIfRegArgs = true;
3002 if (Opcode == AArch64ISD::NEON_CMP && SwapIfRegArgs) {
3003 CC = getSetCCSwappedOperands(CC);
3004 std::swap(Op0, Op1);
3007 // Generate FP Compare Mask instr or FP Compare Mask against Zero instr
3008 SDValue NeonCmp = DAG.getNode(Opcode, DL, VT, Op0, Op1, DAG.getCondCode(CC));
3010 if (NeonCmpAlt.getNode())
3011 NeonCmp = DAG.getNode(ISD::OR, DL, VT, NeonCmp, NeonCmpAlt);
3014 NeonCmp = DAG.getNOT(DL, NeonCmp, VT);
3019 // (SETCC lhs, rhs, condcode)
3021 AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
3023 SDValue LHS = Op.getOperand(0);
3024 SDValue RHS = Op.getOperand(1);
3025 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
3026 EVT VT = Op.getValueType();
3029 return LowerVectorSETCC(Op, DAG);
3031 if (LHS.getValueType() == MVT::f128) {
3032 // f128 comparisons will be lowered to libcalls giving a valid LHS and RHS
3033 // for the rest of the function (some i32 or i64 values).
3034 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3036 // If softenSetCCOperands returned a scalar, use it.
3037 if (!RHS.getNode()) {
3038 assert(LHS.getValueType() == Op.getValueType() &&
3039 "Unexpected setcc expansion!");
3044 if (LHS.getValueType().isInteger()) {
3047 // Integers are handled in a separate function because the combinations of
3048 // immediates and tests can get hairy and we may want to fiddle things.
3049 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
3051 return DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
3052 CmpOp, DAG.getConstant(1, VT), DAG.getConstant(0, VT),
3056 // Note that some LLVM floating-point CondCodes can't be lowered to a single
3057 // conditional branch, hence FPCCToA64CC can set a second test, where either
3058 // passing is sufficient.
3059 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3060 CondCode = FPCCToA64CC(CC, Alternative);
3061 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3062 SDValue CmpOp = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3063 DAG.getCondCode(CC));
3064 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
3065 CmpOp, DAG.getConstant(1, VT),
3066 DAG.getConstant(0, VT), A64cc);
3068 if (Alternative != A64CC::Invalid) {
3069 A64cc = DAG.getConstant(Alternative, MVT::i32);
3070 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
3071 DAG.getConstant(1, VT), A64SELECT_CC, A64cc);
3074 return A64SELECT_CC;
3077 static SDValue LowerVectorSELECT_CC(SDValue Op, SelectionDAG &DAG) {
3079 SDValue LHS = Op.getOperand(0);
3080 SDValue RHS = Op.getOperand(1);
3081 SDValue IfTrue = Op.getOperand(2);
3082 SDValue IfFalse = Op.getOperand(3);
3083 EVT IfTrueVT = IfTrue.getValueType();
3084 EVT CondVT = IfTrueVT.changeVectorElementTypeToInteger();
3085 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3087 // If LHS & RHS are floating point and IfTrue & IfFalse are vectors, we will
3088 // use NEON compare.
3089 if ((LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64)) {
3090 EVT EltVT = LHS.getValueType();
3091 unsigned EltNum = 128 / EltVT.getSizeInBits();
3092 EVT VT = EVT::getVectorVT(*DAG.getContext(), EltVT, EltNum);
3093 unsigned SubConstant =
3094 (LHS.getValueType() == MVT::f32) ? AArch64::sub_32 :AArch64::sub_64;
3095 EVT CEltT = (LHS.getValueType() == MVT::f32) ? MVT::i32 : MVT::i64;
3096 EVT CVT = EVT::getVectorVT(*DAG.getContext(), CEltT, EltNum);
3099 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
3100 VT, DAG.getTargetConstant(0, MVT::i32), LHS,
3101 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3103 = SDValue(DAG.getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
3104 VT, DAG.getTargetConstant(0, MVT::i32), RHS,
3105 DAG.getTargetConstant(SubConstant, MVT::i32)), 0);
3107 SDValue VSetCC = DAG.getSetCC(dl, CVT, LHS, RHS, CC);
3108 SDValue ResCC = LowerVectorSETCC(VSetCC, DAG);
3109 if (CEltT.getSizeInBits() < IfTrueVT.getSizeInBits()) {
3111 EVT::getVectorVT(*DAG.getContext(), CEltT,
3112 IfTrueVT.getSizeInBits() / CEltT.getSizeInBits());
3113 ResCC = DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, DUPVT, ResCC,
3114 DAG.getConstant(0, MVT::i64, false));
3116 ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
3118 // FIXME: If IfTrue & IfFalse hold v1i8, v1i16 or v1i32, this function
3119 // can't handle them and will hit this assert.
3120 assert(CEltT.getSizeInBits() == IfTrueVT.getSizeInBits() &&
3121 "Vector of IfTrue & IfFalse is too small.");
3124 EltNum * IfTrueVT.getSizeInBits() / ResCC.getValueSizeInBits();
3125 EVT ExVT = EVT::getVectorVT(*DAG.getContext(), CEltT, ExEltNum);
3126 ResCC = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ExVT, ResCC,
3127 DAG.getConstant(0, MVT::i64, false));
3128 ResCC = DAG.getNode(ISD::BITCAST, dl, CondVT, ResCC);
3130 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3131 ResCC, IfTrue, IfFalse);
3135 // Here we handle the case that LHS & RHS are integer and IfTrue & IfFalse are
3137 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3138 CondCode = FPCCToA64CC(CC, Alternative);
3139 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3140 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3141 DAG.getCondCode(CC));
3142 EVT SEVT = MVT::i32;
3143 if (IfTrue.getValueType().getVectorElementType().getSizeInBits() > 32)
3145 SDValue AllOne = DAG.getConstant(-1, SEVT);
3146 SDValue AllZero = DAG.getConstant(0, SEVT);
3147 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, SEVT, SetCC,
3148 AllOne, AllZero, A64cc);
3150 if (Alternative != A64CC::Invalid) {
3151 A64cc = DAG.getConstant(Alternative, MVT::i32);
3152 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3153 SetCC, AllOne, A64SELECT_CC, A64cc);
3156 if (IfTrue.getValueType().getVectorNumElements() == 1)
3157 VDup = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, CondVT, A64SELECT_CC);
3159 VDup = DAG.getNode(AArch64ISD::NEON_VDUP, dl, CondVT, A64SELECT_CC);
3160 SDValue VSelect = DAG.getNode(ISD::VSELECT, dl, IfTrue.getValueType(),
3161 VDup, IfTrue, IfFalse);
3165 // (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
3167 AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
3169 SDValue LHS = Op.getOperand(0);
3170 SDValue RHS = Op.getOperand(1);
3171 SDValue IfTrue = Op.getOperand(2);
3172 SDValue IfFalse = Op.getOperand(3);
3173 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
3175 if (IfTrue.getValueType().isVector())
3176 return LowerVectorSELECT_CC(Op, DAG);
3178 if (LHS.getValueType() == MVT::f128) {
3179 // f128 comparisons are lowered to libcalls, but slot in nicely here
3181 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
3183 // If softenSetCCOperands returned a scalar, we need to compare the result
3184 // against zero to select between true and false values.
3185 if (!RHS.getNode()) {
3186 RHS = DAG.getConstant(0, LHS.getValueType());
3191 if (LHS.getValueType().isInteger()) {
3194 // Integers are handled in a separate function because the combinations of
3195 // immediates and tests can get hairy and we may want to fiddle things.
3196 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
3198 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(), CmpOp,
3199 IfTrue, IfFalse, A64cc);
3202 // Note that some LLVM floating-point CondCodes can't be lowered to a single
3203 // conditional branch, hence FPCCToA64CC can set a second test, where either
3204 // passing is sufficient.
3205 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
3206 CondCode = FPCCToA64CC(CC, Alternative);
3207 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
3208 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
3209 DAG.getCondCode(CC));
3210 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
3212 SetCC, IfTrue, IfFalse, A64cc);
3214 if (Alternative != A64CC::Invalid) {
3215 A64cc = DAG.getConstant(Alternative, MVT::i32);
3216 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
3217 SetCC, IfTrue, A64SELECT_CC, A64cc);
3221 return A64SELECT_CC;
3225 AArch64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
3226 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
3227 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
3229 // We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
3230 // rather than just 8.
3231 return DAG.getMemcpy(Op.getOperand(0), SDLoc(Op),
3232 Op.getOperand(1), Op.getOperand(2),
3233 DAG.getConstant(32, MVT::i32), 8, false, false,
3234 MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV));
3238 AArch64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
3239 // The layout of the va_list struct is specified in the AArch64 Procedure Call
3240 // Standard, section B.3.
3241 MachineFunction &MF = DAG.getMachineFunction();
3242 AArch64MachineFunctionInfo *FuncInfo
3243 = MF.getInfo<AArch64MachineFunctionInfo>();
3246 SDValue Chain = Op.getOperand(0);
3247 SDValue VAList = Op.getOperand(1);
3248 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
3249 SmallVector<SDValue, 4> MemOps;
3251 // void *__stack at offset 0
3252 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVariadicStackIdx(),
3254 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
3255 MachinePointerInfo(SV), false, false, 0));
3257 // void *__gr_top at offset 8
3258 int GPRSize = FuncInfo->getVariadicGPRSize();
3260 SDValue GRTop, GRTopAddr;
3262 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3263 DAG.getConstant(8, getPointerTy()));
3265 GRTop = DAG.getFrameIndex(FuncInfo->getVariadicGPRIdx(), getPointerTy());
3266 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
3267 DAG.getConstant(GPRSize, getPointerTy()));
3269 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
3270 MachinePointerInfo(SV, 8),
3274 // void *__vr_top at offset 16
3275 int FPRSize = FuncInfo->getVariadicFPRSize();
3277 SDValue VRTop, VRTopAddr;
3278 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3279 DAG.getConstant(16, getPointerTy()));
3281 VRTop = DAG.getFrameIndex(FuncInfo->getVariadicFPRIdx(), getPointerTy());
3282 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
3283 DAG.getConstant(FPRSize, getPointerTy()));
3285 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
3286 MachinePointerInfo(SV, 16),
3290 // int __gr_offs at offset 24
3291 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3292 DAG.getConstant(24, getPointerTy()));
3293 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
3294 GROffsAddr, MachinePointerInfo(SV, 24),
3297 // int __vr_offs at offset 28
3298 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
3299 DAG.getConstant(28, getPointerTy()));
3300 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
3301 VROffsAddr, MachinePointerInfo(SV, 28),
3304 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
3308 AArch64TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
3309 switch (Op.getOpcode()) {
3310 default: llvm_unreachable("Don't know how to custom lower this!");
3311 case ISD::FADD: return LowerF128ToCall(Op, DAG, RTLIB::ADD_F128);
3312 case ISD::FSUB: return LowerF128ToCall(Op, DAG, RTLIB::SUB_F128);
3313 case ISD::FMUL: return LowerF128ToCall(Op, DAG, RTLIB::MUL_F128);
3314 case ISD::FDIV: return LowerF128ToCall(Op, DAG, RTLIB::DIV_F128);
3315 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, true);
3316 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG, false);
3317 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG, true);
3318 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG, false);
3319 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
3320 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
3321 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
3322 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
3324 case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
3325 case ISD::SRL_PARTS:
3326 case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
3328 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
3329 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
3330 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
3331 case ISD::GlobalAddress: return LowerGlobalAddressELF(Op, DAG);
3332 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
3333 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
3334 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
3335 case ISD::SELECT: return LowerSELECT(Op, DAG);
3336 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
3337 case ISD::SETCC: return LowerSETCC(Op, DAG);
3338 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
3339 case ISD::VASTART: return LowerVASTART(Op, DAG);
3340 case ISD::BUILD_VECTOR:
3341 return LowerBUILD_VECTOR(Op, DAG, getSubtarget());
3342 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
3343 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
3349 /// Check if the specified splat value corresponds to a valid vector constant
3350 /// for a Neon instruction with a "modified immediate" operand (e.g., MOVI). If
3351 /// so, return the encoded 8-bit immediate and the OpCmode instruction fields
3353 static bool isNeonModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
3354 unsigned SplatBitSize, SelectionDAG &DAG,
3355 bool is128Bits, NeonModImmType type, EVT &VT,
3356 unsigned &Imm, unsigned &OpCmode) {
3357 switch (SplatBitSize) {
3359 llvm_unreachable("unexpected size for isNeonModifiedImm");
3361 if (type != Neon_Mov_Imm)
3363 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
3364 // Neon movi per byte: Op=0, Cmode=1110.
3367 VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
3371 // Neon move inst per halfword
3372 VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
3373 if ((SplatBits & ~0xff) == 0) {
3374 // Value = 0x00nn is 0x00nn LSL 0
3375 // movi: Op=0, Cmode=1000; mvni: Op=1, Cmode=1000
3376 // bic: Op=1, Cmode=1001; orr: Op=0, Cmode=1001
3382 if ((SplatBits & ~0xff00) == 0) {
3383 // Value = 0xnn00 is 0x00nn LSL 8
3384 // movi: Op=0, Cmode=1010; mvni: Op=1, Cmode=1010
3385 // bic: Op=1, Cmode=1011; orr: Op=0, Cmode=1011
3387 Imm = SplatBits >> 8;
3391 // can't handle any other
3396 // First the LSL variants (MSL is unusable by some interested instructions).
3398 // Neon move instr per word, shift zeros
3399 VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
3400 if ((SplatBits & ~0xff) == 0) {
3401 // Value = 0x000000nn is 0x000000nn LSL 0
3402 // movi: Op=0, Cmode= 0000; mvni: Op=1, Cmode= 0000
3403 // bic: Op=1, Cmode= 0001; orr: Op=0, Cmode= 0001
3409 if ((SplatBits & ~0xff00) == 0) {
3410 // Value = 0x0000nn00 is 0x000000nn LSL 8
3411 // movi: Op=0, Cmode= 0010; mvni: Op=1, Cmode= 0010
3412 // bic: Op=1, Cmode= 0011; orr : Op=0, Cmode= 0011
3414 Imm = SplatBits >> 8;
3418 if ((SplatBits & ~0xff0000) == 0) {
3419 // Value = 0x00nn0000 is 0x000000nn LSL 16
3420 // movi: Op=0, Cmode= 0100; mvni: Op=1, Cmode= 0100
3421 // bic: Op=1, Cmode= 0101; orr: Op=0, Cmode= 0101
3423 Imm = SplatBits >> 16;
3427 if ((SplatBits & ~0xff000000) == 0) {
3428 // Value = 0xnn000000 is 0x000000nn LSL 24
3429 // movi: Op=0, Cmode= 0110; mvni: Op=1, Cmode= 0110
3430 // bic: Op=1, Cmode= 0111; orr: Op=0, Cmode= 0111
3432 Imm = SplatBits >> 24;
3437 // Now the MSL immediates.
3439 // Neon move instr per word, shift ones
3440 if ((SplatBits & ~0xffff) == 0 &&
3441 ((SplatBits | SplatUndef) & 0xff) == 0xff) {
3442 // Value = 0x0000nnff is 0x000000nn MSL 8
3443 // movi: Op=0, Cmode= 1100; mvni: Op=1, Cmode= 1100
3445 Imm = SplatBits >> 8;
3449 if ((SplatBits & ~0xffffff) == 0 &&
3450 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
3451 // Value = 0x00nnffff is 0x000000nn MSL 16
3452 // movi: Op=1, Cmode= 1101; mvni: Op=1, Cmode= 1101
3454 Imm = SplatBits >> 16;
3458 // can't handle any other
3463 if (type != Neon_Mov_Imm)
3465 // Neon move instr bytemask, where each byte is either 0x00 or 0xff.
3466 // movi Op=1, Cmode=1110.
3468 uint64_t BitMask = 0xff;
3470 unsigned ImmMask = 1;
3472 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
3473 if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
3476 } else if ((SplatBits & BitMask) != 0) {
3483 VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
3491 static SDValue PerformANDCombine(SDNode *N,
3492 TargetLowering::DAGCombinerInfo &DCI) {
3494 SelectionDAG &DAG = DCI.DAG;
3496 EVT VT = N->getValueType(0);
3498 // We're looking for an SRA/SHL pair which form an SBFX.
3500 if (VT != MVT::i32 && VT != MVT::i64)
3503 if (!isa<ConstantSDNode>(N->getOperand(1)))
3506 uint64_t TruncMask = N->getConstantOperandVal(1);
3507 if (!isMask_64(TruncMask))
3510 uint64_t Width = CountPopulation_64(TruncMask);
3511 SDValue Shift = N->getOperand(0);
3513 if (Shift.getOpcode() != ISD::SRL)
3516 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3518 uint64_t LSB = Shift->getConstantOperandVal(1);
3520 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3523 return DAG.getNode(AArch64ISD::UBFX, DL, VT, Shift.getOperand(0),
3524 DAG.getConstant(LSB, MVT::i64),
3525 DAG.getConstant(LSB + Width - 1, MVT::i64));
3528 /// For a true bitfield insert, the bits getting into that contiguous mask
3529 /// should come from the low part of an existing value: they must be formed from
3530 /// a compatible SHL operation (unless they're already low). This function
3531 /// checks that condition and returns the least-significant bit that's
3532 /// intended. If the operation not a field preparation, -1 is returned.
3533 static int32_t getLSBForBFI(SelectionDAG &DAG, SDLoc DL, EVT VT,
3534 SDValue &MaskedVal, uint64_t Mask) {
3535 if (!isShiftedMask_64(Mask))
3538 // Now we need to alter MaskedVal so that it is an appropriate input for a BFI
3539 // instruction. BFI will do a left-shift by LSB before applying the mask we've
3540 // spotted, so in general we should pre-emptively "undo" that by making sure
3541 // the incoming bits have had a right-shift applied to them.
3543 // This right shift, however, will combine with existing left/right shifts. In
3544 // the simplest case of a completely straight bitfield operation, it will be
3545 // expected to completely cancel out with an existing SHL. More complicated
3546 // cases (e.g. bitfield to bitfield copy) may still need a real shift before
3549 uint64_t LSB = countTrailingZeros(Mask);
3550 int64_t ShiftRightRequired = LSB;
3551 if (MaskedVal.getOpcode() == ISD::SHL &&
3552 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3553 ShiftRightRequired -= MaskedVal.getConstantOperandVal(1);
3554 MaskedVal = MaskedVal.getOperand(0);
3555 } else if (MaskedVal.getOpcode() == ISD::SRL &&
3556 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
3557 ShiftRightRequired += MaskedVal.getConstantOperandVal(1);
3558 MaskedVal = MaskedVal.getOperand(0);
3561 if (ShiftRightRequired > 0)
3562 MaskedVal = DAG.getNode(ISD::SRL, DL, VT, MaskedVal,
3563 DAG.getConstant(ShiftRightRequired, MVT::i64));
3564 else if (ShiftRightRequired < 0) {
3565 // We could actually end up with a residual left shift, for example with
3566 // "struc.bitfield = val << 1".
3567 MaskedVal = DAG.getNode(ISD::SHL, DL, VT, MaskedVal,
3568 DAG.getConstant(-ShiftRightRequired, MVT::i64));
3574 /// Searches from N for an existing AArch64ISD::BFI node, possibly surrounded by
3575 /// a mask and an extension. Returns true if a BFI was found and provides
3576 /// information on its surroundings.
3577 static bool findMaskedBFI(SDValue N, SDValue &BFI, uint64_t &Mask,
3580 if (N.getOpcode() == ISD::ZERO_EXTEND) {
3582 N = N.getOperand(0);
3585 if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
3586 Mask = N->getConstantOperandVal(1);
3587 N = N.getOperand(0);
3589 // Mask is the whole width.
3590 Mask = -1ULL >> (64 - N.getValueType().getSizeInBits());
3593 if (N.getOpcode() == AArch64ISD::BFI) {
3601 /// Try to combine a subtree (rooted at an OR) into a "masked BFI" node, which
3602 /// is roughly equivalent to (and (BFI ...), mask). This form is used because it
3603 /// can often be further combined with a larger mask. Ultimately, we want mask
3604 /// to be 2^32-1 or 2^64-1 so the AND can be skipped.
3605 static SDValue tryCombineToBFI(SDNode *N,
3606 TargetLowering::DAGCombinerInfo &DCI,
3607 const AArch64Subtarget *Subtarget) {
3608 SelectionDAG &DAG = DCI.DAG;
3610 EVT VT = N->getValueType(0);
3612 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3614 // We need the LHS to be (and SOMETHING, MASK). Find out what that mask is or
3615 // abandon the effort.
3616 SDValue LHS = N->getOperand(0);
3617 if (LHS.getOpcode() != ISD::AND)
3621 if (isa<ConstantSDNode>(LHS.getOperand(1)))
3622 LHSMask = LHS->getConstantOperandVal(1);
3626 // We also need the RHS to be (and SOMETHING, MASK). Find out what that mask
3627 // is or abandon the effort.
3628 SDValue RHS = N->getOperand(1);
3629 if (RHS.getOpcode() != ISD::AND)
3633 if (isa<ConstantSDNode>(RHS.getOperand(1)))
3634 RHSMask = RHS->getConstantOperandVal(1);
3638 // Can't do anything if the masks are incompatible.
3639 if (LHSMask & RHSMask)
3642 // Now we need one of the masks to be a contiguous field. Without loss of
3643 // generality that should be the RHS one.
3644 SDValue Bitfield = LHS.getOperand(0);
3645 if (getLSBForBFI(DAG, DL, VT, Bitfield, LHSMask) != -1) {
3646 // We know that LHS is a candidate new value, and RHS isn't already a better
3648 std::swap(LHS, RHS);
3649 std::swap(LHSMask, RHSMask);
3652 // We've done our best to put the right operands in the right places, all we
3653 // can do now is check whether a BFI exists.
3654 Bitfield = RHS.getOperand(0);
3655 int32_t LSB = getLSBForBFI(DAG, DL, VT, Bitfield, RHSMask);
3659 uint32_t Width = CountPopulation_64(RHSMask);
3660 assert(Width && "Expected non-zero bitfield width");
3662 SDValue BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3663 LHS.getOperand(0), Bitfield,
3664 DAG.getConstant(LSB, MVT::i64),
3665 DAG.getConstant(Width, MVT::i64));
3668 if ((LHSMask | RHSMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3671 return DAG.getNode(ISD::AND, DL, VT, BFI,
3672 DAG.getConstant(LHSMask | RHSMask, VT));
3675 /// Search for the bitwise combining (with careful masks) of a MaskedBFI and its
3676 /// original input. This is surprisingly common because SROA splits things up
3677 /// into i8 chunks, so the originally detected MaskedBFI may actually only act
3678 /// on the low (say) byte of a word. This is then orred into the rest of the
3679 /// word afterwards.
3681 /// Basic input: (or (and OLDFIELD, MASK1), (MaskedBFI MASK2, OLDFIELD, ...)).
3683 /// If MASK1 and MASK2 are compatible, we can fold the whole thing into the
3684 /// MaskedBFI. We can also deal with a certain amount of extend/truncate being
3686 static SDValue tryCombineToLargerBFI(SDNode *N,
3687 TargetLowering::DAGCombinerInfo &DCI,
3688 const AArch64Subtarget *Subtarget) {
3689 SelectionDAG &DAG = DCI.DAG;
3691 EVT VT = N->getValueType(0);
3693 // First job is to hunt for a MaskedBFI on either the left or right. Swap
3694 // operands if it's actually on the right.
3696 SDValue PossExtraMask;
3697 uint64_t ExistingMask = 0;
3698 bool Extended = false;
3699 if (findMaskedBFI(N->getOperand(0), BFI, ExistingMask, Extended))
3700 PossExtraMask = N->getOperand(1);
3701 else if (findMaskedBFI(N->getOperand(1), BFI, ExistingMask, Extended))
3702 PossExtraMask = N->getOperand(0);
3706 // We can only combine a BFI with another compatible mask.
3707 if (PossExtraMask.getOpcode() != ISD::AND ||
3708 !isa<ConstantSDNode>(PossExtraMask.getOperand(1)))
3711 uint64_t ExtraMask = PossExtraMask->getConstantOperandVal(1);
3713 // Masks must be compatible.
3714 if (ExtraMask & ExistingMask)
3717 SDValue OldBFIVal = BFI.getOperand(0);
3718 SDValue NewBFIVal = BFI.getOperand(1);
3720 // We skipped a ZERO_EXTEND above, so the input to the MaskedBFIs should be
3721 // 32-bit and we'll be forming a 64-bit MaskedBFI. The MaskedBFI arguments
3722 // need to be made compatible.
3723 assert(VT == MVT::i64 && BFI.getValueType() == MVT::i32
3724 && "Invalid types for BFI");
3725 OldBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, OldBFIVal);
3726 NewBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, NewBFIVal);
3729 // We need the MaskedBFI to be combined with a mask of the *same* value.
3730 if (PossExtraMask.getOperand(0) != OldBFIVal)
3733 BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
3734 OldBFIVal, NewBFIVal,
3735 BFI.getOperand(2), BFI.getOperand(3));
3737 // If the masking is trivial, we don't need to create it.
3738 if ((ExtraMask | ExistingMask) == (-1ULL >> (64 - VT.getSizeInBits())))
3741 return DAG.getNode(ISD::AND, DL, VT, BFI,
3742 DAG.getConstant(ExtraMask | ExistingMask, VT));
3745 /// An EXTR instruction is made up of two shifts, ORed together. This helper
3746 /// searches for and classifies those shifts.
3747 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
3749 if (N.getOpcode() == ISD::SHL)
3751 else if (N.getOpcode() == ISD::SRL)
3756 if (!isa<ConstantSDNode>(N.getOperand(1)))
3759 ShiftAmount = N->getConstantOperandVal(1);
3760 Src = N->getOperand(0);
3764 /// EXTR instruction extracts a contiguous chunk of bits from two existing
3765 /// registers viewed as a high/low pair. This function looks for the pattern:
3766 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
3767 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
3769 static SDValue tryCombineToEXTR(SDNode *N,
3770 TargetLowering::DAGCombinerInfo &DCI) {
3771 SelectionDAG &DAG = DCI.DAG;
3773 EVT VT = N->getValueType(0);
3775 assert(N->getOpcode() == ISD::OR && "Unexpected root");
3777 if (VT != MVT::i32 && VT != MVT::i64)
3781 uint32_t ShiftLHS = 0;
3783 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
3787 uint32_t ShiftRHS = 0;
3789 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
3792 // If they're both trying to come from the high part of the register, they're
3793 // not really an EXTR.
3794 if (LHSFromHi == RHSFromHi)
3797 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
3801 std::swap(LHS, RHS);
3802 std::swap(ShiftLHS, ShiftRHS);
3805 return DAG.getNode(AArch64ISD::EXTR, DL, VT,
3807 DAG.getConstant(ShiftRHS, MVT::i64));
3810 /// Target-specific dag combine xforms for ISD::OR
3811 static SDValue PerformORCombine(SDNode *N,
3812 TargetLowering::DAGCombinerInfo &DCI,
3813 const AArch64Subtarget *Subtarget) {
3815 SelectionDAG &DAG = DCI.DAG;
3817 EVT VT = N->getValueType(0);
3819 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
3822 // Attempt to recognise bitfield-insert operations.
3823 SDValue Res = tryCombineToBFI(N, DCI, Subtarget);
3827 // Attempt to combine an existing MaskedBFI operation into one with a larger
3829 Res = tryCombineToLargerBFI(N, DCI, Subtarget);
3833 Res = tryCombineToEXTR(N, DCI);
3837 if (!Subtarget->hasNEON())
3840 // Attempt to use vector immediate-form BSL
3841 // (or (and B, A), (and C, ~A)) => (VBSL A, B, C) when A is a constant.
3843 SDValue N0 = N->getOperand(0);
3844 if (N0.getOpcode() != ISD::AND)
3847 SDValue N1 = N->getOperand(1);
3848 if (N1.getOpcode() != ISD::AND)
3851 if (VT.isVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
3853 unsigned SplatBitSize;
3855 BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(1));
3857 if (BVN0 && BVN0->isConstantSplat(SplatBits0, SplatUndef, SplatBitSize,
3860 BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(1));
3862 if (BVN1 && BVN1->isConstantSplat(SplatBits1, SplatUndef, SplatBitSize,
3863 HasAnyUndefs) && !HasAnyUndefs &&
3864 SplatBits0.getBitWidth() == SplatBits1.getBitWidth() &&
3865 SplatBits0 == ~SplatBits1) {
3867 return DAG.getNode(ISD::VSELECT, DL, VT, N0->getOperand(1),
3868 N0->getOperand(0), N1->getOperand(0));
3876 /// Target-specific dag combine xforms for ISD::SRA
3877 static SDValue PerformSRACombine(SDNode *N,
3878 TargetLowering::DAGCombinerInfo &DCI) {
3880 SelectionDAG &DAG = DCI.DAG;
3882 EVT VT = N->getValueType(0);
3884 // We're looking for an SRA/SHL pair which form an SBFX.
3886 if (VT != MVT::i32 && VT != MVT::i64)
3889 if (!isa<ConstantSDNode>(N->getOperand(1)))
3892 uint64_t ExtraSignBits = N->getConstantOperandVal(1);
3893 SDValue Shift = N->getOperand(0);
3895 if (Shift.getOpcode() != ISD::SHL)
3898 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
3901 uint64_t BitsOnLeft = Shift->getConstantOperandVal(1);
3902 uint64_t Width = VT.getSizeInBits() - ExtraSignBits;
3903 uint64_t LSB = VT.getSizeInBits() - Width - BitsOnLeft;
3905 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
3908 return DAG.getNode(AArch64ISD::SBFX, DL, VT, Shift.getOperand(0),
3909 DAG.getConstant(LSB, MVT::i64),
3910 DAG.getConstant(LSB + Width - 1, MVT::i64));
3913 /// Check if this is a valid build_vector for the immediate operand of
3914 /// a vector shift operation, where all the elements of the build_vector
3915 /// must have the same constant integer value.
3916 static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
3917 // Ignore bit_converts.
3918 while (Op.getOpcode() == ISD::BITCAST)
3919 Op = Op.getOperand(0);
3920 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
3921 APInt SplatBits, SplatUndef;
3922 unsigned SplatBitSize;
3924 if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
3925 HasAnyUndefs, ElementBits) ||
3926 SplatBitSize > ElementBits)
3928 Cnt = SplatBits.getSExtValue();
3932 /// Check if this is a valid build_vector for the immediate operand of
3933 /// a vector shift left operation. That value must be in the range:
3934 /// 0 <= Value < ElementBits
3935 static bool isVShiftLImm(SDValue Op, EVT VT, int64_t &Cnt) {
3936 assert(VT.isVector() && "vector shift count is not a vector type");
3937 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3938 if (!getVShiftImm(Op, ElementBits, Cnt))
3940 return (Cnt >= 0 && Cnt < ElementBits);
3943 /// Check if this is a valid build_vector for the immediate operand of a
3944 /// vector shift right operation. The value must be in the range:
3945 /// 1 <= Value <= ElementBits
3946 static bool isVShiftRImm(SDValue Op, EVT VT, int64_t &Cnt) {
3947 assert(VT.isVector() && "vector shift count is not a vector type");
3948 unsigned ElementBits = VT.getVectorElementType().getSizeInBits();
3949 if (!getVShiftImm(Op, ElementBits, Cnt))
3951 return (Cnt >= 1 && Cnt <= ElementBits);
3954 static SDValue GenForSextInreg(SDNode *N,
3955 TargetLowering::DAGCombinerInfo &DCI,
3956 EVT SrcVT, EVT DestVT, EVT SubRegVT,
3957 const int *Mask, SDValue Src) {
3958 SelectionDAG &DAG = DCI.DAG;
3960 = DAG.getNode(ISD::BITCAST, SDLoc(N), SrcVT, Src);
3962 = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), DestVT, Bitcast);
3964 = DAG.getVectorShuffle(DestVT, SDLoc(N), Sext, DAG.getUNDEF(DestVT), Mask);
3965 SDValue ExtractSubreg
3966 = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, SDLoc(N),
3967 SubRegVT, ShuffleVec,
3968 DAG.getTargetConstant(AArch64::sub_64, MVT::i32)), 0);
3969 return ExtractSubreg;
3972 /// Checks for vector shifts and lowers them.
3973 static SDValue PerformShiftCombine(SDNode *N,
3974 TargetLowering::DAGCombinerInfo &DCI,
3975 const AArch64Subtarget *ST) {
3976 SelectionDAG &DAG = DCI.DAG;
3977 EVT VT = N->getValueType(0);
3978 if (N->getOpcode() == ISD::SRA && (VT == MVT::i32 || VT == MVT::i64))
3979 return PerformSRACombine(N, DCI);
3981 // We're looking for an SRA/SHL pair to help generating instruction
3982 // sshll v0.8h, v0.8b, #0
3983 // The instruction STXL is also the alias of this instruction.
3985 // For example, for DAG like below,
3986 // v2i32 = sra (v2i32 (shl v2i32, 16)), 16
3987 // we can transform it into
3988 // v2i32 = EXTRACT_SUBREG
3989 // (v4i32 (suffle_vector
3990 // (v4i32 (sext (v4i16 (bitcast v2i32))),
3991 // undef, (0, 2, u, u)),
3994 // With this transformation we expect to generate "SSHLL + UZIP1"
3995 // Sometimes UZIP1 can be optimized away by combining with other context.
3996 int64_t ShrCnt, ShlCnt;
3997 if (N->getOpcode() == ISD::SRA
3998 && (VT == MVT::v2i32 || VT == MVT::v4i16)
3999 && isVShiftRImm(N->getOperand(1), VT, ShrCnt)
4000 && N->getOperand(0).getOpcode() == ISD::SHL
4001 && isVShiftRImm(N->getOperand(0).getOperand(1), VT, ShlCnt)) {
4002 SDValue Src = N->getOperand(0).getOperand(0);
4003 if (VT == MVT::v2i32 && ShrCnt == 16 && ShlCnt == 16) {
4004 // sext_inreg(v2i32, v2i16)
4005 // We essentially only care the Mask {0, 2, u, u}
4006 int Mask[4] = {0, 2, 4, 6};
4007 return GenForSextInreg(N, DCI, MVT::v4i16, MVT::v4i32, MVT::v2i32,
4010 else if (VT == MVT::v2i32 && ShrCnt == 24 && ShlCnt == 24) {
4011 // sext_inreg(v2i16, v2i8)
4012 // We essentially only care the Mask {0, u, 4, u, u, u, u, u, u, u, u, u}
4013 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
4014 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v2i32,
4017 else if (VT == MVT::v4i16 && ShrCnt == 8 && ShlCnt == 8) {
4018 // sext_inreg(v4i16, v4i8)
4019 // We essentially only care the Mask {0, 2, 4, 6, u, u, u, u, u, u, u, u}
4020 int Mask[8] = {0, 2, 4, 6, 8, 10, 12, 14};
4021 return GenForSextInreg(N, DCI, MVT::v8i8, MVT::v8i16, MVT::v4i16,
4026 // Nothing to be done for scalar shifts.
4027 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4028 if (!VT.isVector() || !TLI.isTypeLegal(VT))
4031 assert(ST->hasNEON() && "unexpected vector shift");
4034 switch (N->getOpcode()) {
4036 llvm_unreachable("unexpected shift opcode");
4039 if (isVShiftLImm(N->getOperand(1), VT, Cnt)) {
4041 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
4042 DAG.getConstant(Cnt, MVT::i32));
4043 return DAG.getNode(ISD::SHL, SDLoc(N), VT, N->getOperand(0), RHS);
4049 if (isVShiftRImm(N->getOperand(1), VT, Cnt)) {
4051 DAG.getNode(AArch64ISD::NEON_VDUP, SDLoc(N->getOperand(1)), VT,
4052 DAG.getConstant(Cnt, MVT::i32));
4053 return DAG.getNode(N->getOpcode(), SDLoc(N), VT, N->getOperand(0), RHS);
4061 /// ARM-specific DAG combining for intrinsics.
4062 static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) {
4063 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
4067 // Don't do anything for most intrinsics.
4070 case Intrinsic::arm_neon_vqshifts:
4071 case Intrinsic::arm_neon_vqshiftu:
4072 EVT VT = N->getOperand(1).getValueType();
4074 if (!isVShiftLImm(N->getOperand(2), VT, Cnt))
4076 unsigned VShiftOpc = (IntNo == Intrinsic::arm_neon_vqshifts)
4077 ? AArch64ISD::NEON_QSHLs
4078 : AArch64ISD::NEON_QSHLu;
4079 return DAG.getNode(VShiftOpc, SDLoc(N), N->getValueType(0),
4080 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32));
4086 /// Target-specific DAG combine function for NEON load/store intrinsics
4087 /// to merge base address updates.
4088 static SDValue CombineBaseUpdate(SDNode *N,
4089 TargetLowering::DAGCombinerInfo &DCI) {
4090 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
4093 SelectionDAG &DAG = DCI.DAG;
4094 bool isIntrinsic = (N->getOpcode() == ISD::INTRINSIC_VOID ||
4095 N->getOpcode() == ISD::INTRINSIC_W_CHAIN);
4096 unsigned AddrOpIdx = (isIntrinsic ? 2 : 1);
4097 SDValue Addr = N->getOperand(AddrOpIdx);
4099 // Search for a use of the address operand that is an increment.
4100 for (SDNode::use_iterator UI = Addr.getNode()->use_begin(),
4101 UE = Addr.getNode()->use_end(); UI != UE; ++UI) {
4103 if (User->getOpcode() != ISD::ADD ||
4104 UI.getUse().getResNo() != Addr.getResNo())
4107 // Check that the add is independent of the load/store. Otherwise, folding
4108 // it would create a cycle.
4109 if (User->isPredecessorOf(N) || N->isPredecessorOf(User))
4112 // Find the new opcode for the updating load/store.
4114 bool isLaneOp = false;
4115 unsigned NewOpc = 0;
4116 unsigned NumVecs = 0;
4118 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
4120 default: llvm_unreachable("unexpected intrinsic for Neon base update");
4121 case Intrinsic::arm_neon_vld1: NewOpc = AArch64ISD::NEON_LD1_UPD;
4123 case Intrinsic::arm_neon_vld2: NewOpc = AArch64ISD::NEON_LD2_UPD;
4125 case Intrinsic::arm_neon_vld3: NewOpc = AArch64ISD::NEON_LD3_UPD;
4127 case Intrinsic::arm_neon_vld4: NewOpc = AArch64ISD::NEON_LD4_UPD;
4129 case Intrinsic::arm_neon_vst1: NewOpc = AArch64ISD::NEON_ST1_UPD;
4130 NumVecs = 1; isLoad = false; break;
4131 case Intrinsic::arm_neon_vst2: NewOpc = AArch64ISD::NEON_ST2_UPD;
4132 NumVecs = 2; isLoad = false; break;
4133 case Intrinsic::arm_neon_vst3: NewOpc = AArch64ISD::NEON_ST3_UPD;
4134 NumVecs = 3; isLoad = false; break;
4135 case Intrinsic::arm_neon_vst4: NewOpc = AArch64ISD::NEON_ST4_UPD;
4136 NumVecs = 4; isLoad = false; break;
4137 case Intrinsic::aarch64_neon_vld1x2: NewOpc = AArch64ISD::NEON_LD1x2_UPD;
4139 case Intrinsic::aarch64_neon_vld1x3: NewOpc = AArch64ISD::NEON_LD1x3_UPD;
4141 case Intrinsic::aarch64_neon_vld1x4: NewOpc = AArch64ISD::NEON_LD1x4_UPD;
4143 case Intrinsic::aarch64_neon_vst1x2: NewOpc = AArch64ISD::NEON_ST1x2_UPD;
4144 NumVecs = 2; isLoad = false; break;
4145 case Intrinsic::aarch64_neon_vst1x3: NewOpc = AArch64ISD::NEON_ST1x3_UPD;
4146 NumVecs = 3; isLoad = false; break;
4147 case Intrinsic::aarch64_neon_vst1x4: NewOpc = AArch64ISD::NEON_ST1x4_UPD;
4148 NumVecs = 4; isLoad = false; break;
4149 case Intrinsic::arm_neon_vld2lane: NewOpc = AArch64ISD::NEON_LD2LN_UPD;
4150 NumVecs = 2; isLaneOp = true; break;
4151 case Intrinsic::arm_neon_vld3lane: NewOpc = AArch64ISD::NEON_LD3LN_UPD;
4152 NumVecs = 3; isLaneOp = true; break;
4153 case Intrinsic::arm_neon_vld4lane: NewOpc = AArch64ISD::NEON_LD4LN_UPD;
4154 NumVecs = 4; isLaneOp = true; break;
4155 case Intrinsic::arm_neon_vst2lane: NewOpc = AArch64ISD::NEON_ST2LN_UPD;
4156 NumVecs = 2; isLoad = false; isLaneOp = true; break;
4157 case Intrinsic::arm_neon_vst3lane: NewOpc = AArch64ISD::NEON_ST3LN_UPD;
4158 NumVecs = 3; isLoad = false; isLaneOp = true; break;
4159 case Intrinsic::arm_neon_vst4lane: NewOpc = AArch64ISD::NEON_ST4LN_UPD;
4160 NumVecs = 4; isLoad = false; isLaneOp = true; break;
4164 switch (N->getOpcode()) {
4165 default: llvm_unreachable("unexpected opcode for Neon base update");
4166 case AArch64ISD::NEON_LD2DUP: NewOpc = AArch64ISD::NEON_LD2DUP_UPD;
4168 case AArch64ISD::NEON_LD3DUP: NewOpc = AArch64ISD::NEON_LD3DUP_UPD;
4170 case AArch64ISD::NEON_LD4DUP: NewOpc = AArch64ISD::NEON_LD4DUP_UPD;
4175 // Find the size of memory referenced by the load/store.
4178 VecTy = N->getValueType(0);
4180 VecTy = N->getOperand(AddrOpIdx + 1).getValueType();
4181 unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
4183 NumBytes /= VecTy.getVectorNumElements();
4185 // If the increment is a constant, it must match the memory ref size.
4186 SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
4187 if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
4188 uint32_t IncVal = CInc->getZExtValue();
4189 if (IncVal != NumBytes)
4191 Inc = DAG.getTargetConstant(IncVal, MVT::i32);
4194 // Create the new updating load/store node.
4196 unsigned NumResultVecs = (isLoad ? NumVecs : 0);
4198 for (n = 0; n < NumResultVecs; ++n)
4200 Tys[n++] = MVT::i64;
4201 Tys[n] = MVT::Other;
4202 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumResultVecs + 2));
4203 SmallVector<SDValue, 8> Ops;
4204 Ops.push_back(N->getOperand(0)); // incoming chain
4205 Ops.push_back(N->getOperand(AddrOpIdx));
4207 for (unsigned i = AddrOpIdx + 1; i < N->getNumOperands(); ++i) {
4208 Ops.push_back(N->getOperand(i));
4210 MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
4211 SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys,
4212 Ops, MemInt->getMemoryVT(),
4213 MemInt->getMemOperand());
4216 std::vector<SDValue> NewResults;
4217 for (unsigned i = 0; i < NumResultVecs; ++i) {
4218 NewResults.push_back(SDValue(UpdN.getNode(), i));
4220 NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1)); // chain
4221 DCI.CombineTo(N, NewResults);
4222 DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));
4229 /// For a VDUPLANE node N, check if its source operand is a vldN-lane (N > 1)
4230 /// intrinsic, and if all the other uses of that intrinsic are also VDUPLANEs.
4231 /// If so, combine them to a vldN-dup operation and return true.
4232 static SDValue CombineVLDDUP(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
4233 SelectionDAG &DAG = DCI.DAG;
4234 EVT VT = N->getValueType(0);
4236 // Check if the VDUPLANE operand is a vldN-dup intrinsic.
4237 SDNode *VLD = N->getOperand(0).getNode();
4238 if (VLD->getOpcode() != ISD::INTRINSIC_W_CHAIN)
4240 unsigned NumVecs = 0;
4241 unsigned NewOpc = 0;
4242 unsigned IntNo = cast<ConstantSDNode>(VLD->getOperand(1))->getZExtValue();
4243 if (IntNo == Intrinsic::arm_neon_vld2lane) {
4245 NewOpc = AArch64ISD::NEON_LD2DUP;
4246 } else if (IntNo == Intrinsic::arm_neon_vld3lane) {
4248 NewOpc = AArch64ISD::NEON_LD3DUP;
4249 } else if (IntNo == Intrinsic::arm_neon_vld4lane) {
4251 NewOpc = AArch64ISD::NEON_LD4DUP;
4256 // First check that all the vldN-lane uses are VDUPLANEs and that the lane
4257 // numbers match the load.
4258 unsigned VLDLaneNo =
4259 cast<ConstantSDNode>(VLD->getOperand(NumVecs + 3))->getZExtValue();
4260 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4262 // Ignore uses of the chain result.
4263 if (UI.getUse().getResNo() == NumVecs)
4266 if (User->getOpcode() != AArch64ISD::NEON_VDUPLANE ||
4267 VLDLaneNo != cast<ConstantSDNode>(User->getOperand(1))->getZExtValue())
4271 // Create the vldN-dup node.
4274 for (n = 0; n < NumVecs; ++n)
4276 Tys[n] = MVT::Other;
4277 SDVTList SDTys = DAG.getVTList(ArrayRef<EVT>(Tys, NumVecs + 1));
4278 SDValue Ops[] = { VLD->getOperand(0), VLD->getOperand(2) };
4279 MemIntrinsicSDNode *VLDMemInt = cast<MemIntrinsicSDNode>(VLD);
4280 SDValue VLDDup = DAG.getMemIntrinsicNode(NewOpc, SDLoc(VLD), SDTys, Ops,
4281 VLDMemInt->getMemoryVT(),
4282 VLDMemInt->getMemOperand());
4285 for (SDNode::use_iterator UI = VLD->use_begin(), UE = VLD->use_end();
4287 unsigned ResNo = UI.getUse().getResNo();
4288 // Ignore uses of the chain result.
4289 if (ResNo == NumVecs)
4292 DCI.CombineTo(User, SDValue(VLDDup.getNode(), ResNo));
4295 // Now the vldN-lane intrinsic is dead except for its chain result.
4296 // Update uses of the chain.
4297 std::vector<SDValue> VLDDupResults;
4298 for (unsigned n = 0; n < NumVecs; ++n)
4299 VLDDupResults.push_back(SDValue(VLDDup.getNode(), n));
4300 VLDDupResults.push_back(SDValue(VLDDup.getNode(), NumVecs));
4301 DCI.CombineTo(VLD, VLDDupResults);
4303 return SDValue(N, 0);
4306 // vselect (v1i1 setcc) ->
4307 // vselect (v1iXX setcc) (XX is the size of the compared operand type)
4308 // FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
4309 // condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
4311 static SDValue PerformVSelectCombine(SDNode *N, SelectionDAG &DAG) {
4312 SDValue N0 = N->getOperand(0);
4313 EVT CCVT = N0.getValueType();
4315 if (N0.getOpcode() != ISD::SETCC || CCVT.getVectorNumElements() != 1 ||
4316 CCVT.getVectorElementType() != MVT::i1)
4319 EVT ResVT = N->getValueType(0);
4320 EVT CmpVT = N0.getOperand(0).getValueType();
4321 // Only combine when the result type is of the same size as the compared
4323 if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
4326 SDValue IfTrue = N->getOperand(1);
4327 SDValue IfFalse = N->getOperand(2);
4329 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
4330 N0.getOperand(0), N0.getOperand(1),
4331 cast<CondCodeSDNode>(N0.getOperand(2))->get());
4332 return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
4336 // sign_extend (extract_vector_elt (v1i1 setcc)) ->
4337 // extract_vector_elt (v1iXX setcc)
4338 // (XX is the size of the compared operand type)
4339 static SDValue PerformSignExtendCombine(SDNode *N, SelectionDAG &DAG) {
4340 SDValue N0 = N->getOperand(0);
4341 SDValue Vec = N0.getOperand(0);
4343 if (N0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4344 Vec.getOpcode() != ISD::SETCC)
4347 EVT ResVT = N->getValueType(0);
4348 EVT CmpVT = Vec.getOperand(0).getValueType();
4349 // Only optimize when the result type is of the same size as the element
4350 // type of the compared operand.
4351 if (ResVT.getSizeInBits() != CmpVT.getVectorElementType().getSizeInBits())
4354 SDValue Lane = N0.getOperand(1);
4356 DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
4357 Vec.getOperand(0), Vec.getOperand(1),
4358 cast<CondCodeSDNode>(Vec.getOperand(2))->get());
4359 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(N), ResVT,
4364 AArch64TargetLowering::PerformDAGCombine(SDNode *N,
4365 DAGCombinerInfo &DCI) const {
4366 switch (N->getOpcode()) {
4368 case ISD::AND: return PerformANDCombine(N, DCI);
4369 case ISD::OR: return PerformORCombine(N, DCI, getSubtarget());
4373 return PerformShiftCombine(N, DCI, getSubtarget());
4374 case ISD::VSELECT: return PerformVSelectCombine(N, DCI.DAG);
4375 case ISD::SIGN_EXTEND: return PerformSignExtendCombine(N, DCI.DAG);
4376 case ISD::INTRINSIC_WO_CHAIN:
4377 return PerformIntrinsicCombine(N, DCI.DAG);
4378 case AArch64ISD::NEON_VDUPLANE:
4379 return CombineVLDDUP(N, DCI);
4380 case AArch64ISD::NEON_LD2DUP:
4381 case AArch64ISD::NEON_LD3DUP:
4382 case AArch64ISD::NEON_LD4DUP:
4383 return CombineBaseUpdate(N, DCI);
4384 case ISD::INTRINSIC_VOID:
4385 case ISD::INTRINSIC_W_CHAIN:
4386 switch (cast<ConstantSDNode>(N->getOperand(1))->getZExtValue()) {
4387 case Intrinsic::arm_neon_vld1:
4388 case Intrinsic::arm_neon_vld2:
4389 case Intrinsic::arm_neon_vld3:
4390 case Intrinsic::arm_neon_vld4:
4391 case Intrinsic::arm_neon_vst1:
4392 case Intrinsic::arm_neon_vst2:
4393 case Intrinsic::arm_neon_vst3:
4394 case Intrinsic::arm_neon_vst4:
4395 case Intrinsic::arm_neon_vld2lane:
4396 case Intrinsic::arm_neon_vld3lane:
4397 case Intrinsic::arm_neon_vld4lane:
4398 case Intrinsic::aarch64_neon_vld1x2:
4399 case Intrinsic::aarch64_neon_vld1x3:
4400 case Intrinsic::aarch64_neon_vld1x4:
4401 case Intrinsic::aarch64_neon_vst1x2:
4402 case Intrinsic::aarch64_neon_vst1x3:
4403 case Intrinsic::aarch64_neon_vst1x4:
4404 case Intrinsic::arm_neon_vst2lane:
4405 case Intrinsic::arm_neon_vst3lane:
4406 case Intrinsic::arm_neon_vst4lane:
4407 return CombineBaseUpdate(N, DCI);
4416 AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
4417 VT = VT.getScalarType();
4422 switch (VT.getSimpleVT().SimpleTy) {
4436 bool AArch64TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
4439 const AArch64Subtarget *Subtarget = getSubtarget();
4440 // The AllowsUnaliged flag models the SCTLR.A setting in ARM cpus
4441 bool AllowsUnaligned = Subtarget->allowsUnalignedMem();
4443 switch (VT.getSimpleVT().SimpleTy) {
4447 case MVT::i8: case MVT::i16:
4448 case MVT::i32: case MVT::i64:
4449 case MVT::f32: case MVT::f64: {
4450 // Unaligned access can use (for example) LRDB, LRDH, LDRW
4451 if (AllowsUnaligned) {
4458 // 64-bit vector types
4459 case MVT::v8i8: case MVT::v4i16:
4460 case MVT::v2i32: case MVT::v1i64:
4461 case MVT::v2f32: case MVT::v1f64:
4462 // 128-bit vector types
4463 case MVT::v16i8: case MVT::v8i16:
4464 case MVT::v4i32: case MVT::v2i64:
4465 case MVT::v4f32: case MVT::v2f64: {
4466 // For any little-endian targets with neon, we can support unaligned
4467 // load/store of V registers using ld1/st1.
4468 // A big-endian target may also explicitly support unaligned accesses
4469 if (Subtarget->hasNEON() && (AllowsUnaligned || isLittleEndian())) {
4479 // Check whether a shuffle_vector could be presented as concat_vector.
4480 bool AArch64TargetLowering::isConcatVector(SDValue Op, SelectionDAG &DAG,
4481 SDValue V0, SDValue V1,
4483 SDValue &Res) const {
4485 EVT VT = Op.getValueType();
4486 if (VT.getSizeInBits() != 128)
4488 if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
4489 VT.getVectorElementType() != V1.getValueType().getVectorElementType())
4492 unsigned NumElts = VT.getVectorNumElements();
4493 bool isContactVector = true;
4494 bool splitV0 = false;
4495 if (V0.getValueType().getSizeInBits() == 128)
4498 for (int I = 0, E = NumElts / 2; I != E; I++) {
4500 isContactVector = false;
4505 if (isContactVector) {
4506 int offset = NumElts / 2;
4507 for (int I = NumElts / 2, E = NumElts; I != E; I++) {
4508 if (Mask[I] != I + splitV0 * offset) {
4509 isContactVector = false;
4515 if (isContactVector) {
4516 EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4519 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
4520 DAG.getConstant(0, MVT::i64));
4522 if (V1.getValueType().getSizeInBits() == 128) {
4523 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
4524 DAG.getConstant(0, MVT::i64));
4526 Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
4532 // Check whether a Build Vector could be presented as Shuffle Vector.
4533 // This Shuffle Vector maybe not legalized, so the length of its operand and
4534 // the length of result may not equal.
4535 bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
4536 SDValue &V0, SDValue &V1,
4539 EVT VT = Op.getValueType();
4540 unsigned NumElts = VT.getVectorNumElements();
4541 unsigned V0NumElts = 0;
4543 // Check if all elements are extracted from less than 3 vectors.
4544 for (unsigned i = 0; i < NumElts; ++i) {
4545 SDValue Elt = Op.getOperand(i);
4546 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4547 Elt.getOperand(0).getValueType().getVectorElementType() !=
4548 VT.getVectorElementType())
4551 if (!V0.getNode()) {
4552 V0 = Elt.getOperand(0);
4553 V0NumElts = V0.getValueType().getVectorNumElements();
4555 if (Elt.getOperand(0) == V0) {
4556 Mask[i] = (cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue());
4558 } else if (!V1.getNode()) {
4559 V1 = Elt.getOperand(0);
4561 if (Elt.getOperand(0) == V1) {
4562 unsigned Lane = cast<ConstantSDNode>(Elt->getOperand(1))->getZExtValue();
4563 Mask[i] = (Lane + V0NumElts);
4572 // LowerShiftRightParts - Lower SRL_PARTS and SRA_PARTS, which returns two
4573 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4574 SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
4575 SelectionDAG &DAG) const {
4576 assert(Op.getNumOperands() == 3 && "Not a quad-shift!");
4577 EVT VT = Op.getValueType();
4578 unsigned VTBits = VT.getSizeInBits();
4580 SDValue ShOpLo = Op.getOperand(0);
4581 SDValue ShOpHi = Op.getOperand(1);
4582 SDValue ShAmt = Op.getOperand(2);
4583 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
4585 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
4586 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4587 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4588 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
4589 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4590 DAG.getConstant(VTBits, MVT::i64));
4591 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
4592 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4593 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
4594 SDValue Tmp3 = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
4597 SDValue CmpOp = getSelectableIntSetCC(ExtraShAmt,
4598 DAG.getConstant(0, MVT::i64),
4602 SDValue Hi = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
4603 DAG.getConstant(0, Tmp3.getValueType()), Tmp3,
4605 SDValue Lo = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
4606 TrueVal, FalseVal, A64cc);
4608 SDValue Ops[2] = { Lo, Hi };
4609 return DAG.getMergeValues(Ops, dl);
4612 /// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
4613 /// i64 values and take a 2 x i64 value to shift plus a shift amount.
4614 SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
4615 SelectionDAG &DAG) const {
4616 assert(Op.getNumOperands() == 3 && "Not a quad-shift!");
4617 EVT VT = Op.getValueType();
4618 unsigned VTBits = VT.getSizeInBits();
4620 SDValue ShOpLo = Op.getOperand(0);
4621 SDValue ShOpHi = Op.getOperand(1);
4622 SDValue ShAmt = Op.getOperand(2);
4624 assert(Op.getOpcode() == ISD::SHL_PARTS);
4625 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
4626 DAG.getConstant(VTBits, MVT::i64), ShAmt);
4627 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
4628 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
4629 DAG.getConstant(VTBits, MVT::i64));
4630 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
4631 SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
4632 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
4633 SDValue Tmp4 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4636 SDValue CmpOp = getSelectableIntSetCC(ExtraShAmt,
4637 DAG.getConstant(0, MVT::i64),
4641 SDValue Lo = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
4642 DAG.getConstant(0, Tmp4.getValueType()), Tmp4,
4644 SDValue Hi = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
4645 Tmp3, FalseVal, A64cc);
4647 SDValue Ops[2] = { Lo, Hi };
4648 return DAG.getMergeValues(Ops, dl);
4651 // If this is a case we can't handle, return null and let the default
4652 // expansion code take care of it.
4654 AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
4655 const AArch64Subtarget *ST) const {
4657 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
4659 EVT VT = Op.getValueType();
4661 APInt SplatBits, SplatUndef;
4662 unsigned SplatBitSize;
4665 unsigned UseNeonMov = VT.getSizeInBits() >= 64;
4667 // Note we favor lowering MOVI over MVNI.
4668 // This has implications on the definition of patterns in TableGen to select
4669 // BIC immediate instructions but not ORR immediate instructions.
4670 // If this lowering order is changed, TableGen patterns for BIC immediate and
4671 // ORR immediate instructions have to be updated.
4673 BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
4674 if (SplatBitSize <= 64) {
4675 // First attempt to use vector immediate-form MOVI
4678 unsigned OpCmode = 0;
4680 if (isNeonModifiedImm(SplatBits.getZExtValue(), SplatUndef.getZExtValue(),
4681 SplatBitSize, DAG, VT.is128BitVector(),
4682 Neon_Mov_Imm, NeonMovVT, Imm, OpCmode)) {
4683 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4684 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4686 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4687 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MOVIMM, DL, NeonMovVT,
4688 ImmVal, OpCmodeVal);
4689 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4693 // Then attempt to use vector immediate-form MVNI
4694 uint64_t NegatedImm = (~SplatBits).getZExtValue();
4695 if (isNeonModifiedImm(NegatedImm, SplatUndef.getZExtValue(), SplatBitSize,
4696 DAG, VT.is128BitVector(), Neon_Mvn_Imm, NeonMovVT,
4698 SDValue ImmVal = DAG.getTargetConstant(Imm, MVT::i32);
4699 SDValue OpCmodeVal = DAG.getConstant(OpCmode, MVT::i32);
4700 if (ImmVal.getNode() && OpCmodeVal.getNode()) {
4701 SDValue NeonMov = DAG.getNode(AArch64ISD::NEON_MVNIMM, DL, NeonMovVT,
4702 ImmVal, OpCmodeVal);
4703 return DAG.getNode(ISD::BITCAST, DL, VT, NeonMov);
4707 // Attempt to use vector immediate-form FMOV
4708 if (((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) ||
4709 (VT == MVT::v2f64 && SplatBitSize == 64)) {
4711 SplatBitSize == 32 ? APFloat::IEEEsingle : APFloat::IEEEdouble,
4714 if (A64Imms::isFPImm(RealVal, ImmVal)) {
4715 SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
4716 return DAG.getNode(AArch64ISD::NEON_FMOVIMM, DL, VT, Val);
4722 unsigned NumElts = VT.getVectorNumElements();
4723 bool isOnlyLowElement = true;
4724 bool usesOnlyOneValue = true;
4725 bool hasDominantValue = false;
4726 bool isConstant = true;
4728 // Map of the number of times a particular SDValue appears in the
4730 DenseMap<SDValue, unsigned> ValueCounts;
4732 for (unsigned i = 0; i < NumElts; ++i) {
4733 SDValue V = Op.getOperand(i);
4734 if (V.getOpcode() == ISD::UNDEF)
4737 isOnlyLowElement = false;
4738 if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
4741 ValueCounts.insert(std::make_pair(V, 0));
4742 unsigned &Count = ValueCounts[V];
4744 // Is this value dominant? (takes up more than half of the lanes)
4745 if (++Count > (NumElts / 2)) {
4746 hasDominantValue = true;
4750 if (ValueCounts.size() != 1)
4751 usesOnlyOneValue = false;
4752 if (!Value.getNode() && ValueCounts.size() > 0)
4753 Value = ValueCounts.begin()->first;
4755 if (ValueCounts.size() == 0)
4756 return DAG.getUNDEF(VT);
4758 if (isOnlyLowElement)
4759 return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
4761 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
4762 if (hasDominantValue && EltSize <= 64) {
4763 // Use VDUP for non-constant splats.
4767 // If we are DUPing a value that comes directly from a vector, we could
4768 // just use DUPLANE. We can only do this if the lane being extracted
4769 // is at a constant index, as the DUP from lane instructions only have
4770 // constant-index forms.
4772 // If there is a TRUNCATE between EXTRACT_VECTOR_ELT and DUP, we can
4773 // remove TRUNCATE for DUPLANE by apdating the source vector to
4774 // appropriate vector type and lane index.
4776 // FIXME: for now we have v1i8, v1i16, v1i32 legal vector types, if they
4777 // are not legal any more, no need to check the type size in bits should
4778 // be large than 64.
4780 if (Value->getOpcode() == ISD::TRUNCATE)
4781 V = Value->getOperand(0);
4782 if (V->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
4783 isa<ConstantSDNode>(V->getOperand(1)) &&
4784 V->getOperand(0).getValueType().getSizeInBits() >= 64) {
4786 // If the element size of source vector is larger than DUPLANE
4787 // element size, we can do transformation by,
4788 // 1) bitcasting source register to smaller element vector
4789 // 2) mutiplying the lane index by SrcEltSize/ResEltSize
4790 // For example, we can lower
4791 // "v8i16 vdup_lane(v4i32, 1)"
4793 // "v8i16 vdup_lane(v8i16 bitcast(v4i32), 2)".
4794 SDValue SrcVec = V->getOperand(0);
4795 unsigned SrcEltSize =
4796 SrcVec.getValueType().getVectorElementType().getSizeInBits();
4797 unsigned ResEltSize = VT.getVectorElementType().getSizeInBits();
4798 if (SrcEltSize > ResEltSize) {
4799 assert((SrcEltSize % ResEltSize == 0) && "Invalid element size");
4801 unsigned SrcSize = SrcVec.getValueType().getSizeInBits();
4802 unsigned ResSize = VT.getSizeInBits();
4804 if (SrcSize > ResSize) {
4805 assert((SrcSize % ResSize == 0) && "Invalid vector size");
4807 EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
4808 SrcSize / ResEltSize);
4809 BitCast = DAG.getNode(ISD::BITCAST, DL, CastVT, SrcVec);
4811 assert((SrcSize == ResSize) && "Invalid vector size of source vec");
4812 BitCast = DAG.getNode(ISD::BITCAST, DL, VT, SrcVec);
4815 unsigned LaneIdx = V->getConstantOperandVal(1);
4817 DAG.getConstant((SrcEltSize / ResEltSize) * LaneIdx, MVT::i64);
4818 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, BitCast, Lane);
4820 assert((SrcEltSize == ResEltSize) &&
4821 "Invalid element size of source vec");
4822 N = DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, V->getOperand(0),
4826 N = DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4828 if (!usesOnlyOneValue) {
4829 // The dominant value was splatted as 'N', but we now have to insert
4830 // all differing elements.
4831 for (unsigned I = 0; I < NumElts; ++I) {
4832 if (Op.getOperand(I) == Value)
4834 SmallVector<SDValue, 3> Ops;
4836 Ops.push_back(Op.getOperand(I));
4837 Ops.push_back(DAG.getConstant(I, MVT::i64));
4838 N = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Ops);
4843 if (usesOnlyOneValue && isConstant) {
4844 return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Value);
4847 // If all elements are constants and the case above didn't get hit, fall back
4848 // to the default expansion, which will generate a load from the constant
4853 // Try to lower this in lowering ShuffleVector way.
4856 if (isKnownShuffleVector(Op, DAG, V0, V1, Mask)) {
4857 unsigned V0NumElts = V0.getValueType().getVectorNumElements();
4858 if (!V1.getNode() && V0NumElts == NumElts * 2) {
4859 V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4860 DAG.getConstant(NumElts, MVT::i64));
4861 V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
4862 DAG.getConstant(0, MVT::i64));
4863 V0NumElts = V0.getValueType().getVectorNumElements();
4866 if (V1.getNode() && NumElts == V0NumElts &&
4867 V0NumElts == V1.getValueType().getVectorNumElements()) {
4868 SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
4869 if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
4872 return LowerVECTOR_SHUFFLE(Shuffle, DAG);
4875 if (isConcatVector(Op, DAG, V0, V1, Mask, Res))
4880 // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
4881 // know the default expansion would otherwise fall back on something even
4882 // worse. For a vector with one or two non-undef values, that's
4883 // scalar_to_vector for the elements followed by a shuffle (provided the
4884 // shuffle is valid for the target) and materialization element by element
4885 // on the stack followed by a load for everything else.
4886 if (!isConstant && !usesOnlyOneValue) {
4887 SDValue Vec = DAG.getUNDEF(VT);
4888 for (unsigned i = 0 ; i < NumElts; ++i) {
4889 SDValue V = Op.getOperand(i);
4890 if (V.getOpcode() == ISD::UNDEF)
4892 SDValue LaneIdx = DAG.getConstant(i, MVT::i64);
4893 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V, LaneIdx);
4900 /// isREVMask - Check if a vector shuffle corresponds to a REV
4901 /// instruction with the specified blocksize. (The order of the elements
4902 /// within each block of the vector is reversed.)
4903 static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
4904 assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
4905 "Only possible block sizes for REV are: 16, 32, 64");
4907 unsigned EltSz = VT.getVectorElementType().getSizeInBits();
4911 unsigned NumElts = VT.getVectorNumElements();
4912 unsigned BlockElts = M[0] + 1;
4913 // If the first shuffle index is UNDEF, be optimistic.
4915 BlockElts = BlockSize / EltSz;
4917 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
4920 for (unsigned i = 0; i < NumElts; ++i) {
4922 continue; // ignore UNDEF indices
4923 if ((unsigned)M[i] != (i - i % BlockElts) + (BlockElts - 1 - i % BlockElts))
4930 // isPermuteMask - Check whether the vector shuffle matches to UZP, ZIP and
4932 static unsigned isPermuteMask(ArrayRef<int> M, EVT VT, bool isV2undef) {
4933 unsigned NumElts = VT.getVectorNumElements();
4937 bool ismatch = true;
4940 for (unsigned i = 0; i < NumElts; ++i) {
4941 unsigned answer = i * 2;
4942 if (isV2undef && answer >= NumElts)
4944 if (M[i] != -1 && (unsigned)M[i] != answer) {
4950 return AArch64ISD::NEON_UZP1;
4954 for (unsigned i = 0; i < NumElts; ++i) {
4955 unsigned answer = i * 2 + 1;
4956 if (isV2undef && answer >= NumElts)
4958 if (M[i] != -1 && (unsigned)M[i] != answer) {
4964 return AArch64ISD::NEON_UZP2;
4968 for (unsigned i = 0; i < NumElts; ++i) {
4969 unsigned answer = i / 2 + NumElts * (i % 2);
4970 if (isV2undef && answer >= NumElts)
4972 if (M[i] != -1 && (unsigned)M[i] != answer) {
4978 return AArch64ISD::NEON_ZIP1;
4982 for (unsigned i = 0; i < NumElts; ++i) {
4983 unsigned answer = (NumElts + i) / 2 + NumElts * (i % 2);
4984 if (isV2undef && answer >= NumElts)
4986 if (M[i] != -1 && (unsigned)M[i] != answer) {
4992 return AArch64ISD::NEON_ZIP2;
4996 for (unsigned i = 0; i < NumElts; ++i) {
4997 unsigned answer = i + (NumElts - 1) * (i % 2);
4998 if (isV2undef && answer >= NumElts)
5000 if (M[i] != -1 && (unsigned)M[i] != answer) {
5006 return AArch64ISD::NEON_TRN1;
5010 for (unsigned i = 0; i < NumElts; ++i) {
5011 unsigned answer = 1 + i + (NumElts - 1) * (i % 2);
5012 if (isV2undef && answer >= NumElts)
5014 if (M[i] != -1 && (unsigned)M[i] != answer) {
5020 return AArch64ISD::NEON_TRN2;
5026 AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
5027 SelectionDAG &DAG) const {
5028 SDValue V1 = Op.getOperand(0);
5029 SDValue V2 = Op.getOperand(1);
5031 EVT VT = Op.getValueType();
5032 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
5034 // Convert shuffles that are directly supported on NEON to target-specific
5035 // DAG nodes, instead of keeping them as shuffles and matching them again
5036 // during code selection. This is more efficient and avoids the possibility
5037 // of inconsistencies between legalization and selection.
5038 ArrayRef<int> ShuffleMask = SVN->getMask();
5040 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
5044 if (isREVMask(ShuffleMask, VT, 64))
5045 return DAG.getNode(AArch64ISD::NEON_REV64, dl, VT, V1);
5046 if (isREVMask(ShuffleMask, VT, 32))
5047 return DAG.getNode(AArch64ISD::NEON_REV32, dl, VT, V1);
5048 if (isREVMask(ShuffleMask, VT, 16))
5049 return DAG.getNode(AArch64ISD::NEON_REV16, dl, VT, V1);
5052 if (V2.getOpcode() == ISD::UNDEF)
5053 ISDNo = isPermuteMask(ShuffleMask, VT, true);
5055 ISDNo = isPermuteMask(ShuffleMask, VT, false);
5058 if (V2.getOpcode() == ISD::UNDEF)
5059 return DAG.getNode(ISDNo, dl, VT, V1, V1);
5061 return DAG.getNode(ISDNo, dl, VT, V1, V2);
5065 if (isConcatVector(Op, DAG, V1, V2, &ShuffleMask[0], Res))
5068 // If the element of shuffle mask are all the same constant, we can
5069 // transform it into either NEON_VDUP or NEON_VDUPLANE
5070 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
5071 int Lane = SVN->getSplatIndex();
5072 // If this is undef splat, generate it via "just" vdup, if possible.
5073 if (Lane == -1) Lane = 0;
5075 // Test if V1 is a SCALAR_TO_VECTOR.
5076 if (V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5077 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT, V1.getOperand(0));
5079 // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR.
5080 if (V1.getOpcode() == ISD::BUILD_VECTOR) {
5081 bool IsScalarToVector = true;
5082 for (unsigned i = 0, e = V1.getNumOperands(); i != e; ++i)
5083 if (V1.getOperand(i).getOpcode() != ISD::UNDEF &&
5084 i != (unsigned)Lane) {
5085 IsScalarToVector = false;
5088 if (IsScalarToVector)
5089 return DAG.getNode(AArch64ISD::NEON_VDUP, dl, VT,
5090 V1.getOperand(Lane));
5093 // Test if V1 is a EXTRACT_SUBVECTOR.
5094 if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
5095 int ExtLane = cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
5096 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1.getOperand(0),
5097 DAG.getConstant(Lane + ExtLane, MVT::i64));
5099 // Test if V1 is a CONCAT_VECTORS.
5100 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
5101 V1.getOperand(1).getOpcode() == ISD::UNDEF) {
5102 SDValue Op0 = V1.getOperand(0);
5103 assert((unsigned)Lane < Op0.getValueType().getVectorNumElements() &&
5104 "Invalid vector lane access");
5105 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, Op0,
5106 DAG.getConstant(Lane, MVT::i64));
5109 return DAG.getNode(AArch64ISD::NEON_VDUPLANE, dl, VT, V1,
5110 DAG.getConstant(Lane, MVT::i64));
5113 int Length = ShuffleMask.size();
5114 int V1EltNum = V1.getValueType().getVectorNumElements();
5116 // If the number of v1 elements is the same as the number of shuffle mask
5117 // element and the shuffle masks are sequential values, we can transform
5118 // it into NEON_VEXTRACT.
5119 if (V1EltNum == Length) {
5120 // Check if the shuffle mask is sequential.
5122 while (ShuffleMask[SkipUndef] == -1) {
5125 int CurMask = ShuffleMask[SkipUndef];
5126 if (CurMask >= SkipUndef) {
5127 bool IsSequential = true;
5128 for (int I = SkipUndef; I < Length; ++I) {
5129 if (ShuffleMask[I] != -1 && ShuffleMask[I] != CurMask) {
5130 IsSequential = false;
5136 assert((EltSize % 8 == 0) && "Bitsize of vector element is incorrect");
5137 unsigned VecSize = EltSize * V1EltNum;
5138 unsigned Index = (EltSize / 8) * (ShuffleMask[SkipUndef] - SkipUndef);
5139 if (VecSize == 64 || VecSize == 128)
5140 return DAG.getNode(AArch64ISD::NEON_VEXTRACT, dl, VT, V1, V2,
5141 DAG.getConstant(Index, MVT::i64));
5146 // For shuffle mask like "0, 1, 2, 3, 4, 5, 13, 7", try to generate insert
5147 // by element from V2 to V1 .
5148 // If shuffle mask is like "0, 1, 10, 11, 12, 13, 14, 15", V2 would be a
5149 // better choice to be inserted than V1 as less insert needed, so we count
5150 // element to be inserted for both V1 and V2, and select less one as insert
5153 // Collect elements need to be inserted and their index.
5154 SmallVector<int, 8> NV1Elt;
5155 SmallVector<int, 8> N1Index;
5156 SmallVector<int, 8> NV2Elt;
5157 SmallVector<int, 8> N2Index;
5158 for (int I = 0; I != Length; ++I) {
5159 if (ShuffleMask[I] != I) {
5160 NV1Elt.push_back(ShuffleMask[I]);
5161 N1Index.push_back(I);
5164 for (int I = 0; I != Length; ++I) {
5165 if (ShuffleMask[I] != (I + V1EltNum)) {
5166 NV2Elt.push_back(ShuffleMask[I]);
5167 N2Index.push_back(I);
5171 // Decide which to be inserted. If all lanes mismatch, neither V1 nor V2
5172 // will be inserted.
5174 SmallVector<int, 8> InsMasks = NV1Elt;
5175 SmallVector<int, 8> InsIndex = N1Index;
5176 if ((int)NV1Elt.size() != Length || (int)NV2Elt.size() != Length) {
5177 if (NV1Elt.size() > NV2Elt.size()) {
5183 InsV = DAG.getNode(ISD::UNDEF, dl, VT);
5186 for (int I = 0, E = InsMasks.size(); I != E; ++I) {
5188 int Mask = InsMasks[I];
5189 if (Mask >= V1EltNum) {
5193 // Any value type smaller than i32 is illegal in AArch64, and this lower
5194 // function is called after legalize pass, so we need to legalize
5197 if (VT.getVectorElementType().isFloatingPoint())
5198 EltVT = (EltSize == 64) ? MVT::f64 : MVT::f32;
5200 EltVT = (EltSize == 64) ? MVT::i64 : MVT::i32;
5203 ExtV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, ExtV,
5204 DAG.getConstant(Mask, MVT::i64));
5205 InsV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, InsV, ExtV,
5206 DAG.getConstant(InsIndex[I], MVT::i64));
5212 AArch64TargetLowering::ConstraintType
5213 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
5214 if (Constraint.size() == 1) {
5215 switch (Constraint[0]) {
5217 case 'w': // An FP/SIMD vector register
5218 return C_RegisterClass;
5219 case 'I': // Constant that can be used with an ADD instruction
5220 case 'J': // Constant that can be used with a SUB instruction
5221 case 'K': // Constant that can be used with a 32-bit logical instruction
5222 case 'L': // Constant that can be used with a 64-bit logical instruction
5223 case 'M': // Constant that can be used as a 32-bit MOV immediate
5224 case 'N': // Constant that can be used as a 64-bit MOV immediate
5225 case 'Y': // Floating point constant zero
5226 case 'Z': // Integer constant zero
5228 case 'Q': // A memory reference with base register and no offset
5230 case 'S': // A symbolic address
5235 // FIXME: Ump, Utf, Usa, Ush
5236 // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes,
5237 // whatever they may be
5238 // Utf: A memory address suitable for ldp/stp in TF mode, whatever it may be
5239 // Usa: An absolute symbolic address
5240 // Ush: The high part (bits 32:12) of a pc-relative symbolic address
5241 assert(Constraint != "Ump" && Constraint != "Utf" && Constraint != "Usa"
5242 && Constraint != "Ush" && "Unimplemented constraints");
5244 return TargetLowering::getConstraintType(Constraint);
5247 TargetLowering::ConstraintWeight
5248 AArch64TargetLowering::getSingleConstraintMatchWeight(AsmOperandInfo &Info,
5249 const char *Constraint) const {
5251 llvm_unreachable("Constraint weight unimplemented");
5255 AArch64TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
5256 std::string &Constraint,
5257 std::vector<SDValue> &Ops,
5258 SelectionDAG &DAG) const {
5261 // Only length 1 constraints are C_Other.
5262 if (Constraint.size() != 1) return;
5264 // Only C_Other constraints get lowered like this. That means constants for us
5265 // so return early if there's no hope the constraint can be lowered.
5267 switch(Constraint[0]) {
5269 case 'I': case 'J': case 'K': case 'L':
5270 case 'M': case 'N': case 'Z': {
5271 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
5275 uint64_t CVal = C->getZExtValue();
5278 switch (Constraint[0]) {
5280 // FIXME: 'M' and 'N' are MOV pseudo-insts -- unsupported in assembly. 'J'
5281 // is a peculiarly useless SUB constraint.
5282 llvm_unreachable("Unimplemented C_Other constraint");
5288 if (A64Imms::isLogicalImm(32, CVal, Bits))
5292 if (A64Imms::isLogicalImm(64, CVal, Bits))
5301 Result = DAG.getTargetConstant(CVal, Op.getValueType());
5305 // An absolute symbolic address or label reference.
5306 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
5307 Result = DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op),
5308 GA->getValueType(0));
5309 } else if (const BlockAddressSDNode *BA
5310 = dyn_cast<BlockAddressSDNode>(Op)) {
5311 Result = DAG.getTargetBlockAddress(BA->getBlockAddress(),
5312 BA->getValueType(0));
5313 } else if (const ExternalSymbolSDNode *ES
5314 = dyn_cast<ExternalSymbolSDNode>(Op)) {
5315 Result = DAG.getTargetExternalSymbol(ES->getSymbol(),
5316 ES->getValueType(0));
5322 if (const ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) {
5323 if (CFP->isExactlyValue(0.0)) {
5324 Result = DAG.getTargetConstantFP(0.0, CFP->getValueType(0));
5331 if (Result.getNode()) {
5332 Ops.push_back(Result);
5336 // It's an unknown constraint for us. Let generic code have a go.
5337 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
5340 std::pair<unsigned, const TargetRegisterClass*>
5341 AArch64TargetLowering::getRegForInlineAsmConstraint(
5342 const std::string &Constraint,
5344 if (Constraint.size() == 1) {
5345 switch (Constraint[0]) {
5347 if (VT.getSizeInBits() <= 32)
5348 return std::make_pair(0U, &AArch64::GPR32RegClass);
5349 else if (VT == MVT::i64)
5350 return std::make_pair(0U, &AArch64::GPR64RegClass);
5354 return std::make_pair(0U, &AArch64::FPR16RegClass);
5355 else if (VT == MVT::f32)
5356 return std::make_pair(0U, &AArch64::FPR32RegClass);
5357 else if (VT.getSizeInBits() == 64)
5358 return std::make_pair(0U, &AArch64::FPR64RegClass);
5359 else if (VT.getSizeInBits() == 128)
5360 return std::make_pair(0U, &AArch64::FPR128RegClass);
5365 // Use the default implementation in TargetLowering to convert the register
5366 // constraint into a member of a register class.
5367 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
5370 /// Represent NEON load and store intrinsics as MemIntrinsicNodes.
5371 /// The associated MachineMemOperands record the alignment specified
5372 /// in the intrinsic calls.
5373 bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
5375 unsigned Intrinsic) const {
5376 switch (Intrinsic) {
5377 case Intrinsic::arm_neon_vld1:
5378 case Intrinsic::arm_neon_vld2:
5379 case Intrinsic::arm_neon_vld3:
5380 case Intrinsic::arm_neon_vld4:
5381 case Intrinsic::aarch64_neon_vld1x2:
5382 case Intrinsic::aarch64_neon_vld1x3:
5383 case Intrinsic::aarch64_neon_vld1x4:
5384 case Intrinsic::arm_neon_vld2lane:
5385 case Intrinsic::arm_neon_vld3lane:
5386 case Intrinsic::arm_neon_vld4lane: {
5387 Info.opc = ISD::INTRINSIC_W_CHAIN;
5388 // Conservatively set memVT to the entire set of vectors loaded.
5389 uint64_t NumElts = getDataLayout()->getTypeAllocSize(I.getType()) / 8;
5390 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5391 Info.ptrVal = I.getArgOperand(0);
5393 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5394 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5395 Info.vol = false; // volatile loads with NEON intrinsics not supported
5396 Info.readMem = true;
5397 Info.writeMem = false;
5400 case Intrinsic::arm_neon_vst1:
5401 case Intrinsic::arm_neon_vst2:
5402 case Intrinsic::arm_neon_vst3:
5403 case Intrinsic::arm_neon_vst4:
5404 case Intrinsic::aarch64_neon_vst1x2:
5405 case Intrinsic::aarch64_neon_vst1x3:
5406 case Intrinsic::aarch64_neon_vst1x4:
5407 case Intrinsic::arm_neon_vst2lane:
5408 case Intrinsic::arm_neon_vst3lane:
5409 case Intrinsic::arm_neon_vst4lane: {
5410 Info.opc = ISD::INTRINSIC_VOID;
5411 // Conservatively set memVT to the entire set of vectors stored.
5412 unsigned NumElts = 0;
5413 for (unsigned ArgI = 1, ArgE = I.getNumArgOperands(); ArgI < ArgE; ++ArgI) {
5414 Type *ArgTy = I.getArgOperand(ArgI)->getType();
5415 if (!ArgTy->isVectorTy())
5417 NumElts += getDataLayout()->getTypeAllocSize(ArgTy) / 8;
5419 Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
5420 Info.ptrVal = I.getArgOperand(0);
5422 Value *AlignArg = I.getArgOperand(I.getNumArgOperands() - 1);
5423 Info.align = cast<ConstantInt>(AlignArg)->getZExtValue();
5424 Info.vol = false; // volatile stores with NEON intrinsics not supported
5425 Info.readMem = false;
5426 Info.writeMem = true;
5436 // Truncations from 64-bit GPR to 32-bit GPR is free.
5437 bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
5438 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5440 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5441 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5442 if (NumBits1 <= NumBits2)
5447 bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
5448 if (!VT1.isInteger() || !VT2.isInteger())
5450 unsigned NumBits1 = VT1.getSizeInBits();
5451 unsigned NumBits2 = VT2.getSizeInBits();
5452 if (NumBits1 <= NumBits2)
5457 // All 32-bit GPR operations implicitly zero the high-half of the corresponding
5459 bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
5460 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
5462 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
5463 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
5464 if (NumBits1 == 32 && NumBits2 == 64)
5469 bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
5470 if (!VT1.isInteger() || !VT2.isInteger())
5472 unsigned NumBits1 = VT1.getSizeInBits();
5473 unsigned NumBits2 = VT2.getSizeInBits();
5474 if (NumBits1 == 32 && NumBits2 == 64)
5479 bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
5480 EVT VT1 = Val.getValueType();
5481 if (isZExtFree(VT1, VT2)) {
5485 if (Val.getOpcode() != ISD::LOAD)
5488 // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
5489 return (VT1.isSimple() && VT1.isInteger() && VT2.isSimple() &&
5490 VT2.isInteger() && VT1.getSizeInBits() <= 32);
5493 // isLegalAddressingMode - Return true if the addressing mode represented
5494 /// by AM is legal for this target, for a load/store of the specified type.
5495 bool AArch64TargetLowering::isLegalAddressingMode(const AddrMode &AM,
5497 // AArch64 has five basic addressing modes:
5499 // reg + 9-bit signed offset
5500 // reg + SIZE_IN_BYTES * 12-bit unsigned offset
5502 // reg + SIZE_IN_BYTES * reg
5504 // No global is ever allowed as a base.
5508 // No reg+reg+imm addressing.
5509 if (AM.HasBaseReg && AM.BaseOffs && AM.Scale)
5512 // check reg + imm case:
5513 // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
5514 uint64_t NumBytes = 0;
5515 if (Ty->isSized()) {
5516 uint64_t NumBits = getDataLayout()->getTypeSizeInBits(Ty);
5517 NumBytes = NumBits / 8;
5518 if (!isPowerOf2_64(NumBits))
5523 int64_t Offset = AM.BaseOffs;
5525 // 9-bit signed offset
5526 if (Offset >= -(1LL << 9) && Offset <= (1LL << 9) - 1)
5529 // 12-bit unsigned offset
5530 unsigned shift = Log2_64(NumBytes);
5531 if (NumBytes && Offset > 0 && (Offset / NumBytes) <= (1LL << 12) - 1 &&
5532 // Must be a multiple of NumBytes (NumBytes is a power of 2)
5533 (Offset >> shift) << shift == Offset)
5537 if (!AM.Scale || AM.Scale == 1 ||
5538 (AM.Scale > 0 && (uint64_t)AM.Scale == NumBytes))
5543 int AArch64TargetLowering::getScalingFactorCost(const AddrMode &AM,
5545 // Scaling factors are not free at all.
5546 // Operands | Rt Latency
5547 // -------------------------------------------
5549 // -------------------------------------------
5550 // Rt, [Xn, Xm, lsl #imm] | Rn: 4 Rm: 5
5551 // Rt, [Xn, Wm, <extend> #imm] |
5552 if (isLegalAddressingMode(AM, Ty))
5553 // Scale represents reg2 * scale, thus account for 1 if
5554 // it is not equal to 0 or 1.
5555 return AM.Scale != 0 && AM.Scale != 1;
5559 /// getMaximalGlobalOffset - Returns the maximal possible offset which can
5560 /// be used for loads / stores from the global.
5561 unsigned AArch64TargetLowering::getMaximalGlobalOffset() const {