1 //===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that AArch64 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "aarch64-isel"
17 #include "AArch64ISelLowering.h"
18 #include "AArch64MachineFunctionInfo.h"
19 #include "AArch64TargetMachine.h"
20 #include "AArch64TargetObjectFile.h"
21 #include "Utils/AArch64BaseInfo.h"
22 #include "llvm/CodeGen/Analysis.h"
23 #include "llvm/CodeGen/CallingConvLower.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineInstrBuilder.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
28 #include "llvm/IR/CallingConv.h"
32 static TargetLoweringObjectFile *createTLOF(AArch64TargetMachine &TM) {
33 const AArch64Subtarget *Subtarget = &TM.getSubtarget<AArch64Subtarget>();
35 if (Subtarget->isTargetLinux())
36 return new AArch64LinuxTargetObjectFile();
37 if (Subtarget->isTargetELF())
38 return new TargetLoweringObjectFileELF();
39 llvm_unreachable("unknown subtarget type");
43 AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
44 : TargetLowering(TM, createTLOF(TM)),
45 Subtarget(&TM.getSubtarget<AArch64Subtarget>()),
46 RegInfo(TM.getRegisterInfo()),
47 Itins(TM.getInstrItineraryData()) {
49 // SIMD compares set the entire lane's bits to 1
50 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
52 // Scalar register <-> type mapping
53 addRegisterClass(MVT::i32, &AArch64::GPR32RegClass);
54 addRegisterClass(MVT::i64, &AArch64::GPR64RegClass);
55 addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
56 addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
57 addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
58 addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
60 computeRegisterProperties();
62 // We combine OR nodes for bitfield and NEON BSL operations.
63 setTargetDAGCombine(ISD::OR);
65 setTargetDAGCombine(ISD::AND);
66 setTargetDAGCombine(ISD::SRA);
68 // AArch64 does not have i1 loads, or much of anything for i1 really.
69 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
70 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
71 setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
73 setStackPointerRegisterToSaveRestore(AArch64::XSP);
74 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
75 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
76 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
78 // We'll lower globals to wrappers for selection.
79 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
80 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
82 // A64 instructions have the comparison predicate attached to the user of the
83 // result, but having a separate comparison is valuable for matching.
84 setOperationAction(ISD::BR_CC, MVT::i32, Custom);
85 setOperationAction(ISD::BR_CC, MVT::i64, Custom);
86 setOperationAction(ISD::BR_CC, MVT::f32, Custom);
87 setOperationAction(ISD::BR_CC, MVT::f64, Custom);
89 setOperationAction(ISD::SELECT, MVT::i32, Custom);
90 setOperationAction(ISD::SELECT, MVT::i64, Custom);
91 setOperationAction(ISD::SELECT, MVT::f32, Custom);
92 setOperationAction(ISD::SELECT, MVT::f64, Custom);
94 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
95 setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
96 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
97 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
99 setOperationAction(ISD::BRCOND, MVT::Other, Custom);
101 setOperationAction(ISD::SETCC, MVT::i32, Custom);
102 setOperationAction(ISD::SETCC, MVT::i64, Custom);
103 setOperationAction(ISD::SETCC, MVT::f32, Custom);
104 setOperationAction(ISD::SETCC, MVT::f64, Custom);
106 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
107 setOperationAction(ISD::JumpTable, MVT::i32, Custom);
108 setOperationAction(ISD::JumpTable, MVT::i64, Custom);
110 setOperationAction(ISD::VASTART, MVT::Other, Custom);
111 setOperationAction(ISD::VACOPY, MVT::Other, Custom);
112 setOperationAction(ISD::VAEND, MVT::Other, Expand);
113 setOperationAction(ISD::VAARG, MVT::Other, Expand);
115 setOperationAction(ISD::BlockAddress, MVT::i64, Custom);
117 setOperationAction(ISD::ROTL, MVT::i32, Expand);
118 setOperationAction(ISD::ROTL, MVT::i64, Expand);
120 setOperationAction(ISD::UREM, MVT::i32, Expand);
121 setOperationAction(ISD::UREM, MVT::i64, Expand);
122 setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
123 setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
125 setOperationAction(ISD::SREM, MVT::i32, Expand);
126 setOperationAction(ISD::SREM, MVT::i64, Expand);
127 setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
128 setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
130 setOperationAction(ISD::CTPOP, MVT::i32, Expand);
131 setOperationAction(ISD::CTPOP, MVT::i64, Expand);
133 // Legal floating-point operations.
134 setOperationAction(ISD::FABS, MVT::f32, Legal);
135 setOperationAction(ISD::FABS, MVT::f64, Legal);
137 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
138 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
140 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
141 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
143 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
144 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
146 setOperationAction(ISD::FNEG, MVT::f32, Legal);
147 setOperationAction(ISD::FNEG, MVT::f64, Legal);
149 setOperationAction(ISD::FRINT, MVT::f32, Legal);
150 setOperationAction(ISD::FRINT, MVT::f64, Legal);
152 setOperationAction(ISD::FSQRT, MVT::f32, Legal);
153 setOperationAction(ISD::FSQRT, MVT::f64, Legal);
155 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
156 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
158 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
159 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
160 setOperationAction(ISD::ConstantFP, MVT::f128, Legal);
162 // Illegal floating-point operations.
163 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
164 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
166 setOperationAction(ISD::FCOS, MVT::f32, Expand);
167 setOperationAction(ISD::FCOS, MVT::f64, Expand);
169 setOperationAction(ISD::FEXP, MVT::f32, Expand);
170 setOperationAction(ISD::FEXP, MVT::f64, Expand);
172 setOperationAction(ISD::FEXP2, MVT::f32, Expand);
173 setOperationAction(ISD::FEXP2, MVT::f64, Expand);
175 setOperationAction(ISD::FLOG, MVT::f32, Expand);
176 setOperationAction(ISD::FLOG, MVT::f64, Expand);
178 setOperationAction(ISD::FLOG2, MVT::f32, Expand);
179 setOperationAction(ISD::FLOG2, MVT::f64, Expand);
181 setOperationAction(ISD::FLOG10, MVT::f32, Expand);
182 setOperationAction(ISD::FLOG10, MVT::f64, Expand);
184 setOperationAction(ISD::FPOW, MVT::f32, Expand);
185 setOperationAction(ISD::FPOW, MVT::f64, Expand);
187 setOperationAction(ISD::FPOWI, MVT::f32, Expand);
188 setOperationAction(ISD::FPOWI, MVT::f64, Expand);
190 setOperationAction(ISD::FREM, MVT::f32, Expand);
191 setOperationAction(ISD::FREM, MVT::f64, Expand);
193 setOperationAction(ISD::FSIN, MVT::f32, Expand);
194 setOperationAction(ISD::FSIN, MVT::f64, Expand);
196 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
197 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
199 // Virtually no operation on f128 is legal, but LLVM can't expand them when
200 // there's a valid register class, so we need custom operations in most cases.
201 setOperationAction(ISD::FABS, MVT::f128, Expand);
202 setOperationAction(ISD::FADD, MVT::f128, Custom);
203 setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
204 setOperationAction(ISD::FCOS, MVT::f128, Expand);
205 setOperationAction(ISD::FDIV, MVT::f128, Custom);
206 setOperationAction(ISD::FMA, MVT::f128, Expand);
207 setOperationAction(ISD::FMUL, MVT::f128, Custom);
208 setOperationAction(ISD::FNEG, MVT::f128, Expand);
209 setOperationAction(ISD::FP_EXTEND, MVT::f128, Expand);
210 setOperationAction(ISD::FP_ROUND, MVT::f128, Expand);
211 setOperationAction(ISD::FPOW, MVT::f128, Expand);
212 setOperationAction(ISD::FREM, MVT::f128, Expand);
213 setOperationAction(ISD::FRINT, MVT::f128, Expand);
214 setOperationAction(ISD::FSIN, MVT::f128, Expand);
215 setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
216 setOperationAction(ISD::FSQRT, MVT::f128, Expand);
217 setOperationAction(ISD::FSUB, MVT::f128, Custom);
218 setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
219 setOperationAction(ISD::SETCC, MVT::f128, Custom);
220 setOperationAction(ISD::BR_CC, MVT::f128, Custom);
221 setOperationAction(ISD::SELECT, MVT::f128, Expand);
222 setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
223 setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
225 // Lowering for many of the conversions is actually specified by the non-f128
226 // type. The LowerXXX function will be trivial when f128 isn't involved.
227 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
228 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
229 setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
230 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
231 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
232 setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
233 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
234 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
235 setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
236 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
237 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
238 setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
239 setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
240 setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
242 // This prevents LLVM trying to compress double constants into a floating
243 // constant-pool entry and trying to load from there. It's of doubtful benefit
244 // for A64: we'd need LDR followed by FCVT, I believe.
245 setLoadExtAction(ISD::EXTLOAD, MVT::f64, Expand);
246 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
247 setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
249 setTruncStoreAction(MVT::f128, MVT::f64, Expand);
250 setTruncStoreAction(MVT::f128, MVT::f32, Expand);
251 setTruncStoreAction(MVT::f128, MVT::f16, Expand);
252 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
253 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
254 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
256 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
257 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
259 setExceptionPointerRegister(AArch64::X0);
260 setExceptionSelectorRegister(AArch64::X1);
263 EVT AArch64TargetLowering::getSetCCResultType(EVT VT) const {
264 // It's reasonably important that this value matches the "natural" legal
265 // promotion from i1 for scalar types. Otherwise LegalizeTypes can get itself
266 // in a twist (e.g. inserting an any_extend which then becomes i64 -> i64).
267 if (!VT.isVector()) return MVT::i32;
268 return VT.changeVectorElementTypeToInteger();
271 static void getExclusiveOperation(unsigned Size, AtomicOrdering Ord,
274 static unsigned LoadBares[] = {AArch64::LDXR_byte, AArch64::LDXR_hword,
275 AArch64::LDXR_word, AArch64::LDXR_dword};
276 static unsigned LoadAcqs[] = {AArch64::LDAXR_byte, AArch64::LDAXR_hword,
277 AArch64::LDAXR_word, AArch64::LDAXR_dword};
278 static unsigned StoreBares[] = {AArch64::STXR_byte, AArch64::STXR_hword,
279 AArch64::STXR_word, AArch64::STXR_dword};
280 static unsigned StoreRels[] = {AArch64::STLXR_byte, AArch64::STLXR_hword,
281 AArch64::STLXR_word, AArch64::STLXR_dword};
283 unsigned *LoadOps, *StoreOps;
284 if (Ord == Acquire || Ord == AcquireRelease || Ord == SequentiallyConsistent)
289 if (Ord == Release || Ord == AcquireRelease || Ord == SequentiallyConsistent)
290 StoreOps = StoreRels;
292 StoreOps = StoreBares;
294 assert(isPowerOf2_32(Size) && Size <= 8 &&
295 "unsupported size for atomic binary op!");
297 LdrOpc = LoadOps[Log2_32(Size)];
298 StrOpc = StoreOps[Log2_32(Size)];
302 AArch64TargetLowering::emitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
304 unsigned BinOpcode) const {
305 // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
308 const BasicBlock *LLVM_BB = BB->getBasicBlock();
309 MachineFunction *MF = BB->getParent();
310 MachineFunction::iterator It = BB;
313 unsigned dest = MI->getOperand(0).getReg();
314 unsigned ptr = MI->getOperand(1).getReg();
315 unsigned incr = MI->getOperand(2).getReg();
316 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
317 DebugLoc dl = MI->getDebugLoc();
319 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
321 unsigned ldrOpc, strOpc;
322 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
324 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
325 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
326 MF->insert(It, loopMBB);
327 MF->insert(It, exitMBB);
329 // Transfer the remainder of BB and its successor edges to exitMBB.
330 exitMBB->splice(exitMBB->begin(), BB,
331 llvm::next(MachineBasicBlock::iterator(MI)),
333 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
335 const TargetRegisterClass *TRC
336 = Size == 8 ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
337 unsigned scratch = (!BinOpcode) ? incr : MRI.createVirtualRegister(TRC);
341 // fallthrough --> loopMBB
342 BB->addSuccessor(loopMBB);
346 // <binop> scratch, dest, incr
347 // stxr stxr_status, scratch, ptr
348 // cbnz stxr_status, loopMBB
349 // fallthrough --> exitMBB
351 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
353 // All arithmetic operations we'll be creating are designed to take an extra
354 // shift or extend operand, which we can conveniently set to zero.
356 // Operand order needs to go the other way for NAND.
357 if (BinOpcode == AArch64::BICwww_lsl || BinOpcode == AArch64::BICxxx_lsl)
358 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
359 .addReg(incr).addReg(dest).addImm(0);
361 BuildMI(BB, dl, TII->get(BinOpcode), scratch)
362 .addReg(dest).addReg(incr).addImm(0);
365 // From the stxr, the register is GPR32; from the cmp it's GPR32wsp
366 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
367 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
369 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(scratch).addReg(ptr);
370 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
371 .addReg(stxr_status).addMBB(loopMBB);
373 BB->addSuccessor(loopMBB);
374 BB->addSuccessor(exitMBB);
380 MI->eraseFromParent(); // The instruction is gone now.
386 AArch64TargetLowering::emitAtomicBinaryMinMax(MachineInstr *MI,
387 MachineBasicBlock *BB,
390 A64CC::CondCodes Cond) const {
391 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
393 const BasicBlock *LLVM_BB = BB->getBasicBlock();
394 MachineFunction *MF = BB->getParent();
395 MachineFunction::iterator It = BB;
398 unsigned dest = MI->getOperand(0).getReg();
399 unsigned ptr = MI->getOperand(1).getReg();
400 unsigned incr = MI->getOperand(2).getReg();
401 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(3).getImm());
403 unsigned oldval = dest;
404 DebugLoc dl = MI->getDebugLoc();
406 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
407 const TargetRegisterClass *TRC, *TRCsp;
409 TRC = &AArch64::GPR64RegClass;
410 TRCsp = &AArch64::GPR64xspRegClass;
412 TRC = &AArch64::GPR32RegClass;
413 TRCsp = &AArch64::GPR32wspRegClass;
416 unsigned ldrOpc, strOpc;
417 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
419 MachineBasicBlock *loopMBB = MF->CreateMachineBasicBlock(LLVM_BB);
420 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
421 MF->insert(It, loopMBB);
422 MF->insert(It, exitMBB);
424 // Transfer the remainder of BB and its successor edges to exitMBB.
425 exitMBB->splice(exitMBB->begin(), BB,
426 llvm::next(MachineBasicBlock::iterator(MI)),
428 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
430 unsigned scratch = MRI.createVirtualRegister(TRC);
431 MRI.constrainRegClass(scratch, TRCsp);
435 // fallthrough --> loopMBB
436 BB->addSuccessor(loopMBB);
440 // cmp incr, dest (, sign extend if necessary)
441 // csel scratch, dest, incr, cond
442 // stxr stxr_status, scratch, ptr
443 // cbnz stxr_status, loopMBB
444 // fallthrough --> exitMBB
446 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
448 // Build compare and cmov instructions.
449 MRI.constrainRegClass(incr, TRCsp);
450 BuildMI(BB, dl, TII->get(CmpOp))
451 .addReg(incr).addReg(oldval).addImm(0);
453 BuildMI(BB, dl, TII->get(Size == 8 ? AArch64::CSELxxxc : AArch64::CSELwwwc),
455 .addReg(oldval).addReg(incr).addImm(Cond);
457 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
458 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
460 BuildMI(BB, dl, TII->get(strOpc), stxr_status)
461 .addReg(scratch).addReg(ptr);
462 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
463 .addReg(stxr_status).addMBB(loopMBB);
465 BB->addSuccessor(loopMBB);
466 BB->addSuccessor(exitMBB);
472 MI->eraseFromParent(); // The instruction is gone now.
478 AArch64TargetLowering::emitAtomicCmpSwap(MachineInstr *MI,
479 MachineBasicBlock *BB,
480 unsigned Size) const {
481 unsigned dest = MI->getOperand(0).getReg();
482 unsigned ptr = MI->getOperand(1).getReg();
483 unsigned oldval = MI->getOperand(2).getReg();
484 unsigned newval = MI->getOperand(3).getReg();
485 AtomicOrdering Ord = static_cast<AtomicOrdering>(MI->getOperand(4).getImm());
486 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
487 DebugLoc dl = MI->getDebugLoc();
489 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
490 const TargetRegisterClass *TRCsp;
491 TRCsp = Size == 8 ? &AArch64::GPR64xspRegClass : &AArch64::GPR32wspRegClass;
493 unsigned ldrOpc, strOpc;
494 getExclusiveOperation(Size, Ord, ldrOpc, strOpc);
496 MachineFunction *MF = BB->getParent();
497 const BasicBlock *LLVM_BB = BB->getBasicBlock();
498 MachineFunction::iterator It = BB;
499 ++It; // insert the new blocks after the current block
501 MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
502 MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
503 MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
504 MF->insert(It, loop1MBB);
505 MF->insert(It, loop2MBB);
506 MF->insert(It, exitMBB);
508 // Transfer the remainder of BB and its successor edges to exitMBB.
509 exitMBB->splice(exitMBB->begin(), BB,
510 llvm::next(MachineBasicBlock::iterator(MI)),
512 exitMBB->transferSuccessorsAndUpdatePHIs(BB);
516 // fallthrough --> loop1MBB
517 BB->addSuccessor(loop1MBB);
524 BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr);
526 unsigned CmpOp = Size == 8 ? AArch64::CMPxx_lsl : AArch64::CMPww_lsl;
527 MRI.constrainRegClass(dest, TRCsp);
528 BuildMI(BB, dl, TII->get(CmpOp))
529 .addReg(dest).addReg(oldval).addImm(0);
530 BuildMI(BB, dl, TII->get(AArch64::Bcc))
531 .addImm(A64CC::NE).addMBB(exitMBB);
532 BB->addSuccessor(loop2MBB);
533 BB->addSuccessor(exitMBB);
536 // strex stxr_status, newval, [ptr]
537 // cbnz stxr_status, loop1MBB
539 unsigned stxr_status = MRI.createVirtualRegister(&AArch64::GPR32RegClass);
540 MRI.constrainRegClass(stxr_status, &AArch64::GPR32wspRegClass);
542 BuildMI(BB, dl, TII->get(strOpc), stxr_status).addReg(newval).addReg(ptr);
543 BuildMI(BB, dl, TII->get(AArch64::CBNZw))
544 .addReg(stxr_status).addMBB(loop1MBB);
545 BB->addSuccessor(loop1MBB);
546 BB->addSuccessor(exitMBB);
552 MI->eraseFromParent(); // The instruction is gone now.
558 AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
559 MachineBasicBlock *MBB) const {
560 // We materialise the F128CSEL pseudo-instruction using conditional branches
561 // and loads, giving an instruciton sequence like:
570 // Using virtual registers would probably not be beneficial since COPY
571 // instructions are expensive for f128 (there's no actual instruction to
574 // An alternative would be to do an integer-CSEL on some address. E.g.:
579 // csel x0, x0, x1, ne
582 // It's unclear which approach is actually optimal.
583 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
584 MachineFunction *MF = MBB->getParent();
585 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
586 DebugLoc DL = MI->getDebugLoc();
587 MachineFunction::iterator It = MBB;
590 unsigned DestReg = MI->getOperand(0).getReg();
591 unsigned IfTrueReg = MI->getOperand(1).getReg();
592 unsigned IfFalseReg = MI->getOperand(2).getReg();
593 unsigned CondCode = MI->getOperand(3).getImm();
594 bool NZCVKilled = MI->getOperand(4).isKill();
596 MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
597 MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
598 MF->insert(It, TrueBB);
599 MF->insert(It, EndBB);
601 // Transfer rest of current basic-block to EndBB
602 EndBB->splice(EndBB->begin(), MBB,
603 llvm::next(MachineBasicBlock::iterator(MI)),
605 EndBB->transferSuccessorsAndUpdatePHIs(MBB);
607 // We need somewhere to store the f128 value needed.
608 int ScratchFI = MF->getFrameInfo()->CreateSpillStackObject(16, 16);
610 // [... start of incoming MBB ...]
611 // str qIFFALSE, [sp]
614 BuildMI(MBB, DL, TII->get(AArch64::LSFP128_STR))
616 .addFrameIndex(ScratchFI)
618 BuildMI(MBB, DL, TII->get(AArch64::Bcc))
621 BuildMI(MBB, DL, TII->get(AArch64::Bimm))
623 MBB->addSuccessor(TrueBB);
624 MBB->addSuccessor(EndBB);
628 BuildMI(TrueBB, DL, TII->get(AArch64::LSFP128_STR))
630 .addFrameIndex(ScratchFI)
633 // Note: fallthrough. We can rely on LLVM adding a branch if it reorders the
635 TrueBB->addSuccessor(EndBB);
639 // [... rest of incoming MBB ...]
641 EndBB->addLiveIn(AArch64::NZCV);
642 MachineInstr *StartOfEnd = EndBB->begin();
643 BuildMI(*EndBB, StartOfEnd, DL, TII->get(AArch64::LSFP128_LDR), DestReg)
644 .addFrameIndex(ScratchFI)
647 MI->eraseFromParent();
652 AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
653 MachineBasicBlock *MBB) const {
654 switch (MI->getOpcode()) {
655 default: llvm_unreachable("Unhandled instruction with custom inserter");
656 case AArch64::F128CSEL:
657 return EmitF128CSEL(MI, MBB);
658 case AArch64::ATOMIC_LOAD_ADD_I8:
659 return emitAtomicBinary(MI, MBB, 1, AArch64::ADDwww_lsl);
660 case AArch64::ATOMIC_LOAD_ADD_I16:
661 return emitAtomicBinary(MI, MBB, 2, AArch64::ADDwww_lsl);
662 case AArch64::ATOMIC_LOAD_ADD_I32:
663 return emitAtomicBinary(MI, MBB, 4, AArch64::ADDwww_lsl);
664 case AArch64::ATOMIC_LOAD_ADD_I64:
665 return emitAtomicBinary(MI, MBB, 8, AArch64::ADDxxx_lsl);
667 case AArch64::ATOMIC_LOAD_SUB_I8:
668 return emitAtomicBinary(MI, MBB, 1, AArch64::SUBwww_lsl);
669 case AArch64::ATOMIC_LOAD_SUB_I16:
670 return emitAtomicBinary(MI, MBB, 2, AArch64::SUBwww_lsl);
671 case AArch64::ATOMIC_LOAD_SUB_I32:
672 return emitAtomicBinary(MI, MBB, 4, AArch64::SUBwww_lsl);
673 case AArch64::ATOMIC_LOAD_SUB_I64:
674 return emitAtomicBinary(MI, MBB, 8, AArch64::SUBxxx_lsl);
676 case AArch64::ATOMIC_LOAD_AND_I8:
677 return emitAtomicBinary(MI, MBB, 1, AArch64::ANDwww_lsl);
678 case AArch64::ATOMIC_LOAD_AND_I16:
679 return emitAtomicBinary(MI, MBB, 2, AArch64::ANDwww_lsl);
680 case AArch64::ATOMIC_LOAD_AND_I32:
681 return emitAtomicBinary(MI, MBB, 4, AArch64::ANDwww_lsl);
682 case AArch64::ATOMIC_LOAD_AND_I64:
683 return emitAtomicBinary(MI, MBB, 8, AArch64::ANDxxx_lsl);
685 case AArch64::ATOMIC_LOAD_OR_I8:
686 return emitAtomicBinary(MI, MBB, 1, AArch64::ORRwww_lsl);
687 case AArch64::ATOMIC_LOAD_OR_I16:
688 return emitAtomicBinary(MI, MBB, 2, AArch64::ORRwww_lsl);
689 case AArch64::ATOMIC_LOAD_OR_I32:
690 return emitAtomicBinary(MI, MBB, 4, AArch64::ORRwww_lsl);
691 case AArch64::ATOMIC_LOAD_OR_I64:
692 return emitAtomicBinary(MI, MBB, 8, AArch64::ORRxxx_lsl);
694 case AArch64::ATOMIC_LOAD_XOR_I8:
695 return emitAtomicBinary(MI, MBB, 1, AArch64::EORwww_lsl);
696 case AArch64::ATOMIC_LOAD_XOR_I16:
697 return emitAtomicBinary(MI, MBB, 2, AArch64::EORwww_lsl);
698 case AArch64::ATOMIC_LOAD_XOR_I32:
699 return emitAtomicBinary(MI, MBB, 4, AArch64::EORwww_lsl);
700 case AArch64::ATOMIC_LOAD_XOR_I64:
701 return emitAtomicBinary(MI, MBB, 8, AArch64::EORxxx_lsl);
703 case AArch64::ATOMIC_LOAD_NAND_I8:
704 return emitAtomicBinary(MI, MBB, 1, AArch64::BICwww_lsl);
705 case AArch64::ATOMIC_LOAD_NAND_I16:
706 return emitAtomicBinary(MI, MBB, 2, AArch64::BICwww_lsl);
707 case AArch64::ATOMIC_LOAD_NAND_I32:
708 return emitAtomicBinary(MI, MBB, 4, AArch64::BICwww_lsl);
709 case AArch64::ATOMIC_LOAD_NAND_I64:
710 return emitAtomicBinary(MI, MBB, 8, AArch64::BICxxx_lsl);
712 case AArch64::ATOMIC_LOAD_MIN_I8:
713 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::GT);
714 case AArch64::ATOMIC_LOAD_MIN_I16:
715 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::GT);
716 case AArch64::ATOMIC_LOAD_MIN_I32:
717 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::GT);
718 case AArch64::ATOMIC_LOAD_MIN_I64:
719 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::GT);
721 case AArch64::ATOMIC_LOAD_MAX_I8:
722 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_sxtb, A64CC::LT);
723 case AArch64::ATOMIC_LOAD_MAX_I16:
724 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_sxth, A64CC::LT);
725 case AArch64::ATOMIC_LOAD_MAX_I32:
726 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LT);
727 case AArch64::ATOMIC_LOAD_MAX_I64:
728 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LT);
730 case AArch64::ATOMIC_LOAD_UMIN_I8:
731 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::HI);
732 case AArch64::ATOMIC_LOAD_UMIN_I16:
733 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::HI);
734 case AArch64::ATOMIC_LOAD_UMIN_I32:
735 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::HI);
736 case AArch64::ATOMIC_LOAD_UMIN_I64:
737 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::HI);
739 case AArch64::ATOMIC_LOAD_UMAX_I8:
740 return emitAtomicBinaryMinMax(MI, MBB, 1, AArch64::CMPww_uxtb, A64CC::LO);
741 case AArch64::ATOMIC_LOAD_UMAX_I16:
742 return emitAtomicBinaryMinMax(MI, MBB, 2, AArch64::CMPww_uxth, A64CC::LO);
743 case AArch64::ATOMIC_LOAD_UMAX_I32:
744 return emitAtomicBinaryMinMax(MI, MBB, 4, AArch64::CMPww_lsl, A64CC::LO);
745 case AArch64::ATOMIC_LOAD_UMAX_I64:
746 return emitAtomicBinaryMinMax(MI, MBB, 8, AArch64::CMPxx_lsl, A64CC::LO);
748 case AArch64::ATOMIC_SWAP_I8:
749 return emitAtomicBinary(MI, MBB, 1, 0);
750 case AArch64::ATOMIC_SWAP_I16:
751 return emitAtomicBinary(MI, MBB, 2, 0);
752 case AArch64::ATOMIC_SWAP_I32:
753 return emitAtomicBinary(MI, MBB, 4, 0);
754 case AArch64::ATOMIC_SWAP_I64:
755 return emitAtomicBinary(MI, MBB, 8, 0);
757 case AArch64::ATOMIC_CMP_SWAP_I8:
758 return emitAtomicCmpSwap(MI, MBB, 1);
759 case AArch64::ATOMIC_CMP_SWAP_I16:
760 return emitAtomicCmpSwap(MI, MBB, 2);
761 case AArch64::ATOMIC_CMP_SWAP_I32:
762 return emitAtomicCmpSwap(MI, MBB, 4);
763 case AArch64::ATOMIC_CMP_SWAP_I64:
764 return emitAtomicCmpSwap(MI, MBB, 8);
769 const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
771 case AArch64ISD::BR_CC: return "AArch64ISD::BR_CC";
772 case AArch64ISD::Call: return "AArch64ISD::Call";
773 case AArch64ISD::FPMOV: return "AArch64ISD::FPMOV";
774 case AArch64ISD::GOTLoad: return "AArch64ISD::GOTLoad";
775 case AArch64ISD::BFI: return "AArch64ISD::BFI";
776 case AArch64ISD::EXTR: return "AArch64ISD::EXTR";
777 case AArch64ISD::Ret: return "AArch64ISD::Ret";
778 case AArch64ISD::SBFX: return "AArch64ISD::SBFX";
779 case AArch64ISD::SELECT_CC: return "AArch64ISD::SELECT_CC";
780 case AArch64ISD::SETCC: return "AArch64ISD::SETCC";
781 case AArch64ISD::TC_RETURN: return "AArch64ISD::TC_RETURN";
782 case AArch64ISD::THREAD_POINTER: return "AArch64ISD::THREAD_POINTER";
783 case AArch64ISD::TLSDESCCALL: return "AArch64ISD::TLSDESCCALL";
784 case AArch64ISD::WrapperLarge: return "AArch64ISD::WrapperLarge";
785 case AArch64ISD::WrapperSmall: return "AArch64ISD::WrapperSmall";
787 default: return NULL;
791 static const uint16_t AArch64FPRArgRegs[] = {
792 AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
793 AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7
795 static const unsigned NumFPRArgRegs = llvm::array_lengthof(AArch64FPRArgRegs);
797 static const uint16_t AArch64ArgRegs[] = {
798 AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3,
799 AArch64::X4, AArch64::X5, AArch64::X6, AArch64::X7
801 static const unsigned NumArgRegs = llvm::array_lengthof(AArch64ArgRegs);
803 static bool CC_AArch64NoMoreRegs(unsigned ValNo, MVT ValVT, MVT LocVT,
804 CCValAssign::LocInfo LocInfo,
805 ISD::ArgFlagsTy ArgFlags, CCState &State) {
806 // Mark all remaining general purpose registers as allocated. We don't
807 // backtrack: if (for example) an i128 gets put on the stack, no subsequent
808 // i64 will go in registers (C.11).
809 for (unsigned i = 0; i < NumArgRegs; ++i)
810 State.AllocateReg(AArch64ArgRegs[i]);
815 #include "AArch64GenCallingConv.inc"
817 CCAssignFn *AArch64TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
820 default: llvm_unreachable("Unsupported calling convention");
821 case CallingConv::Fast:
828 AArch64TargetLowering::SaveVarArgRegisters(CCState &CCInfo, SelectionDAG &DAG,
829 DebugLoc DL, SDValue &Chain) const {
830 MachineFunction &MF = DAG.getMachineFunction();
831 MachineFrameInfo *MFI = MF.getFrameInfo();
832 AArch64MachineFunctionInfo *FuncInfo
833 = MF.getInfo<AArch64MachineFunctionInfo>();
835 SmallVector<SDValue, 8> MemOps;
837 unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(AArch64ArgRegs,
839 unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(AArch64FPRArgRegs,
842 unsigned GPRSaveSize = 8 * (NumArgRegs - FirstVariadicGPR);
844 if (GPRSaveSize != 0) {
845 GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);
847 SDValue FIN = DAG.getFrameIndex(GPRIdx, getPointerTy());
849 for (unsigned i = FirstVariadicGPR; i < NumArgRegs; ++i) {
850 unsigned VReg = MF.addLiveIn(AArch64ArgRegs[i], &AArch64::GPR64RegClass);
851 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
852 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
853 MachinePointerInfo::getStack(i * 8),
855 MemOps.push_back(Store);
856 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
857 DAG.getConstant(8, getPointerTy()));
861 unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
863 if (FPRSaveSize != 0) {
864 FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);
866 SDValue FIN = DAG.getFrameIndex(FPRIdx, getPointerTy());
868 for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
869 unsigned VReg = MF.addLiveIn(AArch64FPRArgRegs[i],
870 &AArch64::FPR128RegClass);
871 SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);
872 SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
873 MachinePointerInfo::getStack(i * 16),
875 MemOps.push_back(Store);
876 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(), FIN,
877 DAG.getConstant(16, getPointerTy()));
881 int StackIdx = MFI->CreateFixedObject(8, CCInfo.getNextStackOffset(), true);
883 FuncInfo->setVariadicStackIdx(StackIdx);
884 FuncInfo->setVariadicGPRIdx(GPRIdx);
885 FuncInfo->setVariadicGPRSize(GPRSaveSize);
886 FuncInfo->setVariadicFPRIdx(FPRIdx);
887 FuncInfo->setVariadicFPRSize(FPRSaveSize);
889 if (!MemOps.empty()) {
890 Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
897 AArch64TargetLowering::LowerFormalArguments(SDValue Chain,
898 CallingConv::ID CallConv, bool isVarArg,
899 const SmallVectorImpl<ISD::InputArg> &Ins,
900 DebugLoc dl, SelectionDAG &DAG,
901 SmallVectorImpl<SDValue> &InVals) const {
902 MachineFunction &MF = DAG.getMachineFunction();
903 AArch64MachineFunctionInfo *FuncInfo
904 = MF.getInfo<AArch64MachineFunctionInfo>();
905 MachineFrameInfo *MFI = MF.getFrameInfo();
906 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
908 SmallVector<CCValAssign, 16> ArgLocs;
909 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
910 getTargetMachine(), ArgLocs, *DAG.getContext());
911 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
913 SmallVector<SDValue, 16> ArgValues;
916 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
917 CCValAssign &VA = ArgLocs[i];
918 ISD::ArgFlagsTy Flags = Ins[i].Flags;
920 if (Flags.isByVal()) {
921 // Byval is used for small structs and HFAs in the PCS, but the system
922 // should work in a non-compliant manner for larger structs.
923 EVT PtrTy = getPointerTy();
924 int Size = Flags.getByValSize();
925 unsigned NumRegs = (Size + 7) / 8;
927 unsigned FrameIdx = MFI->CreateFixedObject(8 * NumRegs,
928 VA.getLocMemOffset(),
930 SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrTy);
931 InVals.push_back(FrameIdxN);
934 } else if (VA.isRegLoc()) {
935 MVT RegVT = VA.getLocVT();
936 const TargetRegisterClass *RC = getRegClassFor(RegVT);
937 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
939 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
940 } else { // VA.isRegLoc()
941 assert(VA.isMemLoc());
943 int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
944 VA.getLocMemOffset(), true);
946 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
947 ArgValue = DAG.getLoad(VA.getLocVT(), dl, Chain, FIN,
948 MachinePointerInfo::getFixedStack(FI),
949 false, false, false, 0);
954 switch (VA.getLocInfo()) {
955 default: llvm_unreachable("Unknown loc info!");
956 case CCValAssign::Full: break;
957 case CCValAssign::BCvt:
958 ArgValue = DAG.getNode(ISD::BITCAST,dl, VA.getValVT(), ArgValue);
960 case CCValAssign::SExt:
961 case CCValAssign::ZExt:
962 case CCValAssign::AExt: {
963 unsigned DestSize = VA.getValVT().getSizeInBits();
967 case 8: DestSubReg = AArch64::sub_8; break;
968 case 16: DestSubReg = AArch64::sub_16; break;
969 case 32: DestSubReg = AArch64::sub_32; break;
970 case 64: DestSubReg = AArch64::sub_64; break;
971 default: llvm_unreachable("Unexpected argument promotion");
974 ArgValue = SDValue(DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, dl,
975 VA.getValVT(), ArgValue,
976 DAG.getTargetConstant(DestSubReg, MVT::i32)),
982 InVals.push_back(ArgValue);
986 SaveVarArgRegisters(CCInfo, DAG, dl, Chain);
988 unsigned StackArgSize = CCInfo.getNextStackOffset();
989 if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
990 // This is a non-standard ABI so by fiat I say we're allowed to make full
991 // use of the stack area to be popped, which must be aligned to 16 bytes in
993 StackArgSize = RoundUpToAlignment(StackArgSize, 16);
995 // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
997 FuncInfo->setArgumentStackToRestore(StackArgSize);
999 // This realignment carries over to the available bytes below. Our own
1000 // callers will guarantee the space is free by giving an aligned value to
1003 // Even if we're not expected to free up the space, it's useful to know how
1004 // much is there while considering tail calls (because we can reuse it).
1005 FuncInfo->setBytesInStackArgArea(StackArgSize);
1011 AArch64TargetLowering::LowerReturn(SDValue Chain,
1012 CallingConv::ID CallConv, bool isVarArg,
1013 const SmallVectorImpl<ISD::OutputArg> &Outs,
1014 const SmallVectorImpl<SDValue> &OutVals,
1015 DebugLoc dl, SelectionDAG &DAG) const {
1016 // CCValAssign - represent the assignment of the return value to a location.
1017 SmallVector<CCValAssign, 16> RVLocs;
1019 // CCState - Info about the registers and stack slots.
1020 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1021 getTargetMachine(), RVLocs, *DAG.getContext());
1023 // Analyze outgoing return values.
1024 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv));
1027 SmallVector<SDValue, 4> RetOps(1, Chain);
1029 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
1030 // PCS: "If the type, T, of the result of a function is such that
1031 // void func(T arg) would require that arg be passed as a value in a
1032 // register (or set of registers) according to the rules in 5.4, then the
1033 // result is returned in the same registers as would be used for such an
1036 // Otherwise, the caller shall reserve a block of memory of sufficient
1037 // size and alignment to hold the result. The address of the memory block
1038 // shall be passed as an additional argument to the function in x8."
1040 // This is implemented in two places. The register-return values are dealt
1041 // with here, more complex returns are passed as an sret parameter, which
1042 // means we don't have to worry about it during actual return.
1043 CCValAssign &VA = RVLocs[i];
1044 assert(VA.isRegLoc() && "Only register-returns should be created by PCS");
1047 SDValue Arg = OutVals[i];
1049 // There's no convenient note in the ABI about this as there is for normal
1050 // arguments, but it says return values are passed in the same registers as
1051 // an argument would be. I believe that includes the comments about
1052 // unspecified higher bits, putting the burden of widening on the *caller*
1053 // for return values.
1054 switch (VA.getLocInfo()) {
1055 default: llvm_unreachable("Unknown loc info");
1056 case CCValAssign::Full: break;
1057 case CCValAssign::SExt:
1058 case CCValAssign::ZExt:
1059 case CCValAssign::AExt:
1060 // Floating-point values should only be extended when they're going into
1061 // memory, which can't happen here so an integer extend is acceptable.
1062 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1064 case CCValAssign::BCvt:
1065 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1069 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);
1070 Flag = Chain.getValue(1);
1071 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
1074 RetOps[0] = Chain; // Update chain.
1076 // Add the flag if we have it.
1078 RetOps.push_back(Flag);
1080 return DAG.getNode(AArch64ISD::Ret, dl, MVT::Other,
1081 &RetOps[0], RetOps.size());
1085 AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
1086 SmallVectorImpl<SDValue> &InVals) const {
1087 SelectionDAG &DAG = CLI.DAG;
1088 DebugLoc &dl = CLI.DL;
1089 SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
1090 SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
1091 SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
1092 SDValue Chain = CLI.Chain;
1093 SDValue Callee = CLI.Callee;
1094 bool &IsTailCall = CLI.IsTailCall;
1095 CallingConv::ID CallConv = CLI.CallConv;
1096 bool IsVarArg = CLI.IsVarArg;
1098 MachineFunction &MF = DAG.getMachineFunction();
1099 AArch64MachineFunctionInfo *FuncInfo
1100 = MF.getInfo<AArch64MachineFunctionInfo>();
1101 bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
1102 bool IsStructRet = !Outs.empty() && Outs[0].Flags.isSRet();
1103 bool IsSibCall = false;
1106 IsTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1107 IsVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1108 Outs, OutVals, Ins, DAG);
1110 // A sibling call is one where we're under the usual C ABI and not planning
1111 // to change that but can still do a tail call:
1112 if (!TailCallOpt && IsTailCall)
1116 SmallVector<CCValAssign, 16> ArgLocs;
1117 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1118 getTargetMachine(), ArgLocs, *DAG.getContext());
1119 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1121 // On AArch64 (and all other architectures I'm aware of) the most this has to
1122 // do is adjust the stack pointer.
1123 unsigned NumBytes = RoundUpToAlignment(CCInfo.getNextStackOffset(), 16);
1125 // Since we're not changing the ABI to make this a tail call, the memory
1126 // operands are already available in the caller's incoming argument space.
1130 // FPDiff is the byte offset of the call's argument area from the callee's.
1131 // Stores to callee stack arguments will be placed in FixedStackSlots offset
1132 // by this amount for a tail call. In a sibling call it must be 0 because the
1133 // caller will deallocate the entire stack and the callee still expects its
1134 // arguments to begin at SP+0. Completely unused for non-tail calls.
1137 if (IsTailCall && !IsSibCall) {
1138 unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();
1140 // FPDiff will be negative if this tail call requires more space than we
1141 // would automatically have in our incoming argument space. Positive if we
1142 // can actually shrink the stack.
1143 FPDiff = NumReusableBytes - NumBytes;
1145 // The stack pointer must be 16-byte aligned at all times it's used for a
1146 // memory operation, which in practice means at *all* times and in
1147 // particular across call boundaries. Therefore our own arguments started at
1148 // a 16-byte aligned SP and the delta applied for the tail call should
1149 // satisfy the same constraint.
1150 assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
1154 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1156 SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, AArch64::XSP,
1159 SmallVector<SDValue, 8> MemOpChains;
1160 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1162 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1163 CCValAssign &VA = ArgLocs[i];
1164 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1165 SDValue Arg = OutVals[i];
1167 // Callee does the actual widening, so all extensions just use an implicit
1168 // definition of the rest of the Loc. Aesthetically, this would be nicer as
1169 // an ANY_EXTEND, but that isn't valid for floating-point types and this
1170 // alternative works on integer types too.
1171 switch (VA.getLocInfo()) {
1172 default: llvm_unreachable("Unknown loc info!");
1173 case CCValAssign::Full: break;
1174 case CCValAssign::SExt:
1175 case CCValAssign::ZExt:
1176 case CCValAssign::AExt: {
1177 unsigned SrcSize = VA.getValVT().getSizeInBits();
1181 case 8: SrcSubReg = AArch64::sub_8; break;
1182 case 16: SrcSubReg = AArch64::sub_16; break;
1183 case 32: SrcSubReg = AArch64::sub_32; break;
1184 case 64: SrcSubReg = AArch64::sub_64; break;
1185 default: llvm_unreachable("Unexpected argument promotion");
1188 Arg = SDValue(DAG.getMachineNode(TargetOpcode::INSERT_SUBREG, dl,
1190 DAG.getUNDEF(VA.getLocVT()),
1192 DAG.getTargetConstant(SrcSubReg, MVT::i32)),
1197 case CCValAssign::BCvt:
1198 Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
1202 if (VA.isRegLoc()) {
1203 // A normal register (sub-) argument. For now we just note it down because
1204 // we want to copy things into registers as late as possible to avoid
1205 // register-pressure (and possibly worse).
1206 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1210 assert(VA.isMemLoc() && "unexpected argument location");
1213 MachinePointerInfo DstInfo;
1215 uint32_t OpSize = Flags.isByVal() ? Flags.getByValSize() :
1216 VA.getLocVT().getSizeInBits();
1217 OpSize = (OpSize + 7) / 8;
1218 int32_t Offset = VA.getLocMemOffset() + FPDiff;
1219 int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
1221 DstAddr = DAG.getFrameIndex(FI, getPointerTy());
1222 DstInfo = MachinePointerInfo::getFixedStack(FI);
1224 // Make sure any stack arguments overlapping with where we're storing are
1225 // loaded before this eventual operation. Otherwise they'll be clobbered.
1226 Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
1228 SDValue PtrOff = DAG.getIntPtrConstant(VA.getLocMemOffset());
1230 DstAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1231 DstInfo = MachinePointerInfo::getStack(VA.getLocMemOffset());
1234 if (Flags.isByVal()) {
1235 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i64);
1236 SDValue Cpy = DAG.getMemcpy(Chain, dl, DstAddr, Arg, SizeNode,
1237 Flags.getByValAlign(),
1238 /*isVolatile = */ false,
1239 /*alwaysInline = */ false,
1240 DstInfo, MachinePointerInfo(0));
1241 MemOpChains.push_back(Cpy);
1243 // Normal stack argument, put it where it's needed.
1244 SDValue Store = DAG.getStore(Chain, dl, Arg, DstAddr, DstInfo,
1246 MemOpChains.push_back(Store);
1250 // The loads and stores generated above shouldn't clash with each
1251 // other. Combining them with this TokenFactor notes that fact for the rest of
1253 if (!MemOpChains.empty())
1254 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1255 &MemOpChains[0], MemOpChains.size());
1257 // Most of the rest of the instructions need to be glued together; we don't
1258 // want assignments to actual registers used by a call to be rearranged by a
1259 // well-meaning scheduler.
1262 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1263 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1264 RegsToPass[i].second, InFlag);
1265 InFlag = Chain.getValue(1);
1268 // The linker is responsible for inserting veneers when necessary to put a
1269 // function call destination in range, so we don't need to bother with a
1271 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1272 const GlobalValue *GV = G->getGlobal();
1273 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
1274 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1275 const char *Sym = S->getSymbol();
1276 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy());
1279 // We don't usually want to end the call-sequence here because we would tidy
1280 // the frame up *after* the call, however in the ABI-changing tail-call case
1281 // we've carefully laid out the parameters so that when sp is reset they'll be
1282 // in the correct location.
1283 if (IsTailCall && !IsSibCall) {
1284 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1285 DAG.getIntPtrConstant(0, true), InFlag);
1286 InFlag = Chain.getValue(1);
1289 // We produce the following DAG scheme for the actual call instruction:
1290 // (AArch64Call Chain, Callee, reg1, ..., regn, preserveMask, inflag?
1292 // Most arguments aren't going to be used and just keep the values live as
1293 // far as LLVM is concerned. It's expected to be selected as simply "bl
1294 // callee" (for a direct, non-tail call).
1295 std::vector<SDValue> Ops;
1296 Ops.push_back(Chain);
1297 Ops.push_back(Callee);
1300 // Each tail call may have to adjust the stack by a different amount, so
1301 // this information must travel along with the operation for eventual
1302 // consumption by emitEpilogue.
1303 Ops.push_back(DAG.getTargetConstant(FPDiff, MVT::i32));
1306 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1307 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1308 RegsToPass[i].second.getValueType()));
1311 // Add a register mask operand representing the call-preserved registers. This
1312 // is used later in codegen to constrain register-allocation.
1313 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
1314 const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
1315 assert(Mask && "Missing call preserved mask for calling convention");
1316 Ops.push_back(DAG.getRegisterMask(Mask));
1318 // If we needed glue, put it in as the last argument.
1319 if (InFlag.getNode())
1320 Ops.push_back(InFlag);
1322 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1325 return DAG.getNode(AArch64ISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());
1328 Chain = DAG.getNode(AArch64ISD::Call, dl, NodeTys, &Ops[0], Ops.size());
1329 InFlag = Chain.getValue(1);
1331 // Now we can reclaim the stack, just as well do it before working out where
1332 // our return value is.
1334 uint64_t CalleePopBytes
1335 = DoesCalleeRestoreStack(CallConv, TailCallOpt) ? NumBytes : 0;
1337 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1338 DAG.getIntPtrConstant(CalleePopBytes, true),
1340 InFlag = Chain.getValue(1);
1343 return LowerCallResult(Chain, InFlag, CallConv,
1344 IsVarArg, Ins, dl, DAG, InVals);
1348 AArch64TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1349 CallingConv::ID CallConv, bool IsVarArg,
1350 const SmallVectorImpl<ISD::InputArg> &Ins,
1351 DebugLoc dl, SelectionDAG &DAG,
1352 SmallVectorImpl<SDValue> &InVals) const {
1353 // Assign locations to each value returned by this call.
1354 SmallVector<CCValAssign, 16> RVLocs;
1355 CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(),
1356 getTargetMachine(), RVLocs, *DAG.getContext());
1357 CCInfo.AnalyzeCallResult(Ins, CCAssignFnForNode(CallConv));
1359 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1360 CCValAssign VA = RVLocs[i];
1362 // Return values that are too big to fit into registers should use an sret
1363 // pointer, so this can be a lot simpler than the main argument code.
1364 assert(VA.isRegLoc() && "Memory locations not expected for call return");
1366 SDValue Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
1368 Chain = Val.getValue(1);
1369 InFlag = Val.getValue(2);
1371 switch (VA.getLocInfo()) {
1372 default: llvm_unreachable("Unknown loc info!");
1373 case CCValAssign::Full: break;
1374 case CCValAssign::BCvt:
1375 Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
1377 case CCValAssign::ZExt:
1378 case CCValAssign::SExt:
1379 case CCValAssign::AExt:
1380 // Floating-point arguments only get extended/truncated if they're going
1381 // in memory, so using the integer operation is acceptable here.
1382 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
1386 InVals.push_back(Val);
1393 AArch64TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
1394 CallingConv::ID CalleeCC,
1396 bool IsCalleeStructRet,
1397 bool IsCallerStructRet,
1398 const SmallVectorImpl<ISD::OutputArg> &Outs,
1399 const SmallVectorImpl<SDValue> &OutVals,
1400 const SmallVectorImpl<ISD::InputArg> &Ins,
1401 SelectionDAG& DAG) const {
1403 // For CallingConv::C this function knows whether the ABI needs
1404 // changing. That's not true for other conventions so they will have to opt in
1406 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1409 const MachineFunction &MF = DAG.getMachineFunction();
1410 const Function *CallerF = MF.getFunction();
1411 CallingConv::ID CallerCC = CallerF->getCallingConv();
1412 bool CCMatch = CallerCC == CalleeCC;
1414 // Byval parameters hand the function a pointer directly into the stack area
1415 // we want to reuse during a tail call. Working around this *is* possible (see
1416 // X86) but less efficient and uglier in LowerCall.
1417 for (Function::const_arg_iterator i = CallerF->arg_begin(),
1418 e = CallerF->arg_end(); i != e; ++i)
1419 if (i->hasByValAttr())
1422 if (getTargetMachine().Options.GuaranteedTailCallOpt) {
1423 if (IsTailCallConvention(CalleeCC) && CCMatch)
1428 // Now we search for cases where we can use a tail call without changing the
1429 // ABI. Sibcall is used in some places (particularly gcc) to refer to this
1432 // I want anyone implementing a new calling convention to think long and hard
1433 // about this assert.
1434 assert((!IsVarArg || CalleeCC == CallingConv::C)
1435 && "Unexpected variadic calling convention");
1437 if (IsVarArg && !Outs.empty()) {
1438 // At least two cases here: if caller is fastcc then we can't have any
1439 // memory arguments (we'd be expected to clean up the stack afterwards). If
1440 // caller is C then we could potentially use its argument area.
1442 // FIXME: for now we take the most conservative of these in both cases:
1443 // disallow all variadic memory operands.
1444 SmallVector<CCValAssign, 16> ArgLocs;
1445 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1446 getTargetMachine(), ArgLocs, *DAG.getContext());
1448 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1449 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
1450 if (!ArgLocs[i].isRegLoc())
1454 // If the calling conventions do not match, then we'd better make sure the
1455 // results are returned in the same way as what the caller expects.
1457 SmallVector<CCValAssign, 16> RVLocs1;
1458 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
1459 getTargetMachine(), RVLocs1, *DAG.getContext());
1460 CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC));
1462 SmallVector<CCValAssign, 16> RVLocs2;
1463 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
1464 getTargetMachine(), RVLocs2, *DAG.getContext());
1465 CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC));
1467 if (RVLocs1.size() != RVLocs2.size())
1469 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
1470 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
1472 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
1474 if (RVLocs1[i].isRegLoc()) {
1475 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
1478 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
1484 // Nothing more to check if the callee is taking no arguments
1488 SmallVector<CCValAssign, 16> ArgLocs;
1489 CCState CCInfo(CalleeCC, IsVarArg, DAG.getMachineFunction(),
1490 getTargetMachine(), ArgLocs, *DAG.getContext());
1492 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
1494 const AArch64MachineFunctionInfo *FuncInfo
1495 = MF.getInfo<AArch64MachineFunctionInfo>();
1497 // If the stack arguments for this call would fit into our own save area then
1498 // the call can be made tail.
1499 return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
1502 bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
1503 bool TailCallOpt) const {
1504 return CallCC == CallingConv::Fast && TailCallOpt;
1507 bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
1508 return CallCC == CallingConv::Fast;
1511 SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
1513 MachineFrameInfo *MFI,
1514 int ClobberedFI) const {
1515 SmallVector<SDValue, 8> ArgChains;
1516 int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
1517 int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;
1519 // Include the original chain at the beginning of the list. When this is
1520 // used by target LowerCall hooks, this helps legalize find the
1521 // CALLSEQ_BEGIN node.
1522 ArgChains.push_back(Chain);
1524 // Add a chain value for each stack argument corresponding
1525 for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
1526 UE = DAG.getEntryNode().getNode()->use_end(); U != UE; ++U)
1527 if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
1528 if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
1529 if (FI->getIndex() < 0) {
1530 int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
1531 int64_t InLastByte = InFirstByte;
1532 InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;
1534 if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
1535 (FirstByte <= InFirstByte && InFirstByte <= LastByte))
1536 ArgChains.push_back(SDValue(L, 1));
1539 // Build a tokenfactor for all the chains.
1540 return DAG.getNode(ISD::TokenFactor, Chain.getDebugLoc(), MVT::Other,
1541 &ArgChains[0], ArgChains.size());
1544 static A64CC::CondCodes IntCCToA64CC(ISD::CondCode CC) {
1546 case ISD::SETEQ: return A64CC::EQ;
1547 case ISD::SETGT: return A64CC::GT;
1548 case ISD::SETGE: return A64CC::GE;
1549 case ISD::SETLT: return A64CC::LT;
1550 case ISD::SETLE: return A64CC::LE;
1551 case ISD::SETNE: return A64CC::NE;
1552 case ISD::SETUGT: return A64CC::HI;
1553 case ISD::SETUGE: return A64CC::HS;
1554 case ISD::SETULT: return A64CC::LO;
1555 case ISD::SETULE: return A64CC::LS;
1556 default: llvm_unreachable("Unexpected condition code");
1560 bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Val) const {
1561 // icmp is implemented using adds/subs immediate, which take an unsigned
1562 // 12-bit immediate, optionally shifted left by 12 bits.
1564 // Symmetric by using adds/subs
1568 return (Val & ~0xfff) == 0 || (Val & ~0xfff000) == 0;
1571 SDValue AArch64TargetLowering::getSelectableIntSetCC(SDValue LHS, SDValue RHS,
1572 ISD::CondCode CC, SDValue &A64cc,
1573 SelectionDAG &DAG, DebugLoc &dl) const {
1574 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
1576 EVT VT = RHSC->getValueType(0);
1577 bool knownInvalid = false;
1579 // I'm not convinced the rest of LLVM handles these edge cases properly, but
1580 // we can at least get it right.
1581 if (isSignedIntSetCC(CC)) {
1582 C = RHSC->getSExtValue();
1583 } else if (RHSC->getZExtValue() > INT64_MAX) {
1584 // A 64-bit constant not representable by a signed 64-bit integer is far
1585 // too big to fit into a SUBS immediate anyway.
1586 knownInvalid = true;
1588 C = RHSC->getZExtValue();
1591 if (!knownInvalid && !isLegalICmpImmediate(C)) {
1592 // Constant does not fit, try adjusting it by one?
1597 if (isLegalICmpImmediate(C-1)) {
1598 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
1599 RHS = DAG.getConstant(C-1, VT);
1604 if (isLegalICmpImmediate(C-1)) {
1605 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
1606 RHS = DAG.getConstant(C-1, VT);
1611 if (isLegalICmpImmediate(C+1)) {
1612 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
1613 RHS = DAG.getConstant(C+1, VT);
1618 if (isLegalICmpImmediate(C+1)) {
1619 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
1620 RHS = DAG.getConstant(C+1, VT);
1627 A64CC::CondCodes CondCode = IntCCToA64CC(CC);
1628 A64cc = DAG.getConstant(CondCode, MVT::i32);
1629 return DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
1630 DAG.getCondCode(CC));
1633 static A64CC::CondCodes FPCCToA64CC(ISD::CondCode CC,
1634 A64CC::CondCodes &Alternative) {
1635 A64CC::CondCodes CondCode = A64CC::Invalid;
1636 Alternative = A64CC::Invalid;
1639 default: llvm_unreachable("Unknown FP condition!");
1641 case ISD::SETOEQ: CondCode = A64CC::EQ; break;
1643 case ISD::SETOGT: CondCode = A64CC::GT; break;
1645 case ISD::SETOGE: CondCode = A64CC::GE; break;
1646 case ISD::SETOLT: CondCode = A64CC::MI; break;
1647 case ISD::SETOLE: CondCode = A64CC::LS; break;
1648 case ISD::SETONE: CondCode = A64CC::MI; Alternative = A64CC::GT; break;
1649 case ISD::SETO: CondCode = A64CC::VC; break;
1650 case ISD::SETUO: CondCode = A64CC::VS; break;
1651 case ISD::SETUEQ: CondCode = A64CC::EQ; Alternative = A64CC::VS; break;
1652 case ISD::SETUGT: CondCode = A64CC::HI; break;
1653 case ISD::SETUGE: CondCode = A64CC::PL; break;
1655 case ISD::SETULT: CondCode = A64CC::LT; break;
1657 case ISD::SETULE: CondCode = A64CC::LE; break;
1659 case ISD::SETUNE: CondCode = A64CC::NE; break;
1665 AArch64TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
1666 DebugLoc DL = Op.getDebugLoc();
1667 EVT PtrVT = getPointerTy();
1668 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
1670 switch(getTargetMachine().getCodeModel()) {
1671 case CodeModel::Small:
1672 // The most efficient code is PC-relative anyway for the small memory model,
1673 // so we don't need to worry about relocation model.
1674 return DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
1675 DAG.getTargetBlockAddress(BA, PtrVT, 0,
1676 AArch64II::MO_NO_FLAG),
1677 DAG.getTargetBlockAddress(BA, PtrVT, 0,
1678 AArch64II::MO_LO12),
1679 DAG.getConstant(/*Alignment=*/ 4, MVT::i32));
1680 case CodeModel::Large:
1682 AArch64ISD::WrapperLarge, DL, PtrVT,
1683 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G3),
1684 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
1685 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
1686 DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
1688 llvm_unreachable("Only small and large code models supported now");
1693 // (BRCOND chain, val, dest)
1695 AArch64TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
1696 DebugLoc dl = Op.getDebugLoc();
1697 SDValue Chain = Op.getOperand(0);
1698 SDValue TheBit = Op.getOperand(1);
1699 SDValue DestBB = Op.getOperand(2);
1701 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
1702 // that as the consumer we are responsible for ignoring rubbish in higher
1704 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
1705 DAG.getConstant(1, MVT::i32));
1707 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
1708 DAG.getConstant(0, TheBit.getValueType()),
1709 DAG.getCondCode(ISD::SETNE));
1711 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other, Chain,
1712 A64CMP, DAG.getConstant(A64CC::NE, MVT::i32),
1716 // (BR_CC chain, condcode, lhs, rhs, dest)
1718 AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
1719 DebugLoc dl = Op.getDebugLoc();
1720 SDValue Chain = Op.getOperand(0);
1721 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
1722 SDValue LHS = Op.getOperand(2);
1723 SDValue RHS = Op.getOperand(3);
1724 SDValue DestBB = Op.getOperand(4);
1726 if (LHS.getValueType() == MVT::f128) {
1727 // f128 comparisons are lowered to runtime calls by a routine which sets
1728 // LHS, RHS and CC appropriately for the rest of this function to continue.
1729 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
1731 // If softenSetCCOperands returned a scalar, we need to compare the result
1732 // against zero to select between true and false values.
1733 if (RHS.getNode() == 0) {
1734 RHS = DAG.getConstant(0, LHS.getValueType());
1739 if (LHS.getValueType().isInteger()) {
1742 // Integers are handled in a separate function because the combinations of
1743 // immediates and tests can get hairy and we may want to fiddle things.
1744 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
1746 return DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
1747 Chain, CmpOp, A64cc, DestBB);
1750 // Note that some LLVM floating-point CondCodes can't be lowered to a single
1751 // conditional branch, hence FPCCToA64CC can set a second test, where either
1752 // passing is sufficient.
1753 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
1754 CondCode = FPCCToA64CC(CC, Alternative);
1755 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
1756 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
1757 DAG.getCondCode(CC));
1758 SDValue A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
1759 Chain, SetCC, A64cc, DestBB);
1761 if (Alternative != A64CC::Invalid) {
1762 A64cc = DAG.getConstant(Alternative, MVT::i32);
1763 A64BR_CC = DAG.getNode(AArch64ISD::BR_CC, dl, MVT::Other,
1764 A64BR_CC, SetCC, A64cc, DestBB);
1772 AArch64TargetLowering::LowerF128ToCall(SDValue Op, SelectionDAG &DAG,
1773 RTLIB::Libcall Call) const {
1776 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
1777 EVT ArgVT = Op.getOperand(i).getValueType();
1778 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
1779 Entry.Node = Op.getOperand(i); Entry.Ty = ArgTy;
1780 Entry.isSExt = false;
1781 Entry.isZExt = false;
1782 Args.push_back(Entry);
1784 SDValue Callee = DAG.getExternalSymbol(getLibcallName(Call), getPointerTy());
1786 Type *RetTy = Op.getValueType().getTypeForEVT(*DAG.getContext());
1788 // By default, the input chain to this libcall is the entry node of the
1789 // function. If the libcall is going to be emitted as a tail call then
1790 // isUsedByReturnOnly will change it to the right chain if the return
1791 // node which is being folded has a non-entry input chain.
1792 SDValue InChain = DAG.getEntryNode();
1794 // isTailCall may be true since the callee does not reference caller stack
1795 // frame. Check if it's in the right position.
1796 SDValue TCChain = InChain;
1797 bool isTailCall = isInTailCallPosition(DAG, Op.getNode(), TCChain);
1802 CallLoweringInfo CLI(InChain, RetTy, false, false, false, false,
1803 0, getLibcallCallingConv(Call), isTailCall,
1804 /*doesNotReturn=*/false, /*isReturnValueUsed=*/true,
1805 Callee, Args, DAG, Op->getDebugLoc());
1806 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
1808 if (!CallInfo.second.getNode())
1809 // It's a tailcall, return the chain (which is the DAG root).
1810 return DAG.getRoot();
1812 return CallInfo.first;
1816 AArch64TargetLowering::LowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
1817 if (Op.getOperand(0).getValueType() != MVT::f128) {
1818 // It's legal except when f128 is involved
1823 LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());
1825 SDValue SrcVal = Op.getOperand(0);
1826 return makeLibCall(DAG, LC, Op.getValueType(), &SrcVal, 1,
1827 /*isSigned*/ false, Op.getDebugLoc());
1831 AArch64TargetLowering::LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) const {
1832 assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
1835 LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());
1837 return LowerF128ToCall(Op, DAG, LC);
1841 AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
1842 bool IsSigned) const {
1843 if (Op.getOperand(0).getValueType() != MVT::f128) {
1844 // It's legal except when f128 is involved
1850 LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
1852 LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());
1854 return LowerF128ToCall(Op, DAG, LC);
1858 AArch64TargetLowering::LowerGlobalAddressELFLarge(SDValue Op,
1859 SelectionDAG &DAG) const {
1860 assert(getTargetMachine().getCodeModel() == CodeModel::Large);
1861 assert(getTargetMachine().getRelocationModel() == Reloc::Static);
1863 EVT PtrVT = getPointerTy();
1864 DebugLoc dl = Op.getDebugLoc();
1865 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
1866 const GlobalValue *GV = GN->getGlobal();
1868 SDValue GlobalAddr = DAG.getNode(
1869 AArch64ISD::WrapperLarge, dl, PtrVT,
1870 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G3),
1871 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G2_NC),
1872 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G1_NC),
1873 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, AArch64II::MO_ABS_G0_NC));
1875 if (GN->getOffset() != 0)
1876 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
1877 DAG.getConstant(GN->getOffset(), PtrVT));
1883 AArch64TargetLowering::LowerGlobalAddressELFSmall(SDValue Op,
1884 SelectionDAG &DAG) const {
1885 assert(getTargetMachine().getCodeModel() == CodeModel::Small);
1887 EVT PtrVT = getPointerTy();
1888 DebugLoc dl = Op.getDebugLoc();
1889 const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
1890 const GlobalValue *GV = GN->getGlobal();
1891 unsigned Alignment = GV->getAlignment();
1892 Reloc::Model RelocM = getTargetMachine().getRelocationModel();
1893 if (GV->isWeakForLinker() && GV->isDeclaration() && RelocM == Reloc::Static) {
1894 // Weak undefined symbols can't use ADRP/ADD pair since they should evaluate
1895 // to zero when they remain undefined. In PIC mode the GOT can take care of
1896 // this, but in absolute mode we use a constant pool load.
1898 PoolAddr = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
1899 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
1900 AArch64II::MO_NO_FLAG),
1901 DAG.getTargetConstantPool(GV, PtrVT, 0, 0,
1902 AArch64II::MO_LO12),
1903 DAG.getConstant(8, MVT::i32));
1904 SDValue GlobalAddr = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), PoolAddr,
1905 MachinePointerInfo::getConstantPool(),
1906 /*isVolatile=*/ false,
1907 /*isNonTemporal=*/ true,
1908 /*isInvariant=*/ true, 8);
1909 if (GN->getOffset() != 0)
1910 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalAddr,
1911 DAG.getConstant(GN->getOffset(), PtrVT));
1916 if (Alignment == 0) {
1917 const PointerType *GVPtrTy = cast<PointerType>(GV->getType());
1918 if (GVPtrTy->getElementType()->isSized()) {
1920 = getDataLayout()->getABITypeAlignment(GVPtrTy->getElementType());
1922 // Be conservative if we can't guess, not that it really matters:
1923 // functions and labels aren't valid for loads, and the methods used to
1924 // actually calculate an address work with any alignment.
1929 unsigned char HiFixup, LoFixup;
1930 bool UseGOT = Subtarget->GVIsIndirectSymbol(GV, RelocM);
1933 HiFixup = AArch64II::MO_GOT;
1934 LoFixup = AArch64II::MO_GOT_LO12;
1937 HiFixup = AArch64II::MO_NO_FLAG;
1938 LoFixup = AArch64II::MO_LO12;
1941 // AArch64's small model demands the following sequence:
1942 // ADRP x0, somewhere
1943 // ADD x0, x0, #:lo12:somewhere ; (or LDR directly).
1944 SDValue GlobalRef = DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
1945 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1947 DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0,
1949 DAG.getConstant(Alignment, MVT::i32));
1952 GlobalRef = DAG.getNode(AArch64ISD::GOTLoad, dl, PtrVT, DAG.getEntryNode(),
1956 if (GN->getOffset() != 0)
1957 return DAG.getNode(ISD::ADD, dl, PtrVT, GlobalRef,
1958 DAG.getConstant(GN->getOffset(), PtrVT));
1964 AArch64TargetLowering::LowerGlobalAddressELF(SDValue Op,
1965 SelectionDAG &DAG) const {
1966 // TableGen doesn't have easy access to the CodeModel or RelocationModel, so
1967 // we make those distinctions here.
1969 switch (getTargetMachine().getCodeModel()) {
1970 case CodeModel::Small:
1971 return LowerGlobalAddressELFSmall(Op, DAG);
1972 case CodeModel::Large:
1973 return LowerGlobalAddressELFLarge(Op, DAG);
1975 llvm_unreachable("Only small and large code models supported now");
1979 SDValue AArch64TargetLowering::LowerTLSDescCall(SDValue SymAddr,
1982 SelectionDAG &DAG) const {
1983 EVT PtrVT = getPointerTy();
1985 // The function we need to call is simply the first entry in the GOT for this
1986 // descriptor, load it in preparation.
1987 SDValue Func, Chain;
1988 Func = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
1991 // The function takes only one argument: the address of the descriptor itself
1994 Chain = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X0, DescAddr, Glue);
1995 Glue = Chain.getValue(1);
1997 // Finally, there's a special calling-convention which means that the lookup
1998 // must preserve all registers (except X0, obviously).
1999 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
2000 const AArch64RegisterInfo *A64RI
2001 = static_cast<const AArch64RegisterInfo *>(TRI);
2002 const uint32_t *Mask = A64RI->getTLSDescCallPreservedMask();
2004 // We're now ready to populate the argument list, as with a normal call:
2005 std::vector<SDValue> Ops;
2006 Ops.push_back(Chain);
2007 Ops.push_back(Func);
2008 Ops.push_back(SymAddr);
2009 Ops.push_back(DAG.getRegister(AArch64::X0, PtrVT));
2010 Ops.push_back(DAG.getRegisterMask(Mask));
2011 Ops.push_back(Glue);
2013 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2014 Chain = DAG.getNode(AArch64ISD::TLSDESCCALL, DL, NodeTys, &Ops[0],
2016 Glue = Chain.getValue(1);
2018 // After the call, the offset from TPIDR_EL0 is in X0, copy it out and pass it
2019 // back to the generic handling code.
2020 return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
2024 AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
2025 SelectionDAG &DAG) const {
2026 assert(Subtarget->isTargetELF() &&
2027 "TLS not implemented for non-ELF targets");
2028 const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
2030 TLSModel::Model Model = getTargetMachine().getTLSModel(GA->getGlobal());
2033 EVT PtrVT = getPointerTy();
2034 DebugLoc DL = Op.getDebugLoc();
2035 const GlobalValue *GV = GA->getGlobal();
2037 SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
2039 if (Model == TLSModel::InitialExec) {
2040 TPOff = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2041 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2042 AArch64II::MO_GOTTPREL),
2043 DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2044 AArch64II::MO_GOTTPREL_LO12),
2045 DAG.getConstant(8, MVT::i32));
2046 TPOff = DAG.getNode(AArch64ISD::GOTLoad, DL, PtrVT, DAG.getEntryNode(),
2048 } else if (Model == TLSModel::LocalExec) {
2049 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2050 AArch64II::MO_TPREL_G1);
2051 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2052 AArch64II::MO_TPREL_G0_NC);
2054 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2055 DAG.getTargetConstant(0, MVT::i32)), 0);
2056 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2058 DAG.getTargetConstant(0, MVT::i32)), 0);
2059 } else if (Model == TLSModel::GeneralDynamic) {
2060 // Accesses used in this sequence go via the TLS descriptor which lives in
2061 // the GOT. Prepare an address we can use to handle this.
2062 SDValue HiDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2063 AArch64II::MO_TLSDESC);
2064 SDValue LoDesc = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
2065 AArch64II::MO_TLSDESC_LO12);
2066 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2068 DAG.getConstant(8, MVT::i32));
2069 SDValue SymAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0);
2071 TPOff = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2072 } else if (Model == TLSModel::LocalDynamic) {
2073 // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
2074 // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
2075 // the beginning of the module's TLS region, followed by a DTPREL offset
2078 // These accesses will need deduplicating if there's more than one.
2079 AArch64MachineFunctionInfo* MFI = DAG.getMachineFunction()
2080 .getInfo<AArch64MachineFunctionInfo>();
2081 MFI->incNumLocalDynamicTLSAccesses();
2084 // Get the location of _TLS_MODULE_BASE_:
2085 SDValue HiDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2086 AArch64II::MO_TLSDESC);
2087 SDValue LoDesc = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
2088 AArch64II::MO_TLSDESC_LO12);
2089 SDValue DescAddr = DAG.getNode(AArch64ISD::WrapperSmall, DL, PtrVT,
2091 DAG.getConstant(8, MVT::i32));
2092 SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT);
2094 ThreadBase = LowerTLSDescCall(SymAddr, DescAddr, DL, DAG);
2096 // Get the variable's offset from _TLS_MODULE_BASE_
2097 SDValue HiVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2098 AArch64II::MO_DTPREL_G1);
2099 SDValue LoVar = DAG.getTargetGlobalAddress(GV, DL, MVT::i64, 0,
2100 AArch64II::MO_DTPREL_G0_NC);
2102 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZxii, DL, PtrVT, HiVar,
2103 DAG.getTargetConstant(0, MVT::i32)), 0);
2104 TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKxii, DL, PtrVT,
2106 DAG.getTargetConstant(0, MVT::i32)), 0);
2108 llvm_unreachable("Unsupported TLS access model");
2111 return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
2115 AArch64TargetLowering::LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG,
2116 bool IsSigned) const {
2117 if (Op.getValueType() != MVT::f128) {
2118 // Legal for everything except f128.
2124 LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2126 LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
2128 return LowerF128ToCall(Op, DAG, LC);
2133 AArch64TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
2134 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
2135 DebugLoc dl = JT->getDebugLoc();
2136 EVT PtrVT = getPointerTy();
2138 // When compiling PIC, jump tables get put in the code section so a static
2139 // relocation-style is acceptable for both cases.
2140 switch (getTargetMachine().getCodeModel()) {
2141 case CodeModel::Small:
2142 return DAG.getNode(AArch64ISD::WrapperSmall, dl, PtrVT,
2143 DAG.getTargetJumpTable(JT->getIndex(), PtrVT),
2144 DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
2145 AArch64II::MO_LO12),
2146 DAG.getConstant(1, MVT::i32));
2147 case CodeModel::Large:
2149 AArch64ISD::WrapperLarge, dl, PtrVT,
2150 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G3),
2151 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G2_NC),
2152 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G1_NC),
2153 DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_ABS_G0_NC));
2155 llvm_unreachable("Only small and large code models supported now");
2159 // (SELECT_CC lhs, rhs, iftrue, iffalse, condcode)
2161 AArch64TargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
2162 DebugLoc dl = Op.getDebugLoc();
2163 SDValue LHS = Op.getOperand(0);
2164 SDValue RHS = Op.getOperand(1);
2165 SDValue IfTrue = Op.getOperand(2);
2166 SDValue IfFalse = Op.getOperand(3);
2167 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
2169 if (LHS.getValueType() == MVT::f128) {
2170 // f128 comparisons are lowered to libcalls, but slot in nicely here
2172 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2174 // If softenSetCCOperands returned a scalar, we need to compare the result
2175 // against zero to select between true and false values.
2176 if (RHS.getNode() == 0) {
2177 RHS = DAG.getConstant(0, LHS.getValueType());
2182 if (LHS.getValueType().isInteger()) {
2185 // Integers are handled in a separate function because the combinations of
2186 // immediates and tests can get hairy and we may want to fiddle things.
2187 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2189 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2190 CmpOp, IfTrue, IfFalse, A64cc);
2193 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2194 // conditional branch, hence FPCCToA64CC can set a second test, where either
2195 // passing is sufficient.
2196 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2197 CondCode = FPCCToA64CC(CC, Alternative);
2198 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2199 SDValue SetCC = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2200 DAG.getCondCode(CC));
2201 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl,
2203 SetCC, IfTrue, IfFalse, A64cc);
2205 if (Alternative != A64CC::Invalid) {
2206 A64cc = DAG.getConstant(Alternative, MVT::i32);
2207 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2208 SetCC, IfTrue, A64SELECT_CC, A64cc);
2212 return A64SELECT_CC;
2215 // (SELECT testbit, iftrue, iffalse)
2217 AArch64TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
2218 DebugLoc dl = Op.getDebugLoc();
2219 SDValue TheBit = Op.getOperand(0);
2220 SDValue IfTrue = Op.getOperand(1);
2221 SDValue IfFalse = Op.getOperand(2);
2223 // AArch64 BooleanContents is the default UndefinedBooleanContent, which means
2224 // that as the consumer we are responsible for ignoring rubbish in higher
2226 TheBit = DAG.getNode(ISD::AND, dl, MVT::i32, TheBit,
2227 DAG.getConstant(1, MVT::i32));
2228 SDValue A64CMP = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, TheBit,
2229 DAG.getConstant(0, TheBit.getValueType()),
2230 DAG.getCondCode(ISD::SETNE));
2232 return DAG.getNode(AArch64ISD::SELECT_CC, dl, Op.getValueType(),
2233 A64CMP, IfTrue, IfFalse,
2234 DAG.getConstant(A64CC::NE, MVT::i32));
2237 // (SETCC lhs, rhs, condcode)
2239 AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
2240 DebugLoc dl = Op.getDebugLoc();
2241 SDValue LHS = Op.getOperand(0);
2242 SDValue RHS = Op.getOperand(1);
2243 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
2244 EVT VT = Op.getValueType();
2246 if (LHS.getValueType() == MVT::f128) {
2247 // f128 comparisons will be lowered to libcalls giving a valid LHS and RHS
2248 // for the rest of the function (some i32 or i64 values).
2249 softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);
2251 // If softenSetCCOperands returned a scalar, use it.
2252 if (RHS.getNode() == 0) {
2253 assert(LHS.getValueType() == Op.getValueType() &&
2254 "Unexpected setcc expansion!");
2259 if (LHS.getValueType().isInteger()) {
2262 // Integers are handled in a separate function because the combinations of
2263 // immediates and tests can get hairy and we may want to fiddle things.
2264 SDValue CmpOp = getSelectableIntSetCC(LHS, RHS, CC, A64cc, DAG, dl);
2266 return DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
2267 CmpOp, DAG.getConstant(1, VT), DAG.getConstant(0, VT),
2271 // Note that some LLVM floating-point CondCodes can't be lowered to a single
2272 // conditional branch, hence FPCCToA64CC can set a second test, where either
2273 // passing is sufficient.
2274 A64CC::CondCodes CondCode, Alternative = A64CC::Invalid;
2275 CondCode = FPCCToA64CC(CC, Alternative);
2276 SDValue A64cc = DAG.getConstant(CondCode, MVT::i32);
2277 SDValue CmpOp = DAG.getNode(AArch64ISD::SETCC, dl, MVT::i32, LHS, RHS,
2278 DAG.getCondCode(CC));
2279 SDValue A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT,
2280 CmpOp, DAG.getConstant(1, VT),
2281 DAG.getConstant(0, VT), A64cc);
2283 if (Alternative != A64CC::Invalid) {
2284 A64cc = DAG.getConstant(Alternative, MVT::i32);
2285 A64SELECT_CC = DAG.getNode(AArch64ISD::SELECT_CC, dl, VT, CmpOp,
2286 DAG.getConstant(1, VT), A64SELECT_CC, A64cc);
2289 return A64SELECT_CC;
2293 AArch64TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
2294 const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
2295 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
2297 // We have to make sure we copy the entire structure: 8+8+8+4+4 = 32 bytes
2298 // rather than just 8.
2299 return DAG.getMemcpy(Op.getOperand(0), Op.getDebugLoc(),
2300 Op.getOperand(1), Op.getOperand(2),
2301 DAG.getConstant(32, MVT::i32), 8, false, false,
2302 MachinePointerInfo(DestSV), MachinePointerInfo(SrcSV));
2306 AArch64TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
2307 // The layout of the va_list struct is specified in the AArch64 Procedure Call
2308 // Standard, section B.3.
2309 MachineFunction &MF = DAG.getMachineFunction();
2310 AArch64MachineFunctionInfo *FuncInfo
2311 = MF.getInfo<AArch64MachineFunctionInfo>();
2312 DebugLoc DL = Op.getDebugLoc();
2314 SDValue Chain = Op.getOperand(0);
2315 SDValue VAList = Op.getOperand(1);
2316 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
2317 SmallVector<SDValue, 4> MemOps;
2319 // void *__stack at offset 0
2320 SDValue Stack = DAG.getFrameIndex(FuncInfo->getVariadicStackIdx(),
2322 MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
2323 MachinePointerInfo(SV), false, false, 0));
2325 // void *__gr_top at offset 8
2326 int GPRSize = FuncInfo->getVariadicGPRSize();
2328 SDValue GRTop, GRTopAddr;
2330 GRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
2331 DAG.getConstant(8, getPointerTy()));
2333 GRTop = DAG.getFrameIndex(FuncInfo->getVariadicGPRIdx(), getPointerTy());
2334 GRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), GRTop,
2335 DAG.getConstant(GPRSize, getPointerTy()));
2337 MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
2338 MachinePointerInfo(SV, 8),
2342 // void *__vr_top at offset 16
2343 int FPRSize = FuncInfo->getVariadicFPRSize();
2345 SDValue VRTop, VRTopAddr;
2346 VRTopAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
2347 DAG.getConstant(16, getPointerTy()));
2349 VRTop = DAG.getFrameIndex(FuncInfo->getVariadicFPRIdx(), getPointerTy());
2350 VRTop = DAG.getNode(ISD::ADD, DL, getPointerTy(), VRTop,
2351 DAG.getConstant(FPRSize, getPointerTy()));
2353 MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
2354 MachinePointerInfo(SV, 16),
2358 // int __gr_offs at offset 24
2359 SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
2360 DAG.getConstant(24, getPointerTy()));
2361 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-GPRSize, MVT::i32),
2362 GROffsAddr, MachinePointerInfo(SV, 24),
2365 // int __vr_offs at offset 28
2366 SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, getPointerTy(), VAList,
2367 DAG.getConstant(28, getPointerTy()));
2368 MemOps.push_back(DAG.getStore(Chain, DL, DAG.getConstant(-FPRSize, MVT::i32),
2369 VROffsAddr, MachinePointerInfo(SV, 28),
2372 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &MemOps[0],
2377 AArch64TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
2378 switch (Op.getOpcode()) {
2379 default: llvm_unreachable("Don't know how to custom lower this!");
2380 case ISD::FADD: return LowerF128ToCall(Op, DAG, RTLIB::ADD_F128);
2381 case ISD::FSUB: return LowerF128ToCall(Op, DAG, RTLIB::SUB_F128);
2382 case ISD::FMUL: return LowerF128ToCall(Op, DAG, RTLIB::MUL_F128);
2383 case ISD::FDIV: return LowerF128ToCall(Op, DAG, RTLIB::DIV_F128);
2384 case ISD::FP_TO_SINT: return LowerFP_TO_INT(Op, DAG, true);
2385 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG, false);
2386 case ISD::SINT_TO_FP: return LowerINT_TO_FP(Op, DAG, true);
2387 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG, false);
2388 case ISD::FP_ROUND: return LowerFP_ROUND(Op, DAG);
2389 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
2391 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
2392 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
2393 case ISD::BR_CC: return LowerBR_CC(Op, DAG);
2394 case ISD::GlobalAddress: return LowerGlobalAddressELF(Op, DAG);
2395 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
2396 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
2397 case ISD::SELECT: return LowerSELECT(Op, DAG);
2398 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
2399 case ISD::SETCC: return LowerSETCC(Op, DAG);
2400 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
2401 case ISD::VASTART: return LowerVASTART(Op, DAG);
2407 static SDValue PerformANDCombine(SDNode *N,
2408 TargetLowering::DAGCombinerInfo &DCI) {
2410 SelectionDAG &DAG = DCI.DAG;
2411 DebugLoc DL = N->getDebugLoc();
2412 EVT VT = N->getValueType(0);
2414 // We're looking for an SRA/SHL pair which form an SBFX.
2416 if (VT != MVT::i32 && VT != MVT::i64)
2419 if (!isa<ConstantSDNode>(N->getOperand(1)))
2422 uint64_t TruncMask = N->getConstantOperandVal(1);
2423 if (!isMask_64(TruncMask))
2426 uint64_t Width = CountPopulation_64(TruncMask);
2427 SDValue Shift = N->getOperand(0);
2429 if (Shift.getOpcode() != ISD::SRL)
2432 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
2434 uint64_t LSB = Shift->getConstantOperandVal(1);
2436 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
2439 return DAG.getNode(AArch64ISD::UBFX, DL, VT, Shift.getOperand(0),
2440 DAG.getConstant(LSB, MVT::i64),
2441 DAG.getConstant(LSB + Width - 1, MVT::i64));
2444 /// For a true bitfield insert, the bits getting into that contiguous mask
2445 /// should come from the low part of an existing value: they must be formed from
2446 /// a compatible SHL operation (unless they're already low). This function
2447 /// checks that condition and returns the least-significant bit that's
2448 /// intended. If the operation not a field preparation, -1 is returned.
2449 static int32_t getLSBForBFI(SelectionDAG &DAG, DebugLoc DL, EVT VT,
2450 SDValue &MaskedVal, uint64_t Mask) {
2451 if (!isShiftedMask_64(Mask))
2454 // Now we need to alter MaskedVal so that it is an appropriate input for a BFI
2455 // instruction. BFI will do a left-shift by LSB before applying the mask we've
2456 // spotted, so in general we should pre-emptively "undo" that by making sure
2457 // the incoming bits have had a right-shift applied to them.
2459 // This right shift, however, will combine with existing left/right shifts. In
2460 // the simplest case of a completely straight bitfield operation, it will be
2461 // expected to completely cancel out with an existing SHL. More complicated
2462 // cases (e.g. bitfield to bitfield copy) may still need a real shift before
2465 uint64_t LSB = CountTrailingZeros_64(Mask);
2466 int64_t ShiftRightRequired = LSB;
2467 if (MaskedVal.getOpcode() == ISD::SHL &&
2468 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
2469 ShiftRightRequired -= MaskedVal.getConstantOperandVal(1);
2470 MaskedVal = MaskedVal.getOperand(0);
2471 } else if (MaskedVal.getOpcode() == ISD::SRL &&
2472 isa<ConstantSDNode>(MaskedVal.getOperand(1))) {
2473 ShiftRightRequired += MaskedVal.getConstantOperandVal(1);
2474 MaskedVal = MaskedVal.getOperand(0);
2477 if (ShiftRightRequired > 0)
2478 MaskedVal = DAG.getNode(ISD::SRL, DL, VT, MaskedVal,
2479 DAG.getConstant(ShiftRightRequired, MVT::i64));
2480 else if (ShiftRightRequired < 0) {
2481 // We could actually end up with a residual left shift, for example with
2482 // "struc.bitfield = val << 1".
2483 MaskedVal = DAG.getNode(ISD::SHL, DL, VT, MaskedVal,
2484 DAG.getConstant(-ShiftRightRequired, MVT::i64));
2490 /// Searches from N for an existing AArch64ISD::BFI node, possibly surrounded by
2491 /// a mask and an extension. Returns true if a BFI was found and provides
2492 /// information on its surroundings.
2493 static bool findMaskedBFI(SDValue N, SDValue &BFI, uint64_t &Mask,
2496 if (N.getOpcode() == ISD::ZERO_EXTEND) {
2498 N = N.getOperand(0);
2501 if (N.getOpcode() == ISD::AND && isa<ConstantSDNode>(N.getOperand(1))) {
2502 Mask = N->getConstantOperandVal(1);
2503 N = N.getOperand(0);
2505 // Mask is the whole width.
2506 Mask = -1ULL >> (64 - N.getValueType().getSizeInBits());
2509 if (N.getOpcode() == AArch64ISD::BFI) {
2517 /// Try to combine a subtree (rooted at an OR) into a "masked BFI" node, which
2518 /// is roughly equivalent to (and (BFI ...), mask). This form is used because it
2519 /// can often be further combined with a larger mask. Ultimately, we want mask
2520 /// to be 2^32-1 or 2^64-1 so the AND can be skipped.
2521 static SDValue tryCombineToBFI(SDNode *N,
2522 TargetLowering::DAGCombinerInfo &DCI,
2523 const AArch64Subtarget *Subtarget) {
2524 SelectionDAG &DAG = DCI.DAG;
2525 DebugLoc DL = N->getDebugLoc();
2526 EVT VT = N->getValueType(0);
2528 assert(N->getOpcode() == ISD::OR && "Unexpected root");
2530 // We need the LHS to be (and SOMETHING, MASK). Find out what that mask is or
2531 // abandon the effort.
2532 SDValue LHS = N->getOperand(0);
2533 if (LHS.getOpcode() != ISD::AND)
2537 if (isa<ConstantSDNode>(LHS.getOperand(1)))
2538 LHSMask = LHS->getConstantOperandVal(1);
2542 // We also need the RHS to be (and SOMETHING, MASK). Find out what that mask
2543 // is or abandon the effort.
2544 SDValue RHS = N->getOperand(1);
2545 if (RHS.getOpcode() != ISD::AND)
2549 if (isa<ConstantSDNode>(RHS.getOperand(1)))
2550 RHSMask = RHS->getConstantOperandVal(1);
2554 // Can't do anything if the masks are incompatible.
2555 if (LHSMask & RHSMask)
2558 // Now we need one of the masks to be a contiguous field. Without loss of
2559 // generality that should be the RHS one.
2560 SDValue Bitfield = LHS.getOperand(0);
2561 if (getLSBForBFI(DAG, DL, VT, Bitfield, LHSMask) != -1) {
2562 // We know that LHS is a candidate new value, and RHS isn't already a better
2564 std::swap(LHS, RHS);
2565 std::swap(LHSMask, RHSMask);
2568 // We've done our best to put the right operands in the right places, all we
2569 // can do now is check whether a BFI exists.
2570 Bitfield = RHS.getOperand(0);
2571 int32_t LSB = getLSBForBFI(DAG, DL, VT, Bitfield, RHSMask);
2575 uint32_t Width = CountPopulation_64(RHSMask);
2576 assert(Width && "Expected non-zero bitfield width");
2578 SDValue BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
2579 LHS.getOperand(0), Bitfield,
2580 DAG.getConstant(LSB, MVT::i64),
2581 DAG.getConstant(Width, MVT::i64));
2584 if ((LHSMask | RHSMask) == (-1ULL >> (64 - VT.getSizeInBits())))
2587 return DAG.getNode(ISD::AND, DL, VT, BFI,
2588 DAG.getConstant(LHSMask | RHSMask, VT));
2591 /// Search for the bitwise combining (with careful masks) of a MaskedBFI and its
2592 /// original input. This is surprisingly common because SROA splits things up
2593 /// into i8 chunks, so the originally detected MaskedBFI may actually only act
2594 /// on the low (say) byte of a word. This is then orred into the rest of the
2595 /// word afterwards.
2597 /// Basic input: (or (and OLDFIELD, MASK1), (MaskedBFI MASK2, OLDFIELD, ...)).
2599 /// If MASK1 and MASK2 are compatible, we can fold the whole thing into the
2600 /// MaskedBFI. We can also deal with a certain amount of extend/truncate being
2602 static SDValue tryCombineToLargerBFI(SDNode *N,
2603 TargetLowering::DAGCombinerInfo &DCI,
2604 const AArch64Subtarget *Subtarget) {
2605 SelectionDAG &DAG = DCI.DAG;
2606 DebugLoc DL = N->getDebugLoc();
2607 EVT VT = N->getValueType(0);
2609 // First job is to hunt for a MaskedBFI on either the left or right. Swap
2610 // operands if it's actually on the right.
2612 SDValue PossExtraMask;
2613 uint64_t ExistingMask = 0;
2614 bool Extended = false;
2615 if (findMaskedBFI(N->getOperand(0), BFI, ExistingMask, Extended))
2616 PossExtraMask = N->getOperand(1);
2617 else if (findMaskedBFI(N->getOperand(1), BFI, ExistingMask, Extended))
2618 PossExtraMask = N->getOperand(0);
2622 // We can only combine a BFI with another compatible mask.
2623 if (PossExtraMask.getOpcode() != ISD::AND ||
2624 !isa<ConstantSDNode>(PossExtraMask.getOperand(1)))
2627 uint64_t ExtraMask = PossExtraMask->getConstantOperandVal(1);
2629 // Masks must be compatible.
2630 if (ExtraMask & ExistingMask)
2633 SDValue OldBFIVal = BFI.getOperand(0);
2634 SDValue NewBFIVal = BFI.getOperand(1);
2636 // We skipped a ZERO_EXTEND above, so the input to the MaskedBFIs should be
2637 // 32-bit and we'll be forming a 64-bit MaskedBFI. The MaskedBFI arguments
2638 // need to be made compatible.
2639 assert(VT == MVT::i64 && BFI.getValueType() == MVT::i32
2640 && "Invalid types for BFI");
2641 OldBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, OldBFIVal);
2642 NewBFIVal = DAG.getNode(ISD::ANY_EXTEND, DL, VT, NewBFIVal);
2645 // We need the MaskedBFI to be combined with a mask of the *same* value.
2646 if (PossExtraMask.getOperand(0) != OldBFIVal)
2649 BFI = DAG.getNode(AArch64ISD::BFI, DL, VT,
2650 OldBFIVal, NewBFIVal,
2651 BFI.getOperand(2), BFI.getOperand(3));
2653 // If the masking is trivial, we don't need to create it.
2654 if ((ExtraMask | ExistingMask) == (-1ULL >> (64 - VT.getSizeInBits())))
2657 return DAG.getNode(ISD::AND, DL, VT, BFI,
2658 DAG.getConstant(ExtraMask | ExistingMask, VT));
2661 /// An EXTR instruction is made up of two shifts, ORed together. This helper
2662 /// searches for and classifies those shifts.
2663 static bool findEXTRHalf(SDValue N, SDValue &Src, uint32_t &ShiftAmount,
2665 if (N.getOpcode() == ISD::SHL)
2667 else if (N.getOpcode() == ISD::SRL)
2672 if (!isa<ConstantSDNode>(N.getOperand(1)))
2675 ShiftAmount = N->getConstantOperandVal(1);
2676 Src = N->getOperand(0);
2680 /// EXTR instruction extracts a contiguous chunk of bits from two existing
2681 /// registers viewed as a high/low pair. This function looks for the pattern:
2682 /// (or (shl VAL1, #N), (srl VAL2, #RegWidth-N)) and replaces it with an
2683 /// EXTR. Can't quite be done in TableGen because the two immediates aren't
2685 static SDValue tryCombineToEXTR(SDNode *N,
2686 TargetLowering::DAGCombinerInfo &DCI) {
2687 SelectionDAG &DAG = DCI.DAG;
2688 DebugLoc DL = N->getDebugLoc();
2689 EVT VT = N->getValueType(0);
2691 assert(N->getOpcode() == ISD::OR && "Unexpected root");
2693 if (VT != MVT::i32 && VT != MVT::i64)
2697 uint32_t ShiftLHS = 0;
2699 if (!findEXTRHalf(N->getOperand(0), LHS, ShiftLHS, LHSFromHi))
2703 uint32_t ShiftRHS = 0;
2705 if (!findEXTRHalf(N->getOperand(1), RHS, ShiftRHS, RHSFromHi))
2708 // If they're both trying to come from the high part of the register, they're
2709 // not really an EXTR.
2710 if (LHSFromHi == RHSFromHi)
2713 if (ShiftLHS + ShiftRHS != VT.getSizeInBits())
2717 std::swap(LHS, RHS);
2718 std::swap(ShiftLHS, ShiftRHS);
2721 return DAG.getNode(AArch64ISD::EXTR, DL, VT,
2723 DAG.getConstant(ShiftRHS, MVT::i64));
2726 /// Target-specific dag combine xforms for ISD::OR
2727 static SDValue PerformORCombine(SDNode *N,
2728 TargetLowering::DAGCombinerInfo &DCI,
2729 const AArch64Subtarget *Subtarget) {
2731 SelectionDAG &DAG = DCI.DAG;
2732 EVT VT = N->getValueType(0);
2734 if(!DAG.getTargetLoweringInfo().isTypeLegal(VT))
2737 // Attempt to recognise bitfield-insert operations.
2738 SDValue Res = tryCombineToBFI(N, DCI, Subtarget);
2742 // Attempt to combine an existing MaskedBFI operation into one with a larger
2744 Res = tryCombineToLargerBFI(N, DCI, Subtarget);
2748 Res = tryCombineToEXTR(N, DCI);
2755 /// Target-specific dag combine xforms for ISD::SRA
2756 static SDValue PerformSRACombine(SDNode *N,
2757 TargetLowering::DAGCombinerInfo &DCI) {
2759 SelectionDAG &DAG = DCI.DAG;
2760 DebugLoc DL = N->getDebugLoc();
2761 EVT VT = N->getValueType(0);
2763 // We're looking for an SRA/SHL pair which form an SBFX.
2765 if (VT != MVT::i32 && VT != MVT::i64)
2768 if (!isa<ConstantSDNode>(N->getOperand(1)))
2771 uint64_t ExtraSignBits = N->getConstantOperandVal(1);
2772 SDValue Shift = N->getOperand(0);
2774 if (Shift.getOpcode() != ISD::SHL)
2777 if (!isa<ConstantSDNode>(Shift->getOperand(1)))
2780 uint64_t BitsOnLeft = Shift->getConstantOperandVal(1);
2781 uint64_t Width = VT.getSizeInBits() - ExtraSignBits;
2782 uint64_t LSB = VT.getSizeInBits() - Width - BitsOnLeft;
2784 if (LSB > VT.getSizeInBits() || Width > VT.getSizeInBits())
2787 return DAG.getNode(AArch64ISD::SBFX, DL, VT, Shift.getOperand(0),
2788 DAG.getConstant(LSB, MVT::i64),
2789 DAG.getConstant(LSB + Width - 1, MVT::i64));
2794 AArch64TargetLowering::PerformDAGCombine(SDNode *N,
2795 DAGCombinerInfo &DCI) const {
2796 switch (N->getOpcode()) {
2798 case ISD::AND: return PerformANDCombine(N, DCI);
2799 case ISD::OR: return PerformORCombine(N, DCI, Subtarget);
2800 case ISD::SRA: return PerformSRACombine(N, DCI);
2805 AArch64TargetLowering::ConstraintType
2806 AArch64TargetLowering::getConstraintType(const std::string &Constraint) const {
2807 if (Constraint.size() == 1) {
2808 switch (Constraint[0]) {
2810 case 'w': // An FP/SIMD vector register
2811 return C_RegisterClass;
2812 case 'I': // Constant that can be used with an ADD instruction
2813 case 'J': // Constant that can be used with a SUB instruction
2814 case 'K': // Constant that can be used with a 32-bit logical instruction
2815 case 'L': // Constant that can be used with a 64-bit logical instruction
2816 case 'M': // Constant that can be used as a 32-bit MOV immediate
2817 case 'N': // Constant that can be used as a 64-bit MOV immediate
2818 case 'Y': // Floating point constant zero
2819 case 'Z': // Integer constant zero
2821 case 'Q': // A memory reference with base register and no offset
2823 case 'S': // A symbolic address
2828 // FIXME: Ump, Utf, Usa, Ush
2829 // Ump: A memory address suitable for ldp/stp in SI, DI, SF and DF modes,
2830 // whatever they may be
2831 // Utf: A memory address suitable for ldp/stp in TF mode, whatever it may be
2832 // Usa: An absolute symbolic address
2833 // Ush: The high part (bits 32:12) of a pc-relative symbolic address
2834 assert(Constraint != "Ump" && Constraint != "Utf" && Constraint != "Usa"
2835 && Constraint != "Ush" && "Unimplemented constraints");
2837 return TargetLowering::getConstraintType(Constraint);
2840 TargetLowering::ConstraintWeight
2841 AArch64TargetLowering::getSingleConstraintMatchWeight(AsmOperandInfo &Info,
2842 const char *Constraint) const {
2844 llvm_unreachable("Constraint weight unimplemented");
2848 AArch64TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
2849 std::string &Constraint,
2850 std::vector<SDValue> &Ops,
2851 SelectionDAG &DAG) const {
2852 SDValue Result(0, 0);
2854 // Only length 1 constraints are C_Other.
2855 if (Constraint.size() != 1) return;
2857 // Only C_Other constraints get lowered like this. That means constants for us
2858 // so return early if there's no hope the constraint can be lowered.
2860 switch(Constraint[0]) {
2862 case 'I': case 'J': case 'K': case 'L':
2863 case 'M': case 'N': case 'Z': {
2864 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
2868 uint64_t CVal = C->getZExtValue();
2871 switch (Constraint[0]) {
2873 // FIXME: 'M' and 'N' are MOV pseudo-insts -- unsupported in assembly. 'J'
2874 // is a peculiarly useless SUB constraint.
2875 llvm_unreachable("Unimplemented C_Other constraint");
2881 if (A64Imms::isLogicalImm(32, CVal, Bits))
2885 if (A64Imms::isLogicalImm(64, CVal, Bits))
2894 Result = DAG.getTargetConstant(CVal, Op.getValueType());
2898 // An absolute symbolic address or label reference.
2899 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op)) {
2900 Result = DAG.getTargetGlobalAddress(GA->getGlobal(), Op.getDebugLoc(),
2901 GA->getValueType(0));
2902 } else if (const BlockAddressSDNode *BA
2903 = dyn_cast<BlockAddressSDNode>(Op)) {
2904 Result = DAG.getTargetBlockAddress(BA->getBlockAddress(),
2905 BA->getValueType(0));
2906 } else if (const ExternalSymbolSDNode *ES
2907 = dyn_cast<ExternalSymbolSDNode>(Op)) {
2908 Result = DAG.getTargetExternalSymbol(ES->getSymbol(),
2909 ES->getValueType(0));
2915 if (const ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) {
2916 if (CFP->isExactlyValue(0.0)) {
2917 Result = DAG.getTargetConstantFP(0.0, CFP->getValueType(0));
2924 if (Result.getNode()) {
2925 Ops.push_back(Result);
2929 // It's an unknown constraint for us. Let generic code have a go.
2930 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2933 std::pair<unsigned, const TargetRegisterClass*>
2934 AArch64TargetLowering::getRegForInlineAsmConstraint(
2935 const std::string &Constraint,
2937 if (Constraint.size() == 1) {
2938 switch (Constraint[0]) {
2940 if (VT.getSizeInBits() <= 32)
2941 return std::make_pair(0U, &AArch64::GPR32RegClass);
2942 else if (VT == MVT::i64)
2943 return std::make_pair(0U, &AArch64::GPR64RegClass);
2947 return std::make_pair(0U, &AArch64::FPR16RegClass);
2948 else if (VT == MVT::f32)
2949 return std::make_pair(0U, &AArch64::FPR32RegClass);
2950 else if (VT == MVT::f64)
2951 return std::make_pair(0U, &AArch64::FPR64RegClass);
2952 else if (VT.getSizeInBits() == 64)
2953 return std::make_pair(0U, &AArch64::VPR64RegClass);
2954 else if (VT == MVT::f128)
2955 return std::make_pair(0U, &AArch64::FPR128RegClass);
2956 else if (VT.getSizeInBits() == 128)
2957 return std::make_pair(0U, &AArch64::VPR128RegClass);
2962 // Use the default implementation in TargetLowering to convert the register
2963 // constraint into a member of a register class.
2964 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);