1 //===-- AArch64TargetMachine.cpp - Define TargetMachine for AArch64 -------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 //
10 //===----------------------------------------------------------------------===//
11
12 #include "AArch64TargetMachine.h"
13 #include "AArch64.h"
14 #include "AArch64MacroFusion.h"
15 #include "AArch64Subtarget.h"
16 #include "AArch64TargetObjectFile.h"
17 #include "AArch64TargetTransformInfo.h"
18 #include "MCTargetDesc/AArch64MCTargetDesc.h"
19 #include "TargetInfo/AArch64TargetInfo.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/CodeGen/CSEConfigBase.h"
24 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
25 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
26 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
27 #include "llvm/CodeGen/GlobalISel/Localizer.h"
28 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
29 #include "llvm/CodeGen/MachineScheduler.h"
30 #include "llvm/CodeGen/Passes.h"
31 #include "llvm/CodeGen/TargetPassConfig.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/Function.h"
34 #include "llvm/InitializePasses.h"
35 #include "llvm/MC/MCAsmInfo.h"
36 #include "llvm/MC/MCTargetOptions.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CodeGen.h"
39 #include "llvm/Support/CommandLine.h"
40 #include "llvm/Support/TargetRegistry.h"
41 #include "llvm/Target/TargetLoweringObjectFile.h"
42 #include "llvm/Target/TargetOptions.h"
43 #include "llvm/Transforms/CFGuard.h"
44 #include "llvm/Transforms/Scalar.h"
45 #include <memory>
46 #include <string>
47
48 using namespace llvm;
49
50 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp",
51 cl::desc("Enable the CCMP formation pass"),
52 cl::init(true), cl::Hidden);
53
54 static cl::opt<bool>
55 EnableCondBrTuning("aarch64-enable-cond-br-tune",
56 cl::desc("Enable the conditional branch tuning pass"),
57 cl::init(true), cl::Hidden);
58
59 static cl::opt<bool> EnableMCR("aarch64-enable-mcr",
60 cl::desc("Enable the machine combiner pass"),
61 cl::init(true), cl::Hidden);
62
63 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress",
64 cl::desc("Suppress STP for AArch64"),
65 cl::init(true), cl::Hidden);
66
67 static cl::opt<bool> EnableAdvSIMDScalar(
68 "aarch64-enable-simd-scalar",
69 cl::desc("Enable use of AdvSIMD scalar integer instructions"),
70 cl::init(false), cl::Hidden);
71
72 static cl::opt<bool>
73 EnablePromoteConstant("aarch64-enable-promote-const",
74 cl::desc("Enable the promote constant pass"),
75 cl::init(true), cl::Hidden);
76
77 static cl::opt<bool> EnableCollectLOH(
78 "aarch64-enable-collect-loh",
79 cl::desc("Enable the pass that emits the linker optimization hints (LOH)"),
80 cl::init(true), cl::Hidden);
81
82 static cl::opt<bool>
83 EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden,
84 cl::desc("Enable the pass that removes dead"
85 " definitons and replaces stores to"
86 " them with stores to the zero"
87 " register"),
88 cl::init(true));
89
90 static cl::opt<bool> EnableRedundantCopyElimination(
91 "aarch64-enable-copyelim",
92 cl::desc("Enable the redundant copy elimination pass"), cl::init(true),
93 cl::Hidden);
94
95 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt",
96 cl::desc("Enable the load/store pair"
97 " optimization pass"),
98 cl::init(true), cl::Hidden);
99
100 static cl::opt<bool> EnableAtomicTidy(
101 "aarch64-enable-atomic-cfg-tidy", cl::Hidden,
102 cl::desc("Run SimplifyCFG after expanding atomic operations"
103 " to make use of cmpxchg flow-based information"),
104 cl::init(true));
105
106 static cl::opt<bool>
107 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
108 cl::desc("Run early if-conversion"),
109 cl::init(true));
110
111 static cl::opt<bool>
112 EnableCondOpt("aarch64-enable-condopt",
113 cl::desc("Enable the condition optimizer pass"),
114 cl::init(true), cl::Hidden);
115
116 static cl::opt<bool>
117 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
118 cl::desc("Work around Cortex-A53 erratum 835769"),
119 cl::init(false));
120
121 static cl::opt<bool>
122 EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
123 cl::desc("Enable optimizations on complex GEPs"),
124 cl::init(false));
125
126 static cl::opt<bool>
127 BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true),
128 cl::desc("Relax out of range conditional branches"));
129
130 static cl::opt<bool> EnableCompressJumpTables(
131 "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true),
132 cl::desc("Use smallest entry possible for jump tables"));
133
134 // FIXME: Unify control over GlobalMerge.
135 static cl::opt<cl::boolOrDefault>
136 EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden,
137 cl::desc("Enable the global merge pass"));
138
139 static cl::opt<bool>
140 EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden,
141 cl::desc("Enable the loop data prefetch pass"),
142 cl::init(true));
143
144 static cl::opt<int> EnableGlobalISelAtO(
145 "aarch64-enable-global-isel-at-O", cl::Hidden,
146 cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"),
147 cl::init(0));
148
149 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix",
150 cl::init(true), cl::Hidden);
151
152 static cl::opt<bool>
153 EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden,
154 cl::desc("Enable the AAcrh64 branch target pass"),
155 cl::init(true));
156
LLVMInitializeAArch64Target()157 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAArch64Target() {
158 // Register the target.
159 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget());
160 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget());
161 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target());
162 RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target());
163 RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target());
164 auto PR = PassRegistry::getPassRegistry();
165 initializeGlobalISel(*PR);
166 initializeAArch64A53Fix835769Pass(*PR);
167 initializeAArch64A57FPLoadBalancingPass(*PR);
168 initializeAArch64AdvSIMDScalarPass(*PR);
169 initializeAArch64BranchTargetsPass(*PR);
170 initializeAArch64CollectLOHPass(*PR);
171 initializeAArch64CompressJumpTablesPass(*PR);
172 initializeAArch64ConditionalComparesPass(*PR);
173 initializeAArch64ConditionOptimizerPass(*PR);
174 initializeAArch64DeadRegisterDefinitionsPass(*PR);
175 initializeAArch64ExpandPseudoPass(*PR);
176 initializeAArch64LoadStoreOptPass(*PR);
177 initializeAArch64SIMDInstrOptPass(*PR);
178 initializeAArch64PreLegalizerCombinerPass(*PR);
179 initializeAArch64PromoteConstantPass(*PR);
180 initializeAArch64RedundantCopyEliminationPass(*PR);
181 initializeAArch64StorePairSuppressPass(*PR);
182 initializeFalkorHWPFFixPass(*PR);
183 initializeFalkorMarkStridedAccessesLegacyPass(*PR);
184 initializeLDTLSCleanupPass(*PR);
185 initializeAArch64SpeculationHardeningPass(*PR);
186 initializeAArch64StackTaggingPass(*PR);
187 initializeAArch64StackTaggingPreRAPass(*PR);
188 }
189
190 //===----------------------------------------------------------------------===//
191 // AArch64 Lowering public interface.
192 //===----------------------------------------------------------------------===//
createTLOF(const Triple & TT)193 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
194 if (TT.isOSBinFormatMachO())
195 return std::make_unique<AArch64_MachoTargetObjectFile>();
196 if (TT.isOSBinFormatCOFF())
197 return std::make_unique<AArch64_COFFTargetObjectFile>();
198
199 return std::make_unique<AArch64_ELFTargetObjectFile>();
200 }
201
202 // Helper function to build a DataLayout string
computeDataLayout(const Triple & TT,const MCTargetOptions & Options,bool LittleEndian)203 static std::string computeDataLayout(const Triple &TT,
204 const MCTargetOptions &Options,
205 bool LittleEndian) {
206 if (Options.getABIName() == "ilp32")
207 return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128";
208 if (TT.isOSBinFormatMachO()) {
209 if (TT.getArch() == Triple::aarch64_32)
210 return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128";
211 return "e-m:o-i64:64-i128:128-n32:64-S128";
212 }
213 if (TT.isOSBinFormatCOFF())
214 return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128";
215 if (LittleEndian)
216 return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
217 return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
218 }
219
getEffectiveRelocModel(const Triple & TT,Optional<Reloc::Model> RM)220 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
221 Optional<Reloc::Model> RM) {
222 // AArch64 Darwin and Windows are always PIC.
223 if (TT.isOSDarwin() || TT.isOSWindows())
224 return Reloc::PIC_;
225 // On ELF platforms the default static relocation model has a smart enough
226 // linker to cope with referencing external symbols defined in a shared
227 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
228 if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC)
229 return Reloc::Static;
230 return *RM;
231 }
232
233 static CodeModel::Model
getEffectiveAArch64CodeModel(const Triple & TT,Optional<CodeModel::Model> CM,bool JIT)234 getEffectiveAArch64CodeModel(const Triple &TT, Optional<CodeModel::Model> CM,
235 bool JIT) {
236 if (CM) {
237 if (*CM != CodeModel::Small && *CM != CodeModel::Tiny &&
238 *CM != CodeModel::Large) {
239 if (!TT.isOSFuchsia())
240 report_fatal_error(
241 "Only small, tiny and large code models are allowed on AArch64");
242 else if (*CM != CodeModel::Kernel)
243 report_fatal_error("Only small, tiny, kernel, and large code models "
244 "are allowed on AArch64");
245 } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF())
246 report_fatal_error("tiny code model is only supported on ELF");
247 return *CM;
248 }
249 // The default MCJIT memory managers make no guarantees about where they can
250 // find an executable page; JITed code needs to be able to refer to globals
251 // no matter how far away they are.
252 // We should set the CodeModel::Small for Windows ARM64 in JIT mode,
253 // since with large code model LLVM generating 4 MOV instructions, and
254 // Windows doesn't support relocating these long branch (4 MOVs).
255 if (JIT && !TT.isOSWindows())
256 return CodeModel::Large;
257 return CodeModel::Small;
258 }
259
260 /// Create an AArch64 architecture model.
261 ///
AArch64TargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,Optional<Reloc::Model> RM,Optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT,bool LittleEndian)262 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT,
263 StringRef CPU, StringRef FS,
264 const TargetOptions &Options,
265 Optional<Reloc::Model> RM,
266 Optional<CodeModel::Model> CM,
267 CodeGenOpt::Level OL, bool JIT,
268 bool LittleEndian)
269 : LLVMTargetMachine(T,
270 computeDataLayout(TT, Options.MCOptions, LittleEndian),
271 TT, CPU, FS, Options, getEffectiveRelocModel(TT, RM),
272 getEffectiveAArch64CodeModel(TT, CM, JIT), OL),
273 TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) {
274 initAsmInfo();
275
276 if (TT.isOSBinFormatMachO()) {
277 this->Options.TrapUnreachable = true;
278 this->Options.NoTrapAfterNoreturn = true;
279 }
280
281 if (getMCAsmInfo()->usesWindowsCFI()) {
282 // Unwinding can get confused if the last instruction in an
283 // exception-handling region (function, funclet, try block, etc.)
284 // is a call.
285 //
286 // FIXME: We could elide the trap if the next instruction would be in
287 // the same region anyway.
288 this->Options.TrapUnreachable = true;
289 }
290
291 if (this->Options.TLSSize == 0) // default
292 this->Options.TLSSize = 24;
293 if ((getCodeModel() == CodeModel::Small ||
294 getCodeModel() == CodeModel::Kernel) &&
295 this->Options.TLSSize > 32)
296 // for the small (and kernel) code model, the maximum TLS size is 4GiB
297 this->Options.TLSSize = 32;
298 else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24)
299 // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB)
300 this->Options.TLSSize = 24;
301
302 // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
303 // MachO/CodeModel::Large, which GlobalISel does not support.
304 if (getOptLevel() <= EnableGlobalISelAtO &&
305 TT.getArch() != Triple::aarch64_32 &&
306 !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) {
307 setGlobalISel(true);
308 setGlobalISelAbort(GlobalISelAbortMode::Disable);
309 }
310
311 // AArch64 supports the MachineOutliner.
312 setMachineOutliner(true);
313
314 // AArch64 supports default outlining behaviour.
315 setSupportsDefaultOutlining(true);
316 }
317
318 AArch64TargetMachine::~AArch64TargetMachine() = default;
319
320 const AArch64Subtarget *
getSubtargetImpl(const Function & F) const321 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
322 Attribute CPUAttr = F.getFnAttribute("target-cpu");
323 Attribute FSAttr = F.getFnAttribute("target-features");
324
325 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
326 ? CPUAttr.getValueAsString().str()
327 : TargetCPU;
328 std::string FS = !FSAttr.hasAttribute(Attribute::None)
329 ? FSAttr.getValueAsString().str()
330 : TargetFS;
331
332 auto &I = SubtargetMap[CPU + FS];
333 if (!I) {
334 // This needs to be done before we create a new subtarget since any
335 // creation will depend on the TM and the code generation flags on the
336 // function that reside in TargetOptions.
337 resetTargetOptions(F);
338 I = std::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this,
339 isLittle);
340 }
341 return I.get();
342 }
343
anchor()344 void AArch64leTargetMachine::anchor() { }
345
AArch64leTargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,Optional<Reloc::Model> RM,Optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT)346 AArch64leTargetMachine::AArch64leTargetMachine(
347 const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
348 const TargetOptions &Options, Optional<Reloc::Model> RM,
349 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
350 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {}
351
anchor()352 void AArch64beTargetMachine::anchor() { }
353
AArch64beTargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,Optional<Reloc::Model> RM,Optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT)354 AArch64beTargetMachine::AArch64beTargetMachine(
355 const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
356 const TargetOptions &Options, Optional<Reloc::Model> RM,
357 Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
358 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {}
359
360 namespace {
361
362 /// AArch64 Code Generator Pass Configuration Options.
363 class AArch64PassConfig : public TargetPassConfig {
364 public:
AArch64PassConfig(AArch64TargetMachine & TM,PassManagerBase & PM)365 AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
366 : TargetPassConfig(TM, PM) {
367 if (TM.getOptLevel() != CodeGenOpt::None)
368 substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
369 }
370
getAArch64TargetMachine() const371 AArch64TargetMachine &getAArch64TargetMachine() const {
372 return getTM<AArch64TargetMachine>();
373 }
374
375 ScheduleDAGInstrs *
createMachineScheduler(MachineSchedContext * C) const376 createMachineScheduler(MachineSchedContext *C) const override {
377 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
378 ScheduleDAGMILive *DAG = createGenericSchedLive(C);
379 DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
380 DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
381 if (ST.hasFusion())
382 DAG->addMutation(createAArch64MacroFusionDAGMutation());
383 return DAG;
384 }
385
386 ScheduleDAGInstrs *
createPostMachineScheduler(MachineSchedContext * C) const387 createPostMachineScheduler(MachineSchedContext *C) const override {
388 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
389 if (ST.hasFusion()) {
390 // Run the Macro Fusion after RA again since literals are expanded from
391 // pseudos then (v. addPreSched2()).
392 ScheduleDAGMI *DAG = createGenericSchedPostRA(C);
393 DAG->addMutation(createAArch64MacroFusionDAGMutation());
394 return DAG;
395 }
396
397 return nullptr;
398 }
399
400 void addIRPasses() override;
401 bool addPreISel() override;
402 bool addInstSelector() override;
403 bool addIRTranslator() override;
404 void addPreLegalizeMachineIR() override;
405 bool addLegalizeMachineIR() override;
406 bool addRegBankSelect() override;
407 void addPreGlobalInstructionSelect() override;
408 bool addGlobalInstructionSelect() override;
409 bool addILPOpts() override;
410 void addPreRegAlloc() override;
411 void addPostRegAlloc() override;
412 void addPreSched2() override;
413 void addPreEmitPass() override;
414
415 std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
416 };
417
418 } // end anonymous namespace
419
420 TargetTransformInfo
getTargetTransformInfo(const Function & F)421 AArch64TargetMachine::getTargetTransformInfo(const Function &F) {
422 return TargetTransformInfo(AArch64TTIImpl(this, F));
423 }
424
createPassConfig(PassManagerBase & PM)425 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
426 return new AArch64PassConfig(*this, PM);
427 }
428
getCSEConfig() const429 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const {
430 return getStandardCSEConfigForOpt(TM->getOptLevel());
431 }
432
addIRPasses()433 void AArch64PassConfig::addIRPasses() {
434 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
435 // ourselves.
436 addPass(createAtomicExpandPass());
437
438 // Cmpxchg instructions are often used with a subsequent comparison to
439 // determine whether it succeeded. We can exploit existing control-flow in
440 // ldrex/strex loops to simplify this, but it needs tidying up.
441 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
442 addPass(createCFGSimplificationPass(1, true, true, false, true));
443
444 // Run LoopDataPrefetch
445 //
446 // Run this before LSR to remove the multiplies involved in computing the
447 // pointer values N iterations ahead.
448 if (TM->getOptLevel() != CodeGenOpt::None) {
449 if (EnableLoopDataPrefetch)
450 addPass(createLoopDataPrefetchPass());
451 if (EnableFalkorHWPFFix)
452 addPass(createFalkorMarkStridedAccessesPass());
453 }
454
455 TargetPassConfig::addIRPasses();
456
457 // Match interleaved memory accesses to ldN/stN intrinsics.
458 if (TM->getOptLevel() != CodeGenOpt::None) {
459 addPass(createInterleavedLoadCombinePass());
460 addPass(createInterleavedAccessPass());
461 }
462
463 if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
464 // Call SeparateConstOffsetFromGEP pass to extract constants within indices
465 // and lower a GEP with multiple indices to either arithmetic operations or
466 // multiple GEPs with single index.
467 addPass(createSeparateConstOffsetFromGEPPass(true));
468 // Call EarlyCSE pass to find and remove subexpressions in the lowered
469 // result.
470 addPass(createEarlyCSEPass());
471 // Do loop invariant code motion in case part of the lowered result is
472 // invariant.
473 addPass(createLICMPass());
474 }
475
476 addPass(createAArch64StackTaggingPass(/* MergeInit = */ TM->getOptLevel() !=
477 CodeGenOpt::None));
478
479 // Add Control Flow Guard checks.
480 if (TM->getTargetTriple().isOSWindows())
481 addPass(createCFGuardCheckPass());
482 }
483
484 // Pass Pipeline Configuration
addPreISel()485 bool AArch64PassConfig::addPreISel() {
486 // Run promote constant before global merge, so that the promoted constants
487 // get a chance to be merged
488 if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
489 addPass(createAArch64PromoteConstantPass());
490 // FIXME: On AArch64, this depends on the type.
491 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
492 // and the offset has to be a multiple of the related size in bytes.
493 if ((TM->getOptLevel() != CodeGenOpt::None &&
494 EnableGlobalMerge == cl::BOU_UNSET) ||
495 EnableGlobalMerge == cl::BOU_TRUE) {
496 bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
497 (EnableGlobalMerge == cl::BOU_UNSET);
498
499 // Merging of extern globals is enabled by default on non-Mach-O as we
500 // expect it to be generally either beneficial or harmless. On Mach-O it
501 // is disabled as we emit the .subsections_via_symbols directive which
502 // means that merging extern globals is not safe.
503 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
504
505 // FIXME: extern global merging is only enabled when we optimise for size
506 // because there are some regressions with it also enabled for performance.
507 if (!OnlyOptimizeForSize)
508 MergeExternalByDefault = false;
509
510 addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize,
511 MergeExternalByDefault));
512 }
513
514 return false;
515 }
516
addInstSelector()517 bool AArch64PassConfig::addInstSelector() {
518 addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
519
520 // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
521 // references to _TLS_MODULE_BASE_ as possible.
522 if (TM->getTargetTriple().isOSBinFormatELF() &&
523 getOptLevel() != CodeGenOpt::None)
524 addPass(createAArch64CleanupLocalDynamicTLSPass());
525
526 return false;
527 }
528
addIRTranslator()529 bool AArch64PassConfig::addIRTranslator() {
530 addPass(new IRTranslator());
531 return false;
532 }
533
addPreLegalizeMachineIR()534 void AArch64PassConfig::addPreLegalizeMachineIR() {
535 bool IsOptNone = getOptLevel() == CodeGenOpt::None;
536 addPass(createAArch64PreLegalizeCombiner(IsOptNone));
537 }
538
addLegalizeMachineIR()539 bool AArch64PassConfig::addLegalizeMachineIR() {
540 addPass(new Legalizer());
541 return false;
542 }
543
addRegBankSelect()544 bool AArch64PassConfig::addRegBankSelect() {
545 addPass(new RegBankSelect());
546 return false;
547 }
548
addPreGlobalInstructionSelect()549 void AArch64PassConfig::addPreGlobalInstructionSelect() {
550 addPass(new Localizer());
551 }
552
addGlobalInstructionSelect()553 bool AArch64PassConfig::addGlobalInstructionSelect() {
554 addPass(new InstructionSelect());
555 return false;
556 }
557
addILPOpts()558 bool AArch64PassConfig::addILPOpts() {
559 if (EnableCondOpt)
560 addPass(createAArch64ConditionOptimizerPass());
561 if (EnableCCMP)
562 addPass(createAArch64ConditionalCompares());
563 if (EnableMCR)
564 addPass(&MachineCombinerID);
565 if (EnableCondBrTuning)
566 addPass(createAArch64CondBrTuning());
567 if (EnableEarlyIfConversion)
568 addPass(&EarlyIfConverterID);
569 if (EnableStPairSuppress)
570 addPass(createAArch64StorePairSuppressPass());
571 addPass(createAArch64SIMDInstrOptPass());
572 if (TM->getOptLevel() != CodeGenOpt::None)
573 addPass(createAArch64StackTaggingPreRAPass());
574 return true;
575 }
576
addPreRegAlloc()577 void AArch64PassConfig::addPreRegAlloc() {
578 // Change dead register definitions to refer to the zero register.
579 if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
580 addPass(createAArch64DeadRegisterDefinitions());
581
582 // Use AdvSIMD scalar instructions whenever profitable.
583 if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
584 addPass(createAArch64AdvSIMDScalar());
585 // The AdvSIMD pass may produce copies that can be rewritten to
586 // be register coalescer friendly.
587 addPass(&PeepholeOptimizerID);
588 }
589 }
590
addPostRegAlloc()591 void AArch64PassConfig::addPostRegAlloc() {
592 // Remove redundant copy instructions.
593 if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
594 addPass(createAArch64RedundantCopyEliminationPass());
595
596 if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
597 // Improve performance for some FP/SIMD code for A57.
598 addPass(createAArch64A57FPLoadBalancing());
599 }
600
addPreSched2()601 void AArch64PassConfig::addPreSched2() {
602 // Expand some pseudo instructions to allow proper scheduling.
603 addPass(createAArch64ExpandPseudoPass());
604 // Use load/store pair instructions when possible.
605 if (TM->getOptLevel() != CodeGenOpt::None) {
606 if (EnableLoadStoreOpt)
607 addPass(createAArch64LoadStoreOptimizationPass());
608 }
609
610 // The AArch64SpeculationHardeningPass destroys dominator tree and natural
611 // loop info, which is needed for the FalkorHWPFFixPass and also later on.
612 // Therefore, run the AArch64SpeculationHardeningPass before the
613 // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop
614 // info.
615 addPass(createAArch64SpeculationHardeningPass());
616
617 if (TM->getOptLevel() != CodeGenOpt::None) {
618 if (EnableFalkorHWPFFix)
619 addPass(createFalkorHWPFFixPass());
620 }
621 }
622
addPreEmitPass()623 void AArch64PassConfig::addPreEmitPass() {
624 // Machine Block Placement might have created new opportunities when run
625 // at O3, where the Tail Duplication Threshold is set to 4 instructions.
626 // Run the load/store optimizer once more.
627 if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt)
628 addPass(createAArch64LoadStoreOptimizationPass());
629
630 if (EnableA53Fix835769)
631 addPass(createAArch64A53Fix835769());
632
633 if (EnableBranchTargets)
634 addPass(createAArch64BranchTargetsPass());
635
636 // Relax conditional branch instructions if they're otherwise out of
637 // range of their destination.
638 if (BranchRelaxation)
639 addPass(&BranchRelaxationPassID);
640
641 // Identify valid longjmp targets for Windows Control Flow Guard.
642 if (TM->getTargetTriple().isOSWindows())
643 addPass(createCFGuardLongjmpPass());
644
645 if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables)
646 addPass(createAArch64CompressJumpTablesPass());
647
648 if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
649 TM->getTargetTriple().isOSBinFormatMachO())
650 addPass(createAArch64CollectLOHPass());
651 }
652