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 "AArch64MachineFunctionInfo.h"
15 #include "AArch64MacroFusion.h"
16 #include "AArch64Subtarget.h"
17 #include "AArch64TargetObjectFile.h"
18 #include "AArch64TargetTransformInfo.h"
19 #include "MCTargetDesc/AArch64MCTargetDesc.h"
20 #include "TargetInfo/AArch64TargetInfo.h"
21 #include "llvm/ADT/STLExtras.h"
22 #include "llvm/ADT/Triple.h"
23 #include "llvm/Analysis/TargetTransformInfo.h"
24 #include "llvm/CodeGen/CSEConfigBase.h"
25 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
26 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
27 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
28 #include "llvm/CodeGen/GlobalISel/Localizer.h"
29 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
30 #include "llvm/CodeGen/MIRParser/MIParser.h"
31 #include "llvm/CodeGen/MachineScheduler.h"
32 #include "llvm/CodeGen/Passes.h"
33 #include "llvm/CodeGen/TargetPassConfig.h"
34 #include "llvm/IR/Attributes.h"
35 #include "llvm/IR/Function.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCTargetOptions.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CodeGen.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/TargetRegistry.h"
43 #include "llvm/Target/TargetLoweringObjectFile.h"
44 #include "llvm/Target/TargetOptions.h"
45 #include "llvm/Transforms/CFGuard.h"
46 #include "llvm/Transforms/Scalar.h"
47 #include <memory>
48 #include <string>
49 
50 using namespace llvm;
51 
52 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp",
53                                 cl::desc("Enable the CCMP formation pass"),
54                                 cl::init(true), cl::Hidden);
55 
56 static cl::opt<bool>
57     EnableCondBrTuning("aarch64-enable-cond-br-tune",
58                        cl::desc("Enable the conditional branch tuning pass"),
59                        cl::init(true), cl::Hidden);
60 
61 static cl::opt<bool> EnableMCR("aarch64-enable-mcr",
62                                cl::desc("Enable the machine combiner pass"),
63                                cl::init(true), cl::Hidden);
64 
65 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress",
66                                           cl::desc("Suppress STP for AArch64"),
67                                           cl::init(true), cl::Hidden);
68 
69 static cl::opt<bool> EnableAdvSIMDScalar(
70     "aarch64-enable-simd-scalar",
71     cl::desc("Enable use of AdvSIMD scalar integer instructions"),
72     cl::init(false), cl::Hidden);
73 
74 static cl::opt<bool>
75     EnablePromoteConstant("aarch64-enable-promote-const",
76                           cl::desc("Enable the promote constant pass"),
77                           cl::init(true), cl::Hidden);
78 
79 static cl::opt<bool> EnableCollectLOH(
80     "aarch64-enable-collect-loh",
81     cl::desc("Enable the pass that emits the linker optimization hints (LOH)"),
82     cl::init(true), cl::Hidden);
83 
84 static cl::opt<bool>
85     EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden,
86                                   cl::desc("Enable the pass that removes dead"
87                                            " definitons and replaces stores to"
88                                            " them with stores to the zero"
89                                            " register"),
90                                   cl::init(true));
91 
92 static cl::opt<bool> EnableRedundantCopyElimination(
93     "aarch64-enable-copyelim",
94     cl::desc("Enable the redundant copy elimination pass"), cl::init(true),
95     cl::Hidden);
96 
97 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt",
98                                         cl::desc("Enable the load/store pair"
99                                                  " optimization pass"),
100                                         cl::init(true), cl::Hidden);
101 
102 static cl::opt<bool> EnableAtomicTidy(
103     "aarch64-enable-atomic-cfg-tidy", cl::Hidden,
104     cl::desc("Run SimplifyCFG after expanding atomic operations"
105              " to make use of cmpxchg flow-based information"),
106     cl::init(true));
107 
108 static cl::opt<bool>
109 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
110                         cl::desc("Run early if-conversion"),
111                         cl::init(true));
112 
113 static cl::opt<bool>
114     EnableCondOpt("aarch64-enable-condopt",
115                   cl::desc("Enable the condition optimizer pass"),
116                   cl::init(true), cl::Hidden);
117 
118 static cl::opt<bool>
119 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
120                 cl::desc("Work around Cortex-A53 erratum 835769"),
121                 cl::init(false));
122 
123 static cl::opt<bool>
124     EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
125                  cl::desc("Enable optimizations on complex GEPs"),
126                  cl::init(false));
127 
128 static cl::opt<bool>
129     BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true),
130                      cl::desc("Relax out of range conditional branches"));
131 
132 static cl::opt<bool> EnableCompressJumpTables(
133     "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true),
134     cl::desc("Use smallest entry possible for jump tables"));
135 
136 // FIXME: Unify control over GlobalMerge.
137 static cl::opt<cl::boolOrDefault>
138     EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden,
139                       cl::desc("Enable the global merge pass"));
140 
141 static cl::opt<bool>
142     EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden,
143                            cl::desc("Enable the loop data prefetch pass"),
144                            cl::init(true));
145 
146 static cl::opt<int> EnableGlobalISelAtO(
147     "aarch64-enable-global-isel-at-O", cl::Hidden,
148     cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"),
149     cl::init(0));
150 
151 static cl::opt<bool>
152     EnableSVEIntrinsicOpts("aarch64-enable-sve-intrinsic-opts", cl::Hidden,
153                            cl::desc("Enable SVE intrinsic opts"),
154                            cl::init(true));
155 
156 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix",
157                                          cl::init(true), cl::Hidden);
158 
159 static cl::opt<bool>
160     EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden,
161                         cl::desc("Enable the AArch64 branch target pass"),
162                         cl::init(true));
163 
164 static cl::opt<unsigned> SVEVectorBitsMaxOpt(
165     "aarch64-sve-vector-bits-max",
166     cl::desc("Assume SVE vector registers are at most this big, "
167              "with zero meaning no maximum size is assumed."),
168     cl::init(0), cl::Hidden);
169 
170 static cl::opt<unsigned> SVEVectorBitsMinOpt(
171     "aarch64-sve-vector-bits-min",
172     cl::desc("Assume SVE vector registers are at least this big, "
173              "with zero meaning no minimum size is assumed."),
174     cl::init(0), cl::Hidden);
175 
176 extern cl::opt<bool> EnableHomogeneousPrologEpilog;
177 
178 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAArch64Target() {
179   // Register the target.
180   RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget());
181   RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget());
182   RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target());
183   RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target());
184   RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target());
185   auto PR = PassRegistry::getPassRegistry();
186   initializeGlobalISel(*PR);
187   initializeAArch64A53Fix835769Pass(*PR);
188   initializeAArch64A57FPLoadBalancingPass(*PR);
189   initializeAArch64AdvSIMDScalarPass(*PR);
190   initializeAArch64BranchTargetsPass(*PR);
191   initializeAArch64CollectLOHPass(*PR);
192   initializeAArch64CompressJumpTablesPass(*PR);
193   initializeAArch64ConditionalComparesPass(*PR);
194   initializeAArch64ConditionOptimizerPass(*PR);
195   initializeAArch64DeadRegisterDefinitionsPass(*PR);
196   initializeAArch64ExpandPseudoPass(*PR);
197   initializeAArch64LoadStoreOptPass(*PR);
198   initializeAArch64MIPeepholeOptPass(*PR);
199   initializeAArch64SIMDInstrOptPass(*PR);
200   initializeAArch64O0PreLegalizerCombinerPass(*PR);
201   initializeAArch64PreLegalizerCombinerPass(*PR);
202   initializeAArch64PostLegalizerCombinerPass(*PR);
203   initializeAArch64PostLegalizerLoweringPass(*PR);
204   initializeAArch64PostSelectOptimizePass(*PR);
205   initializeAArch64PromoteConstantPass(*PR);
206   initializeAArch64RedundantCopyEliminationPass(*PR);
207   initializeAArch64StorePairSuppressPass(*PR);
208   initializeFalkorHWPFFixPass(*PR);
209   initializeFalkorMarkStridedAccessesLegacyPass(*PR);
210   initializeLDTLSCleanupPass(*PR);
211   initializeSVEIntrinsicOptsPass(*PR);
212   initializeAArch64SpeculationHardeningPass(*PR);
213   initializeAArch64SLSHardeningPass(*PR);
214   initializeAArch64StackTaggingPass(*PR);
215   initializeAArch64StackTaggingPreRAPass(*PR);
216   initializeAArch64LowerHomogeneousPrologEpilogPass(*PR);
217 }
218 
219 //===----------------------------------------------------------------------===//
220 // AArch64 Lowering public interface.
221 //===----------------------------------------------------------------------===//
222 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
223   if (TT.isOSBinFormatMachO())
224     return std::make_unique<AArch64_MachoTargetObjectFile>();
225   if (TT.isOSBinFormatCOFF())
226     return std::make_unique<AArch64_COFFTargetObjectFile>();
227 
228   return std::make_unique<AArch64_ELFTargetObjectFile>();
229 }
230 
231 // Helper function to build a DataLayout string
232 static std::string computeDataLayout(const Triple &TT,
233                                      const MCTargetOptions &Options,
234                                      bool LittleEndian) {
235   if (TT.isOSBinFormatMachO()) {
236     if (TT.getArch() == Triple::aarch64_32)
237       return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128";
238     return "e-m:o-i64:64-i128:128-n32:64-S128";
239   }
240   if (TT.isOSBinFormatCOFF())
241     return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128";
242   std::string Endian = LittleEndian ? "e" : "E";
243   std::string Ptr32 = TT.getEnvironment() == Triple::GNUILP32 ? "-p:32:32" : "";
244   return Endian + "-m:e" + Ptr32 +
245          "-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
246 }
247 
248 static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) {
249   if (CPU.empty() && TT.isArm64e())
250     return "apple-a12";
251   return CPU;
252 }
253 
254 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
255                                            Optional<Reloc::Model> RM) {
256   // AArch64 Darwin and Windows are always PIC.
257   if (TT.isOSDarwin() || TT.isOSWindows())
258     return Reloc::PIC_;
259   // On ELF platforms the default static relocation model has a smart enough
260   // linker to cope with referencing external symbols defined in a shared
261   // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
262   if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC)
263     return Reloc::Static;
264   return *RM;
265 }
266 
267 static CodeModel::Model
268 getEffectiveAArch64CodeModel(const Triple &TT, Optional<CodeModel::Model> CM,
269                              bool JIT) {
270   if (CM) {
271     if (*CM != CodeModel::Small && *CM != CodeModel::Tiny &&
272         *CM != CodeModel::Large) {
273       report_fatal_error(
274           "Only small, tiny and large code models are allowed on AArch64");
275     } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF())
276       report_fatal_error("tiny code model is only supported on ELF");
277     return *CM;
278   }
279   // The default MCJIT memory managers make no guarantees about where they can
280   // find an executable page; JITed code needs to be able to refer to globals
281   // no matter how far away they are.
282   // We should set the CodeModel::Small for Windows ARM64 in JIT mode,
283   // since with large code model LLVM generating 4 MOV instructions, and
284   // Windows doesn't support relocating these long branch (4 MOVs).
285   if (JIT && !TT.isOSWindows())
286     return CodeModel::Large;
287   return CodeModel::Small;
288 }
289 
290 /// Create an AArch64 architecture model.
291 ///
292 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT,
293                                            StringRef CPU, StringRef FS,
294                                            const TargetOptions &Options,
295                                            Optional<Reloc::Model> RM,
296                                            Optional<CodeModel::Model> CM,
297                                            CodeGenOpt::Level OL, bool JIT,
298                                            bool LittleEndian)
299     : LLVMTargetMachine(T,
300                         computeDataLayout(TT, Options.MCOptions, LittleEndian),
301                         TT, computeDefaultCPU(TT, CPU), FS, Options,
302                         getEffectiveRelocModel(TT, RM),
303                         getEffectiveAArch64CodeModel(TT, CM, JIT), OL),
304       TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) {
305   initAsmInfo();
306 
307   if (TT.isOSBinFormatMachO()) {
308     this->Options.TrapUnreachable = true;
309     this->Options.NoTrapAfterNoreturn = true;
310   }
311 
312   if (getMCAsmInfo()->usesWindowsCFI()) {
313     // Unwinding can get confused if the last instruction in an
314     // exception-handling region (function, funclet, try block, etc.)
315     // is a call.
316     //
317     // FIXME: We could elide the trap if the next instruction would be in
318     // the same region anyway.
319     this->Options.TrapUnreachable = true;
320   }
321 
322   if (this->Options.TLSSize == 0) // default
323     this->Options.TLSSize = 24;
324   if ((getCodeModel() == CodeModel::Small ||
325        getCodeModel() == CodeModel::Kernel) &&
326       this->Options.TLSSize > 32)
327     // for the small (and kernel) code model, the maximum TLS size is 4GiB
328     this->Options.TLSSize = 32;
329   else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24)
330     // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB)
331     this->Options.TLSSize = 24;
332 
333   // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
334   // MachO/CodeModel::Large, which GlobalISel does not support.
335   if (getOptLevel() <= EnableGlobalISelAtO &&
336       TT.getArch() != Triple::aarch64_32 &&
337       TT.getEnvironment() != Triple::GNUILP32 &&
338       !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) {
339     setGlobalISel(true);
340     setGlobalISelAbort(GlobalISelAbortMode::Disable);
341   }
342 
343   // AArch64 supports the MachineOutliner.
344   setMachineOutliner(true);
345 
346   // AArch64 supports default outlining behaviour.
347   setSupportsDefaultOutlining(true);
348 
349   // AArch64 supports the debug entry values.
350   setSupportsDebugEntryValues(true);
351 }
352 
353 AArch64TargetMachine::~AArch64TargetMachine() = default;
354 
355 const AArch64Subtarget *
356 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
357   Attribute CPUAttr = F.getFnAttribute("target-cpu");
358   Attribute FSAttr = F.getFnAttribute("target-features");
359 
360   std::string CPU =
361       CPUAttr.isValid() ? CPUAttr.getValueAsString().str() : TargetCPU;
362   std::string FS =
363       FSAttr.isValid() ? FSAttr.getValueAsString().str() : TargetFS;
364 
365   SmallString<512> Key;
366 
367   unsigned MinSVEVectorSize = 0;
368   unsigned MaxSVEVectorSize = 0;
369   Attribute VScaleRangeAttr = F.getFnAttribute(Attribute::VScaleRange);
370   if (VScaleRangeAttr.isValid()) {
371     std::tie(MinSVEVectorSize, MaxSVEVectorSize) =
372         VScaleRangeAttr.getVScaleRangeArgs();
373     MinSVEVectorSize *= 128;
374     MaxSVEVectorSize *= 128;
375   } else {
376     MinSVEVectorSize = SVEVectorBitsMinOpt;
377     MaxSVEVectorSize = SVEVectorBitsMaxOpt;
378   }
379 
380   assert(MinSVEVectorSize % 128 == 0 &&
381          "SVE requires vector length in multiples of 128!");
382   assert(MaxSVEVectorSize % 128 == 0 &&
383          "SVE requires vector length in multiples of 128!");
384   assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) &&
385          "Minimum SVE vector size should not be larger than its maximum!");
386 
387   // Sanitize user input in case of no asserts
388   if (MaxSVEVectorSize == 0)
389     MinSVEVectorSize = (MinSVEVectorSize / 128) * 128;
390   else {
391     MinSVEVectorSize =
392         (std::min(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
393     MaxSVEVectorSize =
394         (std::max(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
395   }
396 
397   Key += "SVEMin";
398   Key += std::to_string(MinSVEVectorSize);
399   Key += "SVEMax";
400   Key += std::to_string(MaxSVEVectorSize);
401   Key += CPU;
402   Key += FS;
403 
404   auto &I = SubtargetMap[Key];
405   if (!I) {
406     // This needs to be done before we create a new subtarget since any
407     // creation will depend on the TM and the code generation flags on the
408     // function that reside in TargetOptions.
409     resetTargetOptions(F);
410     I = std::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this,
411                                            isLittle, MinSVEVectorSize,
412                                            MaxSVEVectorSize);
413   }
414   return I.get();
415 }
416 
417 void AArch64leTargetMachine::anchor() { }
418 
419 AArch64leTargetMachine::AArch64leTargetMachine(
420     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
421     const TargetOptions &Options, Optional<Reloc::Model> RM,
422     Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
423     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {}
424 
425 void AArch64beTargetMachine::anchor() { }
426 
427 AArch64beTargetMachine::AArch64beTargetMachine(
428     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
429     const TargetOptions &Options, Optional<Reloc::Model> RM,
430     Optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
431     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {}
432 
433 namespace {
434 
435 /// AArch64 Code Generator Pass Configuration Options.
436 class AArch64PassConfig : public TargetPassConfig {
437 public:
438   AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
439       : TargetPassConfig(TM, PM) {
440     if (TM.getOptLevel() != CodeGenOpt::None)
441       substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
442   }
443 
444   AArch64TargetMachine &getAArch64TargetMachine() const {
445     return getTM<AArch64TargetMachine>();
446   }
447 
448   ScheduleDAGInstrs *
449   createMachineScheduler(MachineSchedContext *C) const override {
450     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
451     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
452     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
453     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
454     if (ST.hasFusion())
455       DAG->addMutation(createAArch64MacroFusionDAGMutation());
456     return DAG;
457   }
458 
459   ScheduleDAGInstrs *
460   createPostMachineScheduler(MachineSchedContext *C) const override {
461     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
462     if (ST.hasFusion()) {
463       // Run the Macro Fusion after RA again since literals are expanded from
464       // pseudos then (v. addPreSched2()).
465       ScheduleDAGMI *DAG = createGenericSchedPostRA(C);
466       DAG->addMutation(createAArch64MacroFusionDAGMutation());
467       return DAG;
468     }
469 
470     return nullptr;
471   }
472 
473   void addIRPasses()  override;
474   bool addPreISel() override;
475   bool addInstSelector() override;
476   bool addIRTranslator() override;
477   void addPreLegalizeMachineIR() override;
478   bool addLegalizeMachineIR() override;
479   void addPreRegBankSelect() override;
480   bool addRegBankSelect() override;
481   void addPreGlobalInstructionSelect() override;
482   bool addGlobalInstructionSelect() override;
483   void addMachineSSAOptimization() override;
484   bool addILPOpts() override;
485   void addPreRegAlloc() override;
486   void addPostRegAlloc() override;
487   void addPreSched2() override;
488   void addPreEmitPass() override;
489   void addPreEmitPass2() override;
490 
491   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
492 };
493 
494 } // end anonymous namespace
495 
496 TargetTransformInfo
497 AArch64TargetMachine::getTargetTransformInfo(const Function &F) {
498   return TargetTransformInfo(AArch64TTIImpl(this, F));
499 }
500 
501 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
502   return new AArch64PassConfig(*this, PM);
503 }
504 
505 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const {
506   return getStandardCSEConfigForOpt(TM->getOptLevel());
507 }
508 
509 void AArch64PassConfig::addIRPasses() {
510   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
511   // ourselves.
512   addPass(createAtomicExpandPass());
513 
514   // Expand any SVE vector library calls that we can't code generate directly.
515   if (EnableSVEIntrinsicOpts && TM->getOptLevel() == CodeGenOpt::Aggressive)
516     addPass(createSVEIntrinsicOptsPass());
517 
518   // Cmpxchg instructions are often used with a subsequent comparison to
519   // determine whether it succeeded. We can exploit existing control-flow in
520   // ldrex/strex loops to simplify this, but it needs tidying up.
521   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
522     addPass(createCFGSimplificationPass(SimplifyCFGOptions()
523                                             .forwardSwitchCondToPhi(true)
524                                             .convertSwitchToLookupTable(true)
525                                             .needCanonicalLoops(false)
526                                             .hoistCommonInsts(true)
527                                             .sinkCommonInsts(true)));
528 
529   // Run LoopDataPrefetch
530   //
531   // Run this before LSR to remove the multiplies involved in computing the
532   // pointer values N iterations ahead.
533   if (TM->getOptLevel() != CodeGenOpt::None) {
534     if (EnableLoopDataPrefetch)
535       addPass(createLoopDataPrefetchPass());
536     if (EnableFalkorHWPFFix)
537       addPass(createFalkorMarkStridedAccessesPass());
538   }
539 
540   TargetPassConfig::addIRPasses();
541 
542   addPass(createAArch64StackTaggingPass(
543       /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None));
544 
545   // Match interleaved memory accesses to ldN/stN intrinsics.
546   if (TM->getOptLevel() != CodeGenOpt::None) {
547     addPass(createInterleavedLoadCombinePass());
548     addPass(createInterleavedAccessPass());
549   }
550 
551   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
552     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
553     // and lower a GEP with multiple indices to either arithmetic operations or
554     // multiple GEPs with single index.
555     addPass(createSeparateConstOffsetFromGEPPass(true));
556     // Call EarlyCSE pass to find and remove subexpressions in the lowered
557     // result.
558     addPass(createEarlyCSEPass());
559     // Do loop invariant code motion in case part of the lowered result is
560     // invariant.
561     addPass(createLICMPass());
562   }
563 
564   // Add Control Flow Guard checks.
565   if (TM->getTargetTriple().isOSWindows())
566     addPass(createCFGuardCheckPass());
567 }
568 
569 // Pass Pipeline Configuration
570 bool AArch64PassConfig::addPreISel() {
571   // Run promote constant before global merge, so that the promoted constants
572   // get a chance to be merged
573   if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
574     addPass(createAArch64PromoteConstantPass());
575   // FIXME: On AArch64, this depends on the type.
576   // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
577   // and the offset has to be a multiple of the related size in bytes.
578   if ((TM->getOptLevel() != CodeGenOpt::None &&
579        EnableGlobalMerge == cl::BOU_UNSET) ||
580       EnableGlobalMerge == cl::BOU_TRUE) {
581     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
582                                (EnableGlobalMerge == cl::BOU_UNSET);
583 
584     // Merging of extern globals is enabled by default on non-Mach-O as we
585     // expect it to be generally either beneficial or harmless. On Mach-O it
586     // is disabled as we emit the .subsections_via_symbols directive which
587     // means that merging extern globals is not safe.
588     bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
589 
590     // FIXME: extern global merging is only enabled when we optimise for size
591     // because there are some regressions with it also enabled for performance.
592     if (!OnlyOptimizeForSize)
593       MergeExternalByDefault = false;
594 
595     addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize,
596                                   MergeExternalByDefault));
597   }
598 
599   return false;
600 }
601 
602 bool AArch64PassConfig::addInstSelector() {
603   addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
604 
605   // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
606   // references to _TLS_MODULE_BASE_ as possible.
607   if (TM->getTargetTriple().isOSBinFormatELF() &&
608       getOptLevel() != CodeGenOpt::None)
609     addPass(createAArch64CleanupLocalDynamicTLSPass());
610 
611   return false;
612 }
613 
614 bool AArch64PassConfig::addIRTranslator() {
615   addPass(new IRTranslator(getOptLevel()));
616   return false;
617 }
618 
619 void AArch64PassConfig::addPreLegalizeMachineIR() {
620   if (getOptLevel() == CodeGenOpt::None)
621     addPass(createAArch64O0PreLegalizerCombiner());
622   else
623     addPass(createAArch64PreLegalizerCombiner());
624 }
625 
626 bool AArch64PassConfig::addLegalizeMachineIR() {
627   addPass(new Legalizer());
628   return false;
629 }
630 
631 void AArch64PassConfig::addPreRegBankSelect() {
632   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
633   if (!IsOptNone)
634     addPass(createAArch64PostLegalizerCombiner(IsOptNone));
635   addPass(createAArch64PostLegalizerLowering());
636 }
637 
638 bool AArch64PassConfig::addRegBankSelect() {
639   addPass(new RegBankSelect());
640   return false;
641 }
642 
643 void AArch64PassConfig::addPreGlobalInstructionSelect() {
644   addPass(new Localizer());
645 }
646 
647 bool AArch64PassConfig::addGlobalInstructionSelect() {
648   addPass(new InstructionSelect(getOptLevel()));
649   if (getOptLevel() != CodeGenOpt::None)
650     addPass(createAArch64PostSelectOptimize());
651   return false;
652 }
653 
654 void AArch64PassConfig::addMachineSSAOptimization() {
655   // Run default MachineSSAOptimization first.
656   TargetPassConfig::addMachineSSAOptimization();
657 
658   if (TM->getOptLevel() != CodeGenOpt::None)
659     addPass(createAArch64MIPeepholeOptPass());
660 }
661 
662 bool AArch64PassConfig::addILPOpts() {
663   if (EnableCondOpt)
664     addPass(createAArch64ConditionOptimizerPass());
665   if (EnableCCMP)
666     addPass(createAArch64ConditionalCompares());
667   if (EnableMCR)
668     addPass(&MachineCombinerID);
669   if (EnableCondBrTuning)
670     addPass(createAArch64CondBrTuning());
671   if (EnableEarlyIfConversion)
672     addPass(&EarlyIfConverterID);
673   if (EnableStPairSuppress)
674     addPass(createAArch64StorePairSuppressPass());
675   addPass(createAArch64SIMDInstrOptPass());
676   if (TM->getOptLevel() != CodeGenOpt::None)
677     addPass(createAArch64StackTaggingPreRAPass());
678   return true;
679 }
680 
681 void AArch64PassConfig::addPreRegAlloc() {
682   // Change dead register definitions to refer to the zero register.
683   if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
684     addPass(createAArch64DeadRegisterDefinitions());
685 
686   // Use AdvSIMD scalar instructions whenever profitable.
687   if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
688     addPass(createAArch64AdvSIMDScalar());
689     // The AdvSIMD pass may produce copies that can be rewritten to
690     // be register coalescer friendly.
691     addPass(&PeepholeOptimizerID);
692   }
693 }
694 
695 void AArch64PassConfig::addPostRegAlloc() {
696   // Remove redundant copy instructions.
697   if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
698     addPass(createAArch64RedundantCopyEliminationPass());
699 
700   if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
701     // Improve performance for some FP/SIMD code for A57.
702     addPass(createAArch64A57FPLoadBalancing());
703 }
704 
705 void AArch64PassConfig::addPreSched2() {
706   // Lower homogeneous frame instructions
707   if (EnableHomogeneousPrologEpilog)
708     addPass(createAArch64LowerHomogeneousPrologEpilogPass());
709   // Expand some pseudo instructions to allow proper scheduling.
710   addPass(createAArch64ExpandPseudoPass());
711   // Use load/store pair instructions when possible.
712   if (TM->getOptLevel() != CodeGenOpt::None) {
713     if (EnableLoadStoreOpt)
714       addPass(createAArch64LoadStoreOptimizationPass());
715   }
716 
717   // The AArch64SpeculationHardeningPass destroys dominator tree and natural
718   // loop info, which is needed for the FalkorHWPFFixPass and also later on.
719   // Therefore, run the AArch64SpeculationHardeningPass before the
720   // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop
721   // info.
722   addPass(createAArch64SpeculationHardeningPass());
723 
724   addPass(createAArch64IndirectThunks());
725   addPass(createAArch64SLSHardeningPass());
726 
727   if (TM->getOptLevel() != CodeGenOpt::None) {
728     if (EnableFalkorHWPFFix)
729       addPass(createFalkorHWPFFixPass());
730   }
731 }
732 
733 void AArch64PassConfig::addPreEmitPass() {
734   // Machine Block Placement might have created new opportunities when run
735   // at O3, where the Tail Duplication Threshold is set to 4 instructions.
736   // Run the load/store optimizer once more.
737   if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt)
738     addPass(createAArch64LoadStoreOptimizationPass());
739 
740   if (EnableA53Fix835769)
741     addPass(createAArch64A53Fix835769());
742 
743   if (EnableBranchTargets)
744     addPass(createAArch64BranchTargetsPass());
745 
746   // Relax conditional branch instructions if they're otherwise out of
747   // range of their destination.
748   if (BranchRelaxation)
749     addPass(&BranchRelaxationPassID);
750 
751   if (TM->getTargetTriple().isOSWindows()) {
752     // Identify valid longjmp targets for Windows Control Flow Guard.
753     addPass(createCFGuardLongjmpPass());
754     // Identify valid eh continuation targets for Windows EHCont Guard.
755     addPass(createEHContGuardCatchretPass());
756   }
757 
758   if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables)
759     addPass(createAArch64CompressJumpTablesPass());
760 
761   if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
762       TM->getTargetTriple().isOSBinFormatMachO())
763     addPass(createAArch64CollectLOHPass());
764 }
765 
766 void AArch64PassConfig::addPreEmitPass2() {
767   // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo
768   // instructions are lowered to bundles as well.
769   addPass(createUnpackMachineBundles(nullptr));
770 }
771 
772 yaml::MachineFunctionInfo *
773 AArch64TargetMachine::createDefaultFuncInfoYAML() const {
774   return new yaml::AArch64FunctionInfo();
775 }
776 
777 yaml::MachineFunctionInfo *
778 AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
779   const auto *MFI = MF.getInfo<AArch64FunctionInfo>();
780   return new yaml::AArch64FunctionInfo(*MFI);
781 }
782 
783 bool AArch64TargetMachine::parseMachineFunctionInfo(
784     const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS,
785     SMDiagnostic &Error, SMRange &SourceRange) const {
786   const auto &YamlMFI =
787       reinterpret_cast<const yaml::AArch64FunctionInfo &>(MFI);
788   MachineFunction &MF = PFS.MF;
789   MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI);
790   return false;
791 }
792