xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64TargetMachine.cpp (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57) 
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 "AArch64MachineScheduler.h"
16 #include "AArch64MacroFusion.h"
17 #include "AArch64Subtarget.h"
18 #include "AArch64TargetObjectFile.h"
19 #include "AArch64TargetTransformInfo.h"
20 #include "MCTargetDesc/AArch64MCTargetDesc.h"
21 #include "TargetInfo/AArch64TargetInfo.h"
22 #include "llvm/ADT/STLExtras.h"
23 #include "llvm/ADT/Triple.h"
24 #include "llvm/Analysis/TargetTransformInfo.h"
25 #include "llvm/CodeGen/CFIFixup.h"
26 #include "llvm/CodeGen/CSEConfigBase.h"
27 #include "llvm/CodeGen/GlobalISel/CSEInfo.h"
28 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
29 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
30 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
31 #include "llvm/CodeGen/GlobalISel/LoadStoreOpt.h"
32 #include "llvm/CodeGen/GlobalISel/Localizer.h"
33 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
34 #include "llvm/CodeGen/MIRParser/MIParser.h"
35 #include "llvm/CodeGen/MachineScheduler.h"
36 #include "llvm/CodeGen/Passes.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetPassConfig.h"
39 #include "llvm/IR/Attributes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/InitializePasses.h"
42 #include "llvm/MC/MCAsmInfo.h"
43 #include "llvm/MC/MCTargetOptions.h"
44 #include "llvm/MC/TargetRegistry.h"
45 #include "llvm/Pass.h"
46 #include "llvm/Support/CodeGen.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Target/TargetLoweringObjectFile.h"
49 #include "llvm/Target/TargetOptions.h"
50 #include "llvm/Transforms/CFGuard.h"
51 #include "llvm/Transforms/Scalar.h"
52 #include <memory>
53 #include <optional>
54 #include <string>
55 
56 using namespace llvm;
57 
58 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp",
59                                 cl::desc("Enable the CCMP formation pass"),
60                                 cl::init(true), cl::Hidden);
61 
62 static cl::opt<bool>
63     EnableCondBrTuning("aarch64-enable-cond-br-tune",
64                        cl::desc("Enable the conditional branch tuning pass"),
65                        cl::init(true), cl::Hidden);
66 
67 static cl::opt<bool> EnableAArch64CopyPropagation(
68     "aarch64-enable-copy-propagation",
69     cl::desc("Enable the copy propagation with AArch64 copy instr"),
70     cl::init(true), cl::Hidden);
71 
72 static cl::opt<bool> EnableMCR("aarch64-enable-mcr",
73                                cl::desc("Enable the machine combiner pass"),
74                                cl::init(true), cl::Hidden);
75 
76 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress",
77                                           cl::desc("Suppress STP for AArch64"),
78                                           cl::init(true), cl::Hidden);
79 
80 static cl::opt<bool> EnableAdvSIMDScalar(
81     "aarch64-enable-simd-scalar",
82     cl::desc("Enable use of AdvSIMD scalar integer instructions"),
83     cl::init(false), cl::Hidden);
84 
85 static cl::opt<bool>
86     EnablePromoteConstant("aarch64-enable-promote-const",
87                           cl::desc("Enable the promote constant pass"),
88                           cl::init(true), cl::Hidden);
89 
90 static cl::opt<bool> EnableCollectLOH(
91     "aarch64-enable-collect-loh",
92     cl::desc("Enable the pass that emits the linker optimization hints (LOH)"),
93     cl::init(true), cl::Hidden);
94 
95 static cl::opt<bool>
96     EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden,
97                                   cl::desc("Enable the pass that removes dead"
98                                            " definitons and replaces stores to"
99                                            " them with stores to the zero"
100                                            " register"),
101                                   cl::init(true));
102 
103 static cl::opt<bool> EnableRedundantCopyElimination(
104     "aarch64-enable-copyelim",
105     cl::desc("Enable the redundant copy elimination pass"), cl::init(true),
106     cl::Hidden);
107 
108 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt",
109                                         cl::desc("Enable the load/store pair"
110                                                  " optimization pass"),
111                                         cl::init(true), cl::Hidden);
112 
113 static cl::opt<bool> EnableAtomicTidy(
114     "aarch64-enable-atomic-cfg-tidy", cl::Hidden,
115     cl::desc("Run SimplifyCFG after expanding atomic operations"
116              " to make use of cmpxchg flow-based information"),
117     cl::init(true));
118 
119 static cl::opt<bool>
120 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
121                         cl::desc("Run early if-conversion"),
122                         cl::init(true));
123 
124 static cl::opt<bool>
125     EnableCondOpt("aarch64-enable-condopt",
126                   cl::desc("Enable the condition optimizer pass"),
127                   cl::init(true), cl::Hidden);
128 
129 static cl::opt<bool>
130     EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
131                  cl::desc("Enable optimizations on complex GEPs"),
132                  cl::init(false));
133 
134 static cl::opt<bool>
135     EnableSelectOpt("aarch64-select-opt", cl::Hidden,
136                     cl::desc("Enable select to branch optimizations"),
137                     cl::init(true));
138 
139 static cl::opt<bool>
140     BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true),
141                      cl::desc("Relax out of range conditional branches"));
142 
143 static cl::opt<bool> EnableCompressJumpTables(
144     "aarch64-enable-compress-jump-tables", cl::Hidden, cl::init(true),
145     cl::desc("Use smallest entry possible for jump tables"));
146 
147 // FIXME: Unify control over GlobalMerge.
148 static cl::opt<cl::boolOrDefault>
149     EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden,
150                       cl::desc("Enable the global merge pass"));
151 
152 static cl::opt<bool>
153     EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden,
154                            cl::desc("Enable the loop data prefetch pass"),
155                            cl::init(true));
156 
157 static cl::opt<int> EnableGlobalISelAtO(
158     "aarch64-enable-global-isel-at-O", cl::Hidden,
159     cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"),
160     cl::init(0));
161 
162 static cl::opt<bool>
163     EnableSVEIntrinsicOpts("aarch64-enable-sve-intrinsic-opts", cl::Hidden,
164                            cl::desc("Enable SVE intrinsic opts"),
165                            cl::init(true));
166 
167 static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix",
168                                          cl::init(true), cl::Hidden);
169 
170 static cl::opt<bool>
171     EnableBranchTargets("aarch64-enable-branch-targets", cl::Hidden,
172                         cl::desc("Enable the AArch64 branch target pass"),
173                         cl::init(true));
174 
175 static cl::opt<unsigned> SVEVectorBitsMaxOpt(
176     "aarch64-sve-vector-bits-max",
177     cl::desc("Assume SVE vector registers are at most this big, "
178              "with zero meaning no maximum size is assumed."),
179     cl::init(0), cl::Hidden);
180 
181 static cl::opt<unsigned> SVEVectorBitsMinOpt(
182     "aarch64-sve-vector-bits-min",
183     cl::desc("Assume SVE vector registers are at least this big, "
184              "with zero meaning no minimum size is assumed."),
185     cl::init(0), cl::Hidden);
186 
187 extern cl::opt<bool> EnableHomogeneousPrologEpilog;
188 
189 static cl::opt<bool> EnableGISelLoadStoreOptPreLegal(
190     "aarch64-enable-gisel-ldst-prelegal",
191     cl::desc("Enable GlobalISel's pre-legalizer load/store optimization pass"),
192     cl::init(true), cl::Hidden);
193 
194 static cl::opt<bool> EnableGISelLoadStoreOptPostLegal(
195     "aarch64-enable-gisel-ldst-postlegal",
196     cl::desc("Enable GlobalISel's post-legalizer load/store optimization pass"),
197     cl::init(false), cl::Hidden);
198 
199 extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAArch64Target() {
200   // Register the target.
201   RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget());
202   RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget());
203   RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target());
204   RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target());
205   RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target());
206   auto PR = PassRegistry::getPassRegistry();
207   initializeGlobalISel(*PR);
208   initializeAArch64A53Fix835769Pass(*PR);
209   initializeAArch64A57FPLoadBalancingPass(*PR);
210   initializeAArch64AdvSIMDScalarPass(*PR);
211   initializeAArch64BranchTargetsPass(*PR);
212   initializeAArch64CollectLOHPass(*PR);
213   initializeAArch64CompressJumpTablesPass(*PR);
214   initializeAArch64ConditionalComparesPass(*PR);
215   initializeAArch64ConditionOptimizerPass(*PR);
216   initializeAArch64DeadRegisterDefinitionsPass(*PR);
217   initializeAArch64ExpandPseudoPass(*PR);
218   initializeAArch64KCFIPass(*PR);
219   initializeAArch64LoadStoreOptPass(*PR);
220   initializeAArch64MIPeepholeOptPass(*PR);
221   initializeAArch64SIMDInstrOptPass(*PR);
222   initializeAArch64O0PreLegalizerCombinerPass(*PR);
223   initializeAArch64PreLegalizerCombinerPass(*PR);
224   initializeAArch64PostLegalizerCombinerPass(*PR);
225   initializeAArch64PostLegalizerLoweringPass(*PR);
226   initializeAArch64PostSelectOptimizePass(*PR);
227   initializeAArch64PromoteConstantPass(*PR);
228   initializeAArch64RedundantCopyEliminationPass(*PR);
229   initializeAArch64StorePairSuppressPass(*PR);
230   initializeFalkorHWPFFixPass(*PR);
231   initializeFalkorMarkStridedAccessesLegacyPass(*PR);
232   initializeLDTLSCleanupPass(*PR);
233   initializeSMEABIPass(*PR);
234   initializeSVEIntrinsicOptsPass(*PR);
235   initializeAArch64SpeculationHardeningPass(*PR);
236   initializeAArch64SLSHardeningPass(*PR);
237   initializeAArch64StackTaggingPass(*PR);
238   initializeAArch64StackTaggingPreRAPass(*PR);
239   initializeAArch64LowerHomogeneousPrologEpilogPass(*PR);
240   initializeAArch64DAGToDAGISelPass(*PR);
241 }
242 
243 //===----------------------------------------------------------------------===//
244 // AArch64 Lowering public interface.
245 //===----------------------------------------------------------------------===//
246 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
247   if (TT.isOSBinFormatMachO())
248     return std::make_unique<AArch64_MachoTargetObjectFile>();
249   if (TT.isOSBinFormatCOFF())
250     return std::make_unique<AArch64_COFFTargetObjectFile>();
251 
252   return std::make_unique<AArch64_ELFTargetObjectFile>();
253 }
254 
255 // Helper function to build a DataLayout string
256 static std::string computeDataLayout(const Triple &TT,
257                                      const MCTargetOptions &Options,
258                                      bool LittleEndian) {
259   if (TT.isOSBinFormatMachO()) {
260     if (TT.getArch() == Triple::aarch64_32)
261       return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128";
262     return "e-m:o-i64:64-i128:128-n32:64-S128";
263   }
264   if (TT.isOSBinFormatCOFF())
265     return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128";
266   std::string Endian = LittleEndian ? "e" : "E";
267   std::string Ptr32 = TT.getEnvironment() == Triple::GNUILP32 ? "-p:32:32" : "";
268   return Endian + "-m:e" + Ptr32 +
269          "-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
270 }
271 
272 static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) {
273   if (CPU.empty() && TT.isArm64e())
274     return "apple-a12";
275   return CPU;
276 }
277 
278 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
279                                            std::optional<Reloc::Model> RM) {
280   // AArch64 Darwin and Windows are always PIC.
281   if (TT.isOSDarwin() || TT.isOSWindows())
282     return Reloc::PIC_;
283   // On ELF platforms the default static relocation model has a smart enough
284   // linker to cope with referencing external symbols defined in a shared
285   // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
286   if (!RM || *RM == Reloc::DynamicNoPIC)
287     return Reloc::Static;
288   return *RM;
289 }
290 
291 static CodeModel::Model
292 getEffectiveAArch64CodeModel(const Triple &TT,
293                              std::optional<CodeModel::Model> CM, bool JIT) {
294   if (CM) {
295     if (*CM != CodeModel::Small && *CM != CodeModel::Tiny &&
296         *CM != CodeModel::Large) {
297       report_fatal_error(
298           "Only small, tiny and large code models are allowed on AArch64");
299     } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF())
300       report_fatal_error("tiny code model is only supported on ELF");
301     return *CM;
302   }
303   // The default MCJIT memory managers make no guarantees about where they can
304   // find an executable page; JITed code needs to be able to refer to globals
305   // no matter how far away they are.
306   // We should set the CodeModel::Small for Windows ARM64 in JIT mode,
307   // since with large code model LLVM generating 4 MOV instructions, and
308   // Windows doesn't support relocating these long branch (4 MOVs).
309   if (JIT && !TT.isOSWindows())
310     return CodeModel::Large;
311   return CodeModel::Small;
312 }
313 
314 /// Create an AArch64 architecture model.
315 ///
316 AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT,
317                                            StringRef CPU, StringRef FS,
318                                            const TargetOptions &Options,
319                                            std::optional<Reloc::Model> RM,
320                                            std::optional<CodeModel::Model> CM,
321                                            CodeGenOpt::Level OL, bool JIT,
322                                            bool LittleEndian)
323     : LLVMTargetMachine(T,
324                         computeDataLayout(TT, Options.MCOptions, LittleEndian),
325                         TT, computeDefaultCPU(TT, CPU), FS, Options,
326                         getEffectiveRelocModel(TT, RM),
327                         getEffectiveAArch64CodeModel(TT, CM, JIT), OL),
328       TLOF(createTLOF(getTargetTriple())), isLittle(LittleEndian) {
329   initAsmInfo();
330 
331   if (TT.isOSBinFormatMachO()) {
332     this->Options.TrapUnreachable = true;
333     this->Options.NoTrapAfterNoreturn = true;
334   }
335 
336   if (getMCAsmInfo()->usesWindowsCFI()) {
337     // Unwinding can get confused if the last instruction in an
338     // exception-handling region (function, funclet, try block, etc.)
339     // is a call.
340     //
341     // FIXME: We could elide the trap if the next instruction would be in
342     // the same region anyway.
343     this->Options.TrapUnreachable = true;
344   }
345 
346   if (this->Options.TLSSize == 0) // default
347     this->Options.TLSSize = 24;
348   if ((getCodeModel() == CodeModel::Small ||
349        getCodeModel() == CodeModel::Kernel) &&
350       this->Options.TLSSize > 32)
351     // for the small (and kernel) code model, the maximum TLS size is 4GiB
352     this->Options.TLSSize = 32;
353   else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24)
354     // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB)
355     this->Options.TLSSize = 24;
356 
357   // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
358   // MachO/CodeModel::Large, which GlobalISel does not support.
359   if (getOptLevel() <= EnableGlobalISelAtO &&
360       TT.getArch() != Triple::aarch64_32 &&
361       TT.getEnvironment() != Triple::GNUILP32 &&
362       !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) {
363     setGlobalISel(true);
364     setGlobalISelAbort(GlobalISelAbortMode::Disable);
365   }
366 
367   // AArch64 supports the MachineOutliner.
368   setMachineOutliner(true);
369 
370   // AArch64 supports default outlining behaviour.
371   setSupportsDefaultOutlining(true);
372 
373   // AArch64 supports the debug entry values.
374   setSupportsDebugEntryValues(true);
375 
376   // AArch64 supports fixing up the DWARF unwind information.
377   if (!getMCAsmInfo()->usesWindowsCFI())
378     setCFIFixup(true);
379 }
380 
381 AArch64TargetMachine::~AArch64TargetMachine() = default;
382 
383 const AArch64Subtarget *
384 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
385   Attribute CPUAttr = F.getFnAttribute("target-cpu");
386   Attribute TuneAttr = F.getFnAttribute("tune-cpu");
387   Attribute FSAttr = F.getFnAttribute("target-features");
388 
389   StringRef CPU = CPUAttr.isValid() ? CPUAttr.getValueAsString() : TargetCPU;
390   StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString() : CPU;
391   StringRef FS = FSAttr.isValid() ? FSAttr.getValueAsString() : TargetFS;
392 
393   bool StreamingSVEModeDisabled =
394       !F.hasFnAttribute("aarch64_pstate_sm_enabled") &&
395       !F.hasFnAttribute("aarch64_pstate_sm_compatible") &&
396       !F.hasFnAttribute("aarch64_pstate_sm_body");
397 
398   unsigned MinSVEVectorSize = 0;
399   unsigned MaxSVEVectorSize = 0;
400   Attribute VScaleRangeAttr = F.getFnAttribute(Attribute::VScaleRange);
401   if (VScaleRangeAttr.isValid()) {
402     std::optional<unsigned> VScaleMax = VScaleRangeAttr.getVScaleRangeMax();
403     MinSVEVectorSize = VScaleRangeAttr.getVScaleRangeMin() * 128;
404     MaxSVEVectorSize = VScaleMax ? *VScaleMax * 128 : 0;
405   } else {
406     MinSVEVectorSize = SVEVectorBitsMinOpt;
407     MaxSVEVectorSize = SVEVectorBitsMaxOpt;
408   }
409 
410   assert(MinSVEVectorSize % 128 == 0 &&
411          "SVE requires vector length in multiples of 128!");
412   assert(MaxSVEVectorSize % 128 == 0 &&
413          "SVE requires vector length in multiples of 128!");
414   assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) &&
415          "Minimum SVE vector size should not be larger than its maximum!");
416 
417   // Sanitize user input in case of no asserts
418   if (MaxSVEVectorSize == 0)
419     MinSVEVectorSize = (MinSVEVectorSize / 128) * 128;
420   else {
421     MinSVEVectorSize =
422         (std::min(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
423     MaxSVEVectorSize =
424         (std::max(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
425   }
426 
427   SmallString<512> Key;
428   raw_svector_ostream(Key) << "SVEMin" << MinSVEVectorSize << "SVEMax"
429                            << MaxSVEVectorSize << "StreamingSVEModeDisabled="
430                            << StreamingSVEModeDisabled << CPU << TuneCPU << FS;
431 
432   auto &I = SubtargetMap[Key];
433   if (!I) {
434     // This needs to be done before we create a new subtarget since any
435     // creation will depend on the TM and the code generation flags on the
436     // function that reside in TargetOptions.
437     resetTargetOptions(F);
438     I = std::make_unique<AArch64Subtarget>(
439         TargetTriple, CPU, TuneCPU, FS, *this, isLittle, MinSVEVectorSize,
440         MaxSVEVectorSize, StreamingSVEModeDisabled);
441   }
442   return I.get();
443 }
444 
445 void AArch64leTargetMachine::anchor() { }
446 
447 AArch64leTargetMachine::AArch64leTargetMachine(
448     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
449     const TargetOptions &Options, std::optional<Reloc::Model> RM,
450     std::optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
451     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {}
452 
453 void AArch64beTargetMachine::anchor() { }
454 
455 AArch64beTargetMachine::AArch64beTargetMachine(
456     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
457     const TargetOptions &Options, std::optional<Reloc::Model> RM,
458     std::optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
459     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {}
460 
461 namespace {
462 
463 /// AArch64 Code Generator Pass Configuration Options.
464 class AArch64PassConfig : public TargetPassConfig {
465 public:
466   AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
467       : TargetPassConfig(TM, PM) {
468     if (TM.getOptLevel() != CodeGenOpt::None)
469       substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
470   }
471 
472   AArch64TargetMachine &getAArch64TargetMachine() const {
473     return getTM<AArch64TargetMachine>();
474   }
475 
476   ScheduleDAGInstrs *
477   createMachineScheduler(MachineSchedContext *C) const override {
478     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
479     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
480     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
481     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
482     if (ST.hasFusion())
483       DAG->addMutation(createAArch64MacroFusionDAGMutation());
484     return DAG;
485   }
486 
487   ScheduleDAGInstrs *
488   createPostMachineScheduler(MachineSchedContext *C) const override {
489     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
490     ScheduleDAGMI *DAG =
491         new ScheduleDAGMI(C, std::make_unique<AArch64PostRASchedStrategy>(C),
492                           /* RemoveKillFlags=*/true);
493     if (ST.hasFusion()) {
494       // Run the Macro Fusion after RA again since literals are expanded from
495       // pseudos then (v. addPreSched2()).
496       DAG->addMutation(createAArch64MacroFusionDAGMutation());
497       return DAG;
498     }
499 
500     return DAG;
501   }
502 
503   void addIRPasses()  override;
504   bool addPreISel() override;
505   void addCodeGenPrepare() override;
506   bool addInstSelector() override;
507   bool addIRTranslator() override;
508   void addPreLegalizeMachineIR() override;
509   bool addLegalizeMachineIR() override;
510   void addPreRegBankSelect() override;
511   bool addRegBankSelect() override;
512   void addPreGlobalInstructionSelect() override;
513   bool addGlobalInstructionSelect() override;
514   void addMachineSSAOptimization() override;
515   bool addILPOpts() override;
516   void addPreRegAlloc() override;
517   void addPostRegAlloc() override;
518   void addPreSched2() override;
519   void addPreEmitPass() override;
520   void addPreEmitPass2() override;
521 
522   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
523 };
524 
525 } // end anonymous namespace
526 
527 TargetTransformInfo
528 AArch64TargetMachine::getTargetTransformInfo(const Function &F) const {
529   return TargetTransformInfo(AArch64TTIImpl(this, F));
530 }
531 
532 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
533   return new AArch64PassConfig(*this, PM);
534 }
535 
536 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const {
537   return getStandardCSEConfigForOpt(TM->getOptLevel());
538 }
539 
540 void AArch64PassConfig::addIRPasses() {
541   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
542   // ourselves.
543   addPass(createAtomicExpandPass());
544 
545   // Expand any SVE vector library calls that we can't code generate directly.
546   if (EnableSVEIntrinsicOpts && TM->getOptLevel() == CodeGenOpt::Aggressive)
547     addPass(createSVEIntrinsicOptsPass());
548 
549   // Cmpxchg instructions are often used with a subsequent comparison to
550   // determine whether it succeeded. We can exploit existing control-flow in
551   // ldrex/strex loops to simplify this, but it needs tidying up.
552   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
553     addPass(createCFGSimplificationPass(SimplifyCFGOptions()
554                                             .forwardSwitchCondToPhi(true)
555                                             .convertSwitchRangeToICmp(true)
556                                             .convertSwitchToLookupTable(true)
557                                             .needCanonicalLoops(false)
558                                             .hoistCommonInsts(true)
559                                             .sinkCommonInsts(true)));
560 
561   // Run LoopDataPrefetch
562   //
563   // Run this before LSR to remove the multiplies involved in computing the
564   // pointer values N iterations ahead.
565   if (TM->getOptLevel() != CodeGenOpt::None) {
566     if (EnableLoopDataPrefetch)
567       addPass(createLoopDataPrefetchPass());
568     if (EnableFalkorHWPFFix)
569       addPass(createFalkorMarkStridedAccessesPass());
570   }
571 
572   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
573     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
574     // and lower a GEP with multiple indices to either arithmetic operations or
575     // multiple GEPs with single index.
576     addPass(createSeparateConstOffsetFromGEPPass(true));
577     // Call EarlyCSE pass to find and remove subexpressions in the lowered
578     // result.
579     addPass(createEarlyCSEPass());
580     // Do loop invariant code motion in case part of the lowered result is
581     // invariant.
582     addPass(createLICMPass());
583   }
584 
585   TargetPassConfig::addIRPasses();
586 
587   if (getOptLevel() == CodeGenOpt::Aggressive && EnableSelectOpt)
588     addPass(createSelectOptimizePass());
589 
590   addPass(createAArch64StackTaggingPass(
591       /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None));
592 
593   // Match complex arithmetic patterns
594   if (TM->getOptLevel() >= CodeGenOpt::Default)
595     addPass(createComplexDeinterleavingPass(TM));
596 
597   // Match interleaved memory accesses to ldN/stN intrinsics.
598   if (TM->getOptLevel() != CodeGenOpt::None) {
599     addPass(createInterleavedLoadCombinePass());
600     addPass(createInterleavedAccessPass());
601   }
602 
603   // Expand any functions marked with SME attributes which require special
604   // changes for the calling convention or that require the lazy-saving
605   // mechanism specified in the SME ABI.
606   addPass(createSMEABIPass());
607 
608   // Add Control Flow Guard checks.
609   if (TM->getTargetTriple().isOSWindows())
610     addPass(createCFGuardCheckPass());
611 
612   if (TM->Options.JMCInstrument)
613     addPass(createJMCInstrumenterPass());
614 }
615 
616 // Pass Pipeline Configuration
617 bool AArch64PassConfig::addPreISel() {
618   // Run promote constant before global merge, so that the promoted constants
619   // get a chance to be merged
620   if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
621     addPass(createAArch64PromoteConstantPass());
622   // FIXME: On AArch64, this depends on the type.
623   // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
624   // and the offset has to be a multiple of the related size in bytes.
625   if ((TM->getOptLevel() != CodeGenOpt::None &&
626        EnableGlobalMerge == cl::BOU_UNSET) ||
627       EnableGlobalMerge == cl::BOU_TRUE) {
628     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
629                                (EnableGlobalMerge == cl::BOU_UNSET);
630 
631     // Merging of extern globals is enabled by default on non-Mach-O as we
632     // expect it to be generally either beneficial or harmless. On Mach-O it
633     // is disabled as we emit the .subsections_via_symbols directive which
634     // means that merging extern globals is not safe.
635     bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
636 
637     // FIXME: extern global merging is only enabled when we optimise for size
638     // because there are some regressions with it also enabled for performance.
639     if (!OnlyOptimizeForSize)
640       MergeExternalByDefault = false;
641 
642     addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize,
643                                   MergeExternalByDefault));
644   }
645 
646   return false;
647 }
648 
649 void AArch64PassConfig::addCodeGenPrepare() {
650   if (getOptLevel() != CodeGenOpt::None)
651     addPass(createTypePromotionLegacyPass());
652   TargetPassConfig::addCodeGenPrepare();
653 }
654 
655 bool AArch64PassConfig::addInstSelector() {
656   addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
657 
658   // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
659   // references to _TLS_MODULE_BASE_ as possible.
660   if (TM->getTargetTriple().isOSBinFormatELF() &&
661       getOptLevel() != CodeGenOpt::None)
662     addPass(createAArch64CleanupLocalDynamicTLSPass());
663 
664   return false;
665 }
666 
667 bool AArch64PassConfig::addIRTranslator() {
668   addPass(new IRTranslator(getOptLevel()));
669   return false;
670 }
671 
672 void AArch64PassConfig::addPreLegalizeMachineIR() {
673   if (getOptLevel() == CodeGenOpt::None)
674     addPass(createAArch64O0PreLegalizerCombiner());
675   else {
676     addPass(createAArch64PreLegalizerCombiner());
677     if (EnableGISelLoadStoreOptPreLegal)
678       addPass(new LoadStoreOpt());
679   }
680 }
681 
682 bool AArch64PassConfig::addLegalizeMachineIR() {
683   addPass(new Legalizer());
684   return false;
685 }
686 
687 void AArch64PassConfig::addPreRegBankSelect() {
688   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
689   if (!IsOptNone) {
690     addPass(createAArch64PostLegalizerCombiner(IsOptNone));
691     if (EnableGISelLoadStoreOptPostLegal)
692       addPass(new LoadStoreOpt());
693   }
694   addPass(createAArch64PostLegalizerLowering());
695 }
696 
697 bool AArch64PassConfig::addRegBankSelect() {
698   addPass(new RegBankSelect());
699   return false;
700 }
701 
702 void AArch64PassConfig::addPreGlobalInstructionSelect() {
703   addPass(new Localizer());
704 }
705 
706 bool AArch64PassConfig::addGlobalInstructionSelect() {
707   addPass(new InstructionSelect(getOptLevel()));
708   if (getOptLevel() != CodeGenOpt::None)
709     addPass(createAArch64PostSelectOptimize());
710   return false;
711 }
712 
713 void AArch64PassConfig::addMachineSSAOptimization() {
714   // Run default MachineSSAOptimization first.
715   TargetPassConfig::addMachineSSAOptimization();
716 
717   if (TM->getOptLevel() != CodeGenOpt::None)
718     addPass(createAArch64MIPeepholeOptPass());
719 }
720 
721 bool AArch64PassConfig::addILPOpts() {
722   if (EnableCondOpt)
723     addPass(createAArch64ConditionOptimizerPass());
724   if (EnableCCMP)
725     addPass(createAArch64ConditionalCompares());
726   if (EnableMCR)
727     addPass(&MachineCombinerID);
728   if (EnableCondBrTuning)
729     addPass(createAArch64CondBrTuning());
730   if (EnableEarlyIfConversion)
731     addPass(&EarlyIfConverterID);
732   if (EnableStPairSuppress)
733     addPass(createAArch64StorePairSuppressPass());
734   addPass(createAArch64SIMDInstrOptPass());
735   if (TM->getOptLevel() != CodeGenOpt::None)
736     addPass(createAArch64StackTaggingPreRAPass());
737   return true;
738 }
739 
740 void AArch64PassConfig::addPreRegAlloc() {
741   // Change dead register definitions to refer to the zero register.
742   if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
743     addPass(createAArch64DeadRegisterDefinitions());
744 
745   // Use AdvSIMD scalar instructions whenever profitable.
746   if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
747     addPass(createAArch64AdvSIMDScalar());
748     // The AdvSIMD pass may produce copies that can be rewritten to
749     // be register coalescer friendly.
750     addPass(&PeepholeOptimizerID);
751   }
752 }
753 
754 void AArch64PassConfig::addPostRegAlloc() {
755   // Remove redundant copy instructions.
756   if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
757     addPass(createAArch64RedundantCopyEliminationPass());
758 
759   if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
760     // Improve performance for some FP/SIMD code for A57.
761     addPass(createAArch64A57FPLoadBalancing());
762 }
763 
764 void AArch64PassConfig::addPreSched2() {
765   // Lower homogeneous frame instructions
766   if (EnableHomogeneousPrologEpilog)
767     addPass(createAArch64LowerHomogeneousPrologEpilogPass());
768   // Expand some pseudo instructions to allow proper scheduling.
769   addPass(createAArch64ExpandPseudoPass());
770   // Use load/store pair instructions when possible.
771   if (TM->getOptLevel() != CodeGenOpt::None) {
772     if (EnableLoadStoreOpt)
773       addPass(createAArch64LoadStoreOptimizationPass());
774   }
775   // Emit KCFI checks for indirect calls.
776   addPass(createAArch64KCFIPass());
777 
778   // The AArch64SpeculationHardeningPass destroys dominator tree and natural
779   // loop info, which is needed for the FalkorHWPFFixPass and also later on.
780   // Therefore, run the AArch64SpeculationHardeningPass before the
781   // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop
782   // info.
783   addPass(createAArch64SpeculationHardeningPass());
784 
785   addPass(createAArch64IndirectThunks());
786   addPass(createAArch64SLSHardeningPass());
787 
788   if (TM->getOptLevel() != CodeGenOpt::None) {
789     if (EnableFalkorHWPFFix)
790       addPass(createFalkorHWPFFixPass());
791   }
792 }
793 
794 void AArch64PassConfig::addPreEmitPass() {
795   // Machine Block Placement might have created new opportunities when run
796   // at O3, where the Tail Duplication Threshold is set to 4 instructions.
797   // Run the load/store optimizer once more.
798   if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt)
799     addPass(createAArch64LoadStoreOptimizationPass());
800 
801   if (TM->getOptLevel() >= CodeGenOpt::Aggressive &&
802       EnableAArch64CopyPropagation)
803     addPass(createMachineCopyPropagationPass(true));
804 
805   addPass(createAArch64A53Fix835769());
806 
807   if (EnableBranchTargets)
808     addPass(createAArch64BranchTargetsPass());
809 
810   // Relax conditional branch instructions if they're otherwise out of
811   // range of their destination.
812   if (BranchRelaxation)
813     addPass(&BranchRelaxationPassID);
814 
815   if (TM->getTargetTriple().isOSWindows()) {
816     // Identify valid longjmp targets for Windows Control Flow Guard.
817     addPass(createCFGuardLongjmpPass());
818     // Identify valid eh continuation targets for Windows EHCont Guard.
819     addPass(createEHContGuardCatchretPass());
820   }
821 
822   if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables)
823     addPass(createAArch64CompressJumpTablesPass());
824 
825   if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
826       TM->getTargetTriple().isOSBinFormatMachO())
827     addPass(createAArch64CollectLOHPass());
828 }
829 
830 void AArch64PassConfig::addPreEmitPass2() {
831   // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo
832   // instructions are lowered to bundles as well.
833   addPass(createUnpackMachineBundles(nullptr));
834 }
835 
836 MachineFunctionInfo *AArch64TargetMachine::createMachineFunctionInfo(
837     BumpPtrAllocator &Allocator, const Function &F,
838     const TargetSubtargetInfo *STI) const {
839   return AArch64FunctionInfo::create<AArch64FunctionInfo>(
840       Allocator, F, static_cast<const AArch64Subtarget *>(STI));
841 }
842 
843 yaml::MachineFunctionInfo *
844 AArch64TargetMachine::createDefaultFuncInfoYAML() const {
845   return new yaml::AArch64FunctionInfo();
846 }
847 
848 yaml::MachineFunctionInfo *
849 AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
850   const auto *MFI = MF.getInfo<AArch64FunctionInfo>();
851   return new yaml::AArch64FunctionInfo(*MFI);
852 }
853 
854 bool AArch64TargetMachine::parseMachineFunctionInfo(
855     const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS,
856     SMDiagnostic &Error, SMRange &SourceRange) const {
857   const auto &YamlMFI = static_cast<const yaml::AArch64FunctionInfo &>(MFI);
858   MachineFunction &MF = PFS.MF;
859   MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI);
860   return false;
861 }
862