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 "AArch64AsmPrinter.h"
15#include "AArch64MachineFunctionInfo.h"
16#include "AArch64MachineScheduler.h"
17#include "AArch64MacroFusion.h"
18#include "AArch64Subtarget.h"
19#include "AArch64TargetObjectFile.h"
20#include "AArch64TargetTransformInfo.h"
21#include "MCTargetDesc/AArch64MCTargetDesc.h"
22#include "TargetInfo/AArch64TargetInfo.h"
23#include "llvm/ADT/StringExtras.h"
24#include "llvm/Analysis/TargetTransformInfo.h"
25#include "llvm/Analysis/ValueTracking.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/Option/LibraryOptions.h"
46#include "llvm/Pass.h"
47#include "llvm/Passes/PassBuilder.h"
48#include "llvm/Support/CodeGen.h"
49#include "llvm/Support/CommandLine.h"
50#include "llvm/Support/Compiler.h"
51#include "llvm/Target/TargetLoweringObjectFile.h"
52#include "llvm/Target/TargetOptions.h"
53#include "llvm/TargetParser/Triple.h"
54#include "llvm/Transforms/CFGuard.h"
55#include "llvm/Transforms/Scalar.h"
56#include "llvm/Transforms/Utils/LowerIFunc.h"
57#include "llvm/Transforms/Vectorize/LoopIdiomVectorize.h"
58#include <memory>
59
60using namespace llvm;
61
62extern "C" LLVM_ABI LLVM_EXTERNAL_VISIBILITY void
63LLVMInitializeAArch64Target() {
64 // Register the target.
65 RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget());
66 RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget());
67 RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target());
68 RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target());
69 RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target());
70 static opt::RegisterLibraryOptions<AArch64Options> O;
71 auto &PR = *PassRegistry::getPassRegistry();
72 initializeGlobalISel(PR);
73 initializeAArch64A53Fix835769LegacyPass(PR);
74 initializeAArch64A57FPLoadBalancingLegacyPass(PR);
75 initializeAArch64CodeLayoutOptPass(PR);
76 initializeAArch64AdvSIMDScalarLegacyPass(PR);
77 initializeAArch64AsmPrinterPass(PR);
78 initializeAArch64BranchTargetsLegacyPass(PR);
79 initializeAArch64CollectLOHLegacyPass(PR);
80 initializeAArch64CompressJumpTablesLegacyPass(PR);
81 initializeAArch64ConditionalComparesLegacyPass(PR);
82 initializeAArch64ConditionOptimizerLegacyPass(PR);
83 initializeAArch64DeadRegisterDefinitionsLegacyPass(PR);
84 initializeAArch64ExpandPseudoLegacyPass(PR);
85 initializeAArch64LoadStoreOptLegacyPass(PR);
86 initializeAArch64MIPeepholeOptLegacyPass(PR);
87 initializeAArch64PTrueCoalescingLegacyPass(PR);
88 initializeAArch64SIMDInstrOptLegacyPass(PR);
89 initializeAArch64O0PreLegalizerCombinerLegacyPass(PR);
90 initializeAArch64PredicateAsCounterLoopRewritesPass(PR);
91 initializeAArch64PreLegalizerCombinerLegacyPass(PR);
92 initializeAArch64PointerAuthLegacyPass(PR);
93 initializeAArch64PostCoalescerLegacyPass(PR);
94 initializeAArch64PostLegalizerCombinerLegacyPass(PR);
95 initializeAArch64PostSelectOptimizeLegacyPass(PR);
96 initializeAArch64PostLegalizerLoweringLegacyPass(PR);
97 initializeAArch64PromoteConstantPass(PR);
98 initializeAArch64RedundantCopyEliminationLegacyPass(PR);
99 initializeAArch64RedundantCondBranchLegacyPass(PR);
100 initializeAArch64StorePairSuppressPass(PR);
101 initializeFalkorHWPFFixPass(PR);
102 initializeFalkorMarkStridedAccessesLegacyPass(PR);
103 initializeLDTLSCleanupPass(PR);
104 initializeMachineKCFILegacyPass(PR);
105 initializeMachineSMEABIPass(PR);
106 initializeAArch64SRLTDefineSuperRegsLegacyPass(PR);
107 initializeSMEPeepholeOptPass(PR);
108 initializeAArch64SpeculationHardeningPass(PR);
109 initializeAArch64SLSHardeningLegacyPass(PR);
110 initializeAArch64StackTaggingPass(PR);
111 initializeAArch64StackTaggingPreRALegacyPass(PR);
112 initializeAArch64LowerHomogeneousPrologEpilogLegacyPass(PR);
113 initializeAArch64DAGToDAGISelLegacyPass(PR);
114 initializeAArch64CondBrTuningPass(PR);
115 initializeAArch64Arm64ECCallLoweringPass(PR);
116 initializeSVEShuffleOptsPass(PR);
117}
118
119bool AArch64TargetMachine::isGlobalISelOptNone() const {
120 const bool GlobalISelFlag =
121 getCGPassBuilderOption(TM: *this).EnableGlobalISelOption ==
122 cl::boolOrDefault::BOU_TRUE;
123
124 return getOptLevel() == CodeGenOptLevel::None ||
125 (static_cast<unsigned>(getOptLevel()) >
126 static_cast<unsigned>(CLOpts.enable_global_isel_at_O) &&
127 !GlobalISelFlag);
128}
129
130void AArch64TargetMachine::reset() {
131 SubtargetMap.clear();
132 LastSubtarget.clear();
133}
134
135//===----------------------------------------------------------------------===//
136// AArch64 Lowering public interface.
137//===----------------------------------------------------------------------===//
138static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
139 if (TT.isOSBinFormatMachO())
140 return std::make_unique<AArch64_MachoTargetObjectFile>();
141 if (TT.isOSBinFormatCOFF())
142 return std::make_unique<AArch64_COFFTargetObjectFile>();
143
144 return std::make_unique<AArch64_ELFTargetObjectFile>();
145}
146
147static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) {
148 if (CPU.empty() && TT.isArm64e())
149 return "apple-a12";
150 return CPU;
151}
152
153static Reloc::Model getEffectiveRelocModel(const Triple &TT,
154 std::optional<Reloc::Model> RM) {
155 // AArch64 Darwin, Windows and UEFI are always PIC.
156 if (TT.isOSDarwin() || TT.isOSWindowsOrUEFI())
157 return Reloc::PIC_;
158 // On ELF platforms the default static relocation model has a smart enough
159 // linker to cope with referencing external symbols defined in a shared
160 // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
161 if (!RM || *RM == Reloc::DynamicNoPIC)
162 return Reloc::Static;
163 return *RM;
164}
165
166static CodeModel::Model
167getEffectiveAArch64CodeModel(const Triple &TT,
168 std::optional<CodeModel::Model> CM, bool JIT) {
169 if (CM) {
170 if (*CM != CodeModel::Small && *CM != CodeModel::Tiny &&
171 *CM != CodeModel::Large) {
172 report_fatal_error(
173 reason: "Only small, tiny and large code models are allowed on AArch64");
174 } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF()) {
175 report_fatal_error(reason: "tiny code model is only supported on ELF");
176 }
177 return *CM;
178 }
179 // The default MCJIT memory managers make no guarantees about where they can
180 // find an executable page; JITed code needs to be able to refer to globals
181 // no matter how far away they are.
182 // We should set the CodeModel::Small for Windows ARM64 in JIT mode,
183 // since with large code model LLVM generating 4 MOV instructions, and
184 // Windows doesn't support relocating these long branch (4 MOVs).
185 if (JIT && !TT.isOSWindows())
186 return CodeModel::Large;
187 return CodeModel::Small;
188}
189
190/// Create an AArch64 architecture model.
191///
192AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT,
193 StringRef CPU, StringRef FS,
194 const TargetOptions &Options,
195 std::optional<Reloc::Model> RM,
196 std::optional<CodeModel::Model> CM,
197 CodeGenOptLevel OL, bool JIT,
198 bool LittleEndian)
199 : CodeGenTargetMachineImpl(T, TT, computeDefaultCPU(TT, CPU), FS, Options,
200 getEffectiveRelocModel(TT, RM),
201 getEffectiveAArch64CodeModel(TT, CM, JIT), OL),
202 CLOpts(cl::LibraryOptionsContext::get<AArch64Options>(
203 C: Options.MCOptions.LibraryOptions)),
204 TLOF(createTLOF(TT: getTargetTriple())), isLittle(LittleEndian) {
205 initAsmInfo();
206
207 if (TT.isOSBinFormatMachO()) {
208 this->Options.TrapUnreachable = true;
209 this->Options.NoTrapAfterNoreturn = true;
210 }
211
212 if (this->Options.TLSSize == 0) // default
213 this->Options.TLSSize = 24;
214 if ((getCodeModel() == CodeModel::Small ||
215 getCodeModel() == CodeModel::Kernel) &&
216 this->Options.TLSSize > 32)
217 // for the small (and kernel) code model, the maximum TLS size is 4GiB
218 this->Options.TLSSize = 32;
219 else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24)
220 // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB)
221 this->Options.TLSSize = 24;
222
223 const bool TargetSupportsGISel =
224 TT.getArch() != Triple::aarch64_32 &&
225 TT.getEnvironment() != Triple::GNUILP32 &&
226 !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO());
227
228 const bool GlobalISelFlag =
229 getCGPassBuilderOption(TM: *this).EnableGlobalISelOption ==
230 cl::boolOrDefault::BOU_TRUE;
231
232 // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is
233 // MachO/CodeModel::Large, which GlobalISel does not support.
234 if (TargetSupportsGISel && CLOpts.enable_global_isel_at_O != -1 &&
235 (static_cast<int>(getOptLevel()) <= CLOpts.enable_global_isel_at_O ||
236 (!GlobalISelFlag && !Options.EnableGlobalISel))) {
237 setGlobalISel(true);
238 setGlobalISelAbort(GlobalISelAbortMode::Disable);
239 }
240
241 LLT::setUseExtended(true);
242
243 // AArch64 supports the MachineOutliner.
244 setMachineOutliner(true);
245
246 // AArch64 supports default outlining behaviour.
247 setSupportsDefaultOutlining(true);
248
249 // AArch64 supports the debug entry values.
250 setSupportsDebugEntryValues(true);
251
252 // AArch64 supports fixing up the DWARF unwind information.
253 // enableCFIFixup() checks per function for DWARF unwinding.
254 setCFIFixup(true);
255}
256
257unsigned AArch64TargetMachine::getEnableGlobalISelAtO() const {
258 return CLOpts.enable_global_isel_at_O;
259}
260
261AArch64TargetMachine::~AArch64TargetMachine() = default;
262
263const AArch64Subtarget *
264AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
265 AttributeSet FnAttrs = F.getAttributes().getFnAttrs();
266 const Module *M = F.getParent();
267 if (const AArch64Subtarget *ST = LastSubtarget.lookup(NewFnAttrs: FnAttrs, NewM: M))
268 return ST;
269
270 SubtargetAttrs Attrs;
271 Attrs.CPU = FnAttrs.getAttribute(Kind: "target-cpu");
272 Attrs.TuneCPU = FnAttrs.getAttribute(Kind: "tune-cpu");
273 Attrs.FS = FnAttrs.getAttribute(Kind: "target-features");
274 Attrs.SMEnabled = FnAttrs.getAttribute(Kind: "aarch64_pstate_sm_enabled");
275 Attrs.SMBody = FnAttrs.getAttribute(Kind: "aarch64_pstate_sm_body");
276 Attrs.SMCompatible = FnAttrs.getAttribute(Kind: "aarch64_pstate_sm_compatible");
277 Attrs.MinSize = FnAttrs.getAttribute(Kind: Attribute::MinSize);
278 Attrs.VScaleRange = FnAttrs.getAttribute(Kind: Attribute::VScaleRange);
279
280 if (const AArch64Subtarget *ST = LastSubtarget.lookup(NewFnAttrs: FnAttrs, NewM: M, NewAttrs: Attrs))
281 return ST;
282
283 StringRef CPU =
284 Attrs.CPU.isValid() ? Attrs.CPU.getValueAsString() : TargetCPU;
285 StringRef TuneCPU =
286 Attrs.TuneCPU.isValid() ? Attrs.TuneCPU.getValueAsString() : CPU;
287 StringRef FS = Attrs.FS.isValid() ? Attrs.FS.getValueAsString() : TargetFS;
288 bool HasMinSize = Attrs.MinSize.isValid();
289
290 bool IsStreaming = CLOpts.force_streaming || Attrs.SMEnabled.isValid() ||
291 Attrs.SMBody.isValid();
292 bool IsStreamingCompatible =
293 CLOpts.force_streaming_compatible || Attrs.SMCompatible.isValid();
294
295 unsigned MinSVEVectorSize = 0;
296 unsigned MaxSVEVectorSize = 0;
297 if (Attrs.VScaleRange.isValid()) {
298 ConstantRange CR = getVScaleRange(F: &F, BitWidth: 64);
299 MinSVEVectorSize = CR.getUnsignedMin().getZExtValue() * 128;
300 MaxSVEVectorSize = CR.getUnsignedMax().getZExtValue() * 128;
301 } else {
302 MinSVEVectorSize = CLOpts.sve_vector_bits_min;
303 MaxSVEVectorSize = CLOpts.sve_vector_bits_max;
304 }
305
306 assert(MinSVEVectorSize % 128 == 0 &&
307 "SVE requires vector length in multiples of 128!");
308 assert(MaxSVEVectorSize % 128 == 0 &&
309 "SVE requires vector length in multiples of 128!");
310 assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) &&
311 "Minimum SVE vector size should not be larger than its maximum!");
312
313 // Sanitize user input in case of no asserts
314 if (MaxSVEVectorSize != 0) {
315 MinSVEVectorSize = std::min(a: MinSVEVectorSize, b: MaxSVEVectorSize);
316 MaxSVEVectorSize = std::max(a: MinSVEVectorSize, b: MaxSVEVectorSize);
317 }
318
319 SmallString<512> Key;
320 // This lookup is hot during repeated TTI queries, so build the key directly
321 // instead of formatting through raw_svector_ostream.
322 Key += "SVEMin";
323 Key += utostr(X: MinSVEVectorSize);
324 Key += "SVEMax";
325 Key += utostr(X: MaxSVEVectorSize);
326 Key += "IsStreaming=";
327 Key += utostr(X: IsStreaming);
328 Key += "IsStreamingCompatible=";
329 Key += utostr(X: IsStreamingCompatible);
330 Key += CPU;
331 Key += TuneCPU;
332 Key += FS;
333 Key += "HasMinSize=";
334 Key += utostr(X: HasMinSize);
335
336 auto &I = SubtargetMap[Key];
337 if (!I) {
338 I = std::make_unique<AArch64Subtarget>(
339 args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args: *this, args: isLittle, args&: MinSVEVectorSize,
340 args&: MaxSVEVectorSize, args&: IsStreaming, args&: IsStreamingCompatible, args&: HasMinSize,
341 args: CLOpts.srlt_mitigate_sr2r);
342 }
343
344 if (IsStreaming && !I->hasSME())
345 reportFatalUsageError(reason: "streaming SVE functions require SME");
346
347 LastSubtarget.set(NewFnAttrs: FnAttrs, NewM: M, NewAttrs: Attrs, NewSubtarget: I.get());
348 return I.get();
349}
350
351// Encourage placing FORM_TRANSPOSED_REG immediately before the instruction that
352// uses/consumes it. This ensures its def has a short live range, which means
353// we're more likely to allocate registers its operands first (which works best
354// for the hints in AArch64RegisterInfo::getRegAllocationHints).
355static bool scheduleFormTransposedTupleAdjacentToUsers(
356 const TargetInstrInfo &TII, const TargetSubtargetInfo &TSI,
357 const MachineInstr *FirstMI, const MachineInstr &SecondMI,
358 const SDep *Dep) {
359 if (isNonDataDep(Dep))
360 return false;
361 return !FirstMI ||
362 FirstMI->getOpcode() == AArch64::FORM_TRANSPOSED_REG_TUPLE_X2_PSEUDO ||
363 FirstMI->getOpcode() == AArch64::FORM_TRANSPOSED_REG_TUPLE_X4_PSEUDO;
364}
365
366ScheduleDAGInstrs *
367AArch64TargetMachine::createMachineScheduler(MachineSchedContext *C) const {
368 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
369 ScheduleDAGMILive *DAG = createSchedLive(C);
370 DAG->addMutation(Mutation: createLoadClusterDAGMutation(TII: DAG->TII));
371 DAG->addMutation(Mutation: createStoreClusterDAGMutation(TII: DAG->TII));
372 if (ST.hasFusion())
373 DAG->addMutation(Mutation: createAArch64MacroFusionDAGMutation());
374 if (ST.hasSME() && ST.isStreaming())
375 DAG->addMutation(Mutation: createMacroFusionDAGMutation(
376 Predicates: scheduleFormTransposedTupleAdjacentToUsers));
377 return DAG;
378}
379
380ScheduleDAGInstrs *
381AArch64TargetMachine::createPostMachineScheduler(MachineSchedContext *C) const {
382 const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
383 ScheduleDAGMI *DAG = createSchedPostRA<AArch64PostRASchedStrategy>(C);
384 if (ST.hasFusion()) {
385 // Run the Macro Fusion after RA again since literals are expanded from
386 // pseudos then (v. addPreSched2()).
387 DAG->addMutation(Mutation: createAArch64MacroFusionDAGMutation());
388 return DAG;
389 }
390
391 return DAG;
392}
393
394size_t AArch64TargetMachine::clearLinkerOptimizationHints(
395 const SmallPtrSetImpl<MachineInstr *> &MIs) const {
396 if (MIs.empty())
397 return 0;
398 auto *MI = *MIs.begin();
399 auto *FuncInfo = MI->getMF()->getInfo<AArch64FunctionInfo>();
400 return FuncInfo->clearLinkerOptimizationHints(MIs);
401}
402
403void AArch64leTargetMachine::anchor() { }
404
405AArch64leTargetMachine::AArch64leTargetMachine(
406 const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
407 const TargetOptions &Options, std::optional<Reloc::Model> RM,
408 std::optional<CodeModel::Model> CM, CodeGenOptLevel OL, bool JIT)
409 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {}
410
411void AArch64beTargetMachine::anchor() { }
412
413AArch64beTargetMachine::AArch64beTargetMachine(
414 const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
415 const TargetOptions &Options, std::optional<Reloc::Model> RM,
416 std::optional<CodeModel::Model> CM, CodeGenOptLevel OL, bool JIT)
417 : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {}
418
419namespace {
420
421/// AArch64 Code Generator Pass Configuration Options.
422class AArch64PassConfig : public TargetPassConfig {
423 const AArch64Options &CLOpts;
424
425public:
426 AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
427 : TargetPassConfig(TM, PM), CLOpts(TM.getCLOpts()) {
428 if (TM.getOptLevel() != CodeGenOptLevel::None)
429 substitutePass(StandardID: &PostRASchedulerID, TargetID: &PostMachineSchedulerID);
430 setEnableSinkAndFold(CLOpts.enable_sink_fold);
431 }
432
433 AArch64TargetMachine &getAArch64TargetMachine() const {
434 return getTM<AArch64TargetMachine>();
435 }
436
437 void addIRPasses() override;
438 bool addPreISel() override;
439 void addCodeGenPrepare() override;
440 bool addInstSelector() override;
441 bool addIRTranslator() override;
442 void addPreLegalizeMachineIR() override;
443 bool addLegalizeMachineIR() override;
444 void addPreRegBankSelect() override;
445 bool addRegBankSelect() override;
446 bool addGlobalInstructionSelect() override;
447 void addMachineSSAOptimization() override;
448 bool addILPOpts() override;
449 void addPreRegAlloc() override;
450 void addPostRewrite() override;
451 void addPostRegAlloc() override;
452 void addPreSched2() override;
453 void addPreEmitPass() override;
454 void addPostBBSections() override;
455 void addPreEmitPass2() override;
456 bool addRegAssignAndRewriteOptimized() override;
457
458 std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
459};
460
461} // end anonymous namespace
462
463void AArch64TargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) {
464#define GET_PASS_REGISTRY "AArch64PassRegistry.def"
465#include "llvm/Passes/TargetPassRegistry.inc"
466
467 PB.registerLateLoopOptimizationsEPCallback(
468 C: [=](LoopPassManager &LPM, OptimizationLevel Level) {
469 if (Level != OptimizationLevel::O0)
470 LPM.addPass(Pass: LoopIdiomVectorizePass());
471 });
472 if (getTargetTriple().isOSWindowsOrUEFI())
473 PB.registerPipelineEarlySimplificationEPCallback(
474 C: [](ModulePassManager &PM, OptimizationLevel, ThinOrFullLTOPhase) {
475 PM.addPass(Pass: LowerIFuncPass());
476 });
477}
478
479TargetTransformInfo
480AArch64TargetMachine::getTargetTransformInfo(const Function &F) const {
481 return TargetTransformInfo(std::make_unique<AArch64TTIImpl>(args: this, args: F));
482}
483
484TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
485 return new AArch64PassConfig(*this, PM);
486}
487
488std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const {
489 return getStandardCSEConfigForOpt(Level: TM->getOptLevel());
490}
491
492void AArch64PassConfig::addIRPasses() {
493 // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
494 // ourselves.
495 addPass(P: createAtomicExpandLegacyPass());
496
497 if (getOptLevel() >= CodeGenOptLevel::Default &&
498 CLOpts.enable_subreg_liveness_tracking &&
499 CLOpts.enable_predicate_as_counter_loop_rewrites)
500 addPass(P: createAArch64PredicateAsCounterLoopRewritesPass());
501
502 // Cmpxchg instructions are often used with a subsequent comparison to
503 // determine whether it succeeded. We can exploit existing control-flow in
504 // ldrex/strex loops to simplify this, but it needs tidying up.
505 if (TM->getOptLevel() != CodeGenOptLevel::None &&
506 CLOpts.enable_atomic_cfg_tidy)
507 addPass(P: createCFGSimplificationPass(Options: SimplifyCFGOptions()
508 .forwardSwitchCondToPhi(B: true)
509 .convertSwitchRangeToICmp(B: true)
510 .convertSwitchToLookupTable(B: true)
511 .needCanonicalLoops(B: false)
512 .hoistCommonInsts(B: true)
513 .sinkCommonInsts(B: true)));
514
515 // Run LoopDataPrefetch
516 //
517 // Run this before LSR to remove the multiplies involved in computing the
518 // pointer values N iterations ahead.
519 if (TM->getOptLevel() != CodeGenOptLevel::None) {
520 if (CLOpts.enable_loop_data_prefetch)
521 addPass(P: createLoopDataPrefetchPass());
522 if (CLOpts.enable_falkor_hwpf_fix)
523 addPass(P: createFalkorMarkStridedAccessesPass());
524 }
525
526 if (CLOpts.enable_gep_opt) {
527 // Call SeparateConstOffsetFromGEP pass to extract constants within indices
528 // and lower a GEP with multiple indices to either arithmetic operations or
529 // multiple GEPs with single index.
530 addPass(P: createSeparateConstOffsetFromGEPPass(LowerGEP: true));
531 // Call EarlyCSE pass to find and remove subexpressions in the lowered
532 // result.
533 addPass(P: createEarlyCSEPass());
534 // Do loop invariant code motion in case part of the lowered result is
535 // invariant.
536 addPass(P: createLICMPass());
537 }
538
539 TargetPassConfig::addIRPasses();
540
541 if (getOptLevel() == CodeGenOptLevel::Aggressive && CLOpts.select_opt)
542 addPass(P: createSelectOptimizePass());
543
544 addPass(P: createAArch64StackTaggingPass(
545 CLOpts, /*IsOptNone=*/TM->getOptLevel() == CodeGenOptLevel::None));
546
547 // Try to use tbl in place of other shuffling operations if doing so would
548 // reduce the total number of instructions. Shuffle masks for big endian may
549 // be different, so require a little endian target.
550 if (getOptLevel() >= CodeGenOptLevel::Default &&
551 CLOpts.enable_sve_shuffle_opts && TM->getTargetTriple().isLittleEndian())
552 addPass(P: createSVEShuffleOptsPass());
553
554 // Match complex arithmetic patterns
555 if (TM->getOptLevel() >= CodeGenOptLevel::Default)
556 addPass(P: createComplexDeinterleavingPass(TM));
557
558 // Match interleaved memory accesses to ldN/stN intrinsics.
559 if (TM->getOptLevel() != CodeGenOptLevel::None) {
560 addPass(P: createInterleavedLoadCombinePass());
561 addPass(P: createInterleavedAccessPass());
562 }
563
564 // Add Control Flow Guard checks.
565 if (TM->getTargetTriple().isOSWindows()) {
566 if (TM->getTargetTriple().isWindowsArm64EC())
567 addPass(P: createAArch64Arm64ECCallLoweringPass());
568 else
569 addPass(P: createCFGuardPass());
570 }
571
572 if (TM->Options.JMCInstrument)
573 addPass(P: createJMCInstrumenterPass());
574}
575
576// Pass Pipeline Configuration
577bool AArch64PassConfig::addPreISel() {
578 // Run promote constant before global merge, so that the promoted constants
579 // get a chance to be merged
580 if (TM->getOptLevel() != CodeGenOptLevel::None && CLOpts.enable_promote_const)
581 addPass(P: createAArch64PromoteConstantPass());
582 // FIXME: On AArch64, this depends on the type.
583 // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
584 // and the offset has to be a multiple of the related size in bytes.
585 if (valueOr(X: CLOpts.enable_global_merge,
586 Default: TM->getOptLevel() != CodeGenOptLevel::None)) {
587 bool OnlyOptimizeForSize =
588 (TM->getOptLevel() < CodeGenOptLevel::Aggressive) &&
589 (CLOpts.enable_global_merge == BoolOrDefault::Default);
590
591 // Merging of extern globals is enabled by default on non-Mach-O as we
592 // expect it to be generally either beneficial or harmless. On Mach-O it
593 // is disabled as we emit the .subsections_via_symbols directive which
594 // means that merging extern globals is not safe.
595 bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
596 addPass(P: createGlobalMergePass(TM, MaximalOffset: 4095, OnlyOptimizeForSize,
597 MergeExternalByDefault));
598 }
599
600 return false;
601}
602
603void AArch64PassConfig::addCodeGenPrepare() {
604 if (getOptLevel() != CodeGenOptLevel::None)
605 addPass(P: createTypePromotionLegacyPass());
606 TargetPassConfig::addCodeGenPrepare();
607}
608
609bool AArch64PassConfig::addInstSelector() {
610 addPass(P: createAArch64ISelDag(TM&: getAArch64TargetMachine(), OptLevel: getOptLevel()));
611 return false;
612}
613
614bool AArch64PassConfig::addIRTranslator() {
615 addPass(P: new IRTranslatorLegacy(getOptLevel()));
616 return false;
617}
618
619void AArch64PassConfig::addPreLegalizeMachineIR() {
620 if (getAArch64TargetMachine().isGlobalISelOptNone()) {
621 addPass(P: createAArch64O0PreLegalizerCombiner());
622 } else {
623 addPass(P: createAArch64PreLegalizerCombiner());
624 addPass(P: new LocalizerLegacy());
625 if (CLOpts.enable_gisel_ldst_prelegal)
626 addPass(P: new LoadStoreOptLegacy());
627 }
628}
629
630bool AArch64PassConfig::addLegalizeMachineIR() {
631 addPass(P: new LegalizerLegacy());
632 return false;
633}
634
635void AArch64PassConfig::addPreRegBankSelect() {
636 const bool IsGlobalISelOptNone =
637 getAArch64TargetMachine().isGlobalISelOptNone();
638 if (!IsGlobalISelOptNone) {
639 addPass(P: createAArch64PostLegalizerCombinerLegacy(IsOptNone: IsGlobalISelOptNone));
640 if (CLOpts.enable_gisel_ldst_postlegal)
641 addPass(P: new LoadStoreOptLegacy());
642 }
643 addPass(P: createAArch64PostLegalizerLowering());
644}
645
646bool AArch64PassConfig::addRegBankSelect() {
647 addPass(P: new RegBankSelectLegacy());
648 return false;
649}
650
651bool AArch64PassConfig::addGlobalInstructionSelect() {
652 addPass(P: new InstructionSelectLegacy(getOptLevel()));
653 if (!getAArch64TargetMachine().isGlobalISelOptNone())
654 addPass(P: createAArch64PostSelectOptimize());
655
656 return false;
657}
658
659void AArch64PassConfig::addMachineSSAOptimization() {
660 // For ELF, cleanup any local-dynamic TLS accesses
661 // (i.e. combine as many references to _TLS_MODULE_BASE_ as possible.
662 if (TM->getTargetTriple().isOSBinFormatELF() &&
663 getOptLevel() != CodeGenOptLevel::None)
664 addPass(P: createAArch64CleanupLocalDynamicTLSPass());
665
666 if (TM->getOptLevel() != CodeGenOptLevel::None)
667 addPass(P: createMachineSMEABIPass());
668
669 if (TM->getOptLevel() != CodeGenOptLevel::None &&
670 CLOpts.enable_aarch64_sme_peephole_opt)
671 addPass(P: createSMEPeepholeOptPass());
672
673 // Run default MachineSSAOptimization first.
674 TargetPassConfig::addMachineSSAOptimization();
675
676 if (TM->getOptLevel() != CodeGenOptLevel::None) {
677 addPass(P: createAArch64MIPeepholeOptLegacyPass());
678 addPass(P: createAArch64PTrueCoalescingLegacyPass());
679 }
680}
681
682bool AArch64PassConfig::addILPOpts() {
683 if (CLOpts.enable_condopt)
684 addPass(P: createAArch64ConditionOptimizerLegacyPass());
685 if (CLOpts.enable_ccmp)
686 addPass(P: createAArch64ConditionalCompares());
687 if (CLOpts.enable_mcr)
688 addPass(PassID: &MachineCombinerID);
689 if (CLOpts.enable_cond_br_tune)
690 addPass(P: createAArch64CondBrTuning());
691 if (CLOpts.enable_early_ifcvt)
692 addPass(PassID: &EarlyIfConverterLegacyID);
693 if (CLOpts.enable_stp_suppress)
694 addPass(P: createAArch64StorePairSuppressPass());
695 addPass(P: createAArch64SIMDInstrOptPass());
696 if (TM->getOptLevel() != CodeGenOptLevel::None)
697 addPass(P: createAArch64StackTaggingPreRALegacyPass());
698 return true;
699}
700
701void AArch64PassConfig::addPreRegAlloc() {
702 if (TM->getOptLevel() == CodeGenOptLevel::None)
703 addPass(P: createMachineSMEABIPass());
704
705 // Change dead register definitions to refer to the zero register.
706 if (TM->getOptLevel() != CodeGenOptLevel::None && CLOpts.enable_dead_defs)
707 addPass(P: createAArch64DeadRegisterDefinitions());
708
709 // Use AdvSIMD scalar instructions whenever profitable.
710 if (TM->getOptLevel() != CodeGenOptLevel::None && CLOpts.enable_simd_scalar) {
711 addPass(P: createAArch64AdvSIMDScalar());
712 // The AdvSIMD pass may produce copies that can be rewritten to
713 // be register coalescer friendly.
714 addPass(PassID: &PeepholeOptimizerLegacyID);
715 }
716}
717
718void AArch64PassConfig::addPostRewrite() {
719 if (CLOpts.srlt_mitigate_sr2r)
720 addPass(P: createAArch64SRLTDefineSuperRegsLegacyPass());
721}
722
723void AArch64PassConfig::addPostRegAlloc() {
724 // Remove redundant copy instructions.
725 if (TM->getOptLevel() != CodeGenOptLevel::None && CLOpts.enable_copyelim)
726 addPass(P: createAArch64RedundantCopyEliminationPass());
727
728 if (TM->getOptLevel() != CodeGenOptLevel::None && usingDefaultRegAlloc())
729 // Improve performance for some FP/SIMD code for A57.
730 addPass(P: createAArch64A57FPLoadBalancingLegacyPass());
731}
732
733void AArch64PassConfig::addPreSched2() {
734 // Lower homogeneous frame instructions
735 if (CLOpts.homogeneous_prolog_epilog)
736 addPass(P: createAArch64LowerHomogeneousPrologEpilogPass());
737 // Expand some pseudo instructions to allow proper scheduling.
738 addPass(P: createAArch64ExpandPseudoLegacyPass());
739 // Use load/store pair instructions when possible.
740 if (TM->getOptLevel() != CodeGenOptLevel::None) {
741 if (CLOpts.enable_ldst_opt)
742 addPass(P: createAArch64LoadStoreOptLegacyPass());
743 }
744 // Emit KCFI checks for indirect calls.
745 addPass(P: createKCFIPass());
746
747 // The AArch64SpeculationHardeningPass destroys dominator tree and natural
748 // loop info, which is needed for the FalkorHWPFFixPass and also later on.
749 // Therefore, run the AArch64SpeculationHardeningPass before the
750 // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop
751 // info.
752 addPass(P: createAArch64SpeculationHardeningPass());
753
754 if (TM->getOptLevel() != CodeGenOptLevel::None) {
755 if (CLOpts.enable_falkor_hwpf_fix)
756 addPass(P: createFalkorHWPFFixPass());
757 }
758}
759
760void AArch64PassConfig::addPreEmitPass() {
761 // Machine Block Placement might have created new opportunities when run
762 // at O3, where the Tail Duplication Threshold is set to 4 instructions.
763 // Run the load/store optimizer once more.
764 if (TM->getOptLevel() >= CodeGenOptLevel::Aggressive &&
765 CLOpts.enable_ldst_opt)
766 addPass(P: createAArch64LoadStoreOptLegacyPass());
767
768 if (TM->getOptLevel() >= CodeGenOptLevel::Aggressive &&
769 CLOpts.enable_copy_propagation)
770 addPass(P: createMachineCopyPropagationPass(UseCopyInstr: true));
771 if (TM->getOptLevel() != CodeGenOptLevel::None)
772 addPass(P: createAArch64RedundantCondBranchPass());
773
774 addPass(P: createAArch64A53Fix835769LegacyPass());
775
776 if (TM->getTargetTriple().isOSWindows()) {
777 // Identify valid longjmp targets for Windows Control Flow Guard.
778 addPass(P: createCFGuardLongjmpLegacyPass());
779 // Identify valid eh continuation targets for Windows EHCont Guard.
780 addPass(P: createEHContGuardTargetsLegacy());
781 }
782
783 if (TM->getOptLevel() != CodeGenOptLevel::None && CLOpts.enable_collect_loh &&
784 TM->getTargetTriple().isOSBinFormatMachO())
785 addPass(P: createAArch64CollectLOHPass());
786
787 // Apply code layout optimizations. Run late so detection reflects the
788 // final MI stream.
789 if (getOptLevel() != CodeGenOptLevel::None)
790 addPass(P: createAArch64CodeLayoutOptPass());
791}
792
793void AArch64PassConfig::addPostBBSections() {
794 addPass(P: createAArch64SLSHardeningLegacyPass());
795 addPass(P: createAArch64PointerAuthPass());
796 if (CLOpts.enable_branch_targets)
797 addPass(P: createAArch64BranchTargetsPass());
798 // Relax conditional branch instructions if they're otherwise out of
799 // range of their destination.
800 if (CLOpts.enable_branch_relax)
801 addPass(PassID: &BranchRelaxationPassID);
802
803 if (TM->getOptLevel() != CodeGenOptLevel::None &&
804 CLOpts.enable_compress_jump_tables)
805 addPass(P: createAArch64CompressJumpTablesPass());
806}
807
808void AArch64PassConfig::addPreEmitPass2() {
809 // Insert pseudo probe annotation for callsite profiling
810 addPass(P: createPseudoProbeInserterLegacyPass());
811
812 // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo
813 // instructions are lowered to bundles as well.
814 addPass(P: createUnpackMachineBundlesLegacy(Ftor: nullptr));
815}
816
817bool AArch64PassConfig::addRegAssignAndRewriteOptimized() {
818 addPass(P: createAArch64PostCoalescerPass());
819 return TargetPassConfig::addRegAssignAndRewriteOptimized();
820}
821
822MachineFunctionInfo *AArch64TargetMachine::createMachineFunctionInfo(
823 BumpPtrAllocator &Allocator, const Function &F,
824 const TargetSubtargetInfo *STI) const {
825 return AArch64FunctionInfo::create<AArch64FunctionInfo>(
826 Allocator, F, STI: static_cast<const AArch64Subtarget *>(STI));
827}
828
829yaml::MachineFunctionInfo *
830AArch64TargetMachine::createDefaultFuncInfoYAML() const {
831 return new yaml::AArch64FunctionInfo();
832}
833
834yaml::MachineFunctionInfo *
835AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
836 const auto *MFI = MF.getInfo<AArch64FunctionInfo>();
837 return new yaml::AArch64FunctionInfo(*MFI);
838}
839
840bool AArch64TargetMachine::parseMachineFunctionInfo(
841 const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS,
842 SMDiagnostic &Error, SMRange &SourceRange) const {
843 const auto &YamlMFI = static_cast<const yaml::AArch64FunctionInfo &>(MFI);
844 MachineFunction &MF = PFS.MF;
845 MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI);
846 return false;
847}
848