1 //===-- AArch64TargetMachine.cpp - Define TargetMachine for AArch64 -------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "AArch64.h"
14 #include "AArch64CallLowering.h"
15 #include "AArch64InstructionSelector.h"
16 #include "AArch64MachineLegalizer.h"
17 #include "AArch64RegisterBankInfo.h"
18 #include "AArch64TargetMachine.h"
19 #include "AArch64TargetObjectFile.h"
20 #include "AArch64TargetTransformInfo.h"
21 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
22 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
23 #include "llvm/CodeGen/GlobalISel/MachineLegalizePass.h"
24 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/RegAllocRegistry.h"
27 #include "llvm/CodeGen/TargetPassConfig.h"
28 #include "llvm/IR/Function.h"
29 #include "llvm/IR/LegacyPassManager.h"
30 #include "llvm/InitializePasses.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/TargetRegistry.h"
33 #include "llvm/Target/TargetOptions.h"
34 #include "llvm/Transforms/Scalar.h"
35 using namespace llvm;
36 
37 static cl::opt<bool>
38 EnableCCMP("aarch64-ccmp", cl::desc("Enable the CCMP formation pass"),
39            cl::init(true), cl::Hidden);
40 
41 static cl::opt<bool> EnableMCR("aarch64-mcr",
42                                cl::desc("Enable the machine combiner pass"),
43                                cl::init(true), cl::Hidden);
44 
45 static cl::opt<bool>
46 EnableStPairSuppress("aarch64-stp-suppress", cl::desc("Suppress STP for AArch64"),
47                      cl::init(true), cl::Hidden);
48 
49 static cl::opt<bool>
50 EnableAdvSIMDScalar("aarch64-simd-scalar", cl::desc("Enable use of AdvSIMD scalar"
51                     " integer instructions"), cl::init(false), cl::Hidden);
52 
53 static cl::opt<bool>
54 EnablePromoteConstant("aarch64-promote-const", cl::desc("Enable the promote "
55                       "constant pass"), cl::init(true), cl::Hidden);
56 
57 static cl::opt<bool>
58 EnableCollectLOH("aarch64-collect-loh", cl::desc("Enable the pass that emits the"
59                  " linker optimization hints (LOH)"), cl::init(true),
60                  cl::Hidden);
61 
62 static cl::opt<bool>
63 EnableDeadRegisterElimination("aarch64-dead-def-elimination", cl::Hidden,
64                               cl::desc("Enable the pass that removes dead"
65                                        " definitons and replaces stores to"
66                                        " them with stores to the zero"
67                                        " register"),
68                               cl::init(true));
69 
70 static cl::opt<bool>
71 EnableRedundantCopyElimination("aarch64-redundant-copy-elim",
72               cl::desc("Enable the redundant copy elimination pass"),
73               cl::init(true), cl::Hidden);
74 
75 static cl::opt<bool>
76 EnableLoadStoreOpt("aarch64-load-store-opt", cl::desc("Enable the load/store pair"
77                    " optimization pass"), cl::init(true), cl::Hidden);
78 
79 static cl::opt<bool>
80 EnableAtomicTidy("aarch64-atomic-cfg-tidy", cl::Hidden,
81                  cl::desc("Run SimplifyCFG after expanding atomic operations"
82                           " to make use of cmpxchg flow-based information"),
83                  cl::init(true));
84 
85 static cl::opt<bool>
86 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
87                         cl::desc("Run early if-conversion"),
88                         cl::init(true));
89 
90 static cl::opt<bool>
91 EnableCondOpt("aarch64-condopt",
92               cl::desc("Enable the condition optimizer pass"),
93               cl::init(true), cl::Hidden);
94 
95 static cl::opt<bool>
96 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
97                 cl::desc("Work around Cortex-A53 erratum 835769"),
98                 cl::init(false));
99 
100 static cl::opt<bool>
101 EnableGEPOpt("aarch64-gep-opt", cl::Hidden,
102              cl::desc("Enable optimizations on complex GEPs"),
103              cl::init(false));
104 
105 // FIXME: Unify control over GlobalMerge.
106 static cl::opt<cl::boolOrDefault>
107 EnableGlobalMerge("aarch64-global-merge", cl::Hidden,
108                   cl::desc("Enable the global merge pass"));
109 
110 static cl::opt<bool>
111     EnableLoopDataPrefetch("aarch64-loop-data-prefetch", cl::Hidden,
112                            cl::desc("Enable the loop data prefetch pass"),
113                            cl::init(true));
114 
115 extern "C" void LLVMInitializeAArch64Target() {
116   // Register the target.
117   RegisterTargetMachine<AArch64leTargetMachine> X(TheAArch64leTarget);
118   RegisterTargetMachine<AArch64beTargetMachine> Y(TheAArch64beTarget);
119   RegisterTargetMachine<AArch64leTargetMachine> Z(TheARM64Target);
120   auto PR = PassRegistry::getPassRegistry();
121   initializeGlobalISel(*PR);
122   initializeAArch64ExpandPseudoPass(*PR);
123   initializeAArch64LoadStoreOptPass(*PR);
124 }
125 
126 //===----------------------------------------------------------------------===//
127 // AArch64 Lowering public interface.
128 //===----------------------------------------------------------------------===//
129 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
130   if (TT.isOSBinFormatMachO())
131     return make_unique<AArch64_MachoTargetObjectFile>();
132 
133   return make_unique<AArch64_ELFTargetObjectFile>();
134 }
135 
136 // Helper function to build a DataLayout string
137 static std::string computeDataLayout(const Triple &TT, bool LittleEndian) {
138   if (TT.isOSBinFormatMachO())
139     return "e-m:o-i64:64-i128:128-n32:64-S128";
140   if (LittleEndian)
141     return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
142   return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
143 }
144 
145 // Helper function to set up the defaults for reciprocals.
146 static void initReciprocals(AArch64TargetMachine& TM, AArch64Subtarget& ST)
147 {
148   // For the estimates, convergence is quadratic, so essentially the number of
149   // digits is doubled after each iteration. ARMv8, the minimum architected
150   // accuracy of the initial estimate is 2^-8.  Therefore, the number of extra
151   // steps to refine the result for float (23 mantissa bits) and for double
152   // (52 mantissa bits) are 2 and 3, respectively.
153   unsigned ExtraStepsF = 2,
154            ExtraStepsD = ExtraStepsF + 1;
155   bool UseRsqrt = ST.useRSqrt();
156 
157   TM.Options.Reciprocals.setDefaults("sqrtf", UseRsqrt, ExtraStepsF);
158   TM.Options.Reciprocals.setDefaults("sqrtd", UseRsqrt, ExtraStepsD);
159   TM.Options.Reciprocals.setDefaults("vec-sqrtf", UseRsqrt, ExtraStepsF);
160   TM.Options.Reciprocals.setDefaults("vec-sqrtd", UseRsqrt, ExtraStepsD);
161 
162   TM.Options.Reciprocals.setDefaults("divf", false, ExtraStepsF);
163   TM.Options.Reciprocals.setDefaults("divd", false, ExtraStepsD);
164   TM.Options.Reciprocals.setDefaults("vec-divf", false, ExtraStepsF);
165   TM.Options.Reciprocals.setDefaults("vec-divd", false, ExtraStepsD);
166 }
167 
168 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
169                                            Optional<Reloc::Model> RM) {
170   // AArch64 Darwin is always PIC.
171   if (TT.isOSDarwin())
172     return Reloc::PIC_;
173   // On ELF platforms the default static relocation model has a smart enough
174   // linker to cope with referencing external symbols defined in a shared
175   // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
176   if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC)
177     return Reloc::Static;
178   return *RM;
179 }
180 
181 /// Create an AArch64 architecture model.
182 ///
183 AArch64TargetMachine::AArch64TargetMachine(
184     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
185     const TargetOptions &Options, Optional<Reloc::Model> RM,
186     CodeModel::Model CM, CodeGenOpt::Level OL, bool LittleEndian)
187     // This nested ternary is horrible, but DL needs to be properly
188     // initialized before TLInfo is constructed.
189     : LLVMTargetMachine(T, computeDataLayout(TT, LittleEndian), TT, CPU, FS,
190                         Options, getEffectiveRelocModel(TT, RM), CM, OL),
191       TLOF(createTLOF(getTargetTriple())),
192       Subtarget(TT, CPU, FS, *this, LittleEndian) {
193   initReciprocals(*this, Subtarget);
194   initAsmInfo();
195 }
196 
197 AArch64TargetMachine::~AArch64TargetMachine() {}
198 
199 #ifdef LLVM_BUILD_GLOBAL_ISEL
200 namespace {
201 struct AArch64GISelActualAccessor : public GISelAccessor {
202   std::unique_ptr<CallLowering> CallLoweringInfo;
203   std::unique_ptr<InstructionSelector> InstSelector;
204   std::unique_ptr<MachineLegalizer> Legalizer;
205   std::unique_ptr<RegisterBankInfo> RegBankInfo;
206   const CallLowering *getCallLowering() const override {
207     return CallLoweringInfo.get();
208   }
209   const InstructionSelector *getInstructionSelector() const override {
210     return InstSelector.get();
211   }
212   const class MachineLegalizer *getMachineLegalizer() const override {
213     return Legalizer.get();
214   }
215   const RegisterBankInfo *getRegBankInfo() const override {
216     return RegBankInfo.get();
217   }
218 };
219 } // End anonymous namespace.
220 #endif
221 
222 const AArch64Subtarget *
223 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
224   Attribute CPUAttr = F.getFnAttribute("target-cpu");
225   Attribute FSAttr = F.getFnAttribute("target-features");
226 
227   std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
228                         ? CPUAttr.getValueAsString().str()
229                         : TargetCPU;
230   std::string FS = !FSAttr.hasAttribute(Attribute::None)
231                        ? FSAttr.getValueAsString().str()
232                        : TargetFS;
233 
234   auto &I = SubtargetMap[CPU + FS];
235   if (!I) {
236     // This needs to be done before we create a new subtarget since any
237     // creation will depend on the TM and the code generation flags on the
238     // function that reside in TargetOptions.
239     resetTargetOptions(F);
240     I = llvm::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this,
241                                             Subtarget.isLittleEndian());
242 #ifndef LLVM_BUILD_GLOBAL_ISEL
243    GISelAccessor *GISel = new GISelAccessor();
244 #else
245     AArch64GISelActualAccessor *GISel =
246         new AArch64GISelActualAccessor();
247     GISel->CallLoweringInfo.reset(
248         new AArch64CallLowering(*I->getTargetLowering()));
249     GISel->Legalizer.reset(new AArch64MachineLegalizer());
250 
251     auto *RBI = new AArch64RegisterBankInfo(*I->getRegisterInfo());
252 
253     // FIXME: At this point, we can't rely on Subtarget having RBI.
254     // It's awkward to mix passing RBI and the Subtarget; should we pass
255     // TII/TRI as well?
256     GISel->InstSelector.reset(new AArch64InstructionSelector(*I, *RBI));
257 
258     GISel->RegBankInfo.reset(RBI);
259 #endif
260     I->setGISelAccessor(*GISel);
261   }
262   return I.get();
263 }
264 
265 void AArch64leTargetMachine::anchor() { }
266 
267 AArch64leTargetMachine::AArch64leTargetMachine(
268     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
269     const TargetOptions &Options, Optional<Reloc::Model> RM,
270     CodeModel::Model CM, CodeGenOpt::Level OL)
271     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
272 
273 void AArch64beTargetMachine::anchor() { }
274 
275 AArch64beTargetMachine::AArch64beTargetMachine(
276     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
277     const TargetOptions &Options, Optional<Reloc::Model> RM,
278     CodeModel::Model CM, CodeGenOpt::Level OL)
279     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
280 
281 namespace {
282 /// AArch64 Code Generator Pass Configuration Options.
283 class AArch64PassConfig : public TargetPassConfig {
284 public:
285   AArch64PassConfig(AArch64TargetMachine *TM, PassManagerBase &PM)
286       : TargetPassConfig(TM, PM) {
287     if (TM->getOptLevel() != CodeGenOpt::None)
288       substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
289   }
290 
291   AArch64TargetMachine &getAArch64TargetMachine() const {
292     return getTM<AArch64TargetMachine>();
293   }
294 
295   void addIRPasses()  override;
296   bool addPreISel() override;
297   bool addInstSelector() override;
298 #ifdef LLVM_BUILD_GLOBAL_ISEL
299   bool addIRTranslator() override;
300   bool addLegalizeMachineIR() override;
301   bool addRegBankSelect() override;
302   bool addGlobalInstructionSelect() override;
303 #endif
304   bool addILPOpts() override;
305   void addPreRegAlloc() override;
306   void addPostRegAlloc() override;
307   void addPreSched2() override;
308   void addPreEmitPass() override;
309 };
310 } // namespace
311 
312 TargetIRAnalysis AArch64TargetMachine::getTargetIRAnalysis() {
313   return TargetIRAnalysis([this](const Function &F) {
314     return TargetTransformInfo(AArch64TTIImpl(this, F));
315   });
316 }
317 
318 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
319   return new AArch64PassConfig(this, PM);
320 }
321 
322 void AArch64PassConfig::addIRPasses() {
323   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
324   // ourselves.
325   addPass(createAtomicExpandPass(TM));
326 
327   // Cmpxchg instructions are often used with a subsequent comparison to
328   // determine whether it succeeded. We can exploit existing control-flow in
329   // ldrex/strex loops to simplify this, but it needs tidying up.
330   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
331     addPass(createCFGSimplificationPass());
332 
333   // Run LoopDataPrefetch
334   //
335   // Run this before LSR to remove the multiplies involved in computing the
336   // pointer values N iterations ahead.
337   if (TM->getOptLevel() != CodeGenOpt::None && EnableLoopDataPrefetch)
338     addPass(createLoopDataPrefetchPass());
339 
340   TargetPassConfig::addIRPasses();
341 
342   // Match interleaved memory accesses to ldN/stN intrinsics.
343   if (TM->getOptLevel() != CodeGenOpt::None)
344     addPass(createInterleavedAccessPass(TM));
345 
346   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
347     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
348     // and lower a GEP with multiple indices to either arithmetic operations or
349     // multiple GEPs with single index.
350     addPass(createSeparateConstOffsetFromGEPPass(TM, true));
351     // Call EarlyCSE pass to find and remove subexpressions in the lowered
352     // result.
353     addPass(createEarlyCSEPass());
354     // Do loop invariant code motion in case part of the lowered result is
355     // invariant.
356     addPass(createLICMPass());
357   }
358 }
359 
360 // Pass Pipeline Configuration
361 bool AArch64PassConfig::addPreISel() {
362   // Run promote constant before global merge, so that the promoted constants
363   // get a chance to be merged
364   if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
365     addPass(createAArch64PromoteConstantPass());
366   // FIXME: On AArch64, this depends on the type.
367   // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
368   // and the offset has to be a multiple of the related size in bytes.
369   if ((TM->getOptLevel() != CodeGenOpt::None &&
370        EnableGlobalMerge == cl::BOU_UNSET) ||
371       EnableGlobalMerge == cl::BOU_TRUE) {
372     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
373                                (EnableGlobalMerge == cl::BOU_UNSET);
374     addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize));
375   }
376 
377   if (TM->getOptLevel() != CodeGenOpt::None)
378     addPass(createAArch64AddressTypePromotionPass());
379 
380   return false;
381 }
382 
383 bool AArch64PassConfig::addInstSelector() {
384   addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
385 
386   // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
387   // references to _TLS_MODULE_BASE_ as possible.
388   if (TM->getTargetTriple().isOSBinFormatELF() &&
389       getOptLevel() != CodeGenOpt::None)
390     addPass(createAArch64CleanupLocalDynamicTLSPass());
391 
392   return false;
393 }
394 
395 #ifdef LLVM_BUILD_GLOBAL_ISEL
396 bool AArch64PassConfig::addIRTranslator() {
397   addPass(new IRTranslator());
398   return false;
399 }
400 bool AArch64PassConfig::addLegalizeMachineIR() {
401   addPass(new MachineLegalizePass());
402   return false;
403 }
404 bool AArch64PassConfig::addRegBankSelect() {
405   addPass(new RegBankSelect());
406   return false;
407 }
408 bool AArch64PassConfig::addGlobalInstructionSelect() {
409   addPass(new InstructionSelect());
410   return false;
411 }
412 #endif
413 
414 bool AArch64PassConfig::addILPOpts() {
415   if (EnableCondOpt)
416     addPass(createAArch64ConditionOptimizerPass());
417   if (EnableCCMP)
418     addPass(createAArch64ConditionalCompares());
419   if (EnableMCR)
420     addPass(&MachineCombinerID);
421   if (EnableEarlyIfConversion)
422     addPass(&EarlyIfConverterID);
423   if (EnableStPairSuppress)
424     addPass(createAArch64StorePairSuppressPass());
425   return true;
426 }
427 
428 void AArch64PassConfig::addPreRegAlloc() {
429   // Use AdvSIMD scalar instructions whenever profitable.
430   if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
431     addPass(createAArch64AdvSIMDScalar());
432     // The AdvSIMD pass may produce copies that can be rewritten to
433     // be register coaleascer friendly.
434     addPass(&PeepholeOptimizerID);
435   }
436 }
437 
438 void AArch64PassConfig::addPostRegAlloc() {
439   // Remove redundant copy instructions.
440   if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
441     addPass(createAArch64RedundantCopyEliminationPass());
442 
443   // Change dead register definitions to refer to the zero register.
444   if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
445     addPass(createAArch64DeadRegisterDefinitions());
446   if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
447     // Improve performance for some FP/SIMD code for A57.
448     addPass(createAArch64A57FPLoadBalancing());
449 }
450 
451 void AArch64PassConfig::addPreSched2() {
452   // Expand some pseudo instructions to allow proper scheduling.
453   addPass(createAArch64ExpandPseudoPass());
454   // Use load/store pair instructions when possible.
455   if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt)
456     addPass(createAArch64LoadStoreOptimizationPass());
457 }
458 
459 void AArch64PassConfig::addPreEmitPass() {
460   if (EnableA53Fix835769)
461     addPass(createAArch64A53Fix835769());
462   // Relax conditional branch instructions if they're otherwise out of
463   // range of their destination.
464   addPass(createAArch64BranchRelaxation());
465   if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
466       TM->getTargetTriple().isOSBinFormatMachO())
467     addPass(createAArch64CollectLOHPass());
468 }
469