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 "AArch64TargetMachine.h"
14 #include "AArch64.h"
15 #include "AArch64MacroFusion.h"
16 #include "AArch64Subtarget.h"
17 #include "AArch64TargetObjectFile.h"
18 #include "AArch64TargetTransformInfo.h"
19 #include "MCTargetDesc/AArch64MCTargetDesc.h"
20 #include "llvm/ADT/STLExtras.h"
21 #include "llvm/ADT/Triple.h"
22 #include "llvm/Analysis/TargetTransformInfo.h"
23 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
24 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
25 #include "llvm/CodeGen/GlobalISel/Legalizer.h"
26 #include "llvm/CodeGen/GlobalISel/Localizer.h"
27 #include "llvm/CodeGen/GlobalISel/RegBankSelect.h"
28 #include "llvm/CodeGen/MachineScheduler.h"
29 #include "llvm/CodeGen/Passes.h"
30 #include "llvm/CodeGen/TargetPassConfig.h"
31 #include "llvm/IR/Attributes.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/MC/MCTargetOptions.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CodeGen.h"
36 #include "llvm/Support/CommandLine.h"
37 #include "llvm/Support/TargetRegistry.h"
38 #include "llvm/Target/TargetLoweringObjectFile.h"
39 #include "llvm/Target/TargetOptions.h"
40 #include "llvm/Transforms/Scalar.h"
41 #include <memory>
42 #include <string>
43 
44 using namespace llvm;
45 
46 static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp",
47                                 cl::desc("Enable the CCMP formation pass"),
48                                 cl::init(true), cl::Hidden);
49 
50 static cl::opt<bool>
51     EnableCondBrTuning("aarch64-enable-cond-br-tune",
52                        cl::desc("Enable the conditional branch tuning pass"),
53                        cl::init(true), cl::Hidden);
54 
55 static cl::opt<bool> EnableMCR("aarch64-enable-mcr",
56                                cl::desc("Enable the machine combiner pass"),
57                                cl::init(true), cl::Hidden);
58 
59 static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress",
60                                           cl::desc("Suppress STP for AArch64"),
61                                           cl::init(true), cl::Hidden);
62 
63 static cl::opt<bool> EnableAdvSIMDScalar(
64     "aarch64-enable-simd-scalar",
65     cl::desc("Enable use of AdvSIMD scalar integer instructions"),
66     cl::init(false), cl::Hidden);
67 
68 static cl::opt<bool>
69     EnablePromoteConstant("aarch64-enable-promote-const",
70                           cl::desc("Enable the promote constant pass"),
71                           cl::init(true), cl::Hidden);
72 
73 static cl::opt<bool> EnableCollectLOH(
74     "aarch64-enable-collect-loh",
75     cl::desc("Enable the pass that emits the linker optimization hints (LOH)"),
76     cl::init(true), cl::Hidden);
77 
78 static cl::opt<bool>
79     EnableDeadRegisterElimination("aarch64-enable-dead-defs", cl::Hidden,
80                                   cl::desc("Enable the pass that removes dead"
81                                            " definitons and replaces stores to"
82                                            " them with stores to the zero"
83                                            " register"),
84                                   cl::init(true));
85 
86 static cl::opt<bool> EnableRedundantCopyElimination(
87     "aarch64-enable-copyelim",
88     cl::desc("Enable the redundant copy elimination pass"), cl::init(true),
89     cl::Hidden);
90 
91 static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt",
92                                         cl::desc("Enable the load/store pair"
93                                                  " optimization pass"),
94                                         cl::init(true), cl::Hidden);
95 
96 static cl::opt<bool> EnableAtomicTidy(
97     "aarch64-enable-atomic-cfg-tidy", cl::Hidden,
98     cl::desc("Run SimplifyCFG after expanding atomic operations"
99              " to make use of cmpxchg flow-based information"),
100     cl::init(true));
101 
102 static cl::opt<bool>
103 EnableEarlyIfConversion("aarch64-enable-early-ifcvt", cl::Hidden,
104                         cl::desc("Run early if-conversion"),
105                         cl::init(true));
106 
107 static cl::opt<bool>
108     EnableCondOpt("aarch64-enable-condopt",
109                   cl::desc("Enable the condition optimizer pass"),
110                   cl::init(true), cl::Hidden);
111 
112 static cl::opt<bool>
113 EnableA53Fix835769("aarch64-fix-cortex-a53-835769", cl::Hidden,
114                 cl::desc("Work around Cortex-A53 erratum 835769"),
115                 cl::init(false));
116 
117 static cl::opt<bool>
118     EnableGEPOpt("aarch64-enable-gep-opt", cl::Hidden,
119                  cl::desc("Enable optimizations on complex GEPs"),
120                  cl::init(false));
121 
122 static cl::opt<bool>
123     BranchRelaxation("aarch64-enable-branch-relax", cl::Hidden, cl::init(true),
124                      cl::desc("Relax out of range conditional branches"));
125 
126 // FIXME: Unify control over GlobalMerge.
127 static cl::opt<cl::boolOrDefault>
128     EnableGlobalMerge("aarch64-enable-global-merge", cl::Hidden,
129                       cl::desc("Enable the global merge pass"));
130 
131 static cl::opt<bool>
132     EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch", cl::Hidden,
133                            cl::desc("Enable the loop data prefetch pass"),
134                            cl::init(true));
135 
136 static cl::opt<int> EnableGlobalISelAtO(
137     "aarch64-enable-global-isel-at-O", cl::Hidden,
138     cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)"),
139     cl::init(-1));
140 
141 extern "C" void LLVMInitializeAArch64Target() {
142   // Register the target.
143   RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget());
144   RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget());
145   RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target());
146   auto PR = PassRegistry::getPassRegistry();
147   initializeGlobalISel(*PR);
148   initializeAArch64A53Fix835769Pass(*PR);
149   initializeAArch64A57FPLoadBalancingPass(*PR);
150   initializeAArch64AdvSIMDScalarPass(*PR);
151   initializeAArch64CollectLOHPass(*PR);
152   initializeAArch64ConditionalComparesPass(*PR);
153   initializeAArch64ConditionOptimizerPass(*PR);
154   initializeAArch64DeadRegisterDefinitionsPass(*PR);
155   initializeAArch64ExpandPseudoPass(*PR);
156   initializeAArch64LoadStoreOptPass(*PR);
157   initializeAArch64VectorByElementOptPass(*PR);
158   initializeAArch64PromoteConstantPass(*PR);
159   initializeAArch64RedundantCopyEliminationPass(*PR);
160   initializeAArch64StorePairSuppressPass(*PR);
161   initializeLDTLSCleanupPass(*PR);
162 }
163 
164 //===----------------------------------------------------------------------===//
165 // AArch64 Lowering public interface.
166 //===----------------------------------------------------------------------===//
167 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
168   if (TT.isOSBinFormatMachO())
169     return llvm::make_unique<AArch64_MachoTargetObjectFile>();
170   if (TT.isOSBinFormatCOFF())
171     return llvm::make_unique<AArch64_COFFTargetObjectFile>();
172 
173   return llvm::make_unique<AArch64_ELFTargetObjectFile>();
174 }
175 
176 // Helper function to build a DataLayout string
177 static std::string computeDataLayout(const Triple &TT,
178                                      const MCTargetOptions &Options,
179                                      bool LittleEndian) {
180   if (Options.getABIName() == "ilp32")
181     return "e-m:e-p:32:32-i8:8-i16:16-i64:64-S128";
182   if (TT.isOSBinFormatMachO())
183     return "e-m:o-i64:64-i128:128-n32:64-S128";
184   if (TT.isOSBinFormatCOFF())
185     return "e-m:w-i64:64-i128:128-n32:64-S128";
186   if (LittleEndian)
187     return "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
188   return "E-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128";
189 }
190 
191 static Reloc::Model getEffectiveRelocModel(const Triple &TT,
192                                            Optional<Reloc::Model> RM) {
193   // AArch64 Darwin is always PIC.
194   if (TT.isOSDarwin())
195     return Reloc::PIC_;
196   // On ELF platforms the default static relocation model has a smart enough
197   // linker to cope with referencing external symbols defined in a shared
198   // library. Hence DynamicNoPIC doesn't need to be promoted to PIC.
199   if (!RM.hasValue() || *RM == Reloc::DynamicNoPIC)
200     return Reloc::Static;
201   return *RM;
202 }
203 
204 /// Create an AArch64 architecture model.
205 ///
206 AArch64TargetMachine::AArch64TargetMachine(
207     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
208     const TargetOptions &Options, Optional<Reloc::Model> RM,
209     CodeModel::Model CM, CodeGenOpt::Level OL, bool LittleEndian)
210     // This nested ternary is horrible, but DL needs to be properly
211     // initialized before TLInfo is constructed.
212     : LLVMTargetMachine(T, computeDataLayout(TT, Options.MCOptions,
213                                              LittleEndian),
214                         TT, CPU, FS, Options,
215 			getEffectiveRelocModel(TT, RM), CM, OL),
216       TLOF(createTLOF(getTargetTriple())),
217       isLittle(LittleEndian) {
218   initAsmInfo();
219 }
220 
221 AArch64TargetMachine::~AArch64TargetMachine() = default;
222 
223 const AArch64Subtarget *
224 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
225   Attribute CPUAttr = F.getFnAttribute("target-cpu");
226   Attribute FSAttr = F.getFnAttribute("target-features");
227 
228   std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
229                         ? CPUAttr.getValueAsString().str()
230                         : TargetCPU;
231   std::string FS = !FSAttr.hasAttribute(Attribute::None)
232                        ? FSAttr.getValueAsString().str()
233                        : TargetFS;
234 
235   auto &I = SubtargetMap[CPU + FS];
236   if (!I) {
237     // This needs to be done before we create a new subtarget since any
238     // creation will depend on the TM and the code generation flags on the
239     // function that reside in TargetOptions.
240     resetTargetOptions(F);
241     I = llvm::make_unique<AArch64Subtarget>(TargetTriple, CPU, FS, *this,
242                                             isLittle);
243   }
244   return I.get();
245 }
246 
247 void AArch64leTargetMachine::anchor() { }
248 
249 AArch64leTargetMachine::AArch64leTargetMachine(
250     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
251     const TargetOptions &Options, Optional<Reloc::Model> RM,
252     CodeModel::Model CM, CodeGenOpt::Level OL)
253     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
254 
255 void AArch64beTargetMachine::anchor() { }
256 
257 AArch64beTargetMachine::AArch64beTargetMachine(
258     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
259     const TargetOptions &Options, Optional<Reloc::Model> RM,
260     CodeModel::Model CM, CodeGenOpt::Level OL)
261     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
262 
263 namespace {
264 
265 /// AArch64 Code Generator Pass Configuration Options.
266 class AArch64PassConfig : public TargetPassConfig {
267 public:
268   AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
269       : TargetPassConfig(TM, PM) {
270     if (TM.getOptLevel() != CodeGenOpt::None)
271       substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
272   }
273 
274   AArch64TargetMachine &getAArch64TargetMachine() const {
275     return getTM<AArch64TargetMachine>();
276   }
277 
278   ScheduleDAGInstrs *
279   createMachineScheduler(MachineSchedContext *C) const override {
280     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
281     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
282     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
283     DAG->addMutation(createAArch64MacroFusionDAGMutation());
284     return DAG;
285   }
286 
287   ScheduleDAGInstrs *
288   createPostMachineScheduler(MachineSchedContext *C) const override {
289     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
290     if (ST.hasFuseAES() || ST.hasFuseLiterals()) {
291       // Run the Macro Fusion after RA again since literals are expanded from
292       // pseudos then (v. addPreSched2()).
293       ScheduleDAGMI *DAG = createGenericSchedPostRA(C);
294       DAG->addMutation(createAArch64MacroFusionDAGMutation());
295       return DAG;
296     }
297 
298     return nullptr;
299   }
300 
301   void addIRPasses()  override;
302   bool addPreISel() override;
303   bool addInstSelector() override;
304 #ifdef LLVM_BUILD_GLOBAL_ISEL
305   bool addIRTranslator() override;
306   bool addLegalizeMachineIR() override;
307   bool addRegBankSelect() override;
308   void addPreGlobalInstructionSelect() override;
309   bool addGlobalInstructionSelect() override;
310 #endif
311   bool addILPOpts() override;
312   void addPreRegAlloc() override;
313   void addPostRegAlloc() override;
314   void addPreSched2() override;
315   void addPreEmitPass() override;
316 
317   bool isGlobalISelEnabled() const override;
318 };
319 
320 } // end anonymous namespace
321 
322 TargetIRAnalysis AArch64TargetMachine::getTargetIRAnalysis() {
323   return TargetIRAnalysis([this](const Function &F) {
324     return TargetTransformInfo(AArch64TTIImpl(this, F));
325   });
326 }
327 
328 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
329   return new AArch64PassConfig(*this, PM);
330 }
331 
332 void AArch64PassConfig::addIRPasses() {
333   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
334   // ourselves.
335   addPass(createAtomicExpandPass());
336 
337   // Cmpxchg instructions are often used with a subsequent comparison to
338   // determine whether it succeeded. We can exploit existing control-flow in
339   // ldrex/strex loops to simplify this, but it needs tidying up.
340   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
341     addPass(createCFGSimplificationPass());
342 
343   // Run LoopDataPrefetch
344   //
345   // Run this before LSR to remove the multiplies involved in computing the
346   // pointer values N iterations ahead.
347   if (TM->getOptLevel() != CodeGenOpt::None && EnableLoopDataPrefetch)
348     addPass(createLoopDataPrefetchPass());
349 
350   TargetPassConfig::addIRPasses();
351 
352   // Match interleaved memory accesses to ldN/stN intrinsics.
353   if (TM->getOptLevel() != CodeGenOpt::None)
354     addPass(createInterleavedAccessPass());
355 
356   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
357     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
358     // and lower a GEP with multiple indices to either arithmetic operations or
359     // multiple GEPs with single index.
360     addPass(createSeparateConstOffsetFromGEPPass(TM, true));
361     // Call EarlyCSE pass to find and remove subexpressions in the lowered
362     // result.
363     addPass(createEarlyCSEPass());
364     // Do loop invariant code motion in case part of the lowered result is
365     // invariant.
366     addPass(createLICMPass());
367   }
368 }
369 
370 // Pass Pipeline Configuration
371 bool AArch64PassConfig::addPreISel() {
372   // Run promote constant before global merge, so that the promoted constants
373   // get a chance to be merged
374   if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
375     addPass(createAArch64PromoteConstantPass());
376   // FIXME: On AArch64, this depends on the type.
377   // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
378   // and the offset has to be a multiple of the related size in bytes.
379   if ((TM->getOptLevel() != CodeGenOpt::None &&
380        EnableGlobalMerge == cl::BOU_UNSET) ||
381       EnableGlobalMerge == cl::BOU_TRUE) {
382     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
383                                (EnableGlobalMerge == cl::BOU_UNSET);
384     addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize));
385   }
386 
387   return false;
388 }
389 
390 bool AArch64PassConfig::addInstSelector() {
391   addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
392 
393   // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
394   // references to _TLS_MODULE_BASE_ as possible.
395   if (TM->getTargetTriple().isOSBinFormatELF() &&
396       getOptLevel() != CodeGenOpt::None)
397     addPass(createAArch64CleanupLocalDynamicTLSPass());
398 
399   return false;
400 }
401 
402 #ifdef LLVM_BUILD_GLOBAL_ISEL
403 bool AArch64PassConfig::addIRTranslator() {
404   addPass(new IRTranslator());
405   return false;
406 }
407 
408 bool AArch64PassConfig::addLegalizeMachineIR() {
409   addPass(new Legalizer());
410   return false;
411 }
412 
413 bool AArch64PassConfig::addRegBankSelect() {
414   addPass(new RegBankSelect());
415   return false;
416 }
417 
418 void AArch64PassConfig::addPreGlobalInstructionSelect() {
419   // Workaround the deficiency of the fast register allocator.
420   if (TM->getOptLevel() == CodeGenOpt::None)
421     addPass(new Localizer());
422 }
423 
424 bool AArch64PassConfig::addGlobalInstructionSelect() {
425   addPass(new InstructionSelect());
426   return false;
427 }
428 #endif
429 
430 bool AArch64PassConfig::isGlobalISelEnabled() const {
431   return TM->getOptLevel() <= EnableGlobalISelAtO;
432 }
433 
434 bool AArch64PassConfig::addILPOpts() {
435   if (EnableCondOpt)
436     addPass(createAArch64ConditionOptimizerPass());
437   if (EnableCCMP)
438     addPass(createAArch64ConditionalCompares());
439   if (EnableMCR)
440     addPass(&MachineCombinerID);
441   if (EnableCondBrTuning)
442     addPass(createAArch64CondBrTuning());
443   if (EnableEarlyIfConversion)
444     addPass(&EarlyIfConverterID);
445   if (EnableStPairSuppress)
446     addPass(createAArch64StorePairSuppressPass());
447   addPass(createAArch64VectorByElementOptPass());
448   return true;
449 }
450 
451 void AArch64PassConfig::addPreRegAlloc() {
452   // Change dead register definitions to refer to the zero register.
453   if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
454     addPass(createAArch64DeadRegisterDefinitions());
455 
456   // Use AdvSIMD scalar instructions whenever profitable.
457   if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
458     addPass(createAArch64AdvSIMDScalar());
459     // The AdvSIMD pass may produce copies that can be rewritten to
460     // be register coaleascer friendly.
461     addPass(&PeepholeOptimizerID);
462   }
463 }
464 
465 void AArch64PassConfig::addPostRegAlloc() {
466   // Remove redundant copy instructions.
467   if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
468     addPass(createAArch64RedundantCopyEliminationPass());
469 
470   if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
471     // Improve performance for some FP/SIMD code for A57.
472     addPass(createAArch64A57FPLoadBalancing());
473 }
474 
475 void AArch64PassConfig::addPreSched2() {
476   // Expand some pseudo instructions to allow proper scheduling.
477   addPass(createAArch64ExpandPseudoPass());
478   // Use load/store pair instructions when possible.
479   if (TM->getOptLevel() != CodeGenOpt::None && EnableLoadStoreOpt)
480     addPass(createAArch64LoadStoreOptimizationPass());
481 }
482 
483 void AArch64PassConfig::addPreEmitPass() {
484   if (EnableA53Fix835769)
485     addPass(createAArch64A53Fix835769());
486   // Relax conditional branch instructions if they're otherwise out of
487   // range of their destination.
488   if (BranchRelaxation)
489     addPass(&BranchRelaxationPassID);
490 
491   if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
492       TM->getTargetTriple().isOSBinFormatMachO())
493     addPass(createAArch64CollectLOHPass());
494 }
495