xref: /openbsd-src/gnu/llvm/llvm/lib/Target/AArch64/AArch64TargetMachine.cpp (revision a96b36398fcfb4953e8190127da8bf074c7552f1)
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 
LLVMInitializeAArch64Target()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 //===----------------------------------------------------------------------===//
createTLOF(const Triple & TT)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
computeDataLayout(const Triple & TT,const MCTargetOptions & Options,bool LittleEndian)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 
computeDefaultCPU(const Triple & TT,StringRef CPU)272 static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) {
273   if (CPU.empty() && TT.isArm64e())
274     return "apple-a12";
275   return CPU;
276 }
277 
getEffectiveRelocModel(const Triple & TT,std::optional<Reloc::Model> RM)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
getEffectiveAArch64CodeModel(const Triple & TT,std::optional<CodeModel::Model> CM,bool JIT)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 ///
AArch64TargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,std::optional<Reloc::Model> RM,std::optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT,bool LittleEndian)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       !getTargetTriple().isOSOpenBSD() &&
361       TT.getArch() != Triple::aarch64_32 &&
362       TT.getEnvironment() != Triple::GNUILP32 &&
363       !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) {
364     setGlobalISel(true);
365     setGlobalISelAbort(GlobalISelAbortMode::Disable);
366   }
367 
368   // AArch64 supports the MachineOutliner.
369   setMachineOutliner(true);
370 
371   // AArch64 supports default outlining behaviour.
372   setSupportsDefaultOutlining(true);
373 
374   // AArch64 supports the debug entry values.
375   setSupportsDebugEntryValues(true);
376 
377   // AArch64 supports fixing up the DWARF unwind information.
378   if (!getMCAsmInfo()->usesWindowsCFI())
379     setCFIFixup(true);
380 }
381 
382 AArch64TargetMachine::~AArch64TargetMachine() = default;
383 
384 const AArch64Subtarget *
getSubtargetImpl(const Function & F) const385 AArch64TargetMachine::getSubtargetImpl(const Function &F) const {
386   Attribute CPUAttr = F.getFnAttribute("target-cpu");
387   Attribute TuneAttr = F.getFnAttribute("tune-cpu");
388   Attribute FSAttr = F.getFnAttribute("target-features");
389 
390   StringRef CPU = CPUAttr.isValid() ? CPUAttr.getValueAsString() : TargetCPU;
391   StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString() : CPU;
392   StringRef FS = FSAttr.isValid() ? FSAttr.getValueAsString() : TargetFS;
393 
394   bool StreamingSVEModeDisabled =
395       !F.hasFnAttribute("aarch64_pstate_sm_enabled") &&
396       !F.hasFnAttribute("aarch64_pstate_sm_compatible") &&
397       !F.hasFnAttribute("aarch64_pstate_sm_body");
398 
399   unsigned MinSVEVectorSize = 0;
400   unsigned MaxSVEVectorSize = 0;
401   Attribute VScaleRangeAttr = F.getFnAttribute(Attribute::VScaleRange);
402   if (VScaleRangeAttr.isValid()) {
403     std::optional<unsigned> VScaleMax = VScaleRangeAttr.getVScaleRangeMax();
404     MinSVEVectorSize = VScaleRangeAttr.getVScaleRangeMin() * 128;
405     MaxSVEVectorSize = VScaleMax ? *VScaleMax * 128 : 0;
406   } else {
407     MinSVEVectorSize = SVEVectorBitsMinOpt;
408     MaxSVEVectorSize = SVEVectorBitsMaxOpt;
409   }
410 
411   assert(MinSVEVectorSize % 128 == 0 &&
412          "SVE requires vector length in multiples of 128!");
413   assert(MaxSVEVectorSize % 128 == 0 &&
414          "SVE requires vector length in multiples of 128!");
415   assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) &&
416          "Minimum SVE vector size should not be larger than its maximum!");
417 
418   // Sanitize user input in case of no asserts
419   if (MaxSVEVectorSize == 0)
420     MinSVEVectorSize = (MinSVEVectorSize / 128) * 128;
421   else {
422     MinSVEVectorSize =
423         (std::min(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
424     MaxSVEVectorSize =
425         (std::max(MinSVEVectorSize, MaxSVEVectorSize) / 128) * 128;
426   }
427 
428   SmallString<512> Key;
429   raw_svector_ostream(Key) << "SVEMin" << MinSVEVectorSize << "SVEMax"
430                            << MaxSVEVectorSize << "StreamingSVEModeDisabled="
431                            << StreamingSVEModeDisabled << CPU << TuneCPU << FS;
432 
433   auto &I = SubtargetMap[Key];
434   if (!I) {
435     // This needs to be done before we create a new subtarget since any
436     // creation will depend on the TM and the code generation flags on the
437     // function that reside in TargetOptions.
438     resetTargetOptions(F);
439     I = std::make_unique<AArch64Subtarget>(
440         TargetTriple, CPU, TuneCPU, FS, *this, isLittle, MinSVEVectorSize,
441         MaxSVEVectorSize, StreamingSVEModeDisabled);
442   }
443   return I.get();
444 }
445 
anchor()446 void AArch64leTargetMachine::anchor() { }
447 
AArch64leTargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,std::optional<Reloc::Model> RM,std::optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT)448 AArch64leTargetMachine::AArch64leTargetMachine(
449     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
450     const TargetOptions &Options, std::optional<Reloc::Model> RM,
451     std::optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
452     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {}
453 
anchor()454 void AArch64beTargetMachine::anchor() { }
455 
AArch64beTargetMachine(const Target & T,const Triple & TT,StringRef CPU,StringRef FS,const TargetOptions & Options,std::optional<Reloc::Model> RM,std::optional<CodeModel::Model> CM,CodeGenOpt::Level OL,bool JIT)456 AArch64beTargetMachine::AArch64beTargetMachine(
457     const Target &T, const Triple &TT, StringRef CPU, StringRef FS,
458     const TargetOptions &Options, std::optional<Reloc::Model> RM,
459     std::optional<CodeModel::Model> CM, CodeGenOpt::Level OL, bool JIT)
460     : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {}
461 
462 namespace {
463 
464 /// AArch64 Code Generator Pass Configuration Options.
465 class AArch64PassConfig : public TargetPassConfig {
466 public:
AArch64PassConfig(AArch64TargetMachine & TM,PassManagerBase & PM)467   AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM)
468       : TargetPassConfig(TM, PM) {
469     if (TM.getOptLevel() != CodeGenOpt::None)
470       substitutePass(&PostRASchedulerID, &PostMachineSchedulerID);
471   }
472 
getAArch64TargetMachine() const473   AArch64TargetMachine &getAArch64TargetMachine() const {
474     return getTM<AArch64TargetMachine>();
475   }
476 
477   ScheduleDAGInstrs *
createMachineScheduler(MachineSchedContext * C) const478   createMachineScheduler(MachineSchedContext *C) const override {
479     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
480     ScheduleDAGMILive *DAG = createGenericSchedLive(C);
481     DAG->addMutation(createLoadClusterDAGMutation(DAG->TII, DAG->TRI));
482     DAG->addMutation(createStoreClusterDAGMutation(DAG->TII, DAG->TRI));
483     if (ST.hasFusion())
484       DAG->addMutation(createAArch64MacroFusionDAGMutation());
485     return DAG;
486   }
487 
488   ScheduleDAGInstrs *
createPostMachineScheduler(MachineSchedContext * C) const489   createPostMachineScheduler(MachineSchedContext *C) const override {
490     const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>();
491     ScheduleDAGMI *DAG =
492         new ScheduleDAGMI(C, std::make_unique<AArch64PostRASchedStrategy>(C),
493                           /* RemoveKillFlags=*/true);
494     if (ST.hasFusion()) {
495       // Run the Macro Fusion after RA again since literals are expanded from
496       // pseudos then (v. addPreSched2()).
497       DAG->addMutation(createAArch64MacroFusionDAGMutation());
498       return DAG;
499     }
500 
501     return DAG;
502   }
503 
504   void addIRPasses()  override;
505   bool addPreISel() override;
506   void addCodeGenPrepare() override;
507   bool addInstSelector() override;
508   bool addIRTranslator() override;
509   void addPreLegalizeMachineIR() override;
510   bool addLegalizeMachineIR() override;
511   void addPreRegBankSelect() override;
512   bool addRegBankSelect() override;
513   void addPreGlobalInstructionSelect() override;
514   bool addGlobalInstructionSelect() override;
515   void addMachineSSAOptimization() override;
516   bool addILPOpts() override;
517   void addPreRegAlloc() override;
518   void addPostRegAlloc() override;
519   void addPreSched2() override;
520   void addPreEmitPass() override;
521   void addPreEmitPass2() override;
522 
523   std::unique_ptr<CSEConfigBase> getCSEConfig() const override;
524 };
525 
526 } // end anonymous namespace
527 
528 TargetTransformInfo
getTargetTransformInfo(const Function & F) const529 AArch64TargetMachine::getTargetTransformInfo(const Function &F) const {
530   return TargetTransformInfo(AArch64TTIImpl(this, F));
531 }
532 
createPassConfig(PassManagerBase & PM)533 TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) {
534   return new AArch64PassConfig(*this, PM);
535 }
536 
getCSEConfig() const537 std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const {
538   return getStandardCSEConfigForOpt(TM->getOptLevel());
539 }
540 
addIRPasses()541 void AArch64PassConfig::addIRPasses() {
542   // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg
543   // ourselves.
544   addPass(createAtomicExpandPass());
545 
546   // Expand any SVE vector library calls that we can't code generate directly.
547   if (EnableSVEIntrinsicOpts && TM->getOptLevel() == CodeGenOpt::Aggressive)
548     addPass(createSVEIntrinsicOptsPass());
549 
550   // Cmpxchg instructions are often used with a subsequent comparison to
551   // determine whether it succeeded. We can exploit existing control-flow in
552   // ldrex/strex loops to simplify this, but it needs tidying up.
553   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
554     addPass(createCFGSimplificationPass(SimplifyCFGOptions()
555                                             .forwardSwitchCondToPhi(true)
556                                             .convertSwitchRangeToICmp(true)
557                                             .convertSwitchToLookupTable(true)
558                                             .needCanonicalLoops(false)
559                                             .hoistCommonInsts(true)
560                                             .sinkCommonInsts(true)));
561 
562   // Run LoopDataPrefetch
563   //
564   // Run this before LSR to remove the multiplies involved in computing the
565   // pointer values N iterations ahead.
566   if (TM->getOptLevel() != CodeGenOpt::None) {
567     if (EnableLoopDataPrefetch)
568       addPass(createLoopDataPrefetchPass());
569     if (EnableFalkorHWPFFix)
570       addPass(createFalkorMarkStridedAccessesPass());
571   }
572 
573   if (TM->getOptLevel() == CodeGenOpt::Aggressive && EnableGEPOpt) {
574     // Call SeparateConstOffsetFromGEP pass to extract constants within indices
575     // and lower a GEP with multiple indices to either arithmetic operations or
576     // multiple GEPs with single index.
577     addPass(createSeparateConstOffsetFromGEPPass(true));
578     // Call EarlyCSE pass to find and remove subexpressions in the lowered
579     // result.
580     addPass(createEarlyCSEPass());
581     // Do loop invariant code motion in case part of the lowered result is
582     // invariant.
583     addPass(createLICMPass());
584   }
585 
586   TargetPassConfig::addIRPasses();
587 
588   if (getOptLevel() == CodeGenOpt::Aggressive && EnableSelectOpt)
589     addPass(createSelectOptimizePass());
590 
591   addPass(createAArch64StackTaggingPass(
592       /*IsOptNone=*/TM->getOptLevel() == CodeGenOpt::None));
593 
594   // Match complex arithmetic patterns
595   if (TM->getOptLevel() >= CodeGenOpt::Default)
596     addPass(createComplexDeinterleavingPass(TM));
597 
598   // Match interleaved memory accesses to ldN/stN intrinsics.
599   if (TM->getOptLevel() != CodeGenOpt::None) {
600     addPass(createInterleavedLoadCombinePass());
601     addPass(createInterleavedAccessPass());
602   }
603 
604   // Expand any functions marked with SME attributes which require special
605   // changes for the calling convention or that require the lazy-saving
606   // mechanism specified in the SME ABI.
607   addPass(createSMEABIPass());
608 
609   // Add Control Flow Guard checks.
610   if (TM->getTargetTriple().isOSWindows())
611     addPass(createCFGuardCheckPass());
612 
613   if (TM->Options.JMCInstrument)
614     addPass(createJMCInstrumenterPass());
615 }
616 
617 // Pass Pipeline Configuration
addPreISel()618 bool AArch64PassConfig::addPreISel() {
619   // Run promote constant before global merge, so that the promoted constants
620   // get a chance to be merged
621   if (TM->getOptLevel() != CodeGenOpt::None && EnablePromoteConstant)
622     addPass(createAArch64PromoteConstantPass());
623   // FIXME: On AArch64, this depends on the type.
624   // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes().
625   // and the offset has to be a multiple of the related size in bytes.
626   if ((TM->getOptLevel() != CodeGenOpt::None &&
627        EnableGlobalMerge == cl::BOU_UNSET) ||
628       EnableGlobalMerge == cl::BOU_TRUE) {
629     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
630                                (EnableGlobalMerge == cl::BOU_UNSET);
631 
632     // Merging of extern globals is enabled by default on non-Mach-O as we
633     // expect it to be generally either beneficial or harmless. On Mach-O it
634     // is disabled as we emit the .subsections_via_symbols directive which
635     // means that merging extern globals is not safe.
636     bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
637 
638     // FIXME: extern global merging is only enabled when we optimise for size
639     // because there are some regressions with it also enabled for performance.
640     if (!OnlyOptimizeForSize)
641       MergeExternalByDefault = false;
642 
643     addPass(createGlobalMergePass(TM, 4095, OnlyOptimizeForSize,
644                                   MergeExternalByDefault));
645   }
646 
647   return false;
648 }
649 
addCodeGenPrepare()650 void AArch64PassConfig::addCodeGenPrepare() {
651   if (getOptLevel() != CodeGenOpt::None)
652     addPass(createTypePromotionLegacyPass());
653   TargetPassConfig::addCodeGenPrepare();
654 }
655 
addInstSelector()656 bool AArch64PassConfig::addInstSelector() {
657   addPass(createAArch64ISelDag(getAArch64TargetMachine(), getOptLevel()));
658 
659   // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many
660   // references to _TLS_MODULE_BASE_ as possible.
661   if (TM->getTargetTriple().isOSBinFormatELF() &&
662       getOptLevel() != CodeGenOpt::None)
663     addPass(createAArch64CleanupLocalDynamicTLSPass());
664 
665   return false;
666 }
667 
addIRTranslator()668 bool AArch64PassConfig::addIRTranslator() {
669   addPass(new IRTranslator(getOptLevel()));
670   return false;
671 }
672 
addPreLegalizeMachineIR()673 void AArch64PassConfig::addPreLegalizeMachineIR() {
674   if (getOptLevel() == CodeGenOpt::None)
675     addPass(createAArch64O0PreLegalizerCombiner());
676   else {
677     addPass(createAArch64PreLegalizerCombiner());
678     if (EnableGISelLoadStoreOptPreLegal)
679       addPass(new LoadStoreOpt());
680   }
681 }
682 
addLegalizeMachineIR()683 bool AArch64PassConfig::addLegalizeMachineIR() {
684   addPass(new Legalizer());
685   return false;
686 }
687 
addPreRegBankSelect()688 void AArch64PassConfig::addPreRegBankSelect() {
689   bool IsOptNone = getOptLevel() == CodeGenOpt::None;
690   if (!IsOptNone) {
691     addPass(createAArch64PostLegalizerCombiner(IsOptNone));
692     if (EnableGISelLoadStoreOptPostLegal)
693       addPass(new LoadStoreOpt());
694   }
695   addPass(createAArch64PostLegalizerLowering());
696 }
697 
addRegBankSelect()698 bool AArch64PassConfig::addRegBankSelect() {
699   addPass(new RegBankSelect());
700   return false;
701 }
702 
addPreGlobalInstructionSelect()703 void AArch64PassConfig::addPreGlobalInstructionSelect() {
704   addPass(new Localizer());
705 }
706 
addGlobalInstructionSelect()707 bool AArch64PassConfig::addGlobalInstructionSelect() {
708   addPass(new InstructionSelect(getOptLevel()));
709   if (getOptLevel() != CodeGenOpt::None)
710     addPass(createAArch64PostSelectOptimize());
711   return false;
712 }
713 
addMachineSSAOptimization()714 void AArch64PassConfig::addMachineSSAOptimization() {
715   // Run default MachineSSAOptimization first.
716   TargetPassConfig::addMachineSSAOptimization();
717 
718   if (TM->getOptLevel() != CodeGenOpt::None)
719     addPass(createAArch64MIPeepholeOptPass());
720 }
721 
addILPOpts()722 bool AArch64PassConfig::addILPOpts() {
723   if (EnableCondOpt)
724     addPass(createAArch64ConditionOptimizerPass());
725   if (EnableCCMP)
726     addPass(createAArch64ConditionalCompares());
727   if (EnableMCR)
728     addPass(&MachineCombinerID);
729   if (EnableCondBrTuning)
730     addPass(createAArch64CondBrTuning());
731   if (EnableEarlyIfConversion)
732     addPass(&EarlyIfConverterID);
733   if (EnableStPairSuppress)
734     addPass(createAArch64StorePairSuppressPass());
735   addPass(createAArch64SIMDInstrOptPass());
736   if (TM->getOptLevel() != CodeGenOpt::None)
737     addPass(createAArch64StackTaggingPreRAPass());
738   return true;
739 }
740 
addPreRegAlloc()741 void AArch64PassConfig::addPreRegAlloc() {
742   // Change dead register definitions to refer to the zero register.
743   if (TM->getOptLevel() != CodeGenOpt::None && EnableDeadRegisterElimination)
744     addPass(createAArch64DeadRegisterDefinitions());
745 
746   // Use AdvSIMD scalar instructions whenever profitable.
747   if (TM->getOptLevel() != CodeGenOpt::None && EnableAdvSIMDScalar) {
748     addPass(createAArch64AdvSIMDScalar());
749     // The AdvSIMD pass may produce copies that can be rewritten to
750     // be register coalescer friendly.
751     addPass(&PeepholeOptimizerID);
752   }
753 }
754 
addPostRegAlloc()755 void AArch64PassConfig::addPostRegAlloc() {
756   // Remove redundant copy instructions.
757   if (TM->getOptLevel() != CodeGenOpt::None && EnableRedundantCopyElimination)
758     addPass(createAArch64RedundantCopyEliminationPass());
759 
760   if (TM->getOptLevel() != CodeGenOpt::None && usingDefaultRegAlloc())
761     // Improve performance for some FP/SIMD code for A57.
762     addPass(createAArch64A57FPLoadBalancing());
763 }
764 
addPreSched2()765 void AArch64PassConfig::addPreSched2() {
766   // Lower homogeneous frame instructions
767   if (EnableHomogeneousPrologEpilog)
768     addPass(createAArch64LowerHomogeneousPrologEpilogPass());
769   // Expand some pseudo instructions to allow proper scheduling.
770   addPass(createAArch64ExpandPseudoPass());
771   // Use load/store pair instructions when possible.
772   if (TM->getOptLevel() != CodeGenOpt::None) {
773     if (EnableLoadStoreOpt)
774       addPass(createAArch64LoadStoreOptimizationPass());
775   }
776   // Emit KCFI checks for indirect calls.
777   addPass(createAArch64KCFIPass());
778 
779   // The AArch64SpeculationHardeningPass destroys dominator tree and natural
780   // loop info, which is needed for the FalkorHWPFFixPass and also later on.
781   // Therefore, run the AArch64SpeculationHardeningPass before the
782   // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop
783   // info.
784   addPass(createAArch64SpeculationHardeningPass());
785 
786   addPass(createAArch64IndirectThunks());
787   addPass(createAArch64SLSHardeningPass());
788 
789   if (TM->getOptLevel() != CodeGenOpt::None) {
790     if (EnableFalkorHWPFFix)
791       addPass(createFalkorHWPFFixPass());
792   }
793 }
794 
addPreEmitPass()795 void AArch64PassConfig::addPreEmitPass() {
796   // Machine Block Placement might have created new opportunities when run
797   // at O3, where the Tail Duplication Threshold is set to 4 instructions.
798   // Run the load/store optimizer once more.
799   if (TM->getOptLevel() >= CodeGenOpt::Aggressive && EnableLoadStoreOpt)
800     addPass(createAArch64LoadStoreOptimizationPass());
801 
802   if (TM->getOptLevel() >= CodeGenOpt::Aggressive &&
803       EnableAArch64CopyPropagation)
804     addPass(createMachineCopyPropagationPass(true));
805 
806   addPass(createAArch64A53Fix835769());
807 
808   if (EnableBranchTargets)
809     addPass(createAArch64BranchTargetsPass());
810 
811   // Relax conditional branch instructions if they're otherwise out of
812   // range of their destination.
813   if (BranchRelaxation)
814     addPass(&BranchRelaxationPassID);
815 
816   if (TM->getTargetTriple().isOSWindows()) {
817     // Identify valid longjmp targets for Windows Control Flow Guard.
818     addPass(createCFGuardLongjmpPass());
819     // Identify valid eh continuation targets for Windows EHCont Guard.
820     addPass(createEHContGuardCatchretPass());
821   }
822 
823   if (TM->getOptLevel() != CodeGenOpt::None && EnableCompressJumpTables)
824     addPass(createAArch64CompressJumpTablesPass());
825 
826   if (TM->getOptLevel() != CodeGenOpt::None && EnableCollectLOH &&
827       TM->getTargetTriple().isOSBinFormatMachO())
828     addPass(createAArch64CollectLOHPass());
829 }
830 
addPreEmitPass2()831 void AArch64PassConfig::addPreEmitPass2() {
832   // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo
833   // instructions are lowered to bundles as well.
834   addPass(createUnpackMachineBundles(nullptr));
835 }
836 
createMachineFunctionInfo(BumpPtrAllocator & Allocator,const Function & F,const TargetSubtargetInfo * STI) const837 MachineFunctionInfo *AArch64TargetMachine::createMachineFunctionInfo(
838     BumpPtrAllocator &Allocator, const Function &F,
839     const TargetSubtargetInfo *STI) const {
840   return AArch64FunctionInfo::create<AArch64FunctionInfo>(
841       Allocator, F, static_cast<const AArch64Subtarget *>(STI));
842 }
843 
844 yaml::MachineFunctionInfo *
createDefaultFuncInfoYAML() const845 AArch64TargetMachine::createDefaultFuncInfoYAML() const {
846   return new yaml::AArch64FunctionInfo();
847 }
848 
849 yaml::MachineFunctionInfo *
convertFuncInfoToYAML(const MachineFunction & MF) const850 AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const {
851   const auto *MFI = MF.getInfo<AArch64FunctionInfo>();
852   return new yaml::AArch64FunctionInfo(*MFI);
853 }
854 
parseMachineFunctionInfo(const yaml::MachineFunctionInfo & MFI,PerFunctionMIParsingState & PFS,SMDiagnostic & Error,SMRange & SourceRange) const855 bool AArch64TargetMachine::parseMachineFunctionInfo(
856     const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS,
857     SMDiagnostic &Error, SMRange &SourceRange) const {
858   const auto &YamlMFI = static_cast<const yaml::AArch64FunctionInfo &>(MFI);
859   MachineFunction &MF = PFS.MF;
860   MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI);
861   return false;
862 }
863