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/Analysis/TargetTransformInfo.h" |
24 | #include "llvm/Analysis/ValueTracking.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/Passes/PassBuilder.h" |
47 | #include "llvm/Support/CodeGen.h" |
48 | #include "llvm/Support/CommandLine.h" |
49 | #include "llvm/Target/TargetLoweringObjectFile.h" |
50 | #include "llvm/Target/TargetOptions.h" |
51 | #include "llvm/TargetParser/Triple.h" |
52 | #include "llvm/Transforms/CFGuard.h" |
53 | #include "llvm/Transforms/Scalar.h" |
54 | #include "llvm/Transforms/Vectorize/LoopIdiomVectorize.h" |
55 | #include <memory> |
56 | #include <optional> |
57 | #include <string> |
58 | |
59 | using namespace llvm; |
60 | |
61 | static cl::opt<bool> EnableCCMP("aarch64-enable-ccmp" , |
62 | cl::desc("Enable the CCMP formation pass" ), |
63 | cl::init(Val: true), cl::Hidden); |
64 | |
65 | static cl::opt<bool> |
66 | EnableCondBrTuning("aarch64-enable-cond-br-tune" , |
67 | cl::desc("Enable the conditional branch tuning pass" ), |
68 | cl::init(Val: true), cl::Hidden); |
69 | |
70 | static cl::opt<bool> EnableAArch64CopyPropagation( |
71 | "aarch64-enable-copy-propagation" , |
72 | cl::desc("Enable the copy propagation with AArch64 copy instr" ), |
73 | cl::init(Val: true), cl::Hidden); |
74 | |
75 | static cl::opt<bool> EnableMCR("aarch64-enable-mcr" , |
76 | cl::desc("Enable the machine combiner pass" ), |
77 | cl::init(Val: true), cl::Hidden); |
78 | |
79 | static cl::opt<bool> EnableStPairSuppress("aarch64-enable-stp-suppress" , |
80 | cl::desc("Suppress STP for AArch64" ), |
81 | cl::init(Val: true), cl::Hidden); |
82 | |
83 | static cl::opt<bool> EnableAdvSIMDScalar( |
84 | "aarch64-enable-simd-scalar" , |
85 | cl::desc("Enable use of AdvSIMD scalar integer instructions" ), |
86 | cl::init(Val: false), cl::Hidden); |
87 | |
88 | static cl::opt<bool> |
89 | EnablePromoteConstant("aarch64-enable-promote-const" , |
90 | cl::desc("Enable the promote constant pass" ), |
91 | cl::init(Val: true), cl::Hidden); |
92 | |
93 | static cl::opt<bool> EnableCollectLOH( |
94 | "aarch64-enable-collect-loh" , |
95 | cl::desc("Enable the pass that emits the linker optimization hints (LOH)" ), |
96 | cl::init(Val: true), cl::Hidden); |
97 | |
98 | static cl::opt<bool> |
99 | EnableDeadRegisterElimination("aarch64-enable-dead-defs" , cl::Hidden, |
100 | cl::desc("Enable the pass that removes dead" |
101 | " definitons and replaces stores to" |
102 | " them with stores to the zero" |
103 | " register" ), |
104 | cl::init(Val: true)); |
105 | |
106 | static cl::opt<bool> EnableRedundantCopyElimination( |
107 | "aarch64-enable-copyelim" , |
108 | cl::desc("Enable the redundant copy elimination pass" ), cl::init(Val: true), |
109 | cl::Hidden); |
110 | |
111 | static cl::opt<bool> EnableLoadStoreOpt("aarch64-enable-ldst-opt" , |
112 | cl::desc("Enable the load/store pair" |
113 | " optimization pass" ), |
114 | cl::init(Val: true), cl::Hidden); |
115 | |
116 | static cl::opt<bool> EnableAtomicTidy( |
117 | "aarch64-enable-atomic-cfg-tidy" , cl::Hidden, |
118 | cl::desc("Run SimplifyCFG after expanding atomic operations" |
119 | " to make use of cmpxchg flow-based information" ), |
120 | cl::init(Val: true)); |
121 | |
122 | static cl::opt<bool> |
123 | EnableEarlyIfConversion("aarch64-enable-early-ifcvt" , cl::Hidden, |
124 | cl::desc("Run early if-conversion" ), |
125 | cl::init(Val: true)); |
126 | |
127 | static cl::opt<bool> |
128 | EnableCondOpt("aarch64-enable-condopt" , |
129 | cl::desc("Enable the condition optimizer pass" ), |
130 | cl::init(Val: true), cl::Hidden); |
131 | |
132 | static cl::opt<bool> |
133 | EnableGEPOpt("aarch64-enable-gep-opt" , cl::Hidden, |
134 | cl::desc("Enable optimizations on complex GEPs" ), |
135 | cl::init(Val: false)); |
136 | |
137 | static cl::opt<bool> |
138 | EnableSelectOpt("aarch64-select-opt" , cl::Hidden, |
139 | cl::desc("Enable select to branch optimizations" ), |
140 | cl::init(Val: true)); |
141 | |
142 | static cl::opt<bool> |
143 | BranchRelaxation("aarch64-enable-branch-relax" , cl::Hidden, cl::init(Val: true), |
144 | cl::desc("Relax out of range conditional branches" )); |
145 | |
146 | static cl::opt<bool> EnableCompressJumpTables( |
147 | "aarch64-enable-compress-jump-tables" , cl::Hidden, cl::init(Val: true), |
148 | cl::desc("Use smallest entry possible for jump tables" )); |
149 | |
150 | // FIXME: Unify control over GlobalMerge. |
151 | static cl::opt<cl::boolOrDefault> |
152 | EnableGlobalMerge("aarch64-enable-global-merge" , cl::Hidden, |
153 | cl::desc("Enable the global merge pass" )); |
154 | |
155 | static cl::opt<bool> |
156 | EnableLoopDataPrefetch("aarch64-enable-loop-data-prefetch" , cl::Hidden, |
157 | cl::desc("Enable the loop data prefetch pass" ), |
158 | cl::init(Val: true)); |
159 | |
160 | static cl::opt<int> EnableGlobalISelAtO( |
161 | "aarch64-enable-global-isel-at-O" , cl::Hidden, |
162 | cl::desc("Enable GlobalISel at or below an opt level (-1 to disable)" ), |
163 | cl::init(Val: 0)); |
164 | |
165 | static cl::opt<bool> |
166 | EnableSVEIntrinsicOpts("aarch64-enable-sve-intrinsic-opts" , cl::Hidden, |
167 | cl::desc("Enable SVE intrinsic opts" ), |
168 | cl::init(Val: true)); |
169 | |
170 | static cl::opt<bool> EnableFalkorHWPFFix("aarch64-enable-falkor-hwpf-fix" , |
171 | cl::init(Val: true), cl::Hidden); |
172 | |
173 | static cl::opt<bool> |
174 | EnableBranchTargets("aarch64-enable-branch-targets" , cl::Hidden, |
175 | cl::desc("Enable the AArch64 branch target pass" ), |
176 | cl::init(Val: true)); |
177 | |
178 | static cl::opt<unsigned> SVEVectorBitsMaxOpt( |
179 | "aarch64-sve-vector-bits-max" , |
180 | cl::desc("Assume SVE vector registers are at most this big, " |
181 | "with zero meaning no maximum size is assumed." ), |
182 | cl::init(Val: 0), cl::Hidden); |
183 | |
184 | static cl::opt<unsigned> SVEVectorBitsMinOpt( |
185 | "aarch64-sve-vector-bits-min" , |
186 | cl::desc("Assume SVE vector registers are at least this big, " |
187 | "with zero meaning no minimum size is assumed." ), |
188 | cl::init(Val: 0), cl::Hidden); |
189 | |
190 | static cl::opt<bool> ForceStreaming( |
191 | "force-streaming" , |
192 | cl::desc("Force the use of streaming code for all functions" ), |
193 | cl::init(Val: false), cl::Hidden); |
194 | |
195 | static cl::opt<bool> ForceStreamingCompatible( |
196 | "force-streaming-compatible" , |
197 | cl::desc("Force the use of streaming-compatible code for all functions" ), |
198 | cl::init(Val: false), cl::Hidden); |
199 | |
200 | extern cl::opt<bool> EnableHomogeneousPrologEpilog; |
201 | |
202 | static cl::opt<bool> EnableGISelLoadStoreOptPreLegal( |
203 | "aarch64-enable-gisel-ldst-prelegal" , |
204 | cl::desc("Enable GlobalISel's pre-legalizer load/store optimization pass" ), |
205 | cl::init(Val: true), cl::Hidden); |
206 | |
207 | static cl::opt<bool> EnableGISelLoadStoreOptPostLegal( |
208 | "aarch64-enable-gisel-ldst-postlegal" , |
209 | cl::desc("Enable GlobalISel's post-legalizer load/store optimization pass" ), |
210 | cl::init(Val: false), cl::Hidden); |
211 | |
212 | static cl::opt<bool> |
213 | EnableSinkFold("aarch64-enable-sink-fold" , |
214 | cl::desc("Enable sinking and folding of instruction copies" ), |
215 | cl::init(Val: true), cl::Hidden); |
216 | |
217 | static cl::opt<bool> |
218 | EnableMachinePipeliner("aarch64-enable-pipeliner" , |
219 | cl::desc("Enable Machine Pipeliner for AArch64" ), |
220 | cl::init(Val: false), cl::Hidden); |
221 | |
222 | extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeAArch64Target() { |
223 | // Register the target. |
224 | RegisterTargetMachine<AArch64leTargetMachine> X(getTheAArch64leTarget()); |
225 | RegisterTargetMachine<AArch64beTargetMachine> Y(getTheAArch64beTarget()); |
226 | RegisterTargetMachine<AArch64leTargetMachine> Z(getTheARM64Target()); |
227 | RegisterTargetMachine<AArch64leTargetMachine> W(getTheARM64_32Target()); |
228 | RegisterTargetMachine<AArch64leTargetMachine> V(getTheAArch64_32Target()); |
229 | auto PR = PassRegistry::getPassRegistry(); |
230 | initializeGlobalISel(*PR); |
231 | initializeAArch64A53Fix835769Pass(*PR); |
232 | initializeAArch64A57FPLoadBalancingPass(*PR); |
233 | initializeAArch64AdvSIMDScalarPass(*PR); |
234 | initializeAArch64BranchTargetsPass(*PR); |
235 | initializeAArch64CollectLOHPass(*PR); |
236 | initializeAArch64CompressJumpTablesPass(*PR); |
237 | initializeAArch64ConditionalComparesPass(*PR); |
238 | initializeAArch64ConditionOptimizerPass(*PR); |
239 | initializeAArch64DeadRegisterDefinitionsPass(*PR); |
240 | initializeAArch64ExpandPseudoPass(*PR); |
241 | initializeAArch64LoadStoreOptPass(*PR); |
242 | initializeAArch64MIPeepholeOptPass(*PR); |
243 | initializeAArch64SIMDInstrOptPass(*PR); |
244 | initializeAArch64O0PreLegalizerCombinerPass(*PR); |
245 | initializeAArch64PreLegalizerCombinerPass(*PR); |
246 | initializeAArch64PointerAuthPass(*PR); |
247 | initializeAArch64PostCoalescerPass(*PR); |
248 | initializeAArch64PostLegalizerCombinerPass(*PR); |
249 | initializeAArch64PostLegalizerLoweringPass(*PR); |
250 | initializeAArch64PostSelectOptimizePass(*PR); |
251 | initializeAArch64PromoteConstantPass(*PR); |
252 | initializeAArch64RedundantCopyEliminationPass(*PR); |
253 | initializeAArch64StorePairSuppressPass(*PR); |
254 | initializeFalkorHWPFFixPass(*PR); |
255 | initializeFalkorMarkStridedAccessesLegacyPass(*PR); |
256 | initializeLDTLSCleanupPass(*PR); |
257 | initializeKCFIPass(*PR); |
258 | initializeSMEABIPass(*PR); |
259 | initializeSVEIntrinsicOptsPass(*PR); |
260 | initializeAArch64SpeculationHardeningPass(*PR); |
261 | initializeAArch64SLSHardeningPass(*PR); |
262 | initializeAArch64StackTaggingPass(*PR); |
263 | initializeAArch64StackTaggingPreRAPass(*PR); |
264 | initializeAArch64LowerHomogeneousPrologEpilogPass(*PR); |
265 | initializeAArch64DAGToDAGISelLegacyPass(*PR); |
266 | initializeAArch64GlobalsTaggingPass(*PR); |
267 | } |
268 | |
269 | //===----------------------------------------------------------------------===// |
270 | // AArch64 Lowering public interface. |
271 | //===----------------------------------------------------------------------===// |
272 | static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) { |
273 | if (TT.isOSBinFormatMachO()) |
274 | return std::make_unique<AArch64_MachoTargetObjectFile>(); |
275 | if (TT.isOSBinFormatCOFF()) |
276 | return std::make_unique<AArch64_COFFTargetObjectFile>(); |
277 | |
278 | return std::make_unique<AArch64_ELFTargetObjectFile>(); |
279 | } |
280 | |
281 | // Helper function to build a DataLayout string |
282 | static std::string computeDataLayout(const Triple &TT, |
283 | const MCTargetOptions &Options, |
284 | bool LittleEndian) { |
285 | if (TT.isOSBinFormatMachO()) { |
286 | if (TT.getArch() == Triple::aarch64_32) |
287 | return "e-m:o-p:32:32-i64:64-i128:128-n32:64-S128-Fn32" ; |
288 | return "e-m:o-i64:64-i128:128-n32:64-S128-Fn32" ; |
289 | } |
290 | if (TT.isOSBinFormatCOFF()) |
291 | return "e-m:w-p:64:64-i32:32-i64:64-i128:128-n32:64-S128-Fn32" ; |
292 | std::string Endian = LittleEndian ? "e" : "E" ; |
293 | std::string Ptr32 = TT.getEnvironment() == Triple::GNUILP32 ? "-p:32:32" : "" ; |
294 | return Endian + "-m:e" + Ptr32 + |
295 | "-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128-Fn32" ; |
296 | } |
297 | |
298 | static StringRef computeDefaultCPU(const Triple &TT, StringRef CPU) { |
299 | if (CPU.empty() && TT.isArm64e()) |
300 | return "apple-a12" ; |
301 | return CPU; |
302 | } |
303 | |
304 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, |
305 | std::optional<Reloc::Model> RM) { |
306 | // AArch64 Darwin and Windows are always PIC. |
307 | if (TT.isOSDarwin() || TT.isOSWindows()) |
308 | return Reloc::PIC_; |
309 | // On ELF platforms the default static relocation model has a smart enough |
310 | // linker to cope with referencing external symbols defined in a shared |
311 | // library. Hence DynamicNoPIC doesn't need to be promoted to PIC. |
312 | if (!RM || *RM == Reloc::DynamicNoPIC) |
313 | return Reloc::Static; |
314 | return *RM; |
315 | } |
316 | |
317 | static CodeModel::Model |
318 | getEffectiveAArch64CodeModel(const Triple &TT, |
319 | std::optional<CodeModel::Model> CM, bool JIT) { |
320 | if (CM) { |
321 | if (*CM != CodeModel::Small && *CM != CodeModel::Tiny && |
322 | *CM != CodeModel::Large) { |
323 | report_fatal_error( |
324 | reason: "Only small, tiny and large code models are allowed on AArch64" ); |
325 | } else if (*CM == CodeModel::Tiny && !TT.isOSBinFormatELF()) |
326 | report_fatal_error(reason: "tiny code model is only supported on ELF" ); |
327 | return *CM; |
328 | } |
329 | // The default MCJIT memory managers make no guarantees about where they can |
330 | // find an executable page; JITed code needs to be able to refer to globals |
331 | // no matter how far away they are. |
332 | // We should set the CodeModel::Small for Windows ARM64 in JIT mode, |
333 | // since with large code model LLVM generating 4 MOV instructions, and |
334 | // Windows doesn't support relocating these long branch (4 MOVs). |
335 | if (JIT && !TT.isOSWindows()) |
336 | return CodeModel::Large; |
337 | return CodeModel::Small; |
338 | } |
339 | |
340 | /// Create an AArch64 architecture model. |
341 | /// |
342 | AArch64TargetMachine::AArch64TargetMachine(const Target &T, const Triple &TT, |
343 | StringRef CPU, StringRef FS, |
344 | const TargetOptions &Options, |
345 | std::optional<Reloc::Model> RM, |
346 | std::optional<CodeModel::Model> CM, |
347 | CodeGenOptLevel OL, bool JIT, |
348 | bool LittleEndian) |
349 | : LLVMTargetMachine(T, |
350 | computeDataLayout(TT, Options: Options.MCOptions, LittleEndian), |
351 | TT, computeDefaultCPU(TT, CPU), FS, Options, |
352 | getEffectiveRelocModel(TT, RM), |
353 | getEffectiveAArch64CodeModel(TT, CM, JIT), OL), |
354 | TLOF(createTLOF(TT: getTargetTriple())), isLittle(LittleEndian) { |
355 | initAsmInfo(); |
356 | |
357 | if (TT.isOSBinFormatMachO()) { |
358 | this->Options.TrapUnreachable = true; |
359 | this->Options.NoTrapAfterNoreturn = true; |
360 | } |
361 | |
362 | if (getMCAsmInfo()->usesWindowsCFI()) { |
363 | // Unwinding can get confused if the last instruction in an |
364 | // exception-handling region (function, funclet, try block, etc.) |
365 | // is a call. |
366 | // |
367 | // FIXME: We could elide the trap if the next instruction would be in |
368 | // the same region anyway. |
369 | this->Options.TrapUnreachable = true; |
370 | } |
371 | |
372 | if (this->Options.TLSSize == 0) // default |
373 | this->Options.TLSSize = 24; |
374 | if ((getCodeModel() == CodeModel::Small || |
375 | getCodeModel() == CodeModel::Kernel) && |
376 | this->Options.TLSSize > 32) |
377 | // for the small (and kernel) code model, the maximum TLS size is 4GiB |
378 | this->Options.TLSSize = 32; |
379 | else if (getCodeModel() == CodeModel::Tiny && this->Options.TLSSize > 24) |
380 | // for the tiny code model, the maximum TLS size is 1MiB (< 16MiB) |
381 | this->Options.TLSSize = 24; |
382 | |
383 | // Enable GlobalISel at or below EnableGlobalISelAt0, unless this is |
384 | // MachO/CodeModel::Large, which GlobalISel does not support. |
385 | if (static_cast<int>(getOptLevel()) <= EnableGlobalISelAtO && |
386 | TT.getArch() != Triple::aarch64_32 && |
387 | TT.getEnvironment() != Triple::GNUILP32 && |
388 | !(getCodeModel() == CodeModel::Large && TT.isOSBinFormatMachO())) { |
389 | setGlobalISel(true); |
390 | setGlobalISelAbort(GlobalISelAbortMode::Disable); |
391 | } |
392 | |
393 | // AArch64 supports the MachineOutliner. |
394 | setMachineOutliner(true); |
395 | |
396 | // AArch64 supports default outlining behaviour. |
397 | setSupportsDefaultOutlining(true); |
398 | |
399 | // AArch64 supports the debug entry values. |
400 | setSupportsDebugEntryValues(true); |
401 | |
402 | // AArch64 supports fixing up the DWARF unwind information. |
403 | if (!getMCAsmInfo()->usesWindowsCFI()) |
404 | setCFIFixup(true); |
405 | } |
406 | |
407 | AArch64TargetMachine::~AArch64TargetMachine() = default; |
408 | |
409 | const AArch64Subtarget * |
410 | AArch64TargetMachine::getSubtargetImpl(const Function &F) const { |
411 | Attribute CPUAttr = F.getFnAttribute(Kind: "target-cpu" ); |
412 | Attribute TuneAttr = F.getFnAttribute(Kind: "tune-cpu" ); |
413 | Attribute FSAttr = F.getFnAttribute(Kind: "target-features" ); |
414 | |
415 | StringRef CPU = CPUAttr.isValid() ? CPUAttr.getValueAsString() : TargetCPU; |
416 | StringRef TuneCPU = TuneAttr.isValid() ? TuneAttr.getValueAsString() : CPU; |
417 | StringRef FS = FSAttr.isValid() ? FSAttr.getValueAsString() : TargetFS; |
418 | bool HasMinSize = F.hasMinSize(); |
419 | |
420 | bool IsStreaming = ForceStreaming || |
421 | F.hasFnAttribute(Kind: "aarch64_pstate_sm_enabled" ) || |
422 | F.hasFnAttribute(Kind: "aarch64_pstate_sm_body" ); |
423 | bool IsStreamingCompatible = ForceStreamingCompatible || |
424 | F.hasFnAttribute(Kind: "aarch64_pstate_sm_compatible" ); |
425 | |
426 | unsigned MinSVEVectorSize = 0; |
427 | unsigned MaxSVEVectorSize = 0; |
428 | if (F.hasFnAttribute(Kind: Attribute::VScaleRange)) { |
429 | ConstantRange CR = getVScaleRange(F: &F, BitWidth: 64); |
430 | MinSVEVectorSize = CR.getUnsignedMin().getZExtValue() * 128; |
431 | MaxSVEVectorSize = CR.getUnsignedMax().getZExtValue() * 128; |
432 | } else { |
433 | MinSVEVectorSize = SVEVectorBitsMinOpt; |
434 | MaxSVEVectorSize = SVEVectorBitsMaxOpt; |
435 | } |
436 | |
437 | assert(MinSVEVectorSize % 128 == 0 && |
438 | "SVE requires vector length in multiples of 128!" ); |
439 | assert(MaxSVEVectorSize % 128 == 0 && |
440 | "SVE requires vector length in multiples of 128!" ); |
441 | assert((MaxSVEVectorSize >= MinSVEVectorSize || MaxSVEVectorSize == 0) && |
442 | "Minimum SVE vector size should not be larger than its maximum!" ); |
443 | |
444 | // Sanitize user input in case of no asserts |
445 | if (MaxSVEVectorSize != 0) { |
446 | MinSVEVectorSize = std::min(a: MinSVEVectorSize, b: MaxSVEVectorSize); |
447 | MaxSVEVectorSize = std::max(a: MinSVEVectorSize, b: MaxSVEVectorSize); |
448 | } |
449 | |
450 | SmallString<512> Key; |
451 | raw_svector_ostream(Key) << "SVEMin" << MinSVEVectorSize << "SVEMax" |
452 | << MaxSVEVectorSize << "IsStreaming=" << IsStreaming |
453 | << "IsStreamingCompatible=" << IsStreamingCompatible |
454 | << CPU << TuneCPU << FS |
455 | << "HasMinSize=" << HasMinSize; |
456 | |
457 | auto &I = SubtargetMap[Key]; |
458 | if (!I) { |
459 | // This needs to be done before we create a new subtarget since any |
460 | // creation will depend on the TM and the code generation flags on the |
461 | // function that reside in TargetOptions. |
462 | resetTargetOptions(F); |
463 | I = std::make_unique<AArch64Subtarget>( |
464 | args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args: *this, args: isLittle, args&: MinSVEVectorSize, |
465 | args&: MaxSVEVectorSize, args&: IsStreaming, args&: IsStreamingCompatible, args&: HasMinSize); |
466 | } |
467 | |
468 | assert((!IsStreaming || I->hasSME()) && "Expected SME to be available" ); |
469 | |
470 | return I.get(); |
471 | } |
472 | |
473 | void AArch64leTargetMachine::anchor() { } |
474 | |
475 | AArch64leTargetMachine::AArch64leTargetMachine( |
476 | const Target &T, const Triple &TT, StringRef CPU, StringRef FS, |
477 | const TargetOptions &Options, std::optional<Reloc::Model> RM, |
478 | std::optional<CodeModel::Model> CM, CodeGenOptLevel OL, bool JIT) |
479 | : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, true) {} |
480 | |
481 | void AArch64beTargetMachine::anchor() { } |
482 | |
483 | AArch64beTargetMachine::AArch64beTargetMachine( |
484 | const Target &T, const Triple &TT, StringRef CPU, StringRef FS, |
485 | const TargetOptions &Options, std::optional<Reloc::Model> RM, |
486 | std::optional<CodeModel::Model> CM, CodeGenOptLevel OL, bool JIT) |
487 | : AArch64TargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, JIT, false) {} |
488 | |
489 | namespace { |
490 | |
491 | /// AArch64 Code Generator Pass Configuration Options. |
492 | class AArch64PassConfig : public TargetPassConfig { |
493 | public: |
494 | AArch64PassConfig(AArch64TargetMachine &TM, PassManagerBase &PM) |
495 | : TargetPassConfig(TM, PM) { |
496 | if (TM.getOptLevel() != CodeGenOptLevel::None) |
497 | substitutePass(StandardID: &PostRASchedulerID, TargetID: &PostMachineSchedulerID); |
498 | setEnableSinkAndFold(EnableSinkFold); |
499 | } |
500 | |
501 | AArch64TargetMachine &getAArch64TargetMachine() const { |
502 | return getTM<AArch64TargetMachine>(); |
503 | } |
504 | |
505 | ScheduleDAGInstrs * |
506 | createMachineScheduler(MachineSchedContext *C) const override { |
507 | const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); |
508 | ScheduleDAGMILive *DAG = createGenericSchedLive(C); |
509 | DAG->addMutation(Mutation: createLoadClusterDAGMutation(TII: DAG->TII, TRI: DAG->TRI)); |
510 | DAG->addMutation(Mutation: createStoreClusterDAGMutation(TII: DAG->TII, TRI: DAG->TRI)); |
511 | if (ST.hasFusion()) |
512 | DAG->addMutation(Mutation: createAArch64MacroFusionDAGMutation()); |
513 | return DAG; |
514 | } |
515 | |
516 | ScheduleDAGInstrs * |
517 | createPostMachineScheduler(MachineSchedContext *C) const override { |
518 | const AArch64Subtarget &ST = C->MF->getSubtarget<AArch64Subtarget>(); |
519 | ScheduleDAGMI *DAG = |
520 | new ScheduleDAGMI(C, std::make_unique<AArch64PostRASchedStrategy>(args&: C), |
521 | /* RemoveKillFlags=*/true); |
522 | if (ST.hasFusion()) { |
523 | // Run the Macro Fusion after RA again since literals are expanded from |
524 | // pseudos then (v. addPreSched2()). |
525 | DAG->addMutation(Mutation: createAArch64MacroFusionDAGMutation()); |
526 | return DAG; |
527 | } |
528 | |
529 | return DAG; |
530 | } |
531 | |
532 | void addIRPasses() override; |
533 | bool addPreISel() override; |
534 | void addCodeGenPrepare() override; |
535 | bool addInstSelector() override; |
536 | bool addIRTranslator() override; |
537 | void addPreLegalizeMachineIR() override; |
538 | bool addLegalizeMachineIR() override; |
539 | void addPreRegBankSelect() override; |
540 | bool addRegBankSelect() override; |
541 | bool addGlobalInstructionSelect() override; |
542 | void addMachineSSAOptimization() override; |
543 | bool addILPOpts() override; |
544 | void addPreRegAlloc() override; |
545 | void addPostRegAlloc() override; |
546 | void addPreSched2() override; |
547 | void addPreEmitPass() override; |
548 | void addPostBBSections() override; |
549 | void addPreEmitPass2() override; |
550 | bool addRegAssignAndRewriteOptimized() override; |
551 | |
552 | std::unique_ptr<CSEConfigBase> getCSEConfig() const override; |
553 | }; |
554 | |
555 | } // end anonymous namespace |
556 | |
557 | void AArch64TargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) { |
558 | |
559 | PB.registerLateLoopOptimizationsEPCallback( |
560 | C: [=](LoopPassManager &LPM, OptimizationLevel Level) { |
561 | LPM.addPass(Pass: LoopIdiomVectorizePass()); |
562 | }); |
563 | } |
564 | |
565 | TargetTransformInfo |
566 | AArch64TargetMachine::getTargetTransformInfo(const Function &F) const { |
567 | return TargetTransformInfo(AArch64TTIImpl(this, F)); |
568 | } |
569 | |
570 | TargetPassConfig *AArch64TargetMachine::createPassConfig(PassManagerBase &PM) { |
571 | return new AArch64PassConfig(*this, PM); |
572 | } |
573 | |
574 | std::unique_ptr<CSEConfigBase> AArch64PassConfig::getCSEConfig() const { |
575 | return getStandardCSEConfigForOpt(Level: TM->getOptLevel()); |
576 | } |
577 | |
578 | void AArch64PassConfig::addIRPasses() { |
579 | // Always expand atomic operations, we don't deal with atomicrmw or cmpxchg |
580 | // ourselves. |
581 | addPass(P: createAtomicExpandLegacyPass()); |
582 | |
583 | // Expand any SVE vector library calls that we can't code generate directly. |
584 | if (EnableSVEIntrinsicOpts && |
585 | TM->getOptLevel() == CodeGenOptLevel::Aggressive) |
586 | addPass(P: createSVEIntrinsicOptsPass()); |
587 | |
588 | // Cmpxchg instructions are often used with a subsequent comparison to |
589 | // determine whether it succeeded. We can exploit existing control-flow in |
590 | // ldrex/strex loops to simplify this, but it needs tidying up. |
591 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableAtomicTidy) |
592 | addPass(P: createCFGSimplificationPass(Options: SimplifyCFGOptions() |
593 | .forwardSwitchCondToPhi(B: true) |
594 | .convertSwitchRangeToICmp(B: true) |
595 | .convertSwitchToLookupTable(B: true) |
596 | .needCanonicalLoops(B: false) |
597 | .hoistCommonInsts(B: true) |
598 | .sinkCommonInsts(B: true))); |
599 | |
600 | // Run LoopDataPrefetch |
601 | // |
602 | // Run this before LSR to remove the multiplies involved in computing the |
603 | // pointer values N iterations ahead. |
604 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
605 | if (EnableLoopDataPrefetch) |
606 | addPass(P: createLoopDataPrefetchPass()); |
607 | if (EnableFalkorHWPFFix) |
608 | addPass(P: createFalkorMarkStridedAccessesPass()); |
609 | } |
610 | |
611 | if (EnableGEPOpt) { |
612 | // Call SeparateConstOffsetFromGEP pass to extract constants within indices |
613 | // and lower a GEP with multiple indices to either arithmetic operations or |
614 | // multiple GEPs with single index. |
615 | addPass(P: createSeparateConstOffsetFromGEPPass(LowerGEP: true)); |
616 | // Call EarlyCSE pass to find and remove subexpressions in the lowered |
617 | // result. |
618 | addPass(P: createEarlyCSEPass()); |
619 | // Do loop invariant code motion in case part of the lowered result is |
620 | // invariant. |
621 | addPass(P: createLICMPass()); |
622 | } |
623 | |
624 | TargetPassConfig::addIRPasses(); |
625 | |
626 | if (getOptLevel() == CodeGenOptLevel::Aggressive && EnableSelectOpt) |
627 | addPass(P: createSelectOptimizePass()); |
628 | |
629 | addPass(P: createAArch64GlobalsTaggingPass()); |
630 | addPass(P: createAArch64StackTaggingPass( |
631 | /*IsOptNone=*/TM->getOptLevel() == CodeGenOptLevel::None)); |
632 | |
633 | // Match complex arithmetic patterns |
634 | if (TM->getOptLevel() >= CodeGenOptLevel::Default) |
635 | addPass(P: createComplexDeinterleavingPass(TM)); |
636 | |
637 | // Match interleaved memory accesses to ldN/stN intrinsics. |
638 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
639 | addPass(P: createInterleavedLoadCombinePass()); |
640 | addPass(P: createInterleavedAccessPass()); |
641 | } |
642 | |
643 | // Expand any functions marked with SME attributes which require special |
644 | // changes for the calling convention or that require the lazy-saving |
645 | // mechanism specified in the SME ABI. |
646 | addPass(P: createSMEABIPass()); |
647 | |
648 | // Add Control Flow Guard checks. |
649 | if (TM->getTargetTriple().isOSWindows()) { |
650 | if (TM->getTargetTriple().isWindowsArm64EC()) |
651 | addPass(P: createAArch64Arm64ECCallLoweringPass()); |
652 | else |
653 | addPass(P: createCFGuardCheckPass()); |
654 | } |
655 | |
656 | if (TM->Options.JMCInstrument) |
657 | addPass(P: createJMCInstrumenterPass()); |
658 | } |
659 | |
660 | // Pass Pipeline Configuration |
661 | bool AArch64PassConfig::addPreISel() { |
662 | // Run promote constant before global merge, so that the promoted constants |
663 | // get a chance to be merged |
664 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnablePromoteConstant) |
665 | addPass(P: createAArch64PromoteConstantPass()); |
666 | // FIXME: On AArch64, this depends on the type. |
667 | // Basically, the addressable offsets are up to 4095 * Ty.getSizeInBytes(). |
668 | // and the offset has to be a multiple of the related size in bytes. |
669 | if ((TM->getOptLevel() != CodeGenOptLevel::None && |
670 | EnableGlobalMerge == cl::BOU_UNSET) || |
671 | EnableGlobalMerge == cl::BOU_TRUE) { |
672 | bool OnlyOptimizeForSize = |
673 | (TM->getOptLevel() < CodeGenOptLevel::Aggressive) && |
674 | (EnableGlobalMerge == cl::BOU_UNSET); |
675 | |
676 | // Merging of extern globals is enabled by default on non-Mach-O as we |
677 | // expect it to be generally either beneficial or harmless. On Mach-O it |
678 | // is disabled as we emit the .subsections_via_symbols directive which |
679 | // means that merging extern globals is not safe. |
680 | bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO(); |
681 | |
682 | // FIXME: extern global merging is only enabled when we optimise for size |
683 | // because there are some regressions with it also enabled for performance. |
684 | if (!OnlyOptimizeForSize) |
685 | MergeExternalByDefault = false; |
686 | |
687 | addPass(P: createGlobalMergePass(TM, MaximalOffset: 4095, OnlyOptimizeForSize, |
688 | MergeExternalByDefault)); |
689 | } |
690 | |
691 | return false; |
692 | } |
693 | |
694 | void AArch64PassConfig::addCodeGenPrepare() { |
695 | if (getOptLevel() != CodeGenOptLevel::None) |
696 | addPass(P: createTypePromotionLegacyPass()); |
697 | TargetPassConfig::addCodeGenPrepare(); |
698 | } |
699 | |
700 | bool AArch64PassConfig::addInstSelector() { |
701 | addPass(P: createAArch64ISelDag(TM&: getAArch64TargetMachine(), OptLevel: getOptLevel())); |
702 | |
703 | // For ELF, cleanup any local-dynamic TLS accesses (i.e. combine as many |
704 | // references to _TLS_MODULE_BASE_ as possible. |
705 | if (TM->getTargetTriple().isOSBinFormatELF() && |
706 | getOptLevel() != CodeGenOptLevel::None) |
707 | addPass(P: createAArch64CleanupLocalDynamicTLSPass()); |
708 | |
709 | return false; |
710 | } |
711 | |
712 | bool AArch64PassConfig::addIRTranslator() { |
713 | addPass(P: new IRTranslator(getOptLevel())); |
714 | return false; |
715 | } |
716 | |
717 | void AArch64PassConfig::addPreLegalizeMachineIR() { |
718 | if (getOptLevel() == CodeGenOptLevel::None) { |
719 | addPass(P: createAArch64O0PreLegalizerCombiner()); |
720 | addPass(P: new Localizer()); |
721 | } else { |
722 | addPass(P: createAArch64PreLegalizerCombiner()); |
723 | addPass(P: new Localizer()); |
724 | if (EnableGISelLoadStoreOptPreLegal) |
725 | addPass(P: new LoadStoreOpt()); |
726 | } |
727 | } |
728 | |
729 | bool AArch64PassConfig::addLegalizeMachineIR() { |
730 | addPass(P: new Legalizer()); |
731 | return false; |
732 | } |
733 | |
734 | void AArch64PassConfig::addPreRegBankSelect() { |
735 | bool IsOptNone = getOptLevel() == CodeGenOptLevel::None; |
736 | if (!IsOptNone) { |
737 | addPass(P: createAArch64PostLegalizerCombiner(IsOptNone)); |
738 | if (EnableGISelLoadStoreOptPostLegal) |
739 | addPass(P: new LoadStoreOpt()); |
740 | } |
741 | addPass(P: createAArch64PostLegalizerLowering()); |
742 | } |
743 | |
744 | bool AArch64PassConfig::addRegBankSelect() { |
745 | addPass(P: new RegBankSelect()); |
746 | return false; |
747 | } |
748 | |
749 | bool AArch64PassConfig::addGlobalInstructionSelect() { |
750 | addPass(P: new InstructionSelect(getOptLevel())); |
751 | if (getOptLevel() != CodeGenOptLevel::None) |
752 | addPass(P: createAArch64PostSelectOptimize()); |
753 | return false; |
754 | } |
755 | |
756 | void AArch64PassConfig::addMachineSSAOptimization() { |
757 | // Run default MachineSSAOptimization first. |
758 | TargetPassConfig::addMachineSSAOptimization(); |
759 | |
760 | if (TM->getOptLevel() != CodeGenOptLevel::None) |
761 | addPass(P: createAArch64MIPeepholeOptPass()); |
762 | } |
763 | |
764 | bool AArch64PassConfig::addILPOpts() { |
765 | if (EnableCondOpt) |
766 | addPass(P: createAArch64ConditionOptimizerPass()); |
767 | if (EnableCCMP) |
768 | addPass(P: createAArch64ConditionalCompares()); |
769 | if (EnableMCR) |
770 | addPass(PassID: &MachineCombinerID); |
771 | if (EnableCondBrTuning) |
772 | addPass(P: createAArch64CondBrTuning()); |
773 | if (EnableEarlyIfConversion) |
774 | addPass(PassID: &EarlyIfConverterID); |
775 | if (EnableStPairSuppress) |
776 | addPass(P: createAArch64StorePairSuppressPass()); |
777 | addPass(P: createAArch64SIMDInstrOptPass()); |
778 | if (TM->getOptLevel() != CodeGenOptLevel::None) |
779 | addPass(P: createAArch64StackTaggingPreRAPass()); |
780 | return true; |
781 | } |
782 | |
783 | void AArch64PassConfig::addPreRegAlloc() { |
784 | // Change dead register definitions to refer to the zero register. |
785 | if (TM->getOptLevel() != CodeGenOptLevel::None && |
786 | EnableDeadRegisterElimination) |
787 | addPass(P: createAArch64DeadRegisterDefinitions()); |
788 | |
789 | // Use AdvSIMD scalar instructions whenever profitable. |
790 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableAdvSIMDScalar) { |
791 | addPass(P: createAArch64AdvSIMDScalar()); |
792 | // The AdvSIMD pass may produce copies that can be rewritten to |
793 | // be register coalescer friendly. |
794 | addPass(PassID: &PeepholeOptimizerID); |
795 | } |
796 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableMachinePipeliner) |
797 | addPass(PassID: &MachinePipelinerID); |
798 | } |
799 | |
800 | void AArch64PassConfig::addPostRegAlloc() { |
801 | // Remove redundant copy instructions. |
802 | if (TM->getOptLevel() != CodeGenOptLevel::None && |
803 | EnableRedundantCopyElimination) |
804 | addPass(P: createAArch64RedundantCopyEliminationPass()); |
805 | |
806 | if (TM->getOptLevel() != CodeGenOptLevel::None && usingDefaultRegAlloc()) |
807 | // Improve performance for some FP/SIMD code for A57. |
808 | addPass(P: createAArch64A57FPLoadBalancing()); |
809 | } |
810 | |
811 | void AArch64PassConfig::addPreSched2() { |
812 | // Lower homogeneous frame instructions |
813 | if (EnableHomogeneousPrologEpilog) |
814 | addPass(P: createAArch64LowerHomogeneousPrologEpilogPass()); |
815 | // Expand some pseudo instructions to allow proper scheduling. |
816 | addPass(P: createAArch64ExpandPseudoPass()); |
817 | // Use load/store pair instructions when possible. |
818 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
819 | if (EnableLoadStoreOpt) |
820 | addPass(P: createAArch64LoadStoreOptimizationPass()); |
821 | } |
822 | // Emit KCFI checks for indirect calls. |
823 | addPass(P: createKCFIPass()); |
824 | |
825 | // The AArch64SpeculationHardeningPass destroys dominator tree and natural |
826 | // loop info, which is needed for the FalkorHWPFFixPass and also later on. |
827 | // Therefore, run the AArch64SpeculationHardeningPass before the |
828 | // FalkorHWPFFixPass to avoid recomputing dominator tree and natural loop |
829 | // info. |
830 | addPass(P: createAArch64SpeculationHardeningPass()); |
831 | |
832 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
833 | if (EnableFalkorHWPFFix) |
834 | addPass(P: createFalkorHWPFFixPass()); |
835 | } |
836 | } |
837 | |
838 | void AArch64PassConfig::addPreEmitPass() { |
839 | // Machine Block Placement might have created new opportunities when run |
840 | // at O3, where the Tail Duplication Threshold is set to 4 instructions. |
841 | // Run the load/store optimizer once more. |
842 | if (TM->getOptLevel() >= CodeGenOptLevel::Aggressive && EnableLoadStoreOpt) |
843 | addPass(P: createAArch64LoadStoreOptimizationPass()); |
844 | |
845 | if (TM->getOptLevel() >= CodeGenOptLevel::Aggressive && |
846 | EnableAArch64CopyPropagation) |
847 | addPass(P: createMachineCopyPropagationPass(UseCopyInstr: true)); |
848 | |
849 | addPass(P: createAArch64A53Fix835769()); |
850 | |
851 | if (TM->getTargetTriple().isOSWindows()) { |
852 | // Identify valid longjmp targets for Windows Control Flow Guard. |
853 | addPass(P: createCFGuardLongjmpPass()); |
854 | // Identify valid eh continuation targets for Windows EHCont Guard. |
855 | addPass(P: createEHContGuardCatchretPass()); |
856 | } |
857 | |
858 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableCollectLOH && |
859 | TM->getTargetTriple().isOSBinFormatMachO()) |
860 | addPass(P: createAArch64CollectLOHPass()); |
861 | } |
862 | |
863 | void AArch64PassConfig::addPostBBSections() { |
864 | addPass(P: createAArch64SLSHardeningPass()); |
865 | addPass(P: createAArch64PointerAuthPass()); |
866 | if (EnableBranchTargets) |
867 | addPass(P: createAArch64BranchTargetsPass()); |
868 | // Relax conditional branch instructions if they're otherwise out of |
869 | // range of their destination. |
870 | if (BranchRelaxation) |
871 | addPass(PassID: &BranchRelaxationPassID); |
872 | |
873 | if (TM->getOptLevel() != CodeGenOptLevel::None && EnableCompressJumpTables) |
874 | addPass(P: createAArch64CompressJumpTablesPass()); |
875 | } |
876 | |
877 | void AArch64PassConfig::addPreEmitPass2() { |
878 | // SVE bundles move prefixes with destructive operations. BLR_RVMARKER pseudo |
879 | // instructions are lowered to bundles as well. |
880 | addPass(P: createUnpackMachineBundles(Ftor: nullptr)); |
881 | } |
882 | |
883 | bool AArch64PassConfig::addRegAssignAndRewriteOptimized() { |
884 | addPass(P: createAArch64PostCoalescerPass()); |
885 | return TargetPassConfig::addRegAssignAndRewriteOptimized(); |
886 | } |
887 | |
888 | MachineFunctionInfo *AArch64TargetMachine::createMachineFunctionInfo( |
889 | BumpPtrAllocator &Allocator, const Function &F, |
890 | const TargetSubtargetInfo *STI) const { |
891 | return AArch64FunctionInfo::create<AArch64FunctionInfo>( |
892 | Allocator, F, STI: static_cast<const AArch64Subtarget *>(STI)); |
893 | } |
894 | |
895 | yaml::MachineFunctionInfo * |
896 | AArch64TargetMachine::createDefaultFuncInfoYAML() const { |
897 | return new yaml::AArch64FunctionInfo(); |
898 | } |
899 | |
900 | yaml::MachineFunctionInfo * |
901 | AArch64TargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
902 | const auto *MFI = MF.getInfo<AArch64FunctionInfo>(); |
903 | return new yaml::AArch64FunctionInfo(*MFI); |
904 | } |
905 | |
906 | bool AArch64TargetMachine::parseMachineFunctionInfo( |
907 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
908 | SMDiagnostic &Error, SMRange &SourceRange) const { |
909 | const auto &YamlMFI = static_cast<const yaml::AArch64FunctionInfo &>(MFI); |
910 | MachineFunction &MF = PFS.MF; |
911 | MF.getInfo<AArch64FunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
912 | return false; |
913 | } |
914 | |