1 | //===-- RISCVTargetMachine.cpp - Define TargetMachine for RISC-V ----------===// |
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 | // Implements the info about RISC-V target spec. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "RISCVTargetMachine.h" |
14 | #include "MCTargetDesc/RISCVBaseInfo.h" |
15 | #include "RISCV.h" |
16 | #include "RISCVMachineFunctionInfo.h" |
17 | #include "RISCVTargetObjectFile.h" |
18 | #include "RISCVTargetTransformInfo.h" |
19 | #include "TargetInfo/RISCVTargetInfo.h" |
20 | #include "llvm/ADT/STLExtras.h" |
21 | #include "llvm/Analysis/TargetTransformInfo.h" |
22 | #include "llvm/CodeGen/GlobalISel/IRTranslator.h" |
23 | #include "llvm/CodeGen/GlobalISel/InstructionSelect.h" |
24 | #include "llvm/CodeGen/GlobalISel/Legalizer.h" |
25 | #include "llvm/CodeGen/GlobalISel/RegBankSelect.h" |
26 | #include "llvm/CodeGen/MIRParser/MIParser.h" |
27 | #include "llvm/CodeGen/MIRYamlMapping.h" |
28 | #include "llvm/CodeGen/MachineScheduler.h" |
29 | #include "llvm/CodeGen/MacroFusion.h" |
30 | #include "llvm/CodeGen/Passes.h" |
31 | #include "llvm/CodeGen/RegAllocRegistry.h" |
32 | #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" |
33 | #include "llvm/CodeGen/TargetPassConfig.h" |
34 | #include "llvm/InitializePasses.h" |
35 | #include "llvm/MC/TargetRegistry.h" |
36 | #include "llvm/Passes/PassBuilder.h" |
37 | #include "llvm/Support/FormattedStream.h" |
38 | #include "llvm/Target/TargetOptions.h" |
39 | #include "llvm/Transforms/IPO.h" |
40 | #include "llvm/Transforms/Scalar.h" |
41 | #include "llvm/Transforms/Vectorize/LoopIdiomVectorize.h" |
42 | #include <optional> |
43 | using namespace llvm; |
44 | |
45 | static cl::opt<bool> EnableRedundantCopyElimination( |
46 | "riscv-enable-copyelim" , |
47 | cl::desc("Enable the redundant copy elimination pass" ), cl::init(Val: true), |
48 | cl::Hidden); |
49 | |
50 | // FIXME: Unify control over GlobalMerge. |
51 | static cl::opt<cl::boolOrDefault> |
52 | EnableGlobalMerge("riscv-enable-global-merge" , cl::Hidden, |
53 | cl::desc("Enable the global merge pass" )); |
54 | |
55 | static cl::opt<bool> |
56 | EnableMachineCombiner("riscv-enable-machine-combiner" , |
57 | cl::desc("Enable the machine combiner pass" ), |
58 | cl::init(Val: true), cl::Hidden); |
59 | |
60 | static cl::opt<unsigned> RVVVectorBitsMaxOpt( |
61 | "riscv-v-vector-bits-max" , |
62 | cl::desc("Assume V extension vector registers are at most this big, " |
63 | "with zero meaning no maximum size is assumed." ), |
64 | cl::init(Val: 0), cl::Hidden); |
65 | |
66 | static cl::opt<int> RVVVectorBitsMinOpt( |
67 | "riscv-v-vector-bits-min" , |
68 | cl::desc("Assume V extension vector registers are at least this big, " |
69 | "with zero meaning no minimum size is assumed. A value of -1 " |
70 | "means use Zvl*b extension. This is primarily used to enable " |
71 | "autovectorization with fixed width vectors." ), |
72 | cl::init(Val: -1), cl::Hidden); |
73 | |
74 | static cl::opt<bool> EnableRISCVCopyPropagation( |
75 | "riscv-enable-copy-propagation" , |
76 | cl::desc("Enable the copy propagation with RISC-V copy instr" ), |
77 | cl::init(Val: true), cl::Hidden); |
78 | |
79 | static cl::opt<bool> EnableRISCVDeadRegisterElimination( |
80 | "riscv-enable-dead-defs" , cl::Hidden, |
81 | cl::desc("Enable the pass that removes dead" |
82 | " definitons and replaces stores to" |
83 | " them with stores to x0" ), |
84 | cl::init(Val: true)); |
85 | |
86 | static cl::opt<bool> |
87 | EnableSinkFold("riscv-enable-sink-fold" , |
88 | cl::desc("Enable sinking and folding of instruction copies" ), |
89 | cl::init(Val: true), cl::Hidden); |
90 | |
91 | static cl::opt<bool> |
92 | EnableLoopDataPrefetch("riscv-enable-loop-data-prefetch" , cl::Hidden, |
93 | cl::desc("Enable the loop data prefetch pass" ), |
94 | cl::init(Val: true)); |
95 | |
96 | static cl::opt<bool> EnableMISchedLoadClustering( |
97 | "riscv-misched-load-clustering" , cl::Hidden, |
98 | cl::desc("Enable load clustering in the machine scheduler" ), |
99 | cl::init(Val: false)); |
100 | |
101 | static cl::opt<bool> EnableVSETVLIAfterRVVRegAlloc( |
102 | "riscv-vsetvl-after-rvv-regalloc" , cl::Hidden, |
103 | cl::desc("Insert vsetvls after vector register allocation" ), |
104 | cl::init(Val: true)); |
105 | |
106 | extern "C" LLVM_EXTERNAL_VISIBILITY void LLVMInitializeRISCVTarget() { |
107 | RegisterTargetMachine<RISCVTargetMachine> X(getTheRISCV32Target()); |
108 | RegisterTargetMachine<RISCVTargetMachine> Y(getTheRISCV64Target()); |
109 | auto *PR = PassRegistry::getPassRegistry(); |
110 | initializeGlobalISel(*PR); |
111 | initializeRISCVO0PreLegalizerCombinerPass(*PR); |
112 | initializeRISCVPreLegalizerCombinerPass(*PR); |
113 | initializeRISCVPostLegalizerCombinerPass(*PR); |
114 | initializeKCFIPass(*PR); |
115 | initializeRISCVDeadRegisterDefinitionsPass(*PR); |
116 | initializeRISCVMakeCompressibleOptPass(*PR); |
117 | initializeRISCVGatherScatterLoweringPass(*PR); |
118 | initializeRISCVCodeGenPreparePass(*PR); |
119 | initializeRISCVPostRAExpandPseudoPass(*PR); |
120 | initializeRISCVMergeBaseOffsetOptPass(*PR); |
121 | initializeRISCVOptWInstrsPass(*PR); |
122 | initializeRISCVPreRAExpandPseudoPass(*PR); |
123 | initializeRISCVExpandPseudoPass(*PR); |
124 | initializeRISCVVectorPeepholePass(*PR); |
125 | initializeRISCVInsertVSETVLIPass(*PR); |
126 | initializeRISCVInsertReadWriteCSRPass(*PR); |
127 | initializeRISCVInsertWriteVXRMPass(*PR); |
128 | initializeRISCVDAGToDAGISelLegacyPass(*PR); |
129 | initializeRISCVMoveMergePass(*PR); |
130 | initializeRISCVPushPopOptPass(*PR); |
131 | } |
132 | |
133 | static StringRef computeDataLayout(const Triple &TT, |
134 | const TargetOptions &Options) { |
135 | StringRef ABIName = Options.MCOptions.getABIName(); |
136 | if (TT.isArch64Bit()) { |
137 | if (ABIName == "lp64e" ) |
138 | return "e-m:e-p:64:64-i64:64-i128:128-n32:64-S64" ; |
139 | |
140 | return "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128" ; |
141 | } |
142 | assert(TT.isArch32Bit() && "only RV32 and RV64 are currently supported" ); |
143 | |
144 | if (ABIName == "ilp32e" ) |
145 | return "e-m:e-p:32:32-i64:64-n32-S32" ; |
146 | |
147 | return "e-m:e-p:32:32-i64:64-n32-S128" ; |
148 | } |
149 | |
150 | static Reloc::Model getEffectiveRelocModel(const Triple &TT, |
151 | std::optional<Reloc::Model> RM) { |
152 | return RM.value_or(u: Reloc::Static); |
153 | } |
154 | |
155 | RISCVTargetMachine::RISCVTargetMachine(const Target &T, const Triple &TT, |
156 | StringRef CPU, StringRef FS, |
157 | const TargetOptions &Options, |
158 | std::optional<Reloc::Model> RM, |
159 | std::optional<CodeModel::Model> CM, |
160 | CodeGenOptLevel OL, bool JIT) |
161 | : LLVMTargetMachine(T, computeDataLayout(TT, Options), TT, CPU, FS, Options, |
162 | getEffectiveRelocModel(TT, RM), |
163 | getEffectiveCodeModel(CM, Default: CodeModel::Small), OL), |
164 | TLOF(std::make_unique<RISCVELFTargetObjectFile>()) { |
165 | initAsmInfo(); |
166 | |
167 | // RISC-V supports the MachineOutliner. |
168 | setMachineOutliner(true); |
169 | setSupportsDefaultOutlining(true); |
170 | |
171 | if (TT.isOSFuchsia() && !TT.isArch64Bit()) |
172 | report_fatal_error(reason: "Fuchsia is only supported for 64-bit" ); |
173 | } |
174 | |
175 | const RISCVSubtarget * |
176 | RISCVTargetMachine::getSubtargetImpl(const Function &F) const { |
177 | Attribute CPUAttr = F.getFnAttribute(Kind: "target-cpu" ); |
178 | Attribute TuneAttr = F.getFnAttribute(Kind: "tune-cpu" ); |
179 | Attribute FSAttr = F.getFnAttribute(Kind: "target-features" ); |
180 | |
181 | std::string CPU = |
182 | CPUAttr.isValid() ? CPUAttr.getValueAsString().str() : TargetCPU; |
183 | std::string TuneCPU = |
184 | TuneAttr.isValid() ? TuneAttr.getValueAsString().str() : CPU; |
185 | std::string FS = |
186 | FSAttr.isValid() ? FSAttr.getValueAsString().str() : TargetFS; |
187 | |
188 | unsigned RVVBitsMin = RVVVectorBitsMinOpt; |
189 | unsigned RVVBitsMax = RVVVectorBitsMaxOpt; |
190 | |
191 | Attribute VScaleRangeAttr = F.getFnAttribute(Kind: Attribute::VScaleRange); |
192 | if (VScaleRangeAttr.isValid()) { |
193 | if (!RVVVectorBitsMinOpt.getNumOccurrences()) |
194 | RVVBitsMin = VScaleRangeAttr.getVScaleRangeMin() * RISCV::RVVBitsPerBlock; |
195 | std::optional<unsigned> VScaleMax = VScaleRangeAttr.getVScaleRangeMax(); |
196 | if (VScaleMax.has_value() && !RVVVectorBitsMaxOpt.getNumOccurrences()) |
197 | RVVBitsMax = *VScaleMax * RISCV::RVVBitsPerBlock; |
198 | } |
199 | |
200 | if (RVVBitsMin != -1U) { |
201 | // FIXME: Change to >= 32 when VLEN = 32 is supported. |
202 | assert((RVVBitsMin == 0 || (RVVBitsMin >= 64 && RVVBitsMin <= 65536 && |
203 | isPowerOf2_32(RVVBitsMin))) && |
204 | "V or Zve* extension requires vector length to be in the range of " |
205 | "64 to 65536 and a power 2!" ); |
206 | assert((RVVBitsMax >= RVVBitsMin || RVVBitsMax == 0) && |
207 | "Minimum V extension vector length should not be larger than its " |
208 | "maximum!" ); |
209 | } |
210 | assert((RVVBitsMax == 0 || (RVVBitsMax >= 64 && RVVBitsMax <= 65536 && |
211 | isPowerOf2_32(RVVBitsMax))) && |
212 | "V or Zve* extension requires vector length to be in the range of " |
213 | "64 to 65536 and a power 2!" ); |
214 | |
215 | if (RVVBitsMin != -1U) { |
216 | if (RVVBitsMax != 0) { |
217 | RVVBitsMin = std::min(a: RVVBitsMin, b: RVVBitsMax); |
218 | RVVBitsMax = std::max(a: RVVBitsMin, b: RVVBitsMax); |
219 | } |
220 | |
221 | RVVBitsMin = llvm::bit_floor( |
222 | Value: (RVVBitsMin < 64 || RVVBitsMin > 65536) ? 0 : RVVBitsMin); |
223 | } |
224 | RVVBitsMax = |
225 | llvm::bit_floor(Value: (RVVBitsMax < 64 || RVVBitsMax > 65536) ? 0 : RVVBitsMax); |
226 | |
227 | SmallString<512> Key; |
228 | raw_svector_ostream(Key) << "RVVMin" << RVVBitsMin << "RVVMax" << RVVBitsMax |
229 | << CPU << TuneCPU << FS; |
230 | auto &I = SubtargetMap[Key]; |
231 | if (!I) { |
232 | // This needs to be done before we create a new subtarget since any |
233 | // creation will depend on the TM and the code generation flags on the |
234 | // function that reside in TargetOptions. |
235 | resetTargetOptions(F); |
236 | auto ABIName = Options.MCOptions.getABIName(); |
237 | if (const MDString *ModuleTargetABI = dyn_cast_or_null<MDString>( |
238 | Val: F.getParent()->getModuleFlag(Key: "target-abi" ))) { |
239 | auto TargetABI = RISCVABI::getTargetABI(ABIName); |
240 | if (TargetABI != RISCVABI::ABI_Unknown && |
241 | ModuleTargetABI->getString() != ABIName) { |
242 | report_fatal_error(reason: "-target-abi option != target-abi module flag" ); |
243 | } |
244 | ABIName = ModuleTargetABI->getString(); |
245 | } |
246 | I = std::make_unique<RISCVSubtarget>( |
247 | args: TargetTriple, args&: CPU, args&: TuneCPU, args&: FS, args&: ABIName, args&: RVVBitsMin, args&: RVVBitsMax, args: *this); |
248 | } |
249 | return I.get(); |
250 | } |
251 | |
252 | MachineFunctionInfo *RISCVTargetMachine::createMachineFunctionInfo( |
253 | BumpPtrAllocator &Allocator, const Function &F, |
254 | const TargetSubtargetInfo *STI) const { |
255 | return RISCVMachineFunctionInfo::create<RISCVMachineFunctionInfo>(Allocator, |
256 | F, STI); |
257 | } |
258 | |
259 | TargetTransformInfo |
260 | RISCVTargetMachine::getTargetTransformInfo(const Function &F) const { |
261 | return TargetTransformInfo(RISCVTTIImpl(this, F)); |
262 | } |
263 | |
264 | // A RISC-V hart has a single byte-addressable address space of 2^XLEN bytes |
265 | // for all memory accesses, so it is reasonable to assume that an |
266 | // implementation has no-op address space casts. If an implementation makes a |
267 | // change to this, they can override it here. |
268 | bool RISCVTargetMachine::isNoopAddrSpaceCast(unsigned SrcAS, |
269 | unsigned DstAS) const { |
270 | return true; |
271 | } |
272 | |
273 | namespace { |
274 | |
275 | class RVVRegisterRegAlloc : public RegisterRegAllocBase<RVVRegisterRegAlloc> { |
276 | public: |
277 | RVVRegisterRegAlloc(const char *N, const char *D, FunctionPassCtor C) |
278 | : RegisterRegAllocBase(N, D, C) {} |
279 | }; |
280 | |
281 | static bool onlyAllocateRVVReg(const TargetRegisterInfo &TRI, |
282 | const MachineRegisterInfo &MRI, |
283 | const Register Reg) { |
284 | const TargetRegisterClass *RC = MRI.getRegClass(Reg); |
285 | return RISCVRegisterInfo::isRVVRegClass(RC); |
286 | } |
287 | |
288 | static FunctionPass *useDefaultRegisterAllocator() { return nullptr; } |
289 | |
290 | static llvm::once_flag InitializeDefaultRVVRegisterAllocatorFlag; |
291 | |
292 | /// -riscv-rvv-regalloc=<fast|basic|greedy> command line option. |
293 | /// This option could designate the rvv register allocator only. |
294 | /// For example: -riscv-rvv-regalloc=basic |
295 | static cl::opt<RVVRegisterRegAlloc::FunctionPassCtor, false, |
296 | RegisterPassParser<RVVRegisterRegAlloc>> |
297 | RVVRegAlloc("riscv-rvv-regalloc" , cl::Hidden, |
298 | cl::init(Val: &useDefaultRegisterAllocator), |
299 | cl::desc("Register allocator to use for RVV register." )); |
300 | |
301 | static void initializeDefaultRVVRegisterAllocatorOnce() { |
302 | RegisterRegAlloc::FunctionPassCtor Ctor = RVVRegisterRegAlloc::getDefault(); |
303 | |
304 | if (!Ctor) { |
305 | Ctor = RVVRegAlloc; |
306 | RVVRegisterRegAlloc::setDefault(RVVRegAlloc); |
307 | } |
308 | } |
309 | |
310 | static FunctionPass *createBasicRVVRegisterAllocator() { |
311 | return createBasicRegisterAllocator(F: onlyAllocateRVVReg); |
312 | } |
313 | |
314 | static FunctionPass *createGreedyRVVRegisterAllocator() { |
315 | return createGreedyRegisterAllocator(F: onlyAllocateRVVReg); |
316 | } |
317 | |
318 | static FunctionPass *createFastRVVRegisterAllocator() { |
319 | return createFastRegisterAllocator(F: onlyAllocateRVVReg, ClearVirtRegs: false); |
320 | } |
321 | |
322 | static RVVRegisterRegAlloc basicRegAllocRVVReg("basic" , |
323 | "basic register allocator" , |
324 | createBasicRVVRegisterAllocator); |
325 | static RVVRegisterRegAlloc |
326 | greedyRegAllocRVVReg("greedy" , "greedy register allocator" , |
327 | createGreedyRVVRegisterAllocator); |
328 | |
329 | static RVVRegisterRegAlloc fastRegAllocRVVReg("fast" , "fast register allocator" , |
330 | createFastRVVRegisterAllocator); |
331 | |
332 | class RISCVPassConfig : public TargetPassConfig { |
333 | public: |
334 | RISCVPassConfig(RISCVTargetMachine &TM, PassManagerBase &PM) |
335 | : TargetPassConfig(TM, PM) { |
336 | if (TM.getOptLevel() != CodeGenOptLevel::None) |
337 | substitutePass(StandardID: &PostRASchedulerID, TargetID: &PostMachineSchedulerID); |
338 | setEnableSinkAndFold(EnableSinkFold); |
339 | } |
340 | |
341 | RISCVTargetMachine &getRISCVTargetMachine() const { |
342 | return getTM<RISCVTargetMachine>(); |
343 | } |
344 | |
345 | ScheduleDAGInstrs * |
346 | createMachineScheduler(MachineSchedContext *C) const override { |
347 | ScheduleDAGMILive *DAG = nullptr; |
348 | if (EnableMISchedLoadClustering) { |
349 | DAG = createGenericSchedLive(C); |
350 | DAG->addMutation(Mutation: createLoadClusterDAGMutation( |
351 | TII: DAG->TII, TRI: DAG->TRI, /*ReorderWhileClustering=*/true)); |
352 | } |
353 | return DAG; |
354 | } |
355 | |
356 | void addIRPasses() override; |
357 | bool addPreISel() override; |
358 | void addCodeGenPrepare() override; |
359 | bool addInstSelector() override; |
360 | bool addIRTranslator() override; |
361 | void addPreLegalizeMachineIR() override; |
362 | bool addLegalizeMachineIR() override; |
363 | void addPreRegBankSelect() override; |
364 | bool addRegBankSelect() override; |
365 | bool addGlobalInstructionSelect() override; |
366 | void addPreEmitPass() override; |
367 | void addPreEmitPass2() override; |
368 | void addPreSched2() override; |
369 | void addMachineSSAOptimization() override; |
370 | FunctionPass *createRVVRegAllocPass(bool Optimized); |
371 | bool addRegAssignAndRewriteFast() override; |
372 | bool addRegAssignAndRewriteOptimized() override; |
373 | void addPreRegAlloc() override; |
374 | void addPostRegAlloc() override; |
375 | void addFastRegAlloc() override; |
376 | }; |
377 | } // namespace |
378 | |
379 | TargetPassConfig *RISCVTargetMachine::createPassConfig(PassManagerBase &PM) { |
380 | return new RISCVPassConfig(*this, PM); |
381 | } |
382 | |
383 | FunctionPass *RISCVPassConfig::createRVVRegAllocPass(bool Optimized) { |
384 | // Initialize the global default. |
385 | llvm::call_once(flag&: InitializeDefaultRVVRegisterAllocatorFlag, |
386 | F&: initializeDefaultRVVRegisterAllocatorOnce); |
387 | |
388 | RegisterRegAlloc::FunctionPassCtor Ctor = RVVRegisterRegAlloc::getDefault(); |
389 | if (Ctor != useDefaultRegisterAllocator) |
390 | return Ctor(); |
391 | |
392 | if (Optimized) |
393 | return createGreedyRVVRegisterAllocator(); |
394 | |
395 | return createFastRVVRegisterAllocator(); |
396 | } |
397 | |
398 | bool RISCVPassConfig::addRegAssignAndRewriteFast() { |
399 | addPass(P: createRVVRegAllocPass(Optimized: false)); |
400 | if (EnableVSETVLIAfterRVVRegAlloc) |
401 | addPass(P: createRISCVInsertVSETVLIPass()); |
402 | if (TM->getOptLevel() != CodeGenOptLevel::None && |
403 | EnableRISCVDeadRegisterElimination) |
404 | addPass(P: createRISCVDeadRegisterDefinitionsPass()); |
405 | return TargetPassConfig::addRegAssignAndRewriteFast(); |
406 | } |
407 | |
408 | bool RISCVPassConfig::addRegAssignAndRewriteOptimized() { |
409 | addPass(P: createRVVRegAllocPass(Optimized: true)); |
410 | addPass(P: createVirtRegRewriter(ClearVirtRegs: false)); |
411 | if (EnableVSETVLIAfterRVVRegAlloc) |
412 | addPass(P: createRISCVInsertVSETVLIPass()); |
413 | if (TM->getOptLevel() != CodeGenOptLevel::None && |
414 | EnableRISCVDeadRegisterElimination) |
415 | addPass(P: createRISCVDeadRegisterDefinitionsPass()); |
416 | return TargetPassConfig::addRegAssignAndRewriteOptimized(); |
417 | } |
418 | |
419 | void RISCVPassConfig::addIRPasses() { |
420 | addPass(P: createAtomicExpandLegacyPass()); |
421 | |
422 | if (getOptLevel() != CodeGenOptLevel::None) { |
423 | if (EnableLoopDataPrefetch) |
424 | addPass(P: createLoopDataPrefetchPass()); |
425 | |
426 | addPass(P: createRISCVGatherScatterLoweringPass()); |
427 | addPass(P: createInterleavedAccessPass()); |
428 | addPass(P: createRISCVCodeGenPreparePass()); |
429 | } |
430 | |
431 | TargetPassConfig::addIRPasses(); |
432 | } |
433 | |
434 | bool RISCVPassConfig::addPreISel() { |
435 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
436 | // Add a barrier before instruction selection so that we will not get |
437 | // deleted block address after enabling default outlining. See D99707 for |
438 | // more details. |
439 | addPass(P: createBarrierNoopPass()); |
440 | } |
441 | |
442 | if (EnableGlobalMerge == cl::BOU_TRUE) { |
443 | addPass(P: createGlobalMergePass(TM, /* MaxOffset */ MaximalOffset: 2047, |
444 | /* OnlyOptimizeForSize */ false, |
445 | /* MergeExternalByDefault */ true)); |
446 | } |
447 | |
448 | return false; |
449 | } |
450 | |
451 | void RISCVPassConfig::addCodeGenPrepare() { |
452 | if (getOptLevel() != CodeGenOptLevel::None) |
453 | addPass(P: createTypePromotionLegacyPass()); |
454 | TargetPassConfig::addCodeGenPrepare(); |
455 | } |
456 | |
457 | bool RISCVPassConfig::addInstSelector() { |
458 | addPass(P: createRISCVISelDag(TM&: getRISCVTargetMachine(), OptLevel: getOptLevel())); |
459 | |
460 | return false; |
461 | } |
462 | |
463 | bool RISCVPassConfig::addIRTranslator() { |
464 | addPass(P: new IRTranslator(getOptLevel())); |
465 | return false; |
466 | } |
467 | |
468 | void RISCVPassConfig::addPreLegalizeMachineIR() { |
469 | if (getOptLevel() == CodeGenOptLevel::None) { |
470 | addPass(P: createRISCVO0PreLegalizerCombiner()); |
471 | } else { |
472 | addPass(P: createRISCVPreLegalizerCombiner()); |
473 | } |
474 | } |
475 | |
476 | bool RISCVPassConfig::addLegalizeMachineIR() { |
477 | addPass(P: new Legalizer()); |
478 | return false; |
479 | } |
480 | |
481 | void RISCVPassConfig::addPreRegBankSelect() { |
482 | if (getOptLevel() != CodeGenOptLevel::None) |
483 | addPass(P: createRISCVPostLegalizerCombiner()); |
484 | } |
485 | |
486 | bool RISCVPassConfig::addRegBankSelect() { |
487 | addPass(P: new RegBankSelect()); |
488 | return false; |
489 | } |
490 | |
491 | bool RISCVPassConfig::addGlobalInstructionSelect() { |
492 | addPass(P: new InstructionSelect(getOptLevel())); |
493 | return false; |
494 | } |
495 | |
496 | void RISCVPassConfig::addPreSched2() { |
497 | addPass(P: createRISCVPostRAExpandPseudoPass()); |
498 | |
499 | // Emit KCFI checks for indirect calls. |
500 | addPass(P: createKCFIPass()); |
501 | } |
502 | |
503 | void RISCVPassConfig::addPreEmitPass() { |
504 | // TODO: It would potentially be better to schedule copy propagation after |
505 | // expanding pseudos (in addPreEmitPass2). However, performing copy |
506 | // propagation after the machine outliner (which runs after addPreEmitPass) |
507 | // currently leads to incorrect code-gen, where copies to registers within |
508 | // outlined functions are removed erroneously. |
509 | if (TM->getOptLevel() >= CodeGenOptLevel::Default && |
510 | EnableRISCVCopyPropagation) |
511 | addPass(P: createMachineCopyPropagationPass(UseCopyInstr: true)); |
512 | addPass(PassID: &BranchRelaxationPassID); |
513 | addPass(P: createRISCVMakeCompressibleOptPass()); |
514 | } |
515 | |
516 | void RISCVPassConfig::addPreEmitPass2() { |
517 | if (TM->getOptLevel() != CodeGenOptLevel::None) { |
518 | addPass(P: createRISCVMoveMergePass()); |
519 | // Schedule PushPop Optimization before expansion of Pseudo instruction, |
520 | // ensuring return instruction is detected correctly. |
521 | addPass(P: createRISCVPushPopOptimizationPass()); |
522 | } |
523 | addPass(P: createRISCVExpandPseudoPass()); |
524 | |
525 | // Schedule the expansion of AMOs at the last possible moment, avoiding the |
526 | // possibility for other passes to break the requirements for forward |
527 | // progress in the LR/SC block. |
528 | addPass(P: createRISCVExpandAtomicPseudoPass()); |
529 | |
530 | // KCFI indirect call checks are lowered to a bundle. |
531 | addPass(P: createUnpackMachineBundles(Ftor: [&](const MachineFunction &MF) { |
532 | return MF.getFunction().getParent()->getModuleFlag(Key: "kcfi" ); |
533 | })); |
534 | } |
535 | |
536 | void RISCVPassConfig::addMachineSSAOptimization() { |
537 | addPass(P: createRISCVVectorPeepholePass()); |
538 | |
539 | TargetPassConfig::addMachineSSAOptimization(); |
540 | |
541 | if (EnableMachineCombiner) |
542 | addPass(PassID: &MachineCombinerID); |
543 | |
544 | if (TM->getTargetTriple().isRISCV64()) { |
545 | addPass(P: createRISCVOptWInstrsPass()); |
546 | } |
547 | } |
548 | |
549 | void RISCVPassConfig::addPreRegAlloc() { |
550 | addPass(P: createRISCVPreRAExpandPseudoPass()); |
551 | if (TM->getOptLevel() != CodeGenOptLevel::None) |
552 | addPass(P: createRISCVMergeBaseOffsetOptPass()); |
553 | |
554 | addPass(P: createRISCVInsertReadWriteCSRPass()); |
555 | addPass(P: createRISCVInsertWriteVXRMPass()); |
556 | |
557 | // Run RISCVInsertVSETVLI after PHI elimination. On O1 and above do it after |
558 | // register coalescing so needVSETVLIPHI doesn't need to look through COPYs. |
559 | if (!EnableVSETVLIAfterRVVRegAlloc) { |
560 | if (TM->getOptLevel() == CodeGenOptLevel::None) |
561 | insertPass(TargetPassID: &PHIEliminationID, InsertedPassID: &RISCVInsertVSETVLIID); |
562 | else |
563 | insertPass(TargetPassID: &RegisterCoalescerID, InsertedPassID: &RISCVInsertVSETVLIID); |
564 | } |
565 | } |
566 | |
567 | void RISCVPassConfig::addFastRegAlloc() { |
568 | addPass(PassID: &InitUndefID); |
569 | TargetPassConfig::addFastRegAlloc(); |
570 | } |
571 | |
572 | |
573 | void RISCVPassConfig::addPostRegAlloc() { |
574 | if (TM->getOptLevel() != CodeGenOptLevel::None && |
575 | EnableRedundantCopyElimination) |
576 | addPass(P: createRISCVRedundantCopyEliminationPass()); |
577 | } |
578 | |
579 | void RISCVTargetMachine::registerPassBuilderCallbacks(PassBuilder &PB) { |
580 | PB.registerLateLoopOptimizationsEPCallback(C: [=](LoopPassManager &LPM, |
581 | OptimizationLevel Level) { |
582 | LPM.addPass(Pass: LoopIdiomVectorizePass(LoopIdiomVectorizeStyle::Predicated)); |
583 | }); |
584 | } |
585 | |
586 | yaml::MachineFunctionInfo * |
587 | RISCVTargetMachine::createDefaultFuncInfoYAML() const { |
588 | return new yaml::RISCVMachineFunctionInfo(); |
589 | } |
590 | |
591 | yaml::MachineFunctionInfo * |
592 | RISCVTargetMachine::convertFuncInfoToYAML(const MachineFunction &MF) const { |
593 | const auto *MFI = MF.getInfo<RISCVMachineFunctionInfo>(); |
594 | return new yaml::RISCVMachineFunctionInfo(*MFI); |
595 | } |
596 | |
597 | bool RISCVTargetMachine::parseMachineFunctionInfo( |
598 | const yaml::MachineFunctionInfo &MFI, PerFunctionMIParsingState &PFS, |
599 | SMDiagnostic &Error, SMRange &SourceRange) const { |
600 | const auto &YamlMFI = |
601 | static_cast<const yaml::RISCVMachineFunctionInfo &>(MFI); |
602 | PFS.MF.getInfo<RISCVMachineFunctionInfo>()->initializeBaseYamlFields(YamlMFI); |
603 | return false; |
604 | } |
605 | |