1 | //===- ReducerWorkItem.cpp - Wrapper for Module and MachineFunction -------===// |
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 | #include "ReducerWorkItem.h" |
10 | #include "TestRunner.h" |
11 | #include "llvm/Analysis/ModuleSummaryAnalysis.h" |
12 | #include "llvm/Analysis/ProfileSummaryInfo.h" |
13 | #include "llvm/Bitcode/BitcodeReader.h" |
14 | #include "llvm/Bitcode/BitcodeWriter.h" |
15 | #include "llvm/CodeGen/CommandFlags.h" |
16 | #include "llvm/CodeGen/MIRParser/MIRParser.h" |
17 | #include "llvm/CodeGen/MIRPrinter.h" |
18 | #include "llvm/CodeGen/MachineDominators.h" |
19 | #include "llvm/CodeGen/MachineFrameInfo.h" |
20 | #include "llvm/CodeGen/MachineFunction.h" |
21 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
22 | #include "llvm/CodeGen/MachineModuleInfo.h" |
23 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
24 | #include "llvm/CodeGen/PseudoSourceValueManager.h" |
25 | #include "llvm/CodeGen/TargetInstrInfo.h" |
26 | #include "llvm/IR/Constants.h" |
27 | #include "llvm/IR/Instructions.h" |
28 | #include "llvm/IR/ModuleSummaryIndex.h" |
29 | #include "llvm/IR/Operator.h" |
30 | #include "llvm/IR/Verifier.h" |
31 | #include "llvm/IRReader/IRReader.h" |
32 | #include "llvm/MC/TargetRegistry.h" |
33 | #include "llvm/Passes/PassBuilder.h" |
34 | #include "llvm/Support/MemoryBufferRef.h" |
35 | #include "llvm/Support/SourceMgr.h" |
36 | #include "llvm/Support/TargetSelect.h" |
37 | #include "llvm/Support/ToolOutputFile.h" |
38 | #include "llvm/Support/WithColor.h" |
39 | #include "llvm/Target/TargetMachine.h" |
40 | #include "llvm/TargetParser/Host.h" |
41 | #include "llvm/Transforms/IPO/ThinLTOBitcodeWriter.h" |
42 | #include "llvm/Transforms/Utils/Cloning.h" |
43 | #include <optional> |
44 | |
45 | using namespace llvm; |
46 | |
47 | ReducerWorkItem::ReducerWorkItem() = default; |
48 | ReducerWorkItem::~ReducerWorkItem() = default; |
49 | |
50 | extern cl::OptionCategory LLVMReduceOptions; |
51 | static cl::opt<std::string> TargetTriple("mtriple" , |
52 | cl::desc("Set the target triple" ), |
53 | cl::cat(LLVMReduceOptions)); |
54 | static cl::opt<bool> PrintInvalidMachineReductions( |
55 | "print-invalid-reduction-machine-verifier-errors" , |
56 | cl::desc( |
57 | "Print machine verifier errors on invalid reduction attempts triple" ), |
58 | cl::cat(LLVMReduceOptions)); |
59 | |
60 | static cl::opt<bool> TmpFilesAsBitcode( |
61 | "write-tmp-files-as-bitcode" , |
62 | cl::desc("Always write temporary files as bitcode instead of textual IR" ), |
63 | cl::init(Val: false), cl::cat(LLVMReduceOptions)); |
64 | |
65 | static void cloneFrameInfo( |
66 | MachineFrameInfo &DstMFI, const MachineFrameInfo &SrcMFI, |
67 | const DenseMap<MachineBasicBlock *, MachineBasicBlock *> &Src2DstMBB) { |
68 | DstMFI.setFrameAddressIsTaken(SrcMFI.isFrameAddressTaken()); |
69 | DstMFI.setReturnAddressIsTaken(SrcMFI.isReturnAddressTaken()); |
70 | DstMFI.setHasStackMap(SrcMFI.hasStackMap()); |
71 | DstMFI.setHasPatchPoint(SrcMFI.hasPatchPoint()); |
72 | DstMFI.setUseLocalStackAllocationBlock( |
73 | SrcMFI.getUseLocalStackAllocationBlock()); |
74 | DstMFI.setOffsetAdjustment(SrcMFI.getOffsetAdjustment()); |
75 | |
76 | DstMFI.ensureMaxAlignment(Alignment: SrcMFI.getMaxAlign()); |
77 | assert(DstMFI.getMaxAlign() == SrcMFI.getMaxAlign() && |
78 | "we need to set exact alignment" ); |
79 | |
80 | DstMFI.setAdjustsStack(SrcMFI.adjustsStack()); |
81 | DstMFI.setHasCalls(SrcMFI.hasCalls()); |
82 | DstMFI.setHasOpaqueSPAdjustment(SrcMFI.hasOpaqueSPAdjustment()); |
83 | DstMFI.setHasCopyImplyingStackAdjustment( |
84 | SrcMFI.hasCopyImplyingStackAdjustment()); |
85 | DstMFI.setHasVAStart(SrcMFI.hasVAStart()); |
86 | DstMFI.setHasMustTailInVarArgFunc(SrcMFI.hasMustTailInVarArgFunc()); |
87 | DstMFI.setHasTailCall(SrcMFI.hasTailCall()); |
88 | |
89 | if (SrcMFI.isMaxCallFrameSizeComputed()) |
90 | DstMFI.setMaxCallFrameSize(SrcMFI.getMaxCallFrameSize()); |
91 | |
92 | DstMFI.setCVBytesOfCalleeSavedRegisters( |
93 | SrcMFI.getCVBytesOfCalleeSavedRegisters()); |
94 | |
95 | if (MachineBasicBlock *SavePt = SrcMFI.getSavePoint()) |
96 | DstMFI.setSavePoint(Src2DstMBB.find(Val: SavePt)->second); |
97 | if (MachineBasicBlock *RestorePt = SrcMFI.getRestorePoint()) |
98 | DstMFI.setRestorePoint(Src2DstMBB.find(Val: RestorePt)->second); |
99 | |
100 | |
101 | auto CopyObjectProperties = [](MachineFrameInfo &DstMFI, |
102 | const MachineFrameInfo &SrcMFI, int FI) { |
103 | if (SrcMFI.isStatepointSpillSlotObjectIndex(ObjectIdx: FI)) |
104 | DstMFI.markAsStatepointSpillSlotObjectIndex(ObjectIdx: FI); |
105 | DstMFI.setObjectSSPLayout(ObjectIdx: FI, Kind: SrcMFI.getObjectSSPLayout(ObjectIdx: FI)); |
106 | DstMFI.setObjectZExt(ObjectIdx: FI, IsZExt: SrcMFI.isObjectZExt(ObjectIdx: FI)); |
107 | DstMFI.setObjectSExt(ObjectIdx: FI, IsSExt: SrcMFI.isObjectSExt(ObjectIdx: FI)); |
108 | }; |
109 | |
110 | for (int i = 0, e = SrcMFI.getNumObjects() - SrcMFI.getNumFixedObjects(); |
111 | i != e; ++i) { |
112 | int NewFI; |
113 | |
114 | assert(!SrcMFI.isFixedObjectIndex(i)); |
115 | if (SrcMFI.isVariableSizedObjectIndex(ObjectIdx: i)) { |
116 | NewFI = DstMFI.CreateVariableSizedObject(Alignment: SrcMFI.getObjectAlign(ObjectIdx: i), |
117 | Alloca: SrcMFI.getObjectAllocation(ObjectIdx: i)); |
118 | } else { |
119 | NewFI = DstMFI.CreateStackObject( |
120 | Size: SrcMFI.getObjectSize(ObjectIdx: i), Alignment: SrcMFI.getObjectAlign(ObjectIdx: i), |
121 | isSpillSlot: SrcMFI.isSpillSlotObjectIndex(ObjectIdx: i), Alloca: SrcMFI.getObjectAllocation(ObjectIdx: i), |
122 | ID: SrcMFI.getStackID(ObjectIdx: i)); |
123 | DstMFI.setObjectOffset(ObjectIdx: NewFI, SPOffset: SrcMFI.getObjectOffset(ObjectIdx: i)); |
124 | } |
125 | |
126 | CopyObjectProperties(DstMFI, SrcMFI, i); |
127 | |
128 | (void)NewFI; |
129 | assert(i == NewFI && "expected to keep stable frame index numbering" ); |
130 | } |
131 | |
132 | // Copy the fixed frame objects backwards to preserve frame index numbers, |
133 | // since CreateFixedObject uses front insertion. |
134 | for (int i = -1; i >= (int)-SrcMFI.getNumFixedObjects(); --i) { |
135 | assert(SrcMFI.isFixedObjectIndex(i)); |
136 | int NewFI = DstMFI.CreateFixedObject( |
137 | Size: SrcMFI.getObjectSize(ObjectIdx: i), SPOffset: SrcMFI.getObjectOffset(ObjectIdx: i), |
138 | IsImmutable: SrcMFI.isImmutableObjectIndex(ObjectIdx: i), isAliased: SrcMFI.isAliasedObjectIndex(ObjectIdx: i)); |
139 | CopyObjectProperties(DstMFI, SrcMFI, i); |
140 | |
141 | (void)NewFI; |
142 | assert(i == NewFI && "expected to keep stable frame index numbering" ); |
143 | } |
144 | |
145 | for (unsigned I = 0, E = SrcMFI.getLocalFrameObjectCount(); I < E; ++I) { |
146 | auto LocalObject = SrcMFI.getLocalFrameObjectMap(i: I); |
147 | DstMFI.mapLocalFrameObject(ObjectIndex: LocalObject.first, Offset: LocalObject.second); |
148 | } |
149 | |
150 | DstMFI.setCalleeSavedInfo(SrcMFI.getCalleeSavedInfo()); |
151 | |
152 | if (SrcMFI.hasStackProtectorIndex()) { |
153 | DstMFI.setStackProtectorIndex(SrcMFI.getStackProtectorIndex()); |
154 | } |
155 | |
156 | // FIXME: Needs test, missing MIR serialization. |
157 | if (SrcMFI.hasFunctionContextIndex()) { |
158 | DstMFI.setFunctionContextIndex(SrcMFI.getFunctionContextIndex()); |
159 | } |
160 | } |
161 | |
162 | static void cloneJumpTableInfo( |
163 | MachineFunction &DstMF, const MachineJumpTableInfo &SrcJTI, |
164 | const DenseMap<MachineBasicBlock *, MachineBasicBlock *> &Src2DstMBB) { |
165 | |
166 | auto *DstJTI = DstMF.getOrCreateJumpTableInfo(JTEntryKind: SrcJTI.getEntryKind()); |
167 | |
168 | std::vector<MachineBasicBlock *> DstBBs; |
169 | |
170 | for (const MachineJumpTableEntry &Entry : SrcJTI.getJumpTables()) { |
171 | for (MachineBasicBlock *X : Entry.MBBs) |
172 | DstBBs.push_back(x: Src2DstMBB.find(Val: X)->second); |
173 | |
174 | DstJTI->createJumpTableIndex(DestBBs: DstBBs); |
175 | DstBBs.clear(); |
176 | } |
177 | } |
178 | |
179 | static void cloneMemOperands(MachineInstr &DstMI, MachineInstr &SrcMI, |
180 | MachineFunction &SrcMF, MachineFunction &DstMF) { |
181 | // The new MachineMemOperands should be owned by the new function's |
182 | // Allocator. |
183 | PseudoSourceValueManager &PSVMgr = DstMF.getPSVManager(); |
184 | |
185 | // We also need to remap the PseudoSourceValues from the new function's |
186 | // PseudoSourceValueManager. |
187 | SmallVector<MachineMemOperand *, 2> NewMMOs; |
188 | for (MachineMemOperand *OldMMO : SrcMI.memoperands()) { |
189 | MachinePointerInfo NewPtrInfo(OldMMO->getPointerInfo()); |
190 | if (const PseudoSourceValue *PSV = |
191 | dyn_cast_if_present<const PseudoSourceValue *>(Val&: NewPtrInfo.V)) { |
192 | switch (PSV->kind()) { |
193 | case PseudoSourceValue::Stack: |
194 | NewPtrInfo.V = PSVMgr.getStack(); |
195 | break; |
196 | case PseudoSourceValue::GOT: |
197 | NewPtrInfo.V = PSVMgr.getGOT(); |
198 | break; |
199 | case PseudoSourceValue::JumpTable: |
200 | NewPtrInfo.V = PSVMgr.getJumpTable(); |
201 | break; |
202 | case PseudoSourceValue::ConstantPool: |
203 | NewPtrInfo.V = PSVMgr.getConstantPool(); |
204 | break; |
205 | case PseudoSourceValue::FixedStack: |
206 | NewPtrInfo.V = PSVMgr.getFixedStack( |
207 | FI: cast<FixedStackPseudoSourceValue>(Val: PSV)->getFrameIndex()); |
208 | break; |
209 | case PseudoSourceValue::GlobalValueCallEntry: |
210 | NewPtrInfo.V = PSVMgr.getGlobalValueCallEntry( |
211 | GV: cast<GlobalValuePseudoSourceValue>(Val: PSV)->getValue()); |
212 | break; |
213 | case PseudoSourceValue::ExternalSymbolCallEntry: |
214 | NewPtrInfo.V = PSVMgr.getExternalSymbolCallEntry( |
215 | ES: cast<ExternalSymbolPseudoSourceValue>(Val: PSV)->getSymbol()); |
216 | break; |
217 | case PseudoSourceValue::TargetCustom: |
218 | default: |
219 | // FIXME: We have no generic interface for allocating custom PSVs. |
220 | report_fatal_error(reason: "Cloning TargetCustom PSV not handled" ); |
221 | } |
222 | } |
223 | |
224 | MachineMemOperand *NewMMO = DstMF.getMachineMemOperand( |
225 | PtrInfo: NewPtrInfo, f: OldMMO->getFlags(), MemTy: OldMMO->getMemoryType(), |
226 | base_alignment: OldMMO->getBaseAlign(), AAInfo: OldMMO->getAAInfo(), Ranges: OldMMO->getRanges(), |
227 | SSID: OldMMO->getSyncScopeID(), Ordering: OldMMO->getSuccessOrdering(), |
228 | FailureOrdering: OldMMO->getFailureOrdering()); |
229 | NewMMOs.push_back(Elt: NewMMO); |
230 | } |
231 | |
232 | DstMI.setMemRefs(MF&: DstMF, MemRefs: NewMMOs); |
233 | } |
234 | |
235 | static std::unique_ptr<MachineFunction> cloneMF(MachineFunction *SrcMF, |
236 | MachineModuleInfo &DestMMI) { |
237 | auto DstMF = std::make_unique<MachineFunction>( |
238 | args&: SrcMF->getFunction(), args: SrcMF->getTarget(), args: SrcMF->getSubtarget(), |
239 | args&: SrcMF->getContext(), args: SrcMF->getFunctionNumber()); |
240 | DenseMap<MachineBasicBlock *, MachineBasicBlock *> Src2DstMBB; |
241 | |
242 | auto *SrcMRI = &SrcMF->getRegInfo(); |
243 | auto *DstMRI = &DstMF->getRegInfo(); |
244 | |
245 | // Clone blocks. |
246 | for (MachineBasicBlock &SrcMBB : *SrcMF) { |
247 | MachineBasicBlock *DstMBB = |
248 | DstMF->CreateMachineBasicBlock(BB: SrcMBB.getBasicBlock()); |
249 | Src2DstMBB[&SrcMBB] = DstMBB; |
250 | |
251 | DstMBB->setCallFrameSize(SrcMBB.getCallFrameSize()); |
252 | |
253 | if (SrcMBB.isIRBlockAddressTaken()) |
254 | DstMBB->setAddressTakenIRBlock(SrcMBB.getAddressTakenIRBlock()); |
255 | if (SrcMBB.isMachineBlockAddressTaken()) |
256 | DstMBB->setMachineBlockAddressTaken(); |
257 | |
258 | // FIXME: This is not serialized |
259 | if (SrcMBB.hasLabelMustBeEmitted()) |
260 | DstMBB->setLabelMustBeEmitted(); |
261 | |
262 | DstMBB->setAlignment(SrcMBB.getAlignment()); |
263 | |
264 | // FIXME: This is not serialized |
265 | DstMBB->setMaxBytesForAlignment(SrcMBB.getMaxBytesForAlignment()); |
266 | |
267 | DstMBB->setIsEHPad(SrcMBB.isEHPad()); |
268 | DstMBB->setIsEHScopeEntry(SrcMBB.isEHScopeEntry()); |
269 | DstMBB->setIsEHContTarget(SrcMBB.isEHContTarget()); |
270 | DstMBB->setIsEHFuncletEntry(SrcMBB.isEHFuncletEntry()); |
271 | |
272 | // FIXME: These are not serialized |
273 | DstMBB->setIsCleanupFuncletEntry(SrcMBB.isCleanupFuncletEntry()); |
274 | DstMBB->setIsBeginSection(SrcMBB.isBeginSection()); |
275 | DstMBB->setIsEndSection(SrcMBB.isEndSection()); |
276 | |
277 | DstMBB->setSectionID(SrcMBB.getSectionID()); |
278 | DstMBB->setIsInlineAsmBrIndirectTarget( |
279 | SrcMBB.isInlineAsmBrIndirectTarget()); |
280 | |
281 | // FIXME: This is not serialized |
282 | if (std::optional<uint64_t> Weight = SrcMBB.getIrrLoopHeaderWeight()) |
283 | DstMBB->setIrrLoopHeaderWeight(*Weight); |
284 | } |
285 | |
286 | const MachineFrameInfo &SrcMFI = SrcMF->getFrameInfo(); |
287 | MachineFrameInfo &DstMFI = DstMF->getFrameInfo(); |
288 | |
289 | // Copy stack objects and other info |
290 | cloneFrameInfo(DstMFI, SrcMFI, Src2DstMBB); |
291 | |
292 | if (MachineJumpTableInfo *SrcJTI = SrcMF->getJumpTableInfo()) { |
293 | cloneJumpTableInfo(DstMF&: *DstMF, SrcJTI: *SrcJTI, Src2DstMBB); |
294 | } |
295 | |
296 | // Remap the debug info frame index references. |
297 | DstMF->VariableDbgInfos = SrcMF->VariableDbgInfos; |
298 | |
299 | // Clone virtual registers |
300 | for (unsigned I = 0, E = SrcMRI->getNumVirtRegs(); I != E; ++I) { |
301 | Register Reg = Register::index2VirtReg(Index: I); |
302 | Register NewReg = DstMRI->createIncompleteVirtualRegister( |
303 | Name: SrcMRI->getVRegName(Reg)); |
304 | assert(NewReg == Reg && "expected to preserve virtreg number" ); |
305 | |
306 | DstMRI->setRegClassOrRegBank(Reg: NewReg, RCOrRB: SrcMRI->getRegClassOrRegBank(Reg)); |
307 | |
308 | LLT RegTy = SrcMRI->getType(Reg); |
309 | if (RegTy.isValid()) |
310 | DstMRI->setType(VReg: NewReg, Ty: RegTy); |
311 | |
312 | // Copy register allocation hints. |
313 | const auto *Hints = SrcMRI->getRegAllocationHints(VReg: Reg); |
314 | if (Hints) |
315 | for (Register PrefReg : Hints->second) |
316 | DstMRI->addRegAllocationHint(VReg: NewReg, PrefReg); |
317 | } |
318 | |
319 | const TargetSubtargetInfo &STI = DstMF->getSubtarget(); |
320 | const TargetInstrInfo *TII = STI.getInstrInfo(); |
321 | const TargetRegisterInfo *TRI = STI.getRegisterInfo(); |
322 | |
323 | // Link blocks. |
324 | for (auto &SrcMBB : *SrcMF) { |
325 | auto *DstMBB = Src2DstMBB[&SrcMBB]; |
326 | DstMF->push_back(MBB: DstMBB); |
327 | |
328 | for (auto It = SrcMBB.succ_begin(), IterEnd = SrcMBB.succ_end(); |
329 | It != IterEnd; ++It) { |
330 | auto *SrcSuccMBB = *It; |
331 | auto *DstSuccMBB = Src2DstMBB[SrcSuccMBB]; |
332 | DstMBB->addSuccessor(Succ: DstSuccMBB, Prob: SrcMBB.getSuccProbability(Succ: It)); |
333 | } |
334 | |
335 | for (auto &LI : SrcMBB.liveins_dbg()) |
336 | DstMBB->addLiveIn(RegMaskPair: LI); |
337 | |
338 | // Make sure MRI knows about registers clobbered by unwinder. |
339 | if (DstMBB->isEHPad()) { |
340 | if (auto *RegMask = TRI->getCustomEHPadPreservedMask(MF: *DstMF)) |
341 | DstMRI->addPhysRegsUsedFromRegMask(RegMask); |
342 | } |
343 | } |
344 | |
345 | // Track predefined/named regmasks which we ignore. |
346 | DenseSet<const uint32_t *> ConstRegisterMasks(llvm::from_range, |
347 | TRI->getRegMasks()); |
348 | |
349 | // Clone instructions. |
350 | for (auto &SrcMBB : *SrcMF) { |
351 | auto *DstMBB = Src2DstMBB[&SrcMBB]; |
352 | for (auto &SrcMI : SrcMBB) { |
353 | const auto &MCID = TII->get(Opcode: SrcMI.getOpcode()); |
354 | auto *DstMI = DstMF->CreateMachineInstr(MCID, DL: SrcMI.getDebugLoc(), |
355 | /*NoImplicit=*/true); |
356 | DstMI->setFlags(SrcMI.getFlags()); |
357 | DstMI->setAsmPrinterFlag(SrcMI.getAsmPrinterFlags()); |
358 | |
359 | DstMBB->push_back(MI: DstMI); |
360 | for (auto &SrcMO : SrcMI.operands()) { |
361 | MachineOperand DstMO(SrcMO); |
362 | DstMO.clearParent(); |
363 | |
364 | // Update MBB. |
365 | if (DstMO.isMBB()) |
366 | DstMO.setMBB(Src2DstMBB[DstMO.getMBB()]); |
367 | else if (DstMO.isRegMask()) { |
368 | DstMRI->addPhysRegsUsedFromRegMask(RegMask: DstMO.getRegMask()); |
369 | |
370 | if (!ConstRegisterMasks.count(V: DstMO.getRegMask())) { |
371 | uint32_t *DstMask = DstMF->allocateRegMask(); |
372 | std::memcpy(dest: DstMask, src: SrcMO.getRegMask(), |
373 | n: sizeof(*DstMask) * |
374 | MachineOperand::getRegMaskSize(NumRegs: TRI->getNumRegs())); |
375 | DstMO.setRegMask(DstMask); |
376 | } |
377 | } |
378 | |
379 | DstMI->addOperand(Op: DstMO); |
380 | } |
381 | |
382 | cloneMemOperands(DstMI&: *DstMI, SrcMI, SrcMF&: *SrcMF, DstMF&: *DstMF); |
383 | } |
384 | } |
385 | |
386 | DstMF->setAlignment(SrcMF->getAlignment()); |
387 | DstMF->setExposesReturnsTwice(SrcMF->exposesReturnsTwice()); |
388 | DstMF->setHasInlineAsm(SrcMF->hasInlineAsm()); |
389 | DstMF->setHasWinCFI(SrcMF->hasWinCFI()); |
390 | |
391 | DstMF->getProperties().reset().set(SrcMF->getProperties()); |
392 | |
393 | if (!SrcMF->getFrameInstructions().empty() || |
394 | !SrcMF->getLongjmpTargets().empty() || !SrcMF->getEHContTargets().empty()) |
395 | report_fatal_error(reason: "cloning not implemented for machine function property" ); |
396 | |
397 | DstMF->setCallsEHReturn(SrcMF->callsEHReturn()); |
398 | DstMF->setCallsUnwindInit(SrcMF->callsUnwindInit()); |
399 | DstMF->setHasEHContTarget(SrcMF->hasEHContTarget()); |
400 | DstMF->setHasEHScopes(SrcMF->hasEHScopes()); |
401 | DstMF->setHasEHFunclets(SrcMF->hasEHFunclets()); |
402 | DstMF->setHasFakeUses(SrcMF->hasFakeUses()); |
403 | DstMF->setIsOutlined(SrcMF->isOutlined()); |
404 | |
405 | if (!SrcMF->getLandingPads().empty() || |
406 | !SrcMF->getCodeViewAnnotations().empty() || |
407 | !SrcMF->getTypeInfos().empty() || |
408 | !SrcMF->getFilterIds().empty() || |
409 | SrcMF->hasAnyWasmLandingPadIndex() || |
410 | SrcMF->hasAnyCallSiteLandingPad() || |
411 | SrcMF->hasAnyCallSiteLabel() || |
412 | !SrcMF->getCallSitesInfo().empty()) |
413 | report_fatal_error(reason: "cloning not implemented for machine function property" ); |
414 | |
415 | DstMF->setDebugInstrNumberingCount(SrcMF->DebugInstrNumberingCount); |
416 | |
417 | if (!DstMF->cloneInfoFrom(OrigMF: *SrcMF, Src2DstMBB)) |
418 | report_fatal_error(reason: "target does not implement MachineFunctionInfo cloning" ); |
419 | |
420 | DstMRI->freezeReservedRegs(); |
421 | |
422 | DstMF->verify(p: nullptr, Banner: "" , OS: &errs(), /*AbortOnError=*/true); |
423 | return DstMF; |
424 | } |
425 | |
426 | static void initializeTargetInfo() { |
427 | InitializeAllTargets(); |
428 | InitializeAllTargetMCs(); |
429 | InitializeAllAsmPrinters(); |
430 | InitializeAllAsmParsers(); |
431 | } |
432 | |
433 | void ReducerWorkItem::print(raw_ostream &ROS, void *p) const { |
434 | if (MMI) { |
435 | printMIR(OS&: ROS, M: *M); |
436 | for (Function &F : *M) { |
437 | if (auto *MF = MMI->getMachineFunction(F)) |
438 | printMIR(OS&: ROS, MMI: *MMI, MF: *MF); |
439 | } |
440 | } else { |
441 | M->print(OS&: ROS, /*AssemblyAnnotationWriter=*/AAW: nullptr, |
442 | /*ShouldPreserveUseListOrder=*/true); |
443 | } |
444 | } |
445 | |
446 | bool ReducerWorkItem::verify(raw_fd_ostream *OS) const { |
447 | if (verifyModule(M: *M, OS)) |
448 | return true; |
449 | |
450 | if (!MMI) |
451 | return false; |
452 | |
453 | for (const Function &F : getModule()) { |
454 | if (const MachineFunction *MF = MMI->getMachineFunction(F)) { |
455 | // With the current state of quality, most reduction attempts fail the |
456 | // machine verifier. Avoid spamming large function dumps on nearly every |
457 | // attempt until the situation is better. |
458 | if (!MF->verify(p: nullptr, Banner: "" , |
459 | /*OS=*/PrintInvalidMachineReductions ? &errs() : nullptr, |
460 | /*AbortOnError=*/false)) { |
461 | |
462 | if (!PrintInvalidMachineReductions) { |
463 | WithColor::warning(OS&: errs()) |
464 | << "reduction attempt on function '" << MF->getName() |
465 | << "' failed machine verifier (debug with " |
466 | "-print-invalid-reduction-machine-verifier-errors)\n" ; |
467 | } |
468 | return true; |
469 | } |
470 | } |
471 | } |
472 | |
473 | return false; |
474 | } |
475 | |
476 | bool ReducerWorkItem::isReduced(const TestRunner &Test) const { |
477 | const bool UseBitcode = Test.inputIsBitcode() || TmpFilesAsBitcode; |
478 | |
479 | SmallString<128> CurrentFilepath; |
480 | |
481 | // Write ReducerWorkItem to tmp file |
482 | int FD; |
483 | std::error_code EC = sys::fs::createTemporaryFile( |
484 | Prefix: "llvm-reduce" , Suffix: isMIR() ? "mir" : (UseBitcode ? "bc" : "ll" ), ResultFD&: FD, |
485 | ResultPath&: CurrentFilepath, |
486 | Flags: UseBitcode && !isMIR() ? sys::fs::OF_None : sys::fs::OF_Text); |
487 | if (EC) { |
488 | WithColor::error(OS&: errs(), Prefix: Test.getToolName()) |
489 | << "error making unique filename: " << EC.message() << '\n'; |
490 | exit(status: 1); |
491 | } |
492 | |
493 | ToolOutputFile Out(CurrentFilepath, FD); |
494 | |
495 | writeOutput(OS&: Out.os(), EmitBitcode: UseBitcode); |
496 | |
497 | Out.os().close(); |
498 | if (Out.os().has_error()) { |
499 | WithColor::error(OS&: errs(), Prefix: Test.getToolName()) |
500 | << "error emitting bitcode to file '" << CurrentFilepath |
501 | << "': " << Out.os().error().message() << '\n'; |
502 | exit(status: 1); |
503 | } |
504 | |
505 | // Current Chunks aren't interesting |
506 | return Test.run(Filename: CurrentFilepath); |
507 | } |
508 | |
509 | std::unique_ptr<ReducerWorkItem> |
510 | ReducerWorkItem::clone(const TargetMachine *TM) const { |
511 | auto CloneMMM = std::make_unique<ReducerWorkItem>(); |
512 | if (TM) { |
513 | // We're assuming the Module IR contents are always unchanged by MIR |
514 | // reductions, and can share it as a constant. |
515 | CloneMMM->M = M; |
516 | |
517 | // MachineModuleInfo contains a lot of other state used during codegen which |
518 | // we won't be using here, but we should be able to ignore it (although this |
519 | // is pretty ugly). |
520 | CloneMMM->MMI = std::make_unique<MachineModuleInfo>(args&: TM); |
521 | |
522 | for (const Function &F : getModule()) { |
523 | if (auto *MF = MMI->getMachineFunction(F)) |
524 | CloneMMM->MMI->insertFunction(F, MF: cloneMF(SrcMF: MF, DestMMI&: *CloneMMM->MMI)); |
525 | } |
526 | } else { |
527 | CloneMMM->M = CloneModule(M: *M); |
528 | } |
529 | return CloneMMM; |
530 | } |
531 | |
532 | /// Try to produce some number that indicates a function is getting smaller / |
533 | /// simpler. |
534 | static uint64_t computeMIRComplexityScoreImpl(const MachineFunction &MF) { |
535 | uint64_t Score = 0; |
536 | const MachineFrameInfo &MFI = MF.getFrameInfo(); |
537 | |
538 | // Add for stack objects |
539 | Score += MFI.getNumObjects(); |
540 | |
541 | // Add in the block count. |
542 | Score += 2 * MF.size(); |
543 | |
544 | const MachineRegisterInfo &MRI = MF.getRegInfo(); |
545 | for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) { |
546 | Register Reg = Register::index2VirtReg(Index: I); |
547 | if (const auto *Hints = MRI.getRegAllocationHints(VReg: Reg)) |
548 | Score += Hints->second.size(); |
549 | } |
550 | |
551 | for (const MachineBasicBlock &MBB : MF) { |
552 | for (const MachineInstr &MI : MBB) { |
553 | const unsigned Opc = MI.getOpcode(); |
554 | |
555 | // Reductions may want or need to introduce implicit_defs, so don't count |
556 | // them. |
557 | // TODO: These probably should count in some way. |
558 | if (Opc == TargetOpcode::IMPLICIT_DEF || |
559 | Opc == TargetOpcode::G_IMPLICIT_DEF) |
560 | continue; |
561 | |
562 | // Each instruction adds to the score |
563 | Score += 4; |
564 | |
565 | if (Opc == TargetOpcode::PHI || Opc == TargetOpcode::G_PHI || |
566 | Opc == TargetOpcode::INLINEASM || Opc == TargetOpcode::INLINEASM_BR) |
567 | ++Score; |
568 | |
569 | if (MI.getFlags() != 0) |
570 | ++Score; |
571 | |
572 | // Increase weight for more operands. |
573 | for (const MachineOperand &MO : MI.operands()) { |
574 | ++Score; |
575 | |
576 | // Treat registers as more complex. |
577 | if (MO.isReg()) { |
578 | ++Score; |
579 | |
580 | // And subregisters as even more complex. |
581 | if (MO.getSubReg()) { |
582 | ++Score; |
583 | if (MO.isDef()) |
584 | ++Score; |
585 | } |
586 | } else if (MO.isRegMask()) |
587 | ++Score; |
588 | } |
589 | } |
590 | } |
591 | |
592 | return Score; |
593 | } |
594 | |
595 | uint64_t ReducerWorkItem::computeMIRComplexityScore() const { |
596 | uint64_t Score = 0; |
597 | |
598 | for (const Function &F : getModule()) { |
599 | if (auto *MF = MMI->getMachineFunction(F)) |
600 | Score += computeMIRComplexityScoreImpl(MF: *MF); |
601 | } |
602 | |
603 | return Score; |
604 | } |
605 | |
606 | // FIXME: ReduceOperandsSkip has similar function, except it uses larger numbers |
607 | // for more reduced. |
608 | static unsigned classifyReductivePower(const Value *V) { |
609 | if (auto *C = dyn_cast<ConstantData>(Val: V)) { |
610 | if (C->isNullValue()) |
611 | return 0; |
612 | if (C->isOneValue()) |
613 | return 1; |
614 | if (isa<UndefValue>(Val: V)) |
615 | return 2; |
616 | return 3; |
617 | } |
618 | |
619 | if (isa<GlobalValue>(Val: V)) |
620 | return 4; |
621 | |
622 | // TODO: Account for expression size |
623 | if (isa<ConstantExpr>(Val: V)) |
624 | return 5; |
625 | |
626 | if (isa<Constant>(Val: V)) |
627 | return 1; |
628 | |
629 | if (isa<Argument>(Val: V)) |
630 | return 6; |
631 | |
632 | if (isa<Instruction>(Val: V)) |
633 | return 7; |
634 | |
635 | return 0; |
636 | } |
637 | |
638 | // TODO: Additional flags and attributes may be complexity reducing. If we start |
639 | // adding flags and attributes, they could have negative cost. |
640 | static uint64_t computeIRComplexityScoreImpl(const Function &F) { |
641 | uint64_t Score = 1; // Count the function itself |
642 | SmallVector<std::pair<unsigned, MDNode *>> MDs; |
643 | |
644 | AttributeList Attrs = F.getAttributes(); |
645 | for (AttributeSet AttrSet : Attrs) |
646 | Score += AttrSet.getNumAttributes(); |
647 | |
648 | for (const BasicBlock &BB : F) { |
649 | ++Score; |
650 | |
651 | for (const Instruction &I : BB) { |
652 | ++Score; |
653 | |
654 | if (const auto *OverflowOp = dyn_cast<OverflowingBinaryOperator>(Val: &I)) { |
655 | if (OverflowOp->hasNoUnsignedWrap()) |
656 | ++Score; |
657 | if (OverflowOp->hasNoSignedWrap()) |
658 | ++Score; |
659 | } else if (const auto *Trunc = dyn_cast<TruncInst>(Val: &I)) { |
660 | if (Trunc->hasNoSignedWrap()) |
661 | ++Score; |
662 | if (Trunc->hasNoUnsignedWrap()) |
663 | ++Score; |
664 | } else if (const auto *ExactOp = dyn_cast<PossiblyExactOperator>(Val: &I)) { |
665 | if (ExactOp->isExact()) |
666 | ++Score; |
667 | } else if (const auto *NNI = dyn_cast<PossiblyNonNegInst>(Val: &I)) { |
668 | if (NNI->hasNonNeg()) |
669 | ++Score; |
670 | } else if (const auto *PDI = dyn_cast<PossiblyDisjointInst>(Val: &I)) { |
671 | if (PDI->isDisjoint()) |
672 | ++Score; |
673 | } else if (const auto *GEP = dyn_cast<GEPOperator>(Val: &I)) { |
674 | if (GEP->isInBounds()) |
675 | ++Score; |
676 | if (GEP->hasNoUnsignedSignedWrap()) |
677 | ++Score; |
678 | if (GEP->hasNoUnsignedWrap()) |
679 | ++Score; |
680 | } else if (const auto *FPOp = dyn_cast<FPMathOperator>(Val: &I)) { |
681 | FastMathFlags FMF = FPOp->getFastMathFlags(); |
682 | if (FMF.allowReassoc()) |
683 | ++Score; |
684 | if (FMF.noNaNs()) |
685 | ++Score; |
686 | if (FMF.noInfs()) |
687 | ++Score; |
688 | if (FMF.noSignedZeros()) |
689 | ++Score; |
690 | if (FMF.allowReciprocal()) |
691 | ++Score; |
692 | if (FMF.allowContract()) |
693 | ++Score; |
694 | if (FMF.approxFunc()) |
695 | ++Score; |
696 | } |
697 | |
698 | for (const Value *Operand : I.operands()) { |
699 | ++Score; |
700 | Score += classifyReductivePower(V: Operand); |
701 | } |
702 | |
703 | I.getAllMetadata(MDs); |
704 | Score += MDs.size(); |
705 | MDs.clear(); |
706 | } |
707 | } |
708 | |
709 | return Score; |
710 | } |
711 | |
712 | uint64_t ReducerWorkItem::computeIRComplexityScore() const { |
713 | uint64_t Score = 0; |
714 | |
715 | const Module &M = getModule(); |
716 | Score += M.named_metadata_size(); |
717 | |
718 | SmallVector<std::pair<unsigned, MDNode *>, 32> GlobalMetadata; |
719 | for (const GlobalVariable &GV : M.globals()) { |
720 | ++Score; |
721 | |
722 | if (GV.hasInitializer()) |
723 | Score += classifyReductivePower(V: GV.getInitializer()); |
724 | |
725 | // TODO: Account for linkage? |
726 | |
727 | GV.getAllMetadata(MDs&: GlobalMetadata); |
728 | Score += GlobalMetadata.size(); |
729 | GlobalMetadata.clear(); |
730 | } |
731 | |
732 | for (const GlobalAlias &GA : M.aliases()) |
733 | Score += classifyReductivePower(V: GA.getAliasee()); |
734 | |
735 | for (const GlobalIFunc &GI : M.ifuncs()) |
736 | Score += classifyReductivePower(V: GI.getResolver()); |
737 | |
738 | for (const Function &F : M) |
739 | Score += computeIRComplexityScoreImpl(F); |
740 | |
741 | return Score; |
742 | } |
743 | |
744 | void ReducerWorkItem::writeOutput(raw_ostream &OS, bool EmitBitcode) const { |
745 | // Requesting bitcode emission with mir is nonsense, so just ignore it. |
746 | if (EmitBitcode && !isMIR()) |
747 | writeBitcode(OutStream&: OS); |
748 | else |
749 | print(ROS&: OS, /*AnnotationWriter=*/p: nullptr); |
750 | } |
751 | |
752 | void ReducerWorkItem::readBitcode(MemoryBufferRef Data, LLVMContext &Ctx, |
753 | StringRef ToolName) { |
754 | Expected<BitcodeFileContents> IF = llvm::getBitcodeFileContents(Buffer: Data); |
755 | if (!IF) { |
756 | WithColor::error(OS&: errs(), Prefix: ToolName) << IF.takeError(); |
757 | exit(status: 1); |
758 | } |
759 | |
760 | BitcodeModule BM = IF->Mods[0]; |
761 | Expected<BitcodeLTOInfo> LI = BM.getLTOInfo(); |
762 | if (!LI) { |
763 | WithColor::error(OS&: errs(), Prefix: ToolName) << LI.takeError(); |
764 | exit(status: 1); |
765 | } |
766 | |
767 | Expected<std::unique_ptr<Module>> MOrErr = BM.parseModule(Context&: Ctx); |
768 | if (!MOrErr) { |
769 | WithColor::error(OS&: errs(), Prefix: ToolName) << MOrErr.takeError(); |
770 | exit(status: 1); |
771 | } |
772 | |
773 | LTOInfo = std::make_unique<BitcodeLTOInfo>(args&: *LI); |
774 | M = std::move(MOrErr.get()); |
775 | } |
776 | |
777 | void ReducerWorkItem::writeBitcode(raw_ostream &OutStream) const { |
778 | const bool ShouldPreserveUseListOrder = true; |
779 | |
780 | if (LTOInfo && LTOInfo->IsThinLTO && LTOInfo->EnableSplitLTOUnit) { |
781 | PassBuilder PB; |
782 | LoopAnalysisManager LAM; |
783 | FunctionAnalysisManager FAM; |
784 | CGSCCAnalysisManager CGAM; |
785 | ModuleAnalysisManager MAM; |
786 | PB.registerModuleAnalyses(MAM); |
787 | PB.registerCGSCCAnalyses(CGAM); |
788 | PB.registerFunctionAnalyses(FAM); |
789 | PB.registerLoopAnalyses(LAM); |
790 | PB.crossRegisterProxies(LAM, FAM, CGAM, MAM); |
791 | ModulePassManager MPM; |
792 | MPM.addPass(Pass: ThinLTOBitcodeWriterPass(OutStream, nullptr, |
793 | ShouldPreserveUseListOrder)); |
794 | MPM.run(IR&: *M, AM&: MAM); |
795 | } else { |
796 | std::unique_ptr<ModuleSummaryIndex> Index; |
797 | if (LTOInfo && LTOInfo->HasSummary) { |
798 | ProfileSummaryInfo PSI(*M); |
799 | Index = std::make_unique<ModuleSummaryIndex>( |
800 | args: buildModuleSummaryIndex(M: *M, GetBFICallback: nullptr, PSI: &PSI)); |
801 | } |
802 | WriteBitcodeToFile(M: getModule(), Out&: OutStream, ShouldPreserveUseListOrder, |
803 | Index: Index.get()); |
804 | } |
805 | } |
806 | |
807 | std::pair<std::unique_ptr<ReducerWorkItem>, bool> |
808 | llvm::parseReducerWorkItem(StringRef ToolName, StringRef Filename, |
809 | LLVMContext &Ctxt, |
810 | std::unique_ptr<TargetMachine> &TM, bool IsMIR) { |
811 | bool IsBitcode = false; |
812 | Triple TheTriple; |
813 | |
814 | auto MMM = std::make_unique<ReducerWorkItem>(); |
815 | |
816 | if (IsMIR) { |
817 | initializeTargetInfo(); |
818 | |
819 | auto FileOrErr = MemoryBuffer::getFileOrSTDIN(Filename, /*IsText=*/true); |
820 | if (std::error_code EC = FileOrErr.getError()) { |
821 | WithColor::error(OS&: errs(), Prefix: ToolName) << EC.message() << '\n'; |
822 | return {nullptr, false}; |
823 | } |
824 | |
825 | std::unique_ptr<MIRParser> MParser = |
826 | createMIRParser(Contents: std::move(FileOrErr.get()), Context&: Ctxt); |
827 | |
828 | auto SetDataLayout = [&](StringRef DataLayoutTargetTriple, |
829 | StringRef OldDLStr) -> std::optional<std::string> { |
830 | // NB: We always call createTargetMachineForTriple() even if an explicit |
831 | // DataLayout is already set in the module since we want to use this |
832 | // callback to setup the TargetMachine rather than doing it later. |
833 | std::string IRTargetTriple = DataLayoutTargetTriple.str(); |
834 | if (!TargetTriple.empty()) |
835 | IRTargetTriple = Triple::normalize(Str: TargetTriple); |
836 | TheTriple = Triple(IRTargetTriple); |
837 | if (TheTriple.getTriple().empty()) |
838 | TheTriple.setTriple(sys::getDefaultTargetTriple()); |
839 | ExitOnError ExitOnErr(std::string(ToolName) + ": error: " ); |
840 | TM = ExitOnErr(codegen::createTargetMachineForTriple(TargetTriple: TheTriple.str())); |
841 | |
842 | return TM->createDataLayout().getStringRepresentation(); |
843 | }; |
844 | |
845 | std::unique_ptr<Module> M = MParser->parseIRModule(DataLayoutCallback: SetDataLayout); |
846 | |
847 | MMM->MMI = std::make_unique<MachineModuleInfo>(args: TM.get()); |
848 | MParser->parseMachineFunctions(M&: *M, MMI&: *MMM->MMI); |
849 | MMM->M = std::move(M); |
850 | } else { |
851 | SMDiagnostic Err; |
852 | ErrorOr<std::unique_ptr<MemoryBuffer>> MB = |
853 | MemoryBuffer::getFileOrSTDIN(Filename); |
854 | if (std::error_code EC = MB.getError()) { |
855 | WithColor::error(OS&: errs(), Prefix: ToolName) |
856 | << Filename << ": " << EC.message() << "\n" ; |
857 | return {nullptr, false}; |
858 | } |
859 | |
860 | if (!isBitcode(BufPtr: (const unsigned char *)(*MB)->getBufferStart(), |
861 | BufEnd: (const unsigned char *)(*MB)->getBufferEnd())) { |
862 | std::unique_ptr<Module> Result = parseIR(Buffer: **MB, Err, Context&: Ctxt); |
863 | if (!Result) { |
864 | Err.print(ProgName: ToolName.data(), S&: errs()); |
865 | return {nullptr, false}; |
866 | } |
867 | MMM->M = std::move(Result); |
868 | } else { |
869 | IsBitcode = true; |
870 | MMM->readBitcode(Data: MemoryBufferRef(**MB), Ctx&: Ctxt, ToolName); |
871 | |
872 | if (MMM->LTOInfo->IsThinLTO && MMM->LTOInfo->EnableSplitLTOUnit) |
873 | initializeTargetInfo(); |
874 | } |
875 | } |
876 | if (MMM->verify(OS: &errs())) { |
877 | WithColor::error(OS&: errs(), Prefix: ToolName) |
878 | << Filename << " - input module is broken!\n" ; |
879 | return {nullptr, false}; |
880 | } |
881 | return {std::move(MMM), IsBitcode}; |
882 | } |
883 | |