1 //===- SIAnnotateControlFlow.cpp ------------------------------------------===// 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 /// \file 10 /// Annotates the control flow with hardware specific intrinsics. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "AMDGPU.h" 15 #include "GCNSubtarget.h" 16 #include "llvm/Analysis/LegacyDivergenceAnalysis.h" 17 #include "llvm/Analysis/LoopInfo.h" 18 #include "llvm/CodeGen/TargetPassConfig.h" 19 #include "llvm/IR/BasicBlock.h" 20 #include "llvm/IR/Constants.h" 21 #include "llvm/IR/Dominators.h" 22 #include "llvm/IR/IntrinsicsAMDGPU.h" 23 #include "llvm/InitializePasses.h" 24 #include "llvm/Target/TargetMachine.h" 25 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 26 #include "llvm/Transforms/Utils/Local.h" 27 28 using namespace llvm; 29 30 #define DEBUG_TYPE "si-annotate-control-flow" 31 32 namespace { 33 34 // Complex types used in this pass 35 using StackEntry = std::pair<BasicBlock *, Value *>; 36 using StackVector = SmallVector<StackEntry, 16>; 37 38 class SIAnnotateControlFlow : public FunctionPass { 39 LegacyDivergenceAnalysis *DA; 40 41 Type *Boolean; 42 Type *Void; 43 Type *IntMask; 44 Type *ReturnStruct; 45 46 ConstantInt *BoolTrue; 47 ConstantInt *BoolFalse; 48 UndefValue *BoolUndef; 49 Constant *IntMaskZero; 50 51 Function *If; 52 Function *Else; 53 Function *IfBreak; 54 Function *Loop; 55 Function *EndCf; 56 57 DominatorTree *DT; 58 StackVector Stack; 59 60 LoopInfo *LI; 61 62 void initialize(Module &M, const GCNSubtarget &ST); 63 64 bool isUniform(BranchInst *T); 65 66 bool isTopOfStack(BasicBlock *BB); 67 68 Value *popSaved(); 69 70 void push(BasicBlock *BB, Value *Saved); 71 72 bool isElse(PHINode *Phi); 73 74 bool hasKill(const BasicBlock *BB); 75 76 void eraseIfUnused(PHINode *Phi); 77 78 void openIf(BranchInst *Term); 79 80 void insertElse(BranchInst *Term); 81 82 Value * 83 handleLoopCondition(Value *Cond, PHINode *Broken, llvm::Loop *L, 84 BranchInst *Term); 85 86 void handleLoop(BranchInst *Term); 87 88 void closeControlFlow(BasicBlock *BB); 89 90 public: 91 static char ID; 92 93 SIAnnotateControlFlow() : FunctionPass(ID) {} 94 95 bool runOnFunction(Function &F) override; 96 97 StringRef getPassName() const override { return "SI annotate control flow"; } 98 99 void getAnalysisUsage(AnalysisUsage &AU) const override { 100 AU.addRequired<LoopInfoWrapperPass>(); 101 AU.addRequired<DominatorTreeWrapperPass>(); 102 AU.addRequired<LegacyDivergenceAnalysis>(); 103 AU.addPreserved<LoopInfoWrapperPass>(); 104 AU.addPreserved<DominatorTreeWrapperPass>(); 105 AU.addRequired<TargetPassConfig>(); 106 FunctionPass::getAnalysisUsage(AU); 107 } 108 }; 109 110 } // end anonymous namespace 111 112 INITIALIZE_PASS_BEGIN(SIAnnotateControlFlow, DEBUG_TYPE, 113 "Annotate SI Control Flow", false, false) 114 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 115 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis) 116 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig) 117 INITIALIZE_PASS_END(SIAnnotateControlFlow, DEBUG_TYPE, 118 "Annotate SI Control Flow", false, false) 119 120 char SIAnnotateControlFlow::ID = 0; 121 122 /// Initialize all the types and constants used in the pass 123 void SIAnnotateControlFlow::initialize(Module &M, const GCNSubtarget &ST) { 124 LLVMContext &Context = M.getContext(); 125 126 Void = Type::getVoidTy(Context); 127 Boolean = Type::getInt1Ty(Context); 128 IntMask = ST.isWave32() ? Type::getInt32Ty(Context) 129 : Type::getInt64Ty(Context); 130 ReturnStruct = StructType::get(Boolean, IntMask); 131 132 BoolTrue = ConstantInt::getTrue(Context); 133 BoolFalse = ConstantInt::getFalse(Context); 134 BoolUndef = UndefValue::get(Boolean); 135 IntMaskZero = ConstantInt::get(IntMask, 0); 136 137 If = Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_if, { IntMask }); 138 Else = Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_else, 139 { IntMask, IntMask }); 140 IfBreak = Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_if_break, 141 { IntMask }); 142 Loop = Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_loop, { IntMask }); 143 EndCf = Intrinsic::getDeclaration(&M, Intrinsic::amdgcn_end_cf, { IntMask }); 144 } 145 146 /// Is the branch condition uniform or did the StructurizeCFG pass 147 /// consider it as such? 148 bool SIAnnotateControlFlow::isUniform(BranchInst *T) { 149 return DA->isUniform(T) || 150 T->getMetadata("structurizecfg.uniform") != nullptr; 151 } 152 153 /// Is BB the last block saved on the stack ? 154 bool SIAnnotateControlFlow::isTopOfStack(BasicBlock *BB) { 155 return !Stack.empty() && Stack.back().first == BB; 156 } 157 158 /// Pop the last saved value from the control flow stack 159 Value *SIAnnotateControlFlow::popSaved() { 160 return Stack.pop_back_val().second; 161 } 162 163 /// Push a BB and saved value to the control flow stack 164 void SIAnnotateControlFlow::push(BasicBlock *BB, Value *Saved) { 165 Stack.push_back(std::make_pair(BB, Saved)); 166 } 167 168 /// Can the condition represented by this PHI node treated like 169 /// an "Else" block? 170 bool SIAnnotateControlFlow::isElse(PHINode *Phi) { 171 BasicBlock *IDom = DT->getNode(Phi->getParent())->getIDom()->getBlock(); 172 for (unsigned i = 0, e = Phi->getNumIncomingValues(); i != e; ++i) { 173 if (Phi->getIncomingBlock(i) == IDom) { 174 175 if (Phi->getIncomingValue(i) != BoolTrue) 176 return false; 177 178 } else { 179 if (Phi->getIncomingValue(i) != BoolFalse) 180 return false; 181 182 } 183 } 184 return true; 185 } 186 187 bool SIAnnotateControlFlow::hasKill(const BasicBlock *BB) { 188 for (const Instruction &I : *BB) { 189 if (const CallInst *CI = dyn_cast<CallInst>(&I)) 190 if (CI->getIntrinsicID() == Intrinsic::amdgcn_kill) 191 return true; 192 } 193 return false; 194 } 195 196 // Erase "Phi" if it is not used any more 197 void SIAnnotateControlFlow::eraseIfUnused(PHINode *Phi) { 198 if (RecursivelyDeleteDeadPHINode(Phi)) { 199 LLVM_DEBUG(dbgs() << "Erased unused condition phi\n"); 200 } 201 } 202 203 /// Open a new "If" block 204 void SIAnnotateControlFlow::openIf(BranchInst *Term) { 205 if (isUniform(Term)) 206 return; 207 208 Value *Ret = CallInst::Create(If, Term->getCondition(), "", Term); 209 Term->setCondition(ExtractValueInst::Create(Ret, 0, "", Term)); 210 push(Term->getSuccessor(1), ExtractValueInst::Create(Ret, 1, "", Term)); 211 } 212 213 /// Close the last "If" block and open a new "Else" block 214 void SIAnnotateControlFlow::insertElse(BranchInst *Term) { 215 if (isUniform(Term)) { 216 return; 217 } 218 Value *Ret = CallInst::Create(Else, popSaved(), "", Term); 219 Term->setCondition(ExtractValueInst::Create(Ret, 0, "", Term)); 220 push(Term->getSuccessor(1), ExtractValueInst::Create(Ret, 1, "", Term)); 221 } 222 223 /// Recursively handle the condition leading to a loop 224 Value *SIAnnotateControlFlow::handleLoopCondition( 225 Value *Cond, PHINode *Broken, llvm::Loop *L, BranchInst *Term) { 226 if (Instruction *Inst = dyn_cast<Instruction>(Cond)) { 227 BasicBlock *Parent = Inst->getParent(); 228 Instruction *Insert; 229 if (L->contains(Inst)) { 230 Insert = Parent->getTerminator(); 231 } else { 232 Insert = L->getHeader()->getFirstNonPHIOrDbgOrLifetime(); 233 } 234 235 Value *Args[] = { Cond, Broken }; 236 return CallInst::Create(IfBreak, Args, "", Insert); 237 } 238 239 // Insert IfBreak in the loop header TERM for constant COND other than true. 240 if (isa<Constant>(Cond)) { 241 Instruction *Insert = Cond == BoolTrue ? 242 Term : L->getHeader()->getTerminator(); 243 244 Value *Args[] = { Cond, Broken }; 245 return CallInst::Create(IfBreak, Args, "", Insert); 246 } 247 248 if (isa<Argument>(Cond)) { 249 Instruction *Insert = L->getHeader()->getFirstNonPHIOrDbgOrLifetime(); 250 Value *Args[] = { Cond, Broken }; 251 return CallInst::Create(IfBreak, Args, "", Insert); 252 } 253 254 llvm_unreachable("Unhandled loop condition!"); 255 } 256 257 /// Handle a back edge (loop) 258 void SIAnnotateControlFlow::handleLoop(BranchInst *Term) { 259 if (isUniform(Term)) 260 return; 261 262 BasicBlock *BB = Term->getParent(); 263 llvm::Loop *L = LI->getLoopFor(BB); 264 if (!L) 265 return; 266 267 BasicBlock *Target = Term->getSuccessor(1); 268 PHINode *Broken = PHINode::Create(IntMask, 0, "phi.broken", &Target->front()); 269 270 Value *Cond = Term->getCondition(); 271 Term->setCondition(BoolTrue); 272 Value *Arg = handleLoopCondition(Cond, Broken, L, Term); 273 274 for (BasicBlock *Pred : predecessors(Target)) { 275 Value *PHIValue = IntMaskZero; 276 if (Pred == BB) // Remember the value of the previous iteration. 277 PHIValue = Arg; 278 // If the backedge from Pred to Target could be executed before the exit 279 // of the loop at BB, it should not reset or change "Broken", which keeps 280 // track of the number of threads exited the loop at BB. 281 else if (L->contains(Pred) && DT->dominates(Pred, BB)) 282 PHIValue = Broken; 283 Broken->addIncoming(PHIValue, Pred); 284 } 285 286 Term->setCondition(CallInst::Create(Loop, Arg, "", Term)); 287 288 push(Term->getSuccessor(0), Arg); 289 } 290 291 /// Close the last opened control flow 292 void SIAnnotateControlFlow::closeControlFlow(BasicBlock *BB) { 293 llvm::Loop *L = LI->getLoopFor(BB); 294 295 assert(Stack.back().first == BB); 296 297 if (L && L->getHeader() == BB) { 298 // We can't insert an EndCF call into a loop header, because it will 299 // get executed on every iteration of the loop, when it should be 300 // executed only once before the loop. 301 SmallVector <BasicBlock *, 8> Latches; 302 L->getLoopLatches(Latches); 303 304 SmallVector<BasicBlock *, 2> Preds; 305 for (BasicBlock *Pred : predecessors(BB)) { 306 if (!is_contained(Latches, Pred)) 307 Preds.push_back(Pred); 308 } 309 310 BB = SplitBlockPredecessors(BB, Preds, "endcf.split", DT, LI, nullptr, 311 false); 312 } 313 314 Value *Exec = popSaved(); 315 Instruction *FirstInsertionPt = &*BB->getFirstInsertionPt(); 316 if (!isa<UndefValue>(Exec) && !isa<UnreachableInst>(FirstInsertionPt)) { 317 Instruction *ExecDef = cast<Instruction>(Exec); 318 BasicBlock *DefBB = ExecDef->getParent(); 319 if (!DT->dominates(DefBB, BB)) { 320 // Split edge to make Def dominate Use 321 FirstInsertionPt = &*SplitEdge(DefBB, BB, DT, LI)->getFirstInsertionPt(); 322 } 323 CallInst::Create(EndCf, Exec, "", FirstInsertionPt); 324 } 325 } 326 327 /// Annotate the control flow with intrinsics so the backend can 328 /// recognize if/then/else and loops. 329 bool SIAnnotateControlFlow::runOnFunction(Function &F) { 330 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree(); 331 LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo(); 332 DA = &getAnalysis<LegacyDivergenceAnalysis>(); 333 TargetPassConfig &TPC = getAnalysis<TargetPassConfig>(); 334 const TargetMachine &TM = TPC.getTM<TargetMachine>(); 335 336 initialize(*F.getParent(), TM.getSubtarget<GCNSubtarget>(F)); 337 for (df_iterator<BasicBlock *> I = df_begin(&F.getEntryBlock()), 338 E = df_end(&F.getEntryBlock()); I != E; ++I) { 339 BasicBlock *BB = *I; 340 BranchInst *Term = dyn_cast<BranchInst>(BB->getTerminator()); 341 342 if (!Term || Term->isUnconditional()) { 343 if (isTopOfStack(BB)) 344 closeControlFlow(BB); 345 346 continue; 347 } 348 349 if (I.nodeVisited(Term->getSuccessor(1))) { 350 if (isTopOfStack(BB)) 351 closeControlFlow(BB); 352 353 if (DT->dominates(Term->getSuccessor(1), BB)) 354 handleLoop(Term); 355 continue; 356 } 357 358 if (isTopOfStack(BB)) { 359 PHINode *Phi = dyn_cast<PHINode>(Term->getCondition()); 360 if (Phi && Phi->getParent() == BB && isElse(Phi) && !hasKill(BB)) { 361 insertElse(Term); 362 eraseIfUnused(Phi); 363 continue; 364 } 365 366 closeControlFlow(BB); 367 } 368 369 openIf(Term); 370 } 371 372 if (!Stack.empty()) { 373 // CFG was probably not structured. 374 report_fatal_error("failed to annotate CFG"); 375 } 376 377 return true; 378 } 379 380 /// Create the annotation pass 381 FunctionPass *llvm::createSIAnnotateControlFlowPass() { 382 return new SIAnnotateControlFlow(); 383 } 384