1 //===------ CodeGeneration.cpp - Code generate the Scops using ISL. ----======//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // The CodeGeneration pass takes a Scop created by ScopInfo and translates it
11 // back to LLVM-IR using the ISL code generator.
12 //
13 // The Scop describes the high level memory behaviour of a control flow region.
14 // Transformation passes can update the schedule (execution order) of statements
15 // in the Scop. ISL is used to generate an abstract syntax tree that reflects
16 // the updated execution order. This clast is used to create new LLVM-IR that is
17 // computationally equivalent to the original control flow region, but executes
18 // its code in the new execution order defined by the changed schedule.
19 //
20 //===----------------------------------------------------------------------===//
21 
22 #include "polly/CodeGen/IslAst.h"
23 #include "polly/CodeGen/IslNodeBuilder.h"
24 #include "polly/CodeGen/PerfMonitor.h"
25 #include "polly/CodeGen/Utils.h"
26 #include "polly/DependenceInfo.h"
27 #include "polly/LinkAllPasses.h"
28 #include "polly/Options.h"
29 #include "polly/ScopInfo.h"
30 #include "polly/Support/ScopHelper.h"
31 #include "llvm/Analysis/AliasAnalysis.h"
32 #include "llvm/Analysis/BasicAliasAnalysis.h"
33 #include "llvm/Analysis/GlobalsModRef.h"
34 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
35 #include "llvm/IR/Module.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/Support/Debug.h"
38 
39 using namespace polly;
40 using namespace llvm;
41 
42 #define DEBUG_TYPE "polly-codegen"
43 
44 static cl::opt<bool> Verify("polly-codegen-verify",
45                             cl::desc("Verify the function generated by Polly"),
46                             cl::Hidden, cl::init(true), cl::ZeroOrMore,
47                             cl::cat(PollyCategory));
48 
49 static cl::opt<bool>
50     PerfMonitoring("polly-codegen-perf-monitoring",
51                    cl::desc("Add run-time performance monitoring"), cl::Hidden,
52                    cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
53 
54 namespace {
55 class CodeGeneration : public ScopPass {
56 public:
57   static char ID;
58 
59   CodeGeneration() : ScopPass(ID) {}
60 
61   /// The datalayout used
62   const DataLayout *DL;
63 
64   /// @name The analysis passes we need to generate code.
65   ///
66   ///{
67   LoopInfo *LI;
68   IslAstInfo *AI;
69   DominatorTree *DT;
70   ScalarEvolution *SE;
71   RegionInfo *RI;
72   ///}
73 
74   void verifyGeneratedFunction(Scop &S, Function &F) {
75     if (!verifyFunction(F, &errs()) || !Verify)
76       return;
77 
78     DEBUG({
79       errs() << "== ISL Codegen created an invalid function ==\n\n== The "
80                 "SCoP ==\n";
81       S.print(errs());
82       errs() << "\n== The isl AST ==\n";
83       AI->printScop(errs(), S);
84       errs() << "\n== The invalid function ==\n";
85       F.print(errs());
86     });
87 
88     llvm_unreachable("Polly generated function could not be verified. Add "
89                      "-polly-codegen-verify=false to disable this assertion.");
90   }
91 
92   // CodeGeneration adds a lot of BBs without updating the RegionInfo
93   // We make all created BBs belong to the scop's parent region without any
94   // nested structure to keep the RegionInfo verifier happy.
95   void fixRegionInfo(Function *F, Region *ParentRegion) {
96     for (BasicBlock &BB : *F) {
97       if (RI->getRegionFor(&BB))
98         continue;
99 
100       RI->setRegionFor(&BB, ParentRegion);
101     }
102   }
103 
104   /// Mark a basic block unreachable.
105   ///
106   /// Marks the basic block @p Block unreachable by equipping it with an
107   /// UnreachableInst.
108   void markBlockUnreachable(BasicBlock &Block, PollyIRBuilder &Builder) {
109     auto *OrigTerminator = Block.getTerminator();
110     Builder.SetInsertPoint(OrigTerminator);
111     Builder.CreateUnreachable();
112     OrigTerminator->eraseFromParent();
113   }
114 
115   /// Generate LLVM-IR for the SCoP @p S.
116   bool runOnScop(Scop &S) override {
117     AI = &getAnalysis<IslAstInfo>();
118 
119     // Check if we created an isl_ast root node, otherwise exit.
120     isl_ast_node *AstRoot = AI->getAst();
121     if (!AstRoot)
122       return false;
123 
124     LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
125     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
126     SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
127     DL = &S.getFunction().getParent()->getDataLayout();
128     RI = &getAnalysis<RegionInfoPass>().getRegionInfo();
129     Region *R = &S.getRegion();
130     assert(!R->isTopLevelRegion() && "Top level regions are not supported");
131 
132     ScopAnnotator Annotator;
133     Annotator.buildAliasScopes(S);
134 
135     simplifyRegion(R, DT, LI, RI);
136     assert(R->isSimple());
137     BasicBlock *EnteringBB = S.getEnteringBlock();
138     assert(EnteringBB);
139     PollyIRBuilder Builder = createPollyIRBuilder(EnteringBB, Annotator);
140 
141     // Only build the run-time condition and parameters _after_ having
142     // introduced the conditional branch. This is important as the conditional
143     // branch will guard the original scop from new induction variables that
144     // the SCEVExpander may introduce while code generating the parameters and
145     // which may introduce scalar dependences that prevent us from correctly
146     // code generating this scop.
147     BasicBlock *StartBlock =
148         executeScopConditionally(S, this, Builder.getTrue());
149     auto *SplitBlock = StartBlock->getSinglePredecessor();
150 
151     IslNodeBuilder NodeBuilder(Builder, Annotator, this, *DL, *LI, *SE, *DT, S,
152                                StartBlock);
153 
154     if (PerfMonitoring) {
155       PerfMonitor P(EnteringBB->getParent()->getParent());
156       P.initialize();
157       P.insertRegionStart(SplitBlock->getTerminator());
158 
159       BasicBlock *MergeBlock = SplitBlock->getTerminator()
160                                    ->getSuccessor(0)
161                                    ->getUniqueSuccessor()
162                                    ->getUniqueSuccessor();
163       P.insertRegionEnd(MergeBlock->getTerminator());
164     }
165 
166     // First generate code for the hoisted invariant loads and transitively the
167     // parameters they reference. Afterwards, for the remaining parameters that
168     // might reference the hoisted loads. Finally, build the runtime check
169     // that might reference both hoisted loads as well as parameters.
170     // If the hoisting fails we have to bail and execute the original code.
171     Builder.SetInsertPoint(SplitBlock->getTerminator());
172     if (!NodeBuilder.preloadInvariantLoads()) {
173 
174       // Patch the introduced branch condition to ensure that we always execute
175       // the original SCoP.
176       auto *FalseI1 = Builder.getFalse();
177       auto *SplitBBTerm = Builder.GetInsertBlock()->getTerminator();
178       SplitBBTerm->setOperand(0, FalseI1);
179 
180       // Since the other branch is hence ignored we mark it as unreachable and
181       // adjust the dominator tree accordingly.
182       auto *ExitingBlock = StartBlock->getUniqueSuccessor();
183       assert(ExitingBlock);
184       auto *MergeBlock = ExitingBlock->getUniqueSuccessor();
185       assert(MergeBlock);
186       markBlockUnreachable(*StartBlock, Builder);
187       markBlockUnreachable(*ExitingBlock, Builder);
188       auto *ExitingBB = S.getExitingBlock();
189       assert(ExitingBB);
190       DT->changeImmediateDominator(MergeBlock, ExitingBB);
191       DT->eraseNode(ExitingBlock);
192 
193       isl_ast_node_free(AstRoot);
194     } else {
195       NodeBuilder.allocateNewArrays();
196       NodeBuilder.addParameters(S.getContext());
197       Value *RTC = NodeBuilder.createRTC(AI->getRunCondition());
198 
199       Builder.GetInsertBlock()->getTerminator()->setOperand(0, RTC);
200       Builder.SetInsertPoint(&StartBlock->front());
201 
202       NodeBuilder.create(AstRoot);
203       NodeBuilder.finalize();
204       fixRegionInfo(EnteringBB->getParent(), R->getParent());
205     }
206 
207     Function *F = EnteringBB->getParent();
208     verifyGeneratedFunction(S, *F);
209     for (auto *SubF : NodeBuilder.getParallelSubfunctions())
210       verifyGeneratedFunction(S, *SubF);
211 
212     // Mark the function such that we run additional cleanup passes on this
213     // function (e.g. mem2reg to rediscover phi nodes).
214     F->addFnAttr("polly-optimized");
215 
216     return true;
217   }
218 
219   /// Register all analyses and transformation required.
220   void getAnalysisUsage(AnalysisUsage &AU) const override {
221     AU.addRequired<DominatorTreeWrapperPass>();
222     AU.addRequired<IslAstInfo>();
223     AU.addRequired<RegionInfoPass>();
224     AU.addRequired<ScalarEvolutionWrapperPass>();
225     AU.addRequired<ScopDetection>();
226     AU.addRequired<ScopInfoRegionPass>();
227     AU.addRequired<LoopInfoWrapperPass>();
228 
229     AU.addPreserved<DependenceInfo>();
230 
231     AU.addPreserved<AAResultsWrapperPass>();
232     AU.addPreserved<BasicAAWrapperPass>();
233     AU.addPreserved<LoopInfoWrapperPass>();
234     AU.addPreserved<DominatorTreeWrapperPass>();
235     AU.addPreserved<GlobalsAAWrapperPass>();
236     AU.addPreserved<IslAstInfo>();
237     AU.addPreserved<ScopDetection>();
238     AU.addPreserved<ScalarEvolutionWrapperPass>();
239     AU.addPreserved<SCEVAAWrapperPass>();
240 
241     // FIXME: We do not yet add regions for the newly generated code to the
242     //        region tree.
243     AU.addPreserved<RegionInfoPass>();
244     AU.addPreserved<ScopInfoRegionPass>();
245   }
246 };
247 } // namespace
248 
249 char CodeGeneration::ID = 1;
250 
251 Pass *polly::createCodeGenerationPass() { return new CodeGeneration(); }
252 
253 INITIALIZE_PASS_BEGIN(CodeGeneration, "polly-codegen",
254                       "Polly - Create LLVM-IR from SCoPs", false, false);
255 INITIALIZE_PASS_DEPENDENCY(DependenceInfo);
256 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass);
257 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass);
258 INITIALIZE_PASS_DEPENDENCY(RegionInfoPass);
259 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass);
260 INITIALIZE_PASS_DEPENDENCY(ScopDetection);
261 INITIALIZE_PASS_END(CodeGeneration, "polly-codegen",
262                     "Polly - Create LLVM-IR from SCoPs", false, false)
263