• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- DivergenceAnalysis.cpp --------- Divergence Analysis Implementation -==//
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 // This file implements divergence analysis which determines whether a branch
11 // in a GPU program is divergent.It can help branch optimizations such as jump
12 // threading and loop unswitching to make better decisions.
13 //
14 // GPU programs typically use the SIMD execution model, where multiple threads
15 // in the same execution group have to execute in lock-step. Therefore, if the
16 // code contains divergent branches (i.e., threads in a group do not agree on
17 // which path of the branch to take), the group of threads has to execute all
18 // the paths from that branch with different subsets of threads enabled until
19 // they converge at the immediately post-dominating BB of the paths.
20 //
21 // Due to this execution model, some optimizations such as jump
22 // threading and loop unswitching can be unfortunately harmful when performed on
23 // divergent branches. Therefore, an analysis that computes which branches in a
24 // GPU program are divergent can help the compiler to selectively run these
25 // optimizations.
26 //
27 // This file defines divergence analysis which computes a conservative but
28 // non-trivial approximation of all divergent branches in a GPU program. It
29 // partially implements the approach described in
30 //
31 //   Divergence Analysis
32 //   Sampaio, Souza, Collange, Pereira
33 //   TOPLAS '13
34 //
35 // The divergence analysis identifies the sources of divergence (e.g., special
36 // variables that hold the thread ID), and recursively marks variables that are
37 // data or sync dependent on a source of divergence as divergent.
38 //
39 // While data dependency is a well-known concept, the notion of sync dependency
40 // is worth more explanation. Sync dependence characterizes the control flow
41 // aspect of the propagation of branch divergence. For example,
42 //
43 //   %cond = icmp slt i32 %tid, 10
44 //   br i1 %cond, label %then, label %else
45 // then:
46 //   br label %merge
47 // else:
48 //   br label %merge
49 // merge:
50 //   %a = phi i32 [ 0, %then ], [ 1, %else ]
51 //
52 // Suppose %tid holds the thread ID. Although %a is not data dependent on %tid
53 // because %tid is not on its use-def chains, %a is sync dependent on %tid
54 // because the branch "br i1 %cond" depends on %tid and affects which value %a
55 // is assigned to.
56 //
57 // The current implementation has the following limitations:
58 // 1. intra-procedural. It conservatively considers the arguments of a
59 //    non-kernel-entry function and the return value of a function call as
60 //    divergent.
61 // 2. memory as black box. It conservatively considers values loaded from
62 //    generic or local address as divergent. This can be improved by leveraging
63 //    pointer analysis.
64 //
65 //===----------------------------------------------------------------------===//
66 
67 #include "llvm/Analysis/DivergenceAnalysis.h"
68 #include "llvm/Analysis/Passes.h"
69 #include "llvm/Analysis/PostDominators.h"
70 #include "llvm/Analysis/TargetTransformInfo.h"
71 #include "llvm/IR/Dominators.h"
72 #include "llvm/IR/InstIterator.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/IntrinsicInst.h"
75 #include "llvm/IR/Value.h"
76 #include "llvm/Support/Debug.h"
77 #include "llvm/Support/raw_ostream.h"
78 #include <vector>
79 using namespace llvm;
80 
81 namespace {
82 
83 class DivergencePropagator {
84 public:
DivergencePropagator(Function & F,TargetTransformInfo & TTI,DominatorTree & DT,PostDominatorTree & PDT,DenseSet<const Value * > & DV)85   DivergencePropagator(Function &F, TargetTransformInfo &TTI, DominatorTree &DT,
86                        PostDominatorTree &PDT, DenseSet<const Value *> &DV)
87       : F(F), TTI(TTI), DT(DT), PDT(PDT), DV(DV) {}
88   void populateWithSourcesOfDivergence();
89   void propagate();
90 
91 private:
92   // A helper function that explores data dependents of V.
93   void exploreDataDependency(Value *V);
94   // A helper function that explores sync dependents of TI.
95   void exploreSyncDependency(TerminatorInst *TI);
96   // Computes the influence region from Start to End. This region includes all
97   // basic blocks on any simple path from Start to End.
98   void computeInfluenceRegion(BasicBlock *Start, BasicBlock *End,
99                               DenseSet<BasicBlock *> &InfluenceRegion);
100   // Finds all users of I that are outside the influence region, and add these
101   // users to Worklist.
102   void findUsersOutsideInfluenceRegion(
103       Instruction &I, const DenseSet<BasicBlock *> &InfluenceRegion);
104 
105   Function &F;
106   TargetTransformInfo &TTI;
107   DominatorTree &DT;
108   PostDominatorTree &PDT;
109   std::vector<Value *> Worklist; // Stack for DFS.
110   DenseSet<const Value *> &DV;   // Stores all divergent values.
111 };
112 
populateWithSourcesOfDivergence()113 void DivergencePropagator::populateWithSourcesOfDivergence() {
114   Worklist.clear();
115   DV.clear();
116   for (auto &I : instructions(F)) {
117     if (TTI.isSourceOfDivergence(&I)) {
118       Worklist.push_back(&I);
119       DV.insert(&I);
120     }
121   }
122   for (auto &Arg : F.args()) {
123     if (TTI.isSourceOfDivergence(&Arg)) {
124       Worklist.push_back(&Arg);
125       DV.insert(&Arg);
126     }
127   }
128 }
129 
exploreSyncDependency(TerminatorInst * TI)130 void DivergencePropagator::exploreSyncDependency(TerminatorInst *TI) {
131   // Propagation rule 1: if branch TI is divergent, all PHINodes in TI's
132   // immediate post dominator are divergent. This rule handles if-then-else
133   // patterns. For example,
134   //
135   // if (tid < 5)
136   //   a1 = 1;
137   // else
138   //   a2 = 2;
139   // a = phi(a1, a2); // sync dependent on (tid < 5)
140   BasicBlock *ThisBB = TI->getParent();
141 
142   // Unreachable blocks may not be in the dominator tree.
143   if (!DT.isReachableFromEntry(ThisBB))
144     return;
145 
146   // If the function has no exit blocks or doesn't reach any exit blocks, the
147   // post dominator may be null.
148   DomTreeNode *ThisNode = PDT.getNode(ThisBB);
149   if (!ThisNode)
150     return;
151 
152   BasicBlock *IPostDom = ThisNode->getIDom()->getBlock();
153   if (IPostDom == nullptr)
154     return;
155 
156   for (auto I = IPostDom->begin(); isa<PHINode>(I); ++I) {
157     // A PHINode is uniform if it returns the same value no matter which path is
158     // taken.
159     if (!cast<PHINode>(I)->hasConstantOrUndefValue() && DV.insert(&*I).second)
160       Worklist.push_back(&*I);
161   }
162 
163   // Propagation rule 2: if a value defined in a loop is used outside, the user
164   // is sync dependent on the condition of the loop exits that dominate the
165   // user. For example,
166   //
167   // int i = 0;
168   // do {
169   //   i++;
170   //   if (foo(i)) ... // uniform
171   // } while (i < tid);
172   // if (bar(i)) ...   // divergent
173   //
174   // A program may contain unstructured loops. Therefore, we cannot leverage
175   // LoopInfo, which only recognizes natural loops.
176   //
177   // The algorithm used here handles both natural and unstructured loops.  Given
178   // a branch TI, we first compute its influence region, the union of all simple
179   // paths from TI to its immediate post dominator (IPostDom). Then, we search
180   // for all the values defined in the influence region but used outside. All
181   // these users are sync dependent on TI.
182   DenseSet<BasicBlock *> InfluenceRegion;
183   computeInfluenceRegion(ThisBB, IPostDom, InfluenceRegion);
184   // An insight that can speed up the search process is that all the in-region
185   // values that are used outside must dominate TI. Therefore, instead of
186   // searching every basic blocks in the influence region, we search all the
187   // dominators of TI until it is outside the influence region.
188   BasicBlock *InfluencedBB = ThisBB;
189   while (InfluenceRegion.count(InfluencedBB)) {
190     for (auto &I : *InfluencedBB)
191       findUsersOutsideInfluenceRegion(I, InfluenceRegion);
192     DomTreeNode *IDomNode = DT.getNode(InfluencedBB)->getIDom();
193     if (IDomNode == nullptr)
194       break;
195     InfluencedBB = IDomNode->getBlock();
196   }
197 }
198 
findUsersOutsideInfluenceRegion(Instruction & I,const DenseSet<BasicBlock * > & InfluenceRegion)199 void DivergencePropagator::findUsersOutsideInfluenceRegion(
200     Instruction &I, const DenseSet<BasicBlock *> &InfluenceRegion) {
201   for (User *U : I.users()) {
202     Instruction *UserInst = cast<Instruction>(U);
203     if (!InfluenceRegion.count(UserInst->getParent())) {
204       if (DV.insert(UserInst).second)
205         Worklist.push_back(UserInst);
206     }
207   }
208 }
209 
210 // A helper function for computeInfluenceRegion that adds successors of "ThisBB"
211 // to the influence region.
212 static void
addSuccessorsToInfluenceRegion(BasicBlock * ThisBB,BasicBlock * End,DenseSet<BasicBlock * > & InfluenceRegion,std::vector<BasicBlock * > & InfluenceStack)213 addSuccessorsToInfluenceRegion(BasicBlock *ThisBB, BasicBlock *End,
214                                DenseSet<BasicBlock *> &InfluenceRegion,
215                                std::vector<BasicBlock *> &InfluenceStack) {
216   for (BasicBlock *Succ : successors(ThisBB)) {
217     if (Succ != End && InfluenceRegion.insert(Succ).second)
218       InfluenceStack.push_back(Succ);
219   }
220 }
221 
computeInfluenceRegion(BasicBlock * Start,BasicBlock * End,DenseSet<BasicBlock * > & InfluenceRegion)222 void DivergencePropagator::computeInfluenceRegion(
223     BasicBlock *Start, BasicBlock *End,
224     DenseSet<BasicBlock *> &InfluenceRegion) {
225   assert(PDT.properlyDominates(End, Start) &&
226          "End does not properly dominate Start");
227 
228   // The influence region starts from the end of "Start" to the beginning of
229   // "End". Therefore, "Start" should not be in the region unless "Start" is in
230   // a loop that doesn't contain "End".
231   std::vector<BasicBlock *> InfluenceStack;
232   addSuccessorsToInfluenceRegion(Start, End, InfluenceRegion, InfluenceStack);
233   while (!InfluenceStack.empty()) {
234     BasicBlock *BB = InfluenceStack.back();
235     InfluenceStack.pop_back();
236     addSuccessorsToInfluenceRegion(BB, End, InfluenceRegion, InfluenceStack);
237   }
238 }
239 
exploreDataDependency(Value * V)240 void DivergencePropagator::exploreDataDependency(Value *V) {
241   // Follow def-use chains of V.
242   for (User *U : V->users()) {
243     Instruction *UserInst = cast<Instruction>(U);
244     if (DV.insert(UserInst).second)
245       Worklist.push_back(UserInst);
246   }
247 }
248 
propagate()249 void DivergencePropagator::propagate() {
250   // Traverse the dependency graph using DFS.
251   while (!Worklist.empty()) {
252     Value *V = Worklist.back();
253     Worklist.pop_back();
254     if (TerminatorInst *TI = dyn_cast<TerminatorInst>(V)) {
255       // Terminators with less than two successors won't introduce sync
256       // dependency. Ignore them.
257       if (TI->getNumSuccessors() > 1)
258         exploreSyncDependency(TI);
259     }
260     exploreDataDependency(V);
261   }
262 }
263 
264 } /// end namespace anonymous
265 
266 // Register this pass.
267 char DivergenceAnalysis::ID = 0;
268 INITIALIZE_PASS_BEGIN(DivergenceAnalysis, "divergence", "Divergence Analysis",
269                       false, true)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)270 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
271 INITIALIZE_PASS_DEPENDENCY(PostDominatorTreeWrapperPass)
272 INITIALIZE_PASS_END(DivergenceAnalysis, "divergence", "Divergence Analysis",
273                     false, true)
274 
275 FunctionPass *llvm::createDivergenceAnalysisPass() {
276   return new DivergenceAnalysis();
277 }
278 
getAnalysisUsage(AnalysisUsage & AU) const279 void DivergenceAnalysis::getAnalysisUsage(AnalysisUsage &AU) const {
280   AU.addRequired<DominatorTreeWrapperPass>();
281   AU.addRequired<PostDominatorTreeWrapperPass>();
282   AU.setPreservesAll();
283 }
284 
runOnFunction(Function & F)285 bool DivergenceAnalysis::runOnFunction(Function &F) {
286   auto *TTIWP = getAnalysisIfAvailable<TargetTransformInfoWrapperPass>();
287   if (TTIWP == nullptr)
288     return false;
289 
290   TargetTransformInfo &TTI = TTIWP->getTTI(F);
291   // Fast path: if the target does not have branch divergence, we do not mark
292   // any branch as divergent.
293   if (!TTI.hasBranchDivergence())
294     return false;
295 
296   DivergentValues.clear();
297   auto &PDT = getAnalysis<PostDominatorTreeWrapperPass>().getPostDomTree();
298   DivergencePropagator DP(F, TTI,
299                           getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
300                           PDT, DivergentValues);
301   DP.populateWithSourcesOfDivergence();
302   DP.propagate();
303   return false;
304 }
305 
print(raw_ostream & OS,const Module *) const306 void DivergenceAnalysis::print(raw_ostream &OS, const Module *) const {
307   if (DivergentValues.empty())
308     return;
309   const Value *FirstDivergentValue = *DivergentValues.begin();
310   const Function *F;
311   if (const Argument *Arg = dyn_cast<Argument>(FirstDivergentValue)) {
312     F = Arg->getParent();
313   } else if (const Instruction *I =
314                  dyn_cast<Instruction>(FirstDivergentValue)) {
315     F = I->getParent()->getParent();
316   } else {
317     llvm_unreachable("Only arguments and instructions can be divergent");
318   }
319 
320   // Dumps all divergent values in F, arguments and then instructions.
321   for (auto &Arg : F->args()) {
322     if (DivergentValues.count(&Arg))
323       OS << "DIVERGENT:  " << Arg << "\n";
324   }
325   // Iterate instructions using instructions() to ensure a deterministic order.
326   for (auto &I : instructions(F)) {
327     if (DivergentValues.count(&I))
328       OS << "DIVERGENT:" << I << "\n";
329   }
330 }
331