1 //===- Parallelize.cpp - Auto parallelization using DS Graphs -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements a pass that automatically parallelizes a program,
11 // using the Cilk multi-threaded runtime system to execute parallel code.
13 // The pass uses the Program Dependence Graph (class PDGIterator) to
14 // identify parallelizable function calls, i.e., calls whose instances
15 // can be executed in parallel with instances of other function calls.
16 // (In the future, this should also execute different instances of the same
17 // function call in parallel, but that requires parallelizing across
20 // The output of the pass is LLVM code with:
21 // (1) all parallelizable functions renamed to flag them as parallelizable;
22 // (2) calls to a sync() function introduced at synchronization points.
23 // The CWriter recognizes these functions and inserts the appropriate Cilk
24 // keywords when writing out C code. This C code must be compiled with cilk2c.
26 // Current algorithmic limitations:
27 // -- no array dependence analysis
28 // -- no parallelization for function calls in different loop iterations
29 // (except in unlikely trivial cases)
31 // Limitations of using Cilk:
32 // -- No parallelism within a function body, e.g., in a loop;
33 // -- Simplistic synchronization model requiring all parallel threads
34 // created within a function to block at a sync().
35 // -- Excessive overhead at "spawned" function calls, which has no benefit
36 // once all threads are busy (especially common when the degree of
37 // parallelism is low).
39 //===----------------------------------------------------------------------===//
41 #include "llvm/Transforms/Utils/DemoteRegToStack.h"
42 #include "llvm/Analysis/PgmDependenceGraph.h"
43 #include "llvm/Analysis/DataStructure.h"
44 #include "llvm/Analysis/DSGraph.h"
45 #include "llvm/Module.h"
46 #include "llvm/Instructions.h"
47 #include "llvm/DerivedTypes.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "Support/Statistic.h"
50 #include "Support/STLExtras.h"
51 #include "Support/hash_set"
52 #include "Support/hash_map"
56 //----------------------------------------------------------------------------
57 // Global constants used in marking Cilk functions and function calls.
58 //----------------------------------------------------------------------------
60 static const char * const CilkSuffix = ".llvm2cilk";
61 static const char * const DummySyncFuncName = "__sync.llvm2cilk";
63 //----------------------------------------------------------------------------
64 // Routines to identify Cilk functions, calls to Cilk functions, and syncs.
65 //----------------------------------------------------------------------------
67 static bool isCilk(const Function& F) {
68 return (F.getName().rfind(CilkSuffix) ==
69 F.getName().size() - std::strlen(CilkSuffix));
72 static bool isCilkMain(const Function& F) {
73 return F.getName() == "main" + std::string(CilkSuffix);
77 static bool isCilk(const CallInst& CI) {
78 return CI.getCalledFunction() && isCilk(*CI.getCalledFunction());
81 static bool isSync(const CallInst& CI) {
82 return CI.getCalledFunction() &&
83 CI.getCalledFunction()->getName() == DummySyncFuncName;
87 //----------------------------------------------------------------------------
90 // Code generation pass that transforms code to identify where Cilk keywords
91 // should be inserted. This relies on `llvm-dis -c' to print out the keywords.
92 //----------------------------------------------------------------------------
95 class Cilkifier: public InstVisitor<Cilkifier>
97 Function* DummySyncFunc;
99 // Data used when transforming each function.
100 hash_set<const Instruction*> stmtsVisited; // Flags for recursive DFS
101 hash_map<const CallInst*, hash_set<CallInst*> > spawnToSyncsMap;
103 // Input data for the transformation.
104 const hash_set<Function*>* cilkFunctions; // Set of parallel functions
105 PgmDependenceGraph* depGraph;
107 void DFSVisitInstr (Instruction* I,
109 hash_set<const Instruction*>& depsOfRoot);
112 /*ctor*/ Cilkifier (Module& M);
114 // Transform a single function including its name, its call sites, and syncs
116 void TransformFunc (Function* F,
117 const hash_set<Function*>& cilkFunctions,
118 PgmDependenceGraph& _depGraph);
120 // The visitor function that does most of the hard work, via DFSVisitInstr
122 void visitCallInst(CallInst& CI);
126 Cilkifier::Cilkifier(Module& M)
128 // create the dummy Sync function and add it to the Module
129 DummySyncFunc = M.getOrInsertFunction(DummySyncFuncName, Type::VoidTy, 0);
132 void Cilkifier::TransformFunc(Function* F,
133 const hash_set<Function*>& _cilkFunctions,
134 PgmDependenceGraph& _depGraph)
136 // Memoize the information for this function
137 cilkFunctions = &_cilkFunctions;
138 depGraph = &_depGraph;
140 // Add the marker suffix to the Function name
141 // This should automatically mark all calls to the function also!
142 F->setName(F->getName() + CilkSuffix);
144 // Insert sync operations for each separate spawn
147 // Now traverse the CFG in rPostorder and eliminate redundant syncs, i.e.,
148 // two consecutive sync's on a straight-line path with no intervening spawn.
153 void Cilkifier::DFSVisitInstr(Instruction* I,
155 hash_set<const Instruction*>& depsOfRoot)
157 assert(stmtsVisited.find(I) == stmtsVisited.end());
158 stmtsVisited.insert(I);
160 // If there is a dependence from root to I, insert Sync and return
161 if (depsOfRoot.find(I) != depsOfRoot.end())
162 { // Insert a sync before I and stop searching along this path.
163 // If I is a Phi instruction, the dependence can only be an SSA dep.
164 // and we need to insert the sync in the predecessor on the appropriate
167 if (PHINode* phiI = dyn_cast<PHINode>(I))
168 { // check all operands of the Phi and insert before each one
169 for (unsigned i = 0, N = phiI->getNumIncomingValues(); i < N; ++i)
170 if (phiI->getIncomingValue(i) == root)
171 syncI = new CallInst(DummySyncFunc, std::vector<Value*>(), "",
172 phiI->getIncomingBlock(i)->getTerminator());
175 syncI = new CallInst(DummySyncFunc, std::vector<Value*>(), "", I);
177 // Remember the sync for each spawn to eliminate redundant ones later
178 spawnToSyncsMap[cast<CallInst>(root)].insert(syncI);
183 // else visit unvisited successors
184 if (BranchInst* brI = dyn_cast<BranchInst>(I))
185 { // visit first instruction in each successor BB
186 for (unsigned i = 0, N = brI->getNumSuccessors(); i < N; ++i)
187 if (stmtsVisited.find(&brI->getSuccessor(i)->front())
188 == stmtsVisited.end())
189 DFSVisitInstr(&brI->getSuccessor(i)->front(), root, depsOfRoot);
192 if (Instruction* nextI = I->getNext())
193 if (stmtsVisited.find(nextI) == stmtsVisited.end())
194 DFSVisitInstr(nextI, root, depsOfRoot);
198 void Cilkifier::visitCallInst(CallInst& CI)
200 assert(CI.getCalledFunction() != 0 && "Only direct calls can be spawned.");
201 if (cilkFunctions->find(CI.getCalledFunction()) == cilkFunctions->end())
202 return; // not a spawn
204 // Find all the outgoing memory dependences.
205 hash_set<const Instruction*> depsOfRoot;
206 for (PgmDependenceGraph::iterator DI =
207 depGraph->outDepBegin(CI, MemoryDeps); ! DI.fini(); ++DI)
208 depsOfRoot.insert(&DI->getSink()->getInstr());
210 // Now find all outgoing SSA dependences to the eventual non-Phi users of
211 // the call value (i.e., direct users that are not phis, and for any
212 // user that is a Phi, direct non-Phi users of that Phi, and recursively).
213 std::vector<const PHINode*> phiUsers;
214 hash_set<const PHINode*> phisSeen; // ensures we don't visit a phi twice
215 for (Value::use_iterator UI=CI.use_begin(), UE=CI.use_end(); UI != UE; ++UI)
216 if (const PHINode* phiUser = dyn_cast<PHINode>(*UI))
218 if (phisSeen.find(phiUser) == phisSeen.end())
220 phiUsers.push_back(phiUser);
221 phisSeen.insert(phiUser);
225 depsOfRoot.insert(cast<Instruction>(*UI));
227 // Now we've found the non-Phi users and immediate phi users.
228 // Recursively walk the phi users and add their non-phi users.
229 for (const PHINode* phiUser; !phiUsers.empty(); phiUsers.pop_back())
231 phiUser = phiUsers.back();
232 for (Value::use_const_iterator UI=phiUser->use_begin(),
233 UE=phiUser->use_end(); UI != UE; ++UI)
234 if (const PHINode* pn = dyn_cast<PHINode>(*UI))
236 if (phisSeen.find(pn) == phisSeen.end())
238 phiUsers.push_back(pn);
243 depsOfRoot.insert(cast<Instruction>(*UI));
246 // Walk paths of the CFG starting at the call instruction and insert
247 // one sync before the first dependence on each path, if any.
248 if (! depsOfRoot.empty())
250 stmtsVisited.clear(); // start a new DFS for this CallInst
251 assert(CI.getNext() && "Call instruction cannot be a terminator!");
252 DFSVisitInstr(CI.getNext(), &CI, depsOfRoot);
255 // Now, eliminate all users of the SSA value of the CallInst, i.e.,
256 // if the call instruction returns a value, delete the return value
257 // register and replace it by a stack slot.
258 if (CI.getType() != Type::VoidTy)
259 DemoteRegToStack(CI);
263 //----------------------------------------------------------------------------
264 // class FindParallelCalls
266 // Find all CallInst instructions that have at least one other CallInst
267 // that is independent. These are the instructions that can produce
268 // useful parallelism.
269 //----------------------------------------------------------------------------
271 class FindParallelCalls : public InstVisitor<FindParallelCalls> {
272 typedef hash_set<CallInst*> DependentsSet;
273 typedef DependentsSet::iterator Dependents_iterator;
274 typedef DependentsSet::const_iterator Dependents_const_iterator;
276 PgmDependenceGraph& depGraph; // dependence graph for the function
277 hash_set<Instruction*> stmtsVisited; // flags for DFS walk of depGraph
278 hash_map<CallInst*, bool > completed; // flags marking if a CI is done
279 hash_map<CallInst*, DependentsSet> dependents; // dependent CIs for each CI
281 void VisitOutEdges(Instruction* I,
283 DependentsSet& depsOfRoot);
285 FindParallelCalls(const FindParallelCalls &); // DO NOT IMPLEMENT
286 void operator=(const FindParallelCalls&); // DO NOT IMPLEMENT
288 std::vector<CallInst*> parallelCalls;
291 /*ctor*/ FindParallelCalls (Function& F, PgmDependenceGraph& DG);
292 void visitCallInst (CallInst& CI);
296 FindParallelCalls::FindParallelCalls(Function& F,
297 PgmDependenceGraph& DG)
300 // Find all CallInsts reachable from each CallInst using a recursive DFS
303 // Now we've found all CallInsts reachable from each CallInst.
304 // Find those CallInsts that are parallel with at least one other CallInst
305 // by counting total inEdges and outEdges.
307 unsigned long totalNumCalls = completed.size();
309 if (totalNumCalls == 1)
310 { // Check first for the special case of a single call instruction not
311 // in any loop. It is not parallel, even if it has no dependences
312 // (this is why it is a special case).
315 // THIS CASE IS NOT HANDLED RIGHT NOW, I.E., THERE IS NO
316 // PARALLELISM FOR CALLS IN DIFFERENT ITERATIONS OF A LOOP.
321 hash_map<CallInst*, unsigned long> numDeps;
322 for (hash_map<CallInst*, DependentsSet>::iterator II = dependents.begin(),
323 IE = dependents.end(); II != IE; ++II)
325 CallInst* fromCI = II->first;
326 numDeps[fromCI] += II->second.size();
327 for (Dependents_iterator DI = II->second.begin(), DE = II->second.end();
329 numDeps[*DI]++; // *DI can be reached from II->first
332 for (hash_map<CallInst*, DependentsSet>::iterator
333 II = dependents.begin(), IE = dependents.end(); II != IE; ++II)
335 // FIXME: Remove "- 1" when considering parallelism in loops
336 if (numDeps[II->first] < totalNumCalls - 1)
337 parallelCalls.push_back(II->first);
341 void FindParallelCalls::VisitOutEdges(Instruction* I,
343 DependentsSet& depsOfRoot)
345 assert(stmtsVisited.find(I) == stmtsVisited.end() && "Stmt visited twice?");
346 stmtsVisited.insert(I);
348 if (CallInst* CI = dyn_cast<CallInst>(I))
350 // FIXME: Ignoring parallelism in a loop. Here we're actually *ignoring*
351 // a self-dependence in order to get the count comparison right above.
352 // When we include loop parallelism, self-dependences should be included.
356 { // CallInst root has a path to CallInst I and any calls reachable from I
357 depsOfRoot.insert(CI);
359 { // We have already visited I so we know all nodes it can reach!
360 DependentsSet& depsOfI = dependents[CI];
361 depsOfRoot.insert(depsOfI.begin(), depsOfI.end());
366 // If we reach here, we need to visit all children of I
367 for (PgmDependenceGraph::iterator DI = depGraph.outDepBegin(*I);
370 Instruction* sink = &DI->getSink()->getInstr();
371 if (stmtsVisited.find(sink) == stmtsVisited.end())
372 VisitOutEdges(sink, root, depsOfRoot);
377 void FindParallelCalls::visitCallInst(CallInst& CI)
381 stmtsVisited.clear(); // clear flags to do a fresh DFS
383 // Visit all children of CI using a recursive walk through dep graph
384 DependentsSet& depsOfRoot = dependents[&CI];
385 for (PgmDependenceGraph::iterator DI = depGraph.outDepBegin(CI);
388 Instruction* sink = &DI->getSink()->getInstr();
389 if (stmtsVisited.find(sink) == stmtsVisited.end())
390 VisitOutEdges(sink, &CI, depsOfRoot);
393 completed[&CI] = true;
397 //----------------------------------------------------------------------------
400 // (1) Find candidate parallel functions: any function F s.t.
401 // there is a call C1 to the function F that is followed or preceded
402 // by at least one other call C2 that is independent of this one
403 // (i.e., there is no dependence path from C1 to C2 or C2 to C1)
404 // (2) Label such a function F as a cilk function.
405 // (3) Convert every call to F to a spawn
406 // (4) For every function X, insert sync statements so that
407 // every spawn is postdominated by a sync before any statements
408 // with a data dependence to/from the call site for the spawn
410 //----------------------------------------------------------------------------
413 class Parallelize: public Pass
416 /// Driver functions to transform a program
420 /// getAnalysisUsage - Modifies extensively so preserve nothing.
421 /// Uses the DependenceGraph and the Top-down DS Graph (only to find
422 /// all functions called via an indirect call).
424 void getAnalysisUsage(AnalysisUsage &AU) const {
425 AU.addRequired<TDDataStructures>();
426 AU.addRequired<MemoryDepAnalysis>(); // force this not to be released
427 AU.addRequired<PgmDependenceGraph>(); // because it is needed by this
431 RegisterOpt<Parallelize> X("parallel", "Parallelize program using Cilk");
435 static Function* FindMain(Module& M)
437 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI)
438 if (FI->getName() == std::string("main"))
444 bool Parallelize::run(Module& M)
446 hash_set<Function*> parallelFunctions;
447 hash_set<Function*> safeParallelFunctions;
448 hash_set<const GlobalValue*> indirectlyCalled;
450 // If there is no main (i.e., for an incomplete program), we can do nothing.
451 // If there is a main, mark main as a parallel function.
453 Function* mainFunc = FindMain(M);
457 // (1) Find candidate parallel functions and mark them as Cilk functions
459 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI)
460 if (! FI->isExternal())
463 DSGraph& tdg = getAnalysis<TDDataStructures>().getDSGraph(*F);
465 // All the hard analysis work gets done here!
467 FindParallelCalls finder(*F,
468 getAnalysis<PgmDependenceGraph>().getGraph(*F));
469 /* getAnalysis<MemoryDepAnalysis>().getGraph(*F)); */
471 // Now we know which call instructions are useful to parallelize.
472 // Remember those callee functions.
474 for (std::vector<CallInst*>::iterator
475 CII = finder.parallelCalls.begin(),
476 CIE = finder.parallelCalls.end(); CII != CIE; ++CII)
478 // Check if this is a direct call...
479 if ((*CII)->getCalledFunction() != NULL)
480 { // direct call: if this is to a non-external function,
481 // mark it as a parallelizable function
482 if (! (*CII)->getCalledFunction()->isExternal())
483 parallelFunctions.insert((*CII)->getCalledFunction());
486 { // Indirect call: mark all potential callees as bad
487 std::vector<GlobalValue*> callees =
488 tdg.getNodeForValue((*CII)->getCalledValue())
489 .getNode()->getGlobals();
490 indirectlyCalled.insert(callees.begin(), callees.end());
495 // Remove all indirectly called functions from the list of Cilk functions.
497 for (hash_set<Function*>::iterator PFI = parallelFunctions.begin(),
498 PFE = parallelFunctions.end(); PFI != PFE; ++PFI)
499 if (indirectlyCalled.count(*PFI) == 0)
500 safeParallelFunctions.insert(*PFI);
502 #undef CAN_USE_BIND1ST_ON_REFERENCE_TYPE_ARGS
503 #ifdef CAN_USE_BIND1ST_ON_REFERENCE_TYPE_ARGS
504 // Use this indecipherable STLese because erase invalidates iterators.
505 // Otherwise we have to copy sets as above.
506 hash_set<Function*>::iterator extrasBegin =
507 std::remove_if(parallelFunctions.begin(), parallelFunctions.end(),
508 compose1(std::bind2nd(std::greater<int>(), 0),
509 bind_obj(&indirectlyCalled,
510 &hash_set<const GlobalValue*>::count)));
511 parallelFunctions.erase(extrasBegin, parallelFunctions.end());
514 // If there are no parallel functions, we can just give up.
515 if (safeParallelFunctions.empty())
518 // Add main as a parallel function since Cilk requires this.
519 safeParallelFunctions.insert(mainFunc);
521 // (2,3) Transform each Cilk function and all its calls simply by
522 // adding a unique suffix to the function name.
523 // This should identify both functions and calls to such functions
524 // to the code generator.
525 // (4) Also, insert calls to sync at appropriate points.
527 Cilkifier cilkifier(M);
528 for (hash_set<Function*>::iterator CFI = safeParallelFunctions.begin(),
529 CFE = safeParallelFunctions.end(); CFI != CFE; ++CFI)
531 cilkifier.TransformFunc(*CFI, safeParallelFunctions,
532 getAnalysis<PgmDependenceGraph>().getGraph(**CFI));
533 /* getAnalysis<MemoryDepAnalysis>().getGraph(**CFI)); */