1 //===- Parallelize.cpp - Auto parallelization using DS Graphs -------------===//
3 // This file implements a pass that automatically parallelizes a program,
4 // using the Cilk multi-threaded runtime system to execute parallel code.
6 // The pass uses the Program Dependence Graph (class PDGIterator) to
7 // identify parallelizable function calls, i.e., calls whose instances
8 // can be executed in parallel with instances of other function calls.
9 // (In the future, this should also execute different instances of the same
10 // function call in parallel, but that requires parallelizing across
13 // The output of the pass is LLVM code with:
14 // (1) all parallelizable functions renamed to flag them as parallelizable;
15 // (2) calls to a sync() function introduced at synchronization points.
16 // The CWriter recognizes these functions and inserts the appropriate Cilk
17 // keywords when writing out C code. This C code must be compiled with cilk2c.
19 // Current algorithmic limitations:
20 // -- no array dependence analysis
21 // -- no parallelization for function calls in different loop iterations
22 // (except in unlikely trivial cases)
24 // Limitations of using Cilk:
25 // -- No parallelism within a function body, e.g., in a loop;
26 // -- Simplistic synchronization model requiring all parallel threads
27 // created within a function to block at a sync().
28 // -- Excessive overhead at "spawned" function calls, which has no benefit
29 // once all threads are busy (especially common when the degree of
30 // parallelism is low).
32 //===----------------------------------------------------------------------===//
34 #include "llvm/Transforms/Utils/DemoteRegToStack.h"
35 #include "llvm/Analysis/PgmDependenceGraph.h"
36 #include "llvm/Analysis/DataStructure.h"
37 #include "llvm/Analysis/DSGraph.h"
38 #include "llvm/Module.h"
39 #include "llvm/Instructions.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/Support/InstVisitor.h"
42 #include "Support/Statistic.h"
43 #include "Support/STLExtras.h"
44 #include "Support/hash_set"
45 #include "Support/hash_map"
49 //----------------------------------------------------------------------------
50 // Global constants used in marking Cilk functions and function calls.
51 //----------------------------------------------------------------------------
53 static const char * const CilkSuffix = ".llvm2cilk";
54 static const char * const DummySyncFuncName = "__sync.llvm2cilk";
56 //----------------------------------------------------------------------------
57 // Routines to identify Cilk functions, calls to Cilk functions, and syncs.
58 //----------------------------------------------------------------------------
60 static bool isCilk(const Function& F) {
61 return (F.getName().rfind(CilkSuffix) ==
62 F.getName().size() - std::strlen(CilkSuffix));
65 static bool isCilkMain(const Function& F) {
66 return F.getName() == "main" + std::string(CilkSuffix);
70 static bool isCilk(const CallInst& CI) {
71 return CI.getCalledFunction() && isCilk(*CI.getCalledFunction());
74 static bool isSync(const CallInst& CI) {
75 return CI.getCalledFunction() &&
76 CI.getCalledFunction()->getName() == DummySyncFuncName;
80 //----------------------------------------------------------------------------
83 // Code generation pass that transforms code to identify where Cilk keywords
84 // should be inserted. This relies on `llvm-dis -c' to print out the keywords.
85 //----------------------------------------------------------------------------
88 class Cilkifier: public InstVisitor<Cilkifier>
90 Function* DummySyncFunc;
92 // Data used when transforming each function.
93 hash_set<const Instruction*> stmtsVisited; // Flags for recursive DFS
94 hash_map<const CallInst*, hash_set<CallInst*> > spawnToSyncsMap;
96 // Input data for the transformation.
97 const hash_set<Function*>* cilkFunctions; // Set of parallel functions
98 PgmDependenceGraph* depGraph;
100 void DFSVisitInstr (Instruction* I,
102 hash_set<const Instruction*>& depsOfRoot);
105 /*ctor*/ Cilkifier (Module& M);
107 // Transform a single function including its name, its call sites, and syncs
109 void TransformFunc (Function* F,
110 const hash_set<Function*>& cilkFunctions,
111 PgmDependenceGraph& _depGraph);
113 // The visitor function that does most of the hard work, via DFSVisitInstr
115 void visitCallInst(CallInst& CI);
119 Cilkifier::Cilkifier(Module& M)
121 // create the dummy Sync function and add it to the Module
122 DummySyncFunc = M.getOrInsertFunction(DummySyncFuncName, Type::VoidTy, 0);
125 void Cilkifier::TransformFunc(Function* F,
126 const hash_set<Function*>& _cilkFunctions,
127 PgmDependenceGraph& _depGraph)
129 // Memoize the information for this function
130 cilkFunctions = &_cilkFunctions;
131 depGraph = &_depGraph;
133 // Add the marker suffix to the Function name
134 // This should automatically mark all calls to the function also!
135 F->setName(F->getName() + CilkSuffix);
137 // Insert sync operations for each separate spawn
140 // Now traverse the CFG in rPostorder and eliminate redundant syncs, i.e.,
141 // two consecutive sync's on a straight-line path with no intervening spawn.
146 void Cilkifier::DFSVisitInstr(Instruction* I,
148 hash_set<const Instruction*>& depsOfRoot)
150 assert(stmtsVisited.find(I) == stmtsVisited.end());
151 stmtsVisited.insert(I);
153 // If there is a dependence from root to I, insert Sync and return
154 if (depsOfRoot.find(I) != depsOfRoot.end())
155 { // Insert a sync before I and stop searching along this path.
156 // If I is a Phi instruction, the dependence can only be an SSA dep.
157 // and we need to insert the sync in the predecessor on the appropriate
160 if (PHINode* phiI = dyn_cast<PHINode>(I))
161 { // check all operands of the Phi and insert before each one
162 for (unsigned i = 0, N = phiI->getNumIncomingValues(); i < N; ++i)
163 if (phiI->getIncomingValue(i) == root)
164 syncI = new CallInst(DummySyncFunc, std::vector<Value*>(), "",
165 phiI->getIncomingBlock(i)->getTerminator());
168 syncI = new CallInst(DummySyncFunc, std::vector<Value*>(), "", I);
170 // Remember the sync for each spawn to eliminate redundant ones later
171 spawnToSyncsMap[cast<CallInst>(root)].insert(syncI);
176 // else visit unvisited successors
177 if (BranchInst* brI = dyn_cast<BranchInst>(I))
178 { // visit first instruction in each successor BB
179 for (unsigned i = 0, N = brI->getNumSuccessors(); i < N; ++i)
180 if (stmtsVisited.find(&brI->getSuccessor(i)->front())
181 == stmtsVisited.end())
182 DFSVisitInstr(&brI->getSuccessor(i)->front(), root, depsOfRoot);
185 if (Instruction* nextI = I->getNext())
186 if (stmtsVisited.find(nextI) == stmtsVisited.end())
187 DFSVisitInstr(nextI, root, depsOfRoot);
191 void Cilkifier::visitCallInst(CallInst& CI)
193 assert(CI.getCalledFunction() != 0 && "Only direct calls can be spawned.");
194 if (cilkFunctions->find(CI.getCalledFunction()) == cilkFunctions->end())
195 return; // not a spawn
197 // Find all the outgoing memory dependences.
198 hash_set<const Instruction*> depsOfRoot;
199 for (PgmDependenceGraph::iterator DI =
200 depGraph->outDepBegin(CI, MemoryDeps); ! DI.fini(); ++DI)
201 depsOfRoot.insert(&DI->getSink()->getInstr());
203 // Now find all outgoing SSA dependences to the eventual non-Phi users of
204 // the call value (i.e., direct users that are not phis, and for any
205 // user that is a Phi, direct non-Phi users of that Phi, and recursively).
206 std::vector<const PHINode*> phiUsers;
207 hash_set<const PHINode*> phisSeen; // ensures we don't visit a phi twice
208 for (Value::use_iterator UI=CI.use_begin(), UE=CI.use_end(); UI != UE; ++UI)
209 if (const PHINode* phiUser = dyn_cast<PHINode>(*UI))
211 if (phisSeen.find(phiUser) == phisSeen.end())
213 phiUsers.push_back(phiUser);
214 phisSeen.insert(phiUser);
218 depsOfRoot.insert(cast<Instruction>(*UI));
220 // Now we've found the non-Phi users and immediate phi users.
221 // Recursively walk the phi users and add their non-phi users.
222 for (const PHINode* phiUser; !phiUsers.empty(); phiUsers.pop_back())
224 phiUser = phiUsers.back();
225 for (Value::use_const_iterator UI=phiUser->use_begin(),
226 UE=phiUser->use_end(); UI != UE; ++UI)
227 if (const PHINode* pn = dyn_cast<PHINode>(*UI))
229 if (phisSeen.find(pn) == phisSeen.end())
231 phiUsers.push_back(pn);
236 depsOfRoot.insert(cast<Instruction>(*UI));
239 // Walk paths of the CFG starting at the call instruction and insert
240 // one sync before the first dependence on each path, if any.
241 if (! depsOfRoot.empty())
243 stmtsVisited.clear(); // start a new DFS for this CallInst
244 assert(CI.getNext() && "Call instruction cannot be a terminator!");
245 DFSVisitInstr(CI.getNext(), &CI, depsOfRoot);
248 // Now, eliminate all users of the SSA value of the CallInst, i.e.,
249 // if the call instruction returns a value, delete the return value
250 // register and replace it by a stack slot.
251 if (CI.getType() != Type::VoidTy)
252 DemoteRegToStack(CI);
256 //----------------------------------------------------------------------------
257 // class FindParallelCalls
259 // Find all CallInst instructions that have at least one other CallInst
260 // that is independent. These are the instructions that can produce
261 // useful parallelism.
262 //----------------------------------------------------------------------------
264 class FindParallelCalls : public InstVisitor<FindParallelCalls> {
265 typedef hash_set<CallInst*> DependentsSet;
266 typedef DependentsSet::iterator Dependents_iterator;
267 typedef DependentsSet::const_iterator Dependents_const_iterator;
269 PgmDependenceGraph& depGraph; // dependence graph for the function
270 hash_set<Instruction*> stmtsVisited; // flags for DFS walk of depGraph
271 hash_map<CallInst*, bool > completed; // flags marking if a CI is done
272 hash_map<CallInst*, DependentsSet> dependents; // dependent CIs for each CI
274 void VisitOutEdges(Instruction* I,
276 DependentsSet& depsOfRoot);
278 FindParallelCalls(const FindParallelCalls &); // DO NOT IMPLEMENT
279 void operator=(const FindParallelCalls&); // DO NOT IMPLEMENT
281 std::vector<CallInst*> parallelCalls;
284 /*ctor*/ FindParallelCalls (Function& F, PgmDependenceGraph& DG);
285 void visitCallInst (CallInst& CI);
289 FindParallelCalls::FindParallelCalls(Function& F,
290 PgmDependenceGraph& DG)
293 // Find all CallInsts reachable from each CallInst using a recursive DFS
296 // Now we've found all CallInsts reachable from each CallInst.
297 // Find those CallInsts that are parallel with at least one other CallInst
298 // by counting total inEdges and outEdges.
300 unsigned long totalNumCalls = completed.size();
302 if (totalNumCalls == 1)
303 { // Check first for the special case of a single call instruction not
304 // in any loop. It is not parallel, even if it has no dependences
305 // (this is why it is a special case).
308 // THIS CASE IS NOT HANDLED RIGHT NOW, I.E., THERE IS NO
309 // PARALLELISM FOR CALLS IN DIFFERENT ITERATIONS OF A LOOP.
314 hash_map<CallInst*, unsigned long> numDeps;
315 for (hash_map<CallInst*, DependentsSet>::iterator II = dependents.begin(),
316 IE = dependents.end(); II != IE; ++II)
318 CallInst* fromCI = II->first;
319 numDeps[fromCI] += II->second.size();
320 for (Dependents_iterator DI = II->second.begin(), DE = II->second.end();
322 numDeps[*DI]++; // *DI can be reached from II->first
325 for (hash_map<CallInst*, DependentsSet>::iterator
326 II = dependents.begin(), IE = dependents.end(); II != IE; ++II)
328 // FIXME: Remove "- 1" when considering parallelism in loops
329 if (numDeps[II->first] < totalNumCalls - 1)
330 parallelCalls.push_back(II->first);
334 void FindParallelCalls::VisitOutEdges(Instruction* I,
336 DependentsSet& depsOfRoot)
338 assert(stmtsVisited.find(I) == stmtsVisited.end() && "Stmt visited twice?");
339 stmtsVisited.insert(I);
341 if (CallInst* CI = dyn_cast<CallInst>(I))
343 // FIXME: Ignoring parallelism in a loop. Here we're actually *ignoring*
344 // a self-dependence in order to get the count comparison right above.
345 // When we include loop parallelism, self-dependences should be included.
349 { // CallInst root has a path to CallInst I and any calls reachable from I
350 depsOfRoot.insert(CI);
352 { // We have already visited I so we know all nodes it can reach!
353 DependentsSet& depsOfI = dependents[CI];
354 depsOfRoot.insert(depsOfI.begin(), depsOfI.end());
359 // If we reach here, we need to visit all children of I
360 for (PgmDependenceGraph::iterator DI = depGraph.outDepBegin(*I);
363 Instruction* sink = &DI->getSink()->getInstr();
364 if (stmtsVisited.find(sink) == stmtsVisited.end())
365 VisitOutEdges(sink, root, depsOfRoot);
370 void FindParallelCalls::visitCallInst(CallInst& CI)
374 stmtsVisited.clear(); // clear flags to do a fresh DFS
376 // Visit all children of CI using a recursive walk through dep graph
377 DependentsSet& depsOfRoot = dependents[&CI];
378 for (PgmDependenceGraph::iterator DI = depGraph.outDepBegin(CI);
381 Instruction* sink = &DI->getSink()->getInstr();
382 if (stmtsVisited.find(sink) == stmtsVisited.end())
383 VisitOutEdges(sink, &CI, depsOfRoot);
386 completed[&CI] = true;
390 //----------------------------------------------------------------------------
393 // (1) Find candidate parallel functions: any function F s.t.
394 // there is a call C1 to the function F that is followed or preceded
395 // by at least one other call C2 that is independent of this one
396 // (i.e., there is no dependence path from C1 to C2 or C2 to C1)
397 // (2) Label such a function F as a cilk function.
398 // (3) Convert every call to F to a spawn
399 // (4) For every function X, insert sync statements so that
400 // every spawn is postdominated by a sync before any statements
401 // with a data dependence to/from the call site for the spawn
403 //----------------------------------------------------------------------------
406 class Parallelize: public Pass
409 /// Driver functions to transform a program
413 /// getAnalysisUsage - Modifies extensively so preserve nothing.
414 /// Uses the DependenceGraph and the Top-down DS Graph (only to find
415 /// all functions called via an indirect call).
417 void getAnalysisUsage(AnalysisUsage &AU) const {
418 AU.addRequired<TDDataStructures>();
419 AU.addRequired<MemoryDepAnalysis>(); // force this not to be released
420 AU.addRequired<PgmDependenceGraph>(); // because it is needed by this
424 RegisterOpt<Parallelize> X("parallel", "Parallelize program using Cilk");
428 static Function* FindMain(Module& M)
430 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI)
431 if (FI->getName() == std::string("main"))
437 bool Parallelize::run(Module& M)
439 hash_set<Function*> parallelFunctions;
440 hash_set<Function*> safeParallelFunctions;
441 hash_set<const GlobalValue*> indirectlyCalled;
443 // If there is no main (i.e., for an incomplete program), we can do nothing.
444 // If there is a main, mark main as a parallel function.
446 Function* mainFunc = FindMain(M);
450 // (1) Find candidate parallel functions and mark them as Cilk functions
452 for (Module::iterator FI = M.begin(), FE = M.end(); FI != FE; ++FI)
453 if (! FI->isExternal())
456 DSGraph& tdg = getAnalysis<TDDataStructures>().getDSGraph(*F);
458 // All the hard analysis work gets done here!
460 FindParallelCalls finder(*F,
461 getAnalysis<PgmDependenceGraph>().getGraph(*F));
462 /* getAnalysis<MemoryDepAnalysis>().getGraph(*F)); */
464 // Now we know which call instructions are useful to parallelize.
465 // Remember those callee functions.
467 for (std::vector<CallInst*>::iterator
468 CII = finder.parallelCalls.begin(),
469 CIE = finder.parallelCalls.end(); CII != CIE; ++CII)
471 // Check if this is a direct call...
472 if ((*CII)->getCalledFunction() != NULL)
473 { // direct call: if this is to a non-external function,
474 // mark it as a parallelizable function
475 if (! (*CII)->getCalledFunction()->isExternal())
476 parallelFunctions.insert((*CII)->getCalledFunction());
479 { // Indirect call: mark all potential callees as bad
480 std::vector<GlobalValue*> callees =
481 tdg.getNodeForValue((*CII)->getCalledValue())
482 .getNode()->getGlobals();
483 indirectlyCalled.insert(callees.begin(), callees.end());
488 // Remove all indirectly called functions from the list of Cilk functions.
490 for (hash_set<Function*>::iterator PFI = parallelFunctions.begin(),
491 PFE = parallelFunctions.end(); PFI != PFE; ++PFI)
492 if (indirectlyCalled.count(*PFI) == 0)
493 safeParallelFunctions.insert(*PFI);
495 #undef CAN_USE_BIND1ST_ON_REFERENCE_TYPE_ARGS
496 #ifdef CAN_USE_BIND1ST_ON_REFERENCE_TYPE_ARGS
497 // Use this indecipherable STLese because erase invalidates iterators.
498 // Otherwise we have to copy sets as above.
499 hash_set<Function*>::iterator extrasBegin =
500 std::remove_if(parallelFunctions.begin(), parallelFunctions.end(),
501 compose1(std::bind2nd(std::greater<int>(), 0),
502 bind_obj(&indirectlyCalled,
503 &hash_set<const GlobalValue*>::count)));
504 parallelFunctions.erase(extrasBegin, parallelFunctions.end());
507 // If there are no parallel functions, we can just give up.
508 if (safeParallelFunctions.empty())
511 // Add main as a parallel function since Cilk requires this.
512 safeParallelFunctions.insert(mainFunc);
514 // (2,3) Transform each Cilk function and all its calls simply by
515 // adding a unique suffix to the function name.
516 // This should identify both functions and calls to such functions
517 // to the code generator.
518 // (4) Also, insert calls to sync at appropriate points.
520 Cilkifier cilkifier(M);
521 for (hash_set<Function*>::iterator CFI = safeParallelFunctions.begin(),
522 CFE = safeParallelFunctions.end(); CFI != CFE; ++CFI)
524 cilkifier.TransformFunc(*CFI, safeParallelFunctions,
525 getAnalysis<PgmDependenceGraph>().getGraph(**CFI));
526 /* getAnalysis<MemoryDepAnalysis>().getGraph(**CFI)); */