[TSAR, Parallel, Shared, DVMH] Collapse parallel loops for the GPU target.

Author: kaniandr

lib/Transform/Clang/DVMHSMAutoPar.cpp

@@ -28,6 +28,7 @@
 #include "tsar/Analysis/Clang/ASTDependenceAnalysis.h"
 #include "tsar/Analysis/Clang/CanonicalLoop.h"
 #include "tsar/Analysis/Clang/LoopMatcher.h"
+#include "tsar/Analysis/Clang/PerfectLoop.h"
 #include "tsar/Analysis/Passes.h"
 #include "tsar/Analysis/Parallel/Passes.h"
 #include "tsar/Analysis/Parallel/ParallelLoop.h"
@@ -53,7 +54,7 @@ class ClangDVMHSMParallelization : public ClangSMParallelization {
     initializeClangDVMHSMParallelizationPass(*PassRegistry::getPassRegistry());
   }
 private:
-  bool exploitParallelism(const Loop &IR, const clang::ForStmt &AST,
+  bool exploitParallelism(const DFLoop &IR, const clang::ForStmt &AST,
     const ClangSMParallelProvider &Provider,
     tsar::ClangDependenceAnalyzer &ASTDepInfo,
     TransformationContext &TfmCtx) override;
@@ -108,28 +109,39 @@ void addVarList(const ClangDependenceAnalyzer::ReductionVarListT &VarInfoList,
     ParallelFor.push_back(')');
   }
 }
+
+unsigned getPerfectNestSize(const DFLoop &DFL,
+    const PerfectLoopInfo &PerfectInfo,
+    const CanonicalLoopSet &CanonicalLoops) {
+  auto *CurrDFL = &DFL;
+  unsigned PerfectSize = 1;
+  for (; PerfectInfo.count(CurrDFL) && CurrDFL->getNumRegions() > 0;
+       ++PerfectSize) {
+    CurrDFL = dyn_cast<DFLoop>(*CurrDFL->region_begin());
+    if (!CurrDFL)
+      return PerfectSize;
+    auto CanonicalItr = CanonicalLoops.find_as(const_cast<DFLoop *>(CurrDFL));
+    if (CanonicalItr == CanonicalLoops.end() || !(**CanonicalItr).isCanonical())
+      return PerfectSize;
+  }
+  return PerfectSize;
+}
 } // namespace
 
 bool ClangDVMHSMParallelization::exploitParallelism(
-    const Loop &IR, const clang::ForStmt &AST,
+    const DFLoop &IR, const clang::ForStmt &AST,
     const ClangSMParallelProvider &Provider,
     tsar::ClangDependenceAnalyzer &ASTRegionAnalysis,
     TransformationContext &TfmCtx) {
   auto &ASTDepInfo = ASTRegionAnalysis.getDependenceInfo();
   if (!ASTDepInfo.get<trait::FirstPrivate>().empty() ||
       !ASTDepInfo.get<trait::LastPrivate>().empty())
     return false;
-  SmallString<128> ParallelFor("#pragma dvm parallel (1)");
-  if (!ASTDepInfo.get<trait::Private>().empty()) {
-    ParallelFor += " private";
-    addVarList(ASTDepInfo.get<trait::Private>(), ParallelFor);
-  }
-  addVarList(ASTDepInfo.get<trait::Reduction>(), ParallelFor);
-  ParallelFor += '\n';
   SmallString<128> DVMHRegion("#pragma dvm region");
   SmallString<128> DVMHActual, DVMHGetActual;
   auto &PI = Provider.get<ParallelLoopPass>().getParallelLoopInfo();
-  if (!PI[&IR].isHostOnly() && ASTRegionAnalysis.evaluateDefUse()) {
+  bool HostOnly = false;
+  if (!PI[IR.getLoop()].isHostOnly() && ASTRegionAnalysis.evaluateDefUse()) {
     if (!ASTDepInfo.get<trait::ReadOccurred>().empty()) {
       DVMHActual += "#pragma dvm actual";
       addVarList(ASTDepInfo.get<trait::ReadOccurred>(), DVMHActual);
@@ -150,7 +162,23 @@ bool ClangDVMHSMParallelization::exploitParallelism(
     }
   } else {
     DVMHRegion += " targets(HOST)";
+    HostOnly = true;
   }
+  auto &PerfectInfo = Provider.get<ClangPerfectLoopPass>().getPerfectLoopInfo();
+  auto &CanonicalInfo = Provider.get<CanonicalLoopPass>().getCanonicalLoopInfo();
+  SmallString<128> ParallelFor("#pragma dvm parallel (");
+  if (HostOnly)
+    ParallelFor += "1";
+  else
+    Twine(getPerfectNestSize(IR, PerfectInfo, CanonicalInfo))
+        .toStringRef(ParallelFor);
+  ParallelFor += ")";
+  if (!ASTDepInfo.get<trait::Private>().empty()) {
+    ParallelFor += " private";
+    addVarList(ASTDepInfo.get<trait::Private>(), ParallelFor);
+  }
+  addVarList(ASTDepInfo.get<trait::Reduction>(), ParallelFor);
+  ParallelFor += '\n';
   DVMHRegion += "\n{\n";
   // Add directives to the source code.
   auto &Rewriter = TfmCtx.getRewriter();

lib/Transform/Clang/OpenMPAutoPar.cpp

@@ -46,7 +46,7 @@ class ClangOpenMPParallelization : public ClangSMParallelization {
     initializeClangOpenMPParallelizationPass(*PassRegistry::getPassRegistry());
   }
 private:
-  bool exploitParallelism(const Loop &IR, const clang::ForStmt &AST,
+  bool exploitParallelism(const DFLoop &IR, const clang::ForStmt &AST,
     const ClangSMParallelProvider &Provider,
     tsar::ClangDependenceAnalyzer &ASTDepInfo,
     TransformationContext &TfmCtx) override;
@@ -107,7 +107,7 @@ struct ClausePrinter {
 
 
 bool ClangOpenMPParallelization::exploitParallelism(
-    const Loop &IR, const clang::ForStmt &AST,
+    const DFLoop &IR, const clang::ForStmt &AST,
     const ClangSMParallelProvider &Provider,
     tsar::ClangDependenceAnalyzer &ASTDepInfo,
     TransformationContext &TfmCtx) {

lib/Transform/Clang/SharedMemoryAutoPar.cpp

@@ -30,6 +30,7 @@
 #include "tsar/Analysis/Clang/DIMemoryMatcher.h"
 #include "tsar/Analysis/Clang/LoopMatcher.h"
 #include "tsar/Analysis/Clang/MemoryMatcher.h"
+#include "tsar/Analysis/Clang/PerfectLoop.h"
 #include "tsar/Analysis/Clang/RegionDirectiveInfo.h"
 #include "tsar/Analysis/DFRegionInfo.h"
 #include "tsar/Analysis/Memory/ClonedDIMemoryMatcher.h"
@@ -99,7 +100,8 @@ bool ClangSMParallelization::findParallelLoops(
   if (LMatchItr != LM.end())
     toDiag(Diags, LMatchItr->get<AST>()->getLocStart(),
            clang::diag::remark_parallel_loop);
-  auto CanonicalItr = CL.find_as(RI.getRegionFor(&L));
+  auto DFL = cast<DFLoop>(RI.getRegionFor(&L));
+  auto CanonicalItr = CL.find_as(DFL);
   if (CanonicalItr == CL.end() || !(**CanonicalItr).isCanonical()) {
     toDiag(Diags, LMatchItr->get<AST>()->getLocStart(),
            clang::diag::warn_parallel_not_canonical);
@@ -129,7 +131,7 @@ bool ClangSMParallelization::findParallelLoops(
     *mGlobalOpts, Diags, DIAT, DIDepSet, *DIMemoryMatcher, ASTToClient);
   if (!RegionAnalysis.evaluateDependency())
     return findParallelLoops(L.begin(), L.end(), F, Provider);
-  if (!exploitParallelism(L, *ForStmt, Provider, RegionAnalysis, *mTfmCtx))
+  if (!exploitParallelism(*DFL, *ForStmt, Provider, RegionAnalysis, *mTfmCtx))
     return findParallelLoops(L.begin(), L.end(), F, Provider);
   for (auto *BB : L.blocks())
     for (auto &I : *BB) {

lib/Transform/Clang/SharedMemoryAutoPar.h

@@ -41,6 +41,7 @@ class ForStmt;
 }
 
 namespace tsar {
+class DFLoop;
 class AnalysisSocketInfo;
 class ClangDependenceAnalyzer;
 class DIMemoryEnvironment;
@@ -56,6 +57,7 @@ class Function;
 class Loop;
 
 class CanonicalLoopPass;
+class ClangPerfectLoopPass;
 class ClangDIMemoryMatcherPass;
 class DFRegionInfoPass;
 class LoopMatcherPass;
@@ -66,7 +68,8 @@ class ParallelLoopPass;
 using ClangSMParallelProvider =
     FunctionPassAAProvider<AnalysisSocketImmutableWrapper, LoopInfoWrapperPass,
                            ParallelLoopPass, CanonicalLoopPass, LoopMatcherPass,
-                           DFRegionInfoPass, ClangDIMemoryMatcherPass>;
+                           DFRegionInfoPass, ClangDIMemoryMatcherPass,
+                           ClangPerfectLoopPass>;
 
 /// This pass try to insert directives into a source code to obtain
 /// a parallel program for a shared memory.
@@ -112,8 +115,8 @@ class ClangSMParallelization: public ModulePass, private bcl::Uncopyable{
   /// successfully checked.
   /// \return true if a specified loop could be parallelized and inner loops
   /// should not be processed.
-  virtual bool exploitParallelism(const Loop &IR, const clang::ForStmt &AST,
-    const ClangSMParallelProvider &Provider,
+  virtual bool exploitParallelism(const tsar::DFLoop &IR,
+    const clang::ForStmt &AST, const ClangSMParallelProvider &Provider,
     tsar::ClangDependenceAnalyzer &ASTDepInfo,
     tsar::TransformationContext &TfmCtx) = 0;