Commit ecc7e3a3 authored by Patrick Flynn's avatar Patrick Flynn Committed by Patrick McCormick
Browse files

Various fixes against race conditions when extra blocks are generated in the... 

Various fixes against race conditions when extra blocks are generated in the loop body (mainly for Kokkos::View's)
parent b1361a68

clang/lib/CodeGen/CGKokkos.cpp

+21 −27
Original line number Diff line number Diff line
@@ -580,8 +580,15 @@ bool CodeGenFunction::EmitKokkosInnerLoop(const CallExpr *CE, const LambdaExpr * 
      // At this point in the codegen, the body block has been emitted 

      // and we can safely replace the ''sequential`` induction variable 

      // within the detach basic block.

      //

      // When Kokkos::Views (and likely some other structures) are used, we end up with extra blocks

      // between the original for loop body block and the reattach block. Without the loop, it will

      // only modify the last of these block. The loop iterates and updates all the blocks back to the

      // original for loop body block to use the thread-local induction variables

      //

      llvm::BasicBlock *CurrentBlock = Builder.GetInsertBlock();

      for (int i = 0; i<TLIVarList.size(); i++) {

      for (;;) {

        for (unsigned int i = 0; i<TLIVarList.size(); i++) {

          auto TLVar = TLIVarList.at(i).second;

          auto GInductionVar = TLIVarList.at(i).first;
@@ -593,6 +600,10 @@ bool CodeGenFunction::EmitKokkosInnerLoop(const CallExpr *CE, const LambdaExpr * 
              U.set(TLVar);

          }

        }

        

        if (CurrentBlock == PForBody) break;

        else CurrentBlock = CurrentBlock->getSinglePredecessor();

      }

    } else {

      EmitKokkosInnerLoop(CE, Lambda, ConditionBlock, DimQueue, StartQueue, params, TLIVarList, ForallAttrs);

    }
@@ -602,19 +613,6 @@ bool CodeGenFunction::EmitKokkosInnerLoop(const CallExpr *CE, const LambdaExpr * 
  AllocaInsertPt = OldAllocaInsertPt; 

  tmp->removeFromParent(); 



  // Modify the body to use the ''detach''-local induction variable.

  // At this point in the codegen, the body block has been emitted

  // and we can safely replace the ''sequential`` induction variable

  // within the detach basic block.

 /* llvm::BasicBlock *CurrentBlock = Builder.GetInsertBlock();

  for(llvm::Value::use_iterator UI = GInductionVar->use_begin(), UE = GInductionVar->use_end(); 

      UI != UE; ) {

    llvm::Use &U = *UI++;

    llvm::Instruction *I = cast<llvm::Instruction>(U.getUser());

    if (I->getParent() == CurrentBlock)

      U.set(TLInductionVar);

  }*/



  EmitBlock(Reattach.getBlock());

  Builder.CreateReattach(Increment, SRStart);
@@ -632,10 +630,6 @@ bool CodeGenFunction::EmitKokkosInnerLoop(const CallExpr *CE, const LambdaExpr * 
  PForScope.ForceCleanup();

  LoopStack.pop();



  /*if (TopBlock != nullptr) {

    EmitBranch(TopBlock);

  }*/



  EmitBlock(Sync.getBlock());

  Builder.CreateSync(End, SRStart);

  EmitBlock(End, true);