[llvm] 1886aad - [SimplifyCFG] Common code sinking: relax restriction on non-uncond predecessors

[Nikita Popov via llvm-commits]

On Thu, Apr 29, 2021 at 12:01 AM Roman Lebedev via llvm-commits <
llvm-commits at lists.llvm.org> wrote:

>
> Author: Roman Lebedev
> Date: 2021-04-29T01:01:01+03:00
> New Revision: 1886aad9d03b95c35260d6d8013d746bd39dc94a
>
> URL:
> https://github.com/llvm/llvm-project/commit/1886aad9d03b95c35260d6d8013d746bd39dc94a
> DIFF:
> https://github.com/llvm/llvm-project/commit/1886aad9d03b95c35260d6d8013d746bd39dc94a.diff
>
> LOG: [SimplifyCFG] Common code sinking: relax restriction on non-uncond
> predecessors
>
> While we have a known profitability issue for sinking in presence of
> non-unconditional predecessors, there isn't any known issues
> for having multiple such non-unconditional predecessors,
> so said restriction appears to be artificial. Lift it.
>
> Added:
>
>
> Modified:
>     llvm/lib/Transforms/Utils/SimplifyCFG.cpp
>     llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
>
> Removed:
>
>
>
>
> ################################################################################
> diff  --git a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
> b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
> index 77c0890fa91e..0704c5f2321a 100644
> --- a/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
> +++ b/llvm/lib/Transforms/Utils/SimplifyCFG.cpp
> @@ -1988,15 +1988,13 @@ static bool
> SinkCommonCodeFromPredecessors(BasicBlock *BB,
>    //         [ end ]
>    //
>    SmallVector<BasicBlock*,4> UnconditionalPreds;
> -  Instruction *Cond = nullptr;
> -  for (auto *B : predecessors(BB)) {
> -    auto *T = B->getTerminator();
> -    if (isa<BranchInst>(T) && cast<BranchInst>(T)->isUnconditional())
> -      UnconditionalPreds.push_back(B);
> -    else if ((isa<BranchInst>(T) || isa<SwitchInst>(T)) && !Cond)
> -      Cond = T;
> +  bool AllPredsAreUnconditional = false;
> +  for (auto *PredBB : predecessors(BB)) {
> +    auto *PredBr = dyn_cast<BranchInst>(PredBB->getTerminator());
> +    if (PredBr && PredBr->isUnconditional())
> +      UnconditionalPreds.push_back(PredBB);
>      else
> -      return false;
> +      AllPredsAreUnconditional = true;
>

Is the name of this variable inverted? You're setting it to true when there
is a conditional branch.

Nikita


>    }
>    if (UnconditionalPreds.size() < 2)
>      return false;
> @@ -2063,7 +2061,7 @@ static bool
> SinkCommonCodeFromPredecessors(BasicBlock *BB,
>
>    bool Changed = false;
>
> -  if (Cond) {
> +  if (AllPredsAreUnconditional) {
>      // It is always legal to sink common instructions from unconditional
>      // predecessors. However, if not all predecessors are unconditional,
>      // this transformation might be pessimizing. So as a rule of thumb,
>
> diff  --git a/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
> b/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
> index ade36c7019d1..a85ad8e69a20 100644
> --- a/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
> +++ b/llvm/test/Transforms/SimplifyCFG/X86/sink-common-code.ll
> @@ -1538,14 +1538,11 @@ define void @multiple_cond_preds(i1 %c0, i1 %c1,
> i1 %c2) {
>  ; CHECK-NEXT:    br i1 [[C0:%.*]], label [[DISPATCH1:%.*]], label
> [[DISPATCH2:%.*]]
>  ; CHECK:       dispatch1:
>  ; CHECK-NEXT:    call void @direct_callee2()
> -; CHECK-NEXT:    br i1 [[C1:%.*]], label [[UNCOND_PRED0:%.*]], label
> [[END:%.*]]
> +; CHECK-NEXT:    br i1 [[C1:%.*]], label [[END_SINK_SPLIT:%.*]], label
> [[END:%.*]]
>  ; CHECK:       dispatch2:
>  ; CHECK-NEXT:    call void @direct_callee3()
> -; CHECK-NEXT:    br i1 [[C2:%.*]], label [[UNCOND_PRED1:%.*]], label
> [[END]]
> -; CHECK:       uncond_pred0:
> -; CHECK-NEXT:    call void @direct_callee()
> -; CHECK-NEXT:    br label [[END]]
> -; CHECK:       uncond_pred1:
> +; CHECK-NEXT:    br i1 [[C2:%.*]], label [[END_SINK_SPLIT]], label [[END]]
> +; CHECK:       end.sink.split:
>  ; CHECK-NEXT:    call void @direct_callee()
>  ; CHECK-NEXT:    br label [[END]]
>  ; CHECK:       end:
>
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210429/84b638c2/attachment.html>