[llvm] r321402 - [SCCP] Manually fold branches on undef.
Davide Italiano via llvm-commits
llvm-commits at lists.llvm.org
Sun Jan 7 14:15:30 PST 2018
Sorry for the delay, reverted in r321975 while I investigate the root cause.
On Tue, Jan 2, 2018 at 2:39 PM, Davide Italiano <davide at freebsd.org> wrote:
> The problem is likely elsewhere, I'll revert this and commit a proper
> fix once I get access to my workstation (in ~24 hours).
>
> On Sat, Dec 23, 2017 at 4:06 PM, Davide Italiano via llvm-commits
> <llvm-commits at lists.llvm.org> wrote:
>> Author: davide
>> Date: Sat Dec 23 07:06:30 2017
>> New Revision: 321402
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=321402&view=rev
>> Log:
>> [SCCP] Manually fold branches on undef.
>>
>> This code was originally removed and replace with an assertion
>> because believed unnecessary. It turns out there was simply
>> no test coverage for this case, and the constant folder doesn't
>> yet know about patterns like `br undef %label1, %label2`.
>> Presumably at some point the constant folder might learn about
>> these patterns, but it's a broader change.
>> A testcase will be added to make sure this doesn't regress again
>> in the future.
>>
>> Fixes PR35723.
>>
>> Modified:
>> llvm/trunk/lib/Transforms/Scalar/SCCP.cpp
>>
>> Modified: llvm/trunk/lib/Transforms/Scalar/SCCP.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/SCCP.cpp?rev=321402&r1=321401&r2=321402&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/Transforms/Scalar/SCCP.cpp (original)
>> +++ llvm/trunk/lib/Transforms/Scalar/SCCP.cpp Sat Dec 23 07:06:30 2017
>> @@ -1929,9 +1929,32 @@ static bool runIPSCCP(Module &M, const D
>> if (!I) continue;
>>
>> bool Folded = ConstantFoldTerminator(I->getParent());
>> - assert(Folded &&
>> - "Expect TermInst on constantint or blockaddress to be folded");
>> - (void) Folded;
>> + if (!Folded) {
>> + // The constant folder may not have been able to fold the terminator
>> + // if this is a branch or switch on undef. Fold it manually as a
>> + // branch to the first successor.
>> +#ifndef NDEBUG
>> + if (auto *BI = dyn_cast<BranchInst>(I)) {
>> + assert(BI->isConditional() && isa<UndefValue>(BI->getCondition()) &&
>> + "Branch should be foldable!");
>> + } else if (auto *SI = dyn_cast<SwitchInst>(I)) {
>> + assert(isa<UndefValue>(SI->getCondition()) && "Switch should fold");
>> + } else {
>> + llvm_unreachable("Didn't fold away reference to block!");
>> + }
>> +#endif
>> +
>> + // Make this an uncond branch to the first successor.
>> + TerminatorInst *TI = I->getParent()->getTerminator();
>> + BranchInst::Create(TI->getSuccessor(0), TI);
>> +
>> + // Remove entries in successor phi nodes to remove edges.
>> + for (unsigned i = 1, e = TI->getNumSuccessors(); i != e; ++i)
>> + TI->getSuccessor(i)->removePredecessor(TI->getParent());
>> +
>> + // Remove the old terminator.
>> + TI->eraseFromParent();
>> + }
>> }
>>
>> // Finally, delete the basic block.
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
>
>
> --
> Davide
>
> "There are no solved problems; there are only problems that are more
> or less solved" -- Henri Poincare
--
Davide
"There are no solved problems; there are only problems that are more
or less solved" -- Henri Poincare
More information about the llvm-commits
mailing list