| Issue |
244873
|
| Summary |
[VPlan] Compile-time hang at -O3: exponential recursion in `vputils::isSingleScalar` from `narrowToSingleScalarRecipes`
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
haoxintu
|
## Description
```c
unsigned y;
void f(unsigned s1, unsigned i1)
{
unsigned s2 = 0x5555u, i0;
switch (y)
{
case 33:
for (;;)
{
do
{
i1++;
s2 ^= s2;
i0 = 0;
do
{
i0++;
if ((s2 + s1) & 15)
{
unsigned t = y ^ s2;
s2 -= t ^ (t >> (t * 0x9e3779b9u & 31));
}
}
while (i0 < 17u);
}
while (i1 < 5u);
case 15:
s1 -= s2;
}
}
}
```
```
$ time clang -O3 -c repro.c # does not finish within 300 s
$ time clang -O2 -c repro.c # 0.1 s
```
`-ftime-report` on a larger variant: LoopVectorizePass takes 127.5 s of 128.2 s. `opt -passes=loop-vectorize` on the IR before the vectorizer (`repro.ll`) does not finish within 60 s, and llvm 21.1.8's `opt` finishes it at once.
A stack sample shows unbounded recursion of `vputils::isSingleScalar` over the operands. It is called from `narrowToSingleScalarRecipes`:
```
llvm::vputils::isSingleScalar(llvm::VPValue const*)
std::__find_if<..., _Iter_negate<bool (*)(llvm::VPValue const*)>>(...)
llvm::vputils::isSingleScalar(llvm::VPValue const*)
... (repeated)
narrowToSingleScalarRecipes(llvm::VPlan&)
llvm::VPlanTransforms::optimize(llvm::VPlan&)
llvm::LoopVectorizationPlanner::buildVPlans(...)
llvm::LoopVectorizePass::processLoop(llvm::Loop*)
```
Reproduced in Godbolt (x86-64 clang assertions trunk, -O3): https://godbolt.org/z/e95aoMaWG (the compile times out there)
## Checked Versions
- Hangs at -O3 with clang trunk 2012344e5da8 (2026-11-19); -O0/-O1/-O2/-Os are fine.
- clang 21.1.8 compiles it at -O3 in under 1 s, so this is a regression.
Thanks.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs