| Issue |
242584
|
| Summary |
[Clang][OpenMP] Incorrect trip-count codegen skips loop execution
|
| Labels |
clang
|
| Assignees |
|
| Reporter |
amad00r
|
Clang generates an incorrect trip count for this loop, causing it to be skipped:
```c
// clang -fopenmp --offload-arch=native -O3 %s
int main(void) {
#pragma omp target teams distribute parallel for
for (unsigned long long i = 0; i < ~0ULL; i += 0x1111111111111111ULL)
__builtin_trap();
return 0;
}
```
The loop condition is true initially, so execution should reach __builtin_trap() on the first iteration. Instead, the generated trip count is zero and the program returns normally.
Without the trap, the loop would execute 15 iterations. Its final increment reaches ~0ULL exactly, so the original loop counter does not overflow.
The overflow occurs in the generated trip-count calculation: `((Upper - Lower - 1) + Step) / Step`
For these bounds, adding Step wraps the unsigned 64-bit numerator to a value smaller than Step. The subsequent division therefore produces zero.
For a nonempty loop, computing the following equivalent _expression_ would avoid the overflowing addition: `((Upper - Lower - 1) / Step) + 1`
The affected calculation is in calculateNumIters in clang/lib/Sema/SemaOpenMP.cpp, used by ordinary OpenMP lowering.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs