https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122435
Bug ID: 122435
Summary: Optimization failure with trivial loop
Product: gcc
Version: 15.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: tobi at gcc dot gnu.org
Target Milestone: ---
I stumbled over code like this in one of our projects. The function g is just
a wordy no-op.
https://godbolt.org/z/4crT5xrc7
======================= code follows =================
using size_t = unsigned long int;
template<class T>
struct Span {
T *data;
size_t size;
};
Span<int> g(Span<unsigned> sp)
{
auto* p = reinterpret_cast<int *>(sp.data);
Span sp2{p, sp.size};
for (size_t i = 0; i < sp.size; ++i)
sp2.data[i] = static_cast<int>(sp.data[i]);
return sp2;
}
bool f() // Returns true
{
unsigned a[] = {1, 2, 3, 4};
Span sp{ a, 4 };
Span sp2 = g(sp);
for (int i = 0; i < 4; ++i) {
if (sp2.data[i] != i + 1)
return false;
}
return true;
}
===================== Code ends ==================
Observations: clang manages to trivialize the code for both functions. gcc
fails for g() no matter the optimization settings. The code for f() becomes
worse at -O3 than at -O. Nevertheless, gcc with -O3 succeeds at simplifying
f().
