clang emits useless guard variable for static const value

[Matheus28]

Please see https://godbolt.org/z/K6bM5qfx3 and compare the gcc and clang output

The compiler is able to calculate the pow results at compile time, however it still issues a __cxa_guard_acquire/__cxa_guard_release pair just to move those values around in memory. Meanwhile, gcc is able to use the .rodata section directly.

[frederick-vs-ja]

Related to #139885, possibly a duplicate. FWIW, Clang's __builtin_pow is not yet constexpr.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Matheus Valadares (Matheus28)

Please see https://godbolt.org/z/K6bM5qfx3 and compare the gcc and clang output

The compiler is able to calculate the pow results at compile time, however it still issues a __cxa_guard_acquire/__cxa_guard_release pair just to move those values around in memory. Meanwhile, gcc is able to use the .rodata section directly.

[Matheus28]

@frederick-vs-ja I think it seems unrelated to it not being constexpr, as clang is already successfully calculating its value at compile time. The problem is that it's just not baking those const values into .rodata.

[frederick-vs-ja]

Clang can constant-fold the results of pow, but such optimization seems happening after the phase that determines constant initialization. Another example shows that constexpr matters, for both Clang and GCC.

[Matheus28]

Ah I see. Wouldn't this also present itself as an issue in non-constexpr stuff if a later optimization pass can transform an expression into a constant, but it's too late to bake it into .rodata?

[tbaederr]

Clang can constant-fold the results of pow, but such optimization seems happening after the phase that determines constant initialization. Another example shows that constexpr matters, for both Clang and GCC.

Exactly. If you want to make sure, just add constinit to the declaration and you get a proper error message.