[llvm-dev] float vs. mcmodel=large

[Fangrui Song via llvm-dev]

On 2021-01-04, Ákos Tompos via llvm-dev wrote:
>Hi All,
>
>First time poster here. I ran into an interesting issue today where clang
>generates a code which crashes. I know it's a very common thing among
>programmers to blame the compiler for their own mistakes, but I think it's
>not applicable for this case. I have managed to narrow down the problem to
>a few lines of code and two compiler options. Here is the code:
>
>int main()
>{
>    float f = 1.0f;
>    return 0;
>}
>
>And if you compile it like this:
>clang  -fPIC -mcmodel=large -o main main.c
>
>Then the resulting executable will sigsev on the line trying to assing 1.0
>to f;
>This affects all versions of clang 7-11 on a PC 64bit. I have managed to
>try a clang-12 on a mac as well, that was also affected.
>Clang-6.0 works.
>It also works if you replace float with int.
>
>I am puzzled now as this seems like a very obvious error. Is -mcmodel=large
>supported at all? Am I doing something wrong?
>
>Thanks for your help
>Akos

The unreleased clang-12 should be good.

I have fixed the bug in https://reviews.llvm.org/D86024
It should work on ELF platforms, but I don't know much about macOS.