When -ffast-math was specified on Clang command line, it seems to be impossible to later turn off FP contraction via pragmas

[rygorous]

Example C code:

float g(float a, float b, float c) {
    return a * b + c;
}
#pragma float_control(precise, on)
#pragma STDC FP_CONTRACT OFF

float f(float a, float b, float c) {
    return a * b + c;
}

Compile for x86-64 with -mfma -O2 -ffast-math. Expectation is that g may use FMAs but f may not, but in fact both produce FMAs in Clang-14.0.0. Compiler Explorer link with repro: https://godbolt.org/z/TacbzWzxn

It works when there is no -ffast-math on the command line. Presumably the presence of that flag on the command line triggers some global override behavior?

I would love to just remove fast-math from the command line for the relevant TU, but this turns out to be problematic in the build setup for the project I need this in. I was asked to use pragmas for my use case instead, but can't properly because of the above problem.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[fhahn]

@zahiraam I think you have been working in this area. Do you have any ideas what could be going wrong?

[fhahn]

Still an issue.

[llvmbot]

@llvm/issue-subscribers-clang-codegen