'#pragma float_control' is not supported on this target (AArch64)

[rygorous]

#pragma float_control is documented in the Clang extensions manual but doesn't seem to be supported on AArch64.

https://godbolt.org/z/ejYc1nYjd

I think that's the check at the top of PragmaFloatControlHandler::HandlePragma. I think that logic is checking for the wrong thing there.

I'm assuming the issue is that the AArch64 target doesn't have hasStrictFP() because it doesn't have exception trap support (or at least makes it an optional feature). This affects #pragma float_control(except, ...) but I don't see why precise should be excluded as well. As-is there seems to be no way on AArch64 to turn off fast-math like behavior via pragmas at all.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

[llvmbot]

@llvm/issue-subscribers-clang-frontend

[nfrechette]

I have just bumped into determinism issues caused fast math in clang14 (but not earlier versions) on arm64 and because of the lack of float_control support, I am stuck. I maintain header only C++ libraries where I do not have control over the compilation environment and must control things like fast math through pragma support. This ensures determinism across compiler releases and tool chains.