[cfe-dev] Documentation on -ffp-contract=fast vs pragma contract(off)

[Harald van Dijk via cfe-dev]

On 23/09/2020 15:53, Marius Hillenbrand via cfe-dev wrote:
> Hi,
> 
> The documentation of the #pragma clang fp contract appears to
> misrepresent the interaction of -ffp-contract=fast with the #pragma
> clang fp contract(off). The documentation states:
> 
>> The pragma can also be used with off which turns FP contraction off for a section of the code. This can be useful when fast contraction is otherwise enabled for the translation unit with the -ffp-contract=fast flag.
> 
> (https://clang.llvm.org/docs/LanguageExtensions.html#extensions-to-specify-floating-point-flags)
> 
> 
> However, a minimal example contradicts that:
> 
> float abc_pragma_off(float a, float b, float c) {
> #pragma clang fp contract(off)
> 	return a * b + c;
> }
> 
> ... is compiled to a fused multiply-add:
> $ clang -o - -S -march=haswell -O3 -ffp-contract=fast min.c
> 
> abc_pragma_off:
>    vfmadd213ss	%xmm2, %xmm1, %xmm0 # xmm0 = (xmm1 * xmm0) + xmm2
>    retq

Hi,

This looks very similar to what I had just found in 
<http://lists.llvm.org/pipermail/cfe-dev/2020-September/066870.html> for 
#pragma float_control. There are probably more #pragmas with the same 
problem.

> In my understanding, setting -ffp-contract=fast allows the backend to
> always contract FP operations. Since setting the pragma to off only
> controls clang's frontend and the pragma is not expressed in the IR, it
> cannot override the backend's behavior.

For anything that starts with #pragma clang, that could be reasonable, 
but I had thought this to be a backend bug because it affects pragmas 
that are not specific to clang, where it would seem very strange to me 
to adjust the documentation to match the implementation.

Cheers,
Harald van Dijk