[Mesa-dev] [PATCH 08/10] radeonsi: re-enable unsafe-fp-math for LLVM 3.8

[Connor Abbott]

On Sun, Oct 11, 2015 at 7:08 AM, Marek Olšák <maraeo at gmail.com> wrote:
> On Sun, Oct 11, 2015 at 4:12 AM, Connor Abbott <cwabbott0 at gmail.com> wrote:
>> FWIW, this isn't quite correct with ARB_shader_precision or GL4.1 --
>> it specifies that infinities should be correctly generated through
>> division by 0, which unsafe-fp-math doesn't guarantee. At least,
>> that's assuming this is similar to the "fast" per-instruction flag
>> (http://llvm.org/docs/LangRef.html#fast-math-flags) which says "This
>> flag implies all the others."
>
> We don't use the per-instruction flags yet.
>
> Sadly, we need this flag to be able to get (1/sqrt) -> RSQ and (1/x)
> -> RCP. LLVM doesn't have standard intrinsics for those instructions,
> so we have to unwind them and rely on LLVM to combine them. Without
> it, the TGSI->LLVM conversion would produce worse code.
>
> Marek

Right, but you're also allowing transforms which GLSL explicitly
disallows. It sounds like you need more precise control than what LLVM
currently offers.

Connor