Re: [RFC PATCH] i386: Do not sanitize upper part of V2SFmode reg with -fno-trapping-math [PR110832]

[Richard Biener <rguenther@suse.de>]

On Tue, 8 Aug 2023, Uros Bizjak wrote:

> On Tue, Aug 8, 2023 at 10:07?AM Richard Biener <rguenther@suse.de> wrote:
> >
> > On Mon, 7 Aug 2023, Uros Bizjak wrote:
> >
> > > On Mon, Jul 31, 2023 at 11:40?AM Richard Biener <rguenther@suse.de> wrote:
> > > >
> > > > On Sun, 30 Jul 2023, Uros Bizjak wrote:
> > > >
> > > > > Also introduce -m[no-]mmxfp-with-sse option to disable trapping V2SF
> > > > > named patterns in order to avoid generation of partial vector V4SFmode
> > > > > trapping instructions.
> > > > >
> > > > > The new option is enabled by default, because even with sanitization,
> > > > > a small but consistent speed up of 2 to 3% with Polyhedron capacita
> > > > > benchmark can be achieved vs. scalar code.
> > > > >
> > > > > Using -fno-trapping-math improves Polyhedron capacita runtime 8 to 9%
> > > > > vs. scalar code.  This is what clang does by default, as it defaults
> > > > > to -fno-trapping-math.
> > > >
> > > > I like the new option, note you lack invoke.texi documentation where
> > > > I'd also elaborate a bit on the interaction with -fno-trapping-math
> > > > and the possible performance impact then NaNs or denormals leak
> > > > into the upper halves and cross-reference -mdaz-ftz.
> > >
> > > The attached doc patch is invoke.texi entry for -mmmxfp-with-sse
> > > option. It is written in a way to also cover half-float vectors. WDYT?
> >
> > "generate trapping floating-point operations"
> >
> > I'd say "generate floating-point operations that might affect the
> > set of floating point status flags", the word "trapping" is IMHO
> > misleading.
> > Not sure if "set of floating point status flags" is the correct term,
> > but it's what the C standard seems to refer to when talking about
> > things you get with fegetexceptflag.  feraieexcept refers to
> > "floating-point exceptions".  Unfortunately the -fno-trapping-math
> > documentation is similarly confusing (and maybe even wrong, I read
> > it to conform to 'non-stop' IEEE arithmetic).
> 
> Thanks for suggesting the right terminology. I think that:
> 
> +@opindex mpartial-vector-math
> +@item -mpartial-vector-math
> +This option enables GCC to generate floating-point operations that might
> +affect the set of floating point status flags on partial vectors, where
> +vector elements reside in the low part of the 128-bit SSE register.  Unless
> +@option{-fno-trapping-math} is specified, the compiler guarantees correct
> +behavior by sanitizing all input operands to have zeroes in the unused
> +upper part of the vector register.  Note that by using built-in functions
> +or inline assembly with partial vector arguments, NaNs, denormal or invalid
> +values can leak into the upper part of the vector, causing possible
> +performance issues when @option{-fno-trapping-math} is in effect.  These
> +issues can be mitigated by manually sanitizing the upper part of the partial
> +vector argument register or by using @option{-mdaz-ftz} to set
> +denormals-are-zero (DAZ) flag in the MXCSR register.
> 
> Now explain in adequate detail what the option does. IMO, the
> "floating-point operations that might affect the set of floating point
> status flags" correctly identifies affected operations, so an example,
> as suggested below, is not necessary.
> 
> > I'd maybe give an example of a FP operation that's _not_ affected
> > by the flag (copysign?).
> 
> Please note that I have renamed the option to "-mpartial-vector-math"
> with a short target-specific description:

Ah yes, that's a less confusing name but then it might suggest
that -mno-partial-vector-math would disable all of that, including
integer ops, not only the patterns possibly affecting the exception
flags?  Note I don't have a better suggestion and this is clearly
better than the one mentioning mmx.

> +partial-vector-math
> +Target Var(ix86_partial_vec_math) Init(1)
> +Enable floating-point status flags setting SSE vector operations on
> partial vectors
> 
> which I think summarises the option (without the word "trapping"). The
> same approach will be taken for Float16 operations, so the approach is
> not specific to MMX vectors.
> 
> > Otherwise it looks OK to me.
> 
> Thanks, I have attached the RFC V2 patch; I plan to submit a formal
> patch later today.

Thanks.  With AVX512VL there might also be the option to use
a mask (with the penalty of a very much larger instruction encoding).

Richard.