On Tue, May 10, 2022 at 02:56:58PM -0400, Michael Meissner wrote:
> On Tue, May 10, 2022 at 07:27:30AM -0500, Segher Boessenkool wrote:
> > > IMHO, it's something we want to fix as well, based on the reasons:
> > > 1) bif names have the corresponding mnemonics, users would expect 1-1
> > > mapping here.
> > > 2) clang emits xs{min,max}dp all the time, with cpu type power7/8/9/10.
> > > 3) according to uarch info, xs{min,max}cdp use the same units and have
> > > the same latency,
> > > no benefits to replace with xs{min,max}cdp.
> >
> > I never understood any of this. Mike? Why do we do those "c" things
> > at all, ever?
>
> In the power7, we only had x{s,v}{min,max}{sp,dp}. But those aren't useful
> for
> optimizing normal (a > b) ? a : b without using -ffast-math.
But RTL smin (as well as Gimple min_expr) is *undefined* without
-ffast-math (well, -ffinite-math-only and -fno-signed-zeros at least).
The only place we generate xs{min,max}[c]dp uses s{min,max}. So the
much saner xs{min,max}dp are fine always.
> Power9 added the
> 'c' and 'j' versions of the insns. GCC never generates the 'j' version.
>
> Basically for ?: we generate:
>
> * Code = power8, no -ffast-math: Generate compare, move;
> * Code = power8, -ffast-math: Generate xsmaxdp/xsmindp;
> * Code = power9, no -ffast-mth: Generate xsmaxcdp/xsmincdp; (and)
This one uses broken RTL (and broken Gimple before that): s{min,max}
cannot be used for FP without -ffast-math.
> * Code = power9, -ffast-math: Generate xsmaxcdp/xsmincdp.
xs{min,max}dp will work just as well.
> For the __builtin_fmax and __builtin_fmin functions:
>
> * Code = power8, no -ffast-math: Generate call to fmax/fmin;
> * Code = power8, -ffast-math: Generate xsmaxdp/xsmindp;
> * Code = power9, no -ffast-mth: Generate call to fmax/fmin; (and)
> * Code = power9, -ffast-math: Generate xsmaxcdp/xsmincdp.
Same brokenness here.
> For IEEE 128-bit floating point, we only have xs{min,max}cqp. We do not have
> the version without 'c' nor do we have the 'j' version.
And here.
Why would we ever prefer xsmincdp over xsmindp, other than for machine
code for some not-so-smart C code that wil not do useful things for
signed zeros or NaNs (but using the "c" insns generates faster, smaller
code that has those silly semantics)?
Segher
