Re: [PATCH] math: optimize lrint on 32bit targets

[Rich Felker <dalias@libc.org>]

On Mon, Sep 23, 2019 at 08:38:19PM +0200, Szabolcs Nagy wrote:
> * Rich Felker <dalias@libc.org> [2019-09-23 13:40:29 -0400]:
> > On Sun, Sep 22, 2019 at 10:43:35PM +0200, Szabolcs Nagy wrote:
> > > +long lrint(double x)
> > > +{
> > > +	uint32_t abstop = asuint64(x)>>32 & 0x7fffffff;
> > > +	uint64_t sign = asuint64(x) & (1ULL << 63);
> > > +
> > > +	if (abstop < 0x41dfffff) {
> > > +		/* |x| < 0x7ffffc00, no overflow */
> > > +		double_t toint = asdouble(asuint64(1/EPS) | sign);
> > > +		double_t y = x + toint - toint;
> > > +		return (long)y;
> > > +	}
> > > +	return lrint_slow(x);
> > > +}
> > >  #else
> > >  long lrint(double x)
> > >  {
> > 
> > This code should be considerably faster than calling rint on 64-bit
> > archs too, no? I wonder if it should be something like (untested,
> > written inline here):
> 
> yeah i'd expect it to be a bit faster, but e.g. a
> target may prefer sign?-1/EPS:1/EPS to 1/EPS|sign,

Yes, perhaps. But this will be a lot less of a difference than an
extra call frame I think.

> and you need a threshold check even if there is no
> inexact overflow issue:

Ah yes.

> > long lrint(double x)
> > {
> > 	uint32_t abstop = asuint64(x)>>32 & 0x7fffffff;
> > 	uint64_t sign = asuint64(x) & (1ULL << 63);
> > 
> > #if LONG_MAX < 1U<<53 && defined(FE_INEXACT)
> > 	if (abstop >= 0x41dfffff) return lrint_slow(x);
> 
> #else
> 	if (abstop >= 0x43300000) return (long)x;
> 	/* |x| < 2^52 <= 1/EPS */
> 
> > #endif
> > 	/* |x| < 0x7ffffc00, no overflow */
> > 	double_t toint = asdouble(asuint64(1/EPS) | sign);
> > 	double_t y = x + toint - toint;
> > 	return (long)y;
> > }
> 
> i can try to benchmark this (although on x86_64 and
> aarch64 there is single instruction lrint so i can
> only benchmark machines where this is not relevant).

Yeah I think it's mainly riscv64 and s390x that might benefit right
now. Everything else 64-bit seems to have an optimized one.

Rich