
Date: Sun, 8 Dec 2019 09:54:15 0500 From: Rich Felker <dalias@...c.org> To: musl@...ts.openwall.com Subject: Re: Patches for math subtree On Sat, Dec 07, 2019 at 03:38:04PM 0500, Rich Felker wrote: > On Sat, Dec 07, 2019 at 09:15:34PM +0100, Stefan Kanthak wrote: > > Just some optimisations. > > > >  /src/math/i386/remquo.s > > +++ +/src/math/i386/remquo.s > > @@ 23,23 +23,17 @@ > > remquo: > > mov 20(%esp),%ecx > > fldl 12(%esp) > > fldl 4(%esp) > > mov 19(%esp),%dh > > xor 11(%esp),%dh > > 1: fprem1 > > fnstsw %ax > > sahf > > jp 1b > > fstp %st(1) > >  mov %ah,%dl > >  shr %dl > >  and $1,%dl > >  mov %ah,%al > >  shr $5,%al > >  and $2,%al > >  or %al,%dl > >  mov %ah,%al > >  shl $2,%al > >  and $4,%al > >  or %al,%dl > > + setc %dl > > + shl $2,%ah > > + adc %dl,%dl > > + shl $5,%ah > > + adc %dl,%dl > > test %dh,%dh > > > >  /src/math/ceil.c > > +++ +/src/math/ceil.c > > @@ 18,10 +18,10 @@ > > + /* special case because of nonnearest rounding modes */ > > + if (e < 0x3ff) { > > + FORCE_EVAL(y); > > + return u.i >> 63 ? 0.0 : 1.0; > > + } > > /* y = int(x)  x, where int(x) is an integer neighbor of x */ > > if (u.i >> 63) > > y = x  toint + toint  x; > > else > > y = x + toint  toint  x; > >  /* special case because of nonnearest rounding modes */ > >  if (e <= 0x3ff1) { > >  FORCE_EVAL(y); > >  return u.i >> 63 ? 0.0 : 1; > >  } > > > >  /src/math/floor.c > > +++ +/src/math/floor.c > > @@ 18,10 +18,10 @@ > > + /* special case because of nonnearest rounding modes */ > > + if (e < 0x3ff) { > > + FORCE_EVAL(y); > > + return u.i >> 63 ? 1.0 : 0.0; > > + } > > /* y = int(x)  x, where int(x) is an integer neighbor of x */ > > if (u.i >> 63) > > y = x  toint + toint  x; > > else > > y = x + toint  toint  x; > >  /* special case because of nonnearest rounding modes */ > >  if (e <= 0x3ff1) { > >  FORCE_EVAL(y); > >  return u.i >> 63 ? 1 : 0; > >  } > > Do you have any explanation of why these are optimizations? Specifically, the x86 asm one looks like it probably is, but I haven't read closely enough to verify. If it's a measurable improvement I'll try to take a look at it soon, but at some point all of the .s files here are slated for removal and replacement with inline asm in .c files that avoids all of the delicate flow/logic in asm and just uses the x87 instructions needed, so I don't want to spend a lot of effort on improving and validating improvements to them. For the latter two, the patches as written are wrong. They evaluate an uninitialized variable y. And I think these functions are required to set the status flags, so you can't just remove that. Maybe there's some alternate way to do it that would be faster, like just evaluating x±toint rather than the whole expression, but I'm not sure it helps. Rich
Powered by blists  more mailing lists
Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.