Re: [PATCH] Optimize logarithm addition and subtraction

[Marc Glisse <marc dot glisse at inria dot fr>]

On Sun, 5 Aug 2018, MCC CS wrote:

> Besides, if you think optimizing
> /* logN(a) + logN(b) -> logN(a * b) */
> would be enough, without the coefficients,
> here's a simpler patch:
>
> Index: gcc/match.pd
> ===================================================================
> --- gcc/match.pd    (revision 263305)
> +++ gcc/match.pd    (working copy)
> @@ -4015,6 +4015,18 @@ DEFINE_INT_AND_FLOAT_ROUND_FN (RINT)
>    (exps (logs @0))
>    @0))
>
> +(for logs (LOG LOG2 LOG10)

You seem to be missing some indentation, see neighboring
transformations.

> + /* logN(a) + logN(b) -> logN(a * b). */
> + (simplify
> +  (plus:c (logs:s (@0)) (logs:s (@1)))
> +  (logs (mult (@0) (@1))))

No parentheses around @0 and @1. No :c here, the pattern is symmetric
and a permutation would be redundant.

> +
> +(for logs (LOG LOG2 LOG10)
> + /* logN(a) - logN(b) -> logN(a / b). */
> + (simplify
> +  (sub (logs:s (@0)) (logs:s (@1)))
> +  (logs (rdiv (@0) (@1))))

It would be possible to merge the 2 with a second 'for', but that's not a 
requirement.

> +
>  /* Optimize logN(func()) for various exponential functions.  We
>     want to determine the value "x" and the power "exponent" in
>     order to transform logN(x**exponent) into exponent*logN(x).  */
> Index: gcc/testsuite/gcc.dg/pr86710.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/pr86710.c    (nonexistent)
> +++ gcc/testsuite/gcc.dg/pr86710.c    (working copy)
> @@ -0,0 +1,15 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2" } */

The transformation should not happen at -O2, it should require
-funsafe-math-optimizations or similar.

> +/* { dg-final { scan-assembler-not "log(.*)log(.*)log" } } */

I think I'd rather scan-tree-dump-times the "optimized" dump.

> +
> +double c;
> +
> +void f1 (double x, double s)
> +{
> +  c = __builtin_log(x) + __builtin_log(s);
> +}
> +
> +void f2 (double x, double s)
> +{
> +  c = __builtin_log(x) - __builtin_log(s);
> +}

You need to specify how you tested the patch (bootstrap on what platform, 
run make check, whatever). And then we need to see if other people think 
-funsafe-math-optimization is enough protection or the transformation is 
too dangerous...