Re: [PATCH] middle-end: convert builtin finite -> MINUS/ORD

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 6/12/07, Kaveh R. GHAZI <ghazi@caip.rutgers.edu> wrote:
> This patch adds a transformation of builtin finite(x) into:
>
>         tmp = x - x;
>         result = tmp ORD tmp;
>
> We already do something similar (but simpler) for builtin isnan.
>
> This particular transformation works because "x - x" turns Inf into NaN,
> while preserving an existing NaN.  Then the ORD returns true if we don't
> have NaN, i.e. a finite original value.  (I got the idea from libgcc2.c.)
>
> One glich was that the "x - x" got folded to zero with only the flag
> -funsafe-math-optimizations.  I added an additional check on !HONOR_NANS
> and !HONOR_INFINITIES to those transformations.  These macros check (among
> other things) for -ffinite-math-only before folding "x - x".  I had to
> leave the "unsafe" check in there because the comments indicate that there
> are some wierd corner cases where folding "x - x" to zero is actually
> invalid for non-ieee finite values.
>
> I also filled out the existing two tests with more cases to cover all
> floating point types.
>
> Tested on sparc-sun-solaris2.10, no regressions and all the testcases
> pass.
>
> Okay for mainline?

The fold-const.c and simplify-rtx.c changes are ok.

Andrew is right about x - x trapping if x is a signalling NaN or +Inf
or -Inf (invalid
operation exception), x is a denormal (denormal operand exception).  So the
transformation is only valid for non-trapping math.  Btw. - is it
actually faster to
use the new sequence than fxam in x87 math?  How do code size compare
to a function call?
(Just to see if we should add a finite optab as well to allow the
target to override
the general expansion and conditionalize on OPTIMIZE_SIZE).

Thanks,
Richard.

>                 Thanks,
>                 --Kaveh
>
> 2007-06-11  Kaveh R. Ghazi  <ghazi@caip.rutgers.edu>
>
>         * builtins.c (fold_builtin_classify): Transform builtin finite.
>
>         * fold-const.c (fold_binary): Guard (X-X) -> 0 transformation
>         with !HONOR_NANS and !HONOR_INFINITIES.
>         * simplify-rtx.c (simplify_binary_operation_1): Likewise.
>
> testsuite:
>         * gcc.dg/pr28796-1.c: Add more cases.
>         * gcc.dg/pr28796-2.c: Likewise.