This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] PR 168: More tree_expr_nonnegative_p improvements
Roger Sayle <roger@eyesopen.com> writes:
> Many thanks to Joseph Myers for bringing it to my attention that
> the long-standing PR middle-end/168 is caused by a defficiency of
> GCC's tree_expr_nonnegative_p. This patch contains several small
> improvements to tree_expr_nonnegative_p, which include recognizing
> that floating point conversions and the ceil and floor builtins are
> always sign preserving, that a zero extension is always non-negative,
> and that sign extension of a non-negative operand is non-negative.
>
> Additionally, to actually fix PR middle-end/168, it now knows that
> zero_extend(x) + zero_extend(y) and zero_extend(x) * zero_extend(y)
> are unsigned, iff the result isn't wide enough to overflow into the
> sign bit of the result, using x+y is atmost max(bits(x),bits(y))+1
> wide, and x*y is atmost bits(x)+bits(y) wide.
>
> The following patch has been tested on i686-pc-linux-gnu with a full
> "make bootstrap", all languages except treelang, and regression tested
> with a top-level "make -k check" with no new failures.
>
> Ok for mainline?
OK.
>
> 2003-06-12 Roger Sayle <roger@eyesopen.com>
>
> PR middle-end/168
> * fold-const.c (tree_expr_nonnegative_p): Handle addition
> and multiplication of zero extensions, floating point division,
> and integer<->fp, fp<->fp and zero extension conversions.
> The built-in ceil and floor functions preserve signedness.
>
> * gcc.dg/20030612-1.c: New test case.
--
- Geoffrey Keating <geoffk@geoffk.org>