Re: [PATCH, midlevel]: Fix PR tree-optimization/21004

[Roger Sayle <roger at eyesopen dot com>]

On Thu, 14 Apr 2005, Uros Bizjak wrote:
> 2005-04-14  Uros Bizjak  <uros@kss-loka.si>
>
>     * convert.c (convert_to_integer): Convert ceilf, ceill, floorf
>     and floorl in c99 mode only.

This on it's own isn't sufficient to fix the failure of builtins-53.c
on non-C99 platforms, and although it prevents the segmentation fault,
we should really explicitly catch this failure in future.

In addition to your above patch, could you modify builtins-53.c to
#include "builtins-config.c" and then only test the float and long
double variants if HAVE_C99_RUNTIME is defined.  This follows the
example of builtins-20.c, i.e. dejagnu tests shouldn't call the
float/long double versions of the math functions on all platforms.

Then secondly, in builtins.c's expand_builtin_int_roundingfn, could
you add an assert after calling mathfn_built_in, i.e.

    fallback_fndecl = mathfn_built_in (TREE_TYPE (arg), fallback_fn);
+   /* We shouldn't get here on targets without TARGET_C99_FUNCTIONS.
+      ??? Perhaps convert (int)floorf(x) into (int)floor((double)x).  */
+   gcc_assert (fallback_fndecl != NULL_TREE);
    exp = build_function_call_expr (fallback_fndecl, arglist);


I'll preapprove your patch together with the above two refinements
if it passes bootstrap and regression testing without problems.

Thanks in advance, and sorry for not spotting this during review.

Roger
--