On Thu, 24 Oct 2024, Jakub Jelinek wrote:
> + if (TYPE_UNSIGNED (TREE_TYPE (arg2))
> + || sanitize_flags_p (SANITIZE_SHIFT))
> + arg2 = build2_loc (loc, TRUNC_MOD_EXPR, TREE_TYPE (arg2),
> + arg2, build_int_cst (TREE_TYPE (arg2),
> + prec));
> + else
> + {
> + /* When not sanitizing and second argument is signed,
> + just do the modulo in unsigned type, that results
> + in better generated code (for power of 2 precisions
> + bitwise AND). */
> + tree utype = c_common_unsigned_type (TREE_TYPE (arg2));
> + arg2 = build2_loc (loc, TRUNC_MOD_EXPR, utype,
> + fold_convert (utype, arg2),
> + build_int_cst (utype, prec));
> + }
If sanitizing makes sense for these built-in functions, surely it should
check for all negative shifts, including those that are multiples of the
width (and there should be tests for it in the testsuite). So sanitizing
would require more complicated logic to avoid reducing a negative shift
modulo the width at all. Or in the absence of support for sanitizing
these functions, the logic for modulo reduction shouldn't need to depend
on whether sanitizing.
> @@ -409,7 +416,9 @@ c_fully_fold_internal (tree expr, bool i
> warning_at (loc, OPT_Wshift_count_overflow,
> (code == LSHIFT_EXPR
> ? G_("left shift count >= width of type")
> - : G_("right shift count >= width of type")));
> + : code == RSHIFT_EXPR
> + ? G_("right shift count >= width of type")
> + : G_("rotate count >= width of type")));
This shouldn't be reachable, since larger rotate counts are valid and
should have been reduced modulo the width.
--
Joseph S. Myers
josmy...@redhat.com
