On Fri, 19 May 2017, Marek Polacek wrote:
> extract_muldiv folds
>
> (n * 10000 * z) * 50
>
> to
>
> (n * 500000) * z
>
> which is a wrong transformation to do, because it may introduce an overflow.
> This resulted in a ubsan false positive. So we should just disable this
> folding altogether. Does the approach I took make sense?
>
> Bootstrapped/regtested on x86_64-linux, ok for trunk?
Didn't dig very far to identify extract_muldiv, but I guess it's either
of the following recursions that trigger?
/* If we can extract our operation from the LHS, do so and return a
new operation. Likewise for the RHS from a MULT_EXPR.
Otherwise,
do something only if the second operand is a constant. */
if (same_p
&& (t1 = extract_muldiv (op0, c, code, wide_type,
strict_overflow_p)) != 0)
return fold_build2 (tcode, ctype, fold_convert (ctype, t1),
fold_convert (ctype, op1));
else if (tcode == MULT_EXPR && code == MULT_EXPR
&& (t1 = extract_muldiv (op1, c, code, wide_type,
strict_overflow_p)) != 0)
return fold_build2 (tcode, ctype, fold_convert (ctype, op0),
fold_convert (ctype, t1));
thus I'd simply guard them with TYPE_OVERFLOW_WRAPS ().
In the end I think the whole extract_muldiv mess should be truncated
down to what its name suggest - identifying and removing mul-div
cancellations.
It's for example not clear whether the recursion above assumes
TYPE_OVERFLOW_UNDEFINED (it passes a wide_type .. widening is only
ok if there's no overflow).
Richard.
> 2017-05-19 Marek Polacek <pola...@redhat.com>
>
> PR sanitizer/80800
> * fold-const.c (extract_muldiv_1): Don't fold ((X * C1) * Y) * C
> to (X * C2) * Y.
>
> * c-c++-common/ubsan/pr80800.c: New test.
> * c-c++-common/Wduplicated-branches-1.c: Adjust an expression.
>
> diff --git gcc/fold-const.c gcc/fold-const.c
> index 19aa722..e525c2d 100644
> --- gcc/fold-const.c
> +++ gcc/fold-const.c
> @@ -6260,6 +6260,17 @@ extract_muldiv_1 (tree t, tree c, enum tree_code code,
> tree wide_type,
> break;
>
> case MULT_EXPR:
> + /* ((X * C1) * Y) * C
> + cannot be reduced to
> + (X * C2) * Y (where C2 == C * C1)
> + because that can introduce an overflow. */
> + if (same_p
> + && op0 != NULL_TREE
> + && TREE_CODE (op0) == MULT_EXPR
> + && TREE_CODE (TREE_OPERAND (op0, 1)) == INTEGER_CST
> + && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (t)))
> + break;
> +
> /* We have a special case here if we are doing something like
> (C * 8) % 4 since we know that's zero. */
> if ((code == TRUNC_MOD_EXPR || code == CEIL_MOD_EXPR
> diff --git gcc/testsuite/c-c++-common/Wduplicated-branches-1.c
> gcc/testsuite/c-c++-common/Wduplicated-branches-1.c
> index c0b93fc..7c5062d 100644
> --- gcc/testsuite/c-c++-common/Wduplicated-branches-1.c
> +++ gcc/testsuite/c-c++-common/Wduplicated-branches-1.c
> @@ -89,7 +89,7 @@ f (int i, int *p)
> if (i == 8) /* { dg-warning "this condition has identical branches" } */
> return i * 8 * i * 8;
> else
> - return 8 * i * 8 * i;
> + return i * 8 * i * 8;
>
>
> if (i == 9) /* { dg-warning "this condition has identical branches" } */
> diff --git gcc/testsuite/c-c++-common/ubsan/pr80800.c
> gcc/testsuite/c-c++-common/ubsan/pr80800.c
> index e69de29..992c136 100644
> --- gcc/testsuite/c-c++-common/ubsan/pr80800.c
> +++ gcc/testsuite/c-c++-common/ubsan/pr80800.c
> @@ -0,0 +1,25 @@
> +/* PR sanitizer/80800 */
> +/* { dg-do run } */
> +/* { dg-options "-fsanitize=undefined -fsanitize-undefined-trap-on-error" }
> */
> +
> +int n = 20000;
> +int z = 0;
> +
> +int
> +fn1 (void)
> +{
> + return (n * 10000 * z) * 50;
> +}
> +
> +int
> +fn2 (void)
> +{
> + return (10000 * n * z) * 50;
> +}
> +
> +int
> +main ()
> +{
> + fn1 ();
> + fn2 ();
> +}
>
> Marek
>
>
--
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB
21284 (AG Nuernberg)
