On Mon, 2020-03-16 at 09:23 +0100, Jakub Jelinek via Gcc-patches wrote:
> Hi!
>
> The recent match.pd changes can generate a MEM_REF which can be seen by the
> C FE folding routines. Unlike the C++ FE, they weren't expected in the C FE
> yet. MEM_REF should be handled like INDIRECT_REF, except that it has two
> operands rather than just one and that we should preserve the type of the
> second operand. Given that it already has to be an INTEGER_CST with pointer
> type, I think we are fine, the recursive call should return the INTEGER_CST
> unmodified and STRIP_TYPE_NOPS will not strip anything.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2020-03-16 Jakub Jelinek <ja...@redhat.com>
>
> PR c/94179
> * c-fold.c (c_fully_fold_internal): Handle MEM_REF.
>
> * gcc.c-torture/compile/pr94179.c: New test.
>
> --- gcc/c/c-fold.c.jj 2020-01-12 11:54:36.208416501 +0100
> +++ gcc/c/c-fold.c 2020-03-15 17:26:42.863491563 +0100
> @@ -346,6 +346,7 @@ c_fully_fold_internal (tree expr, bool i
> case UNGT_EXPR:
> case UNGE_EXPR:
> case UNEQ_EXPR:
> + case MEM_REF:
> /* Binary operations evaluating both arguments (increment and
> decrement are binary internally in GCC). */
> orig_op0 = op0 = TREE_OPERAND (expr, 0);
> @@ -435,6 +436,14 @@ c_fully_fold_internal (tree expr, bool i
> || TREE_CODE (TREE_TYPE (orig_op0)) == FIXED_POINT_TYPE)
> && TREE_CODE (TREE_TYPE (orig_op1)) == INTEGER_TYPE)
> warn_for_div_by_zero (loc, op1);
> + if (code == MEM_REF
> + && ret != expr
> + && TREE_CODE (ret) == MEM_REF)
> + {
> + TREE_READONLY (ret) = TREE_READONLY (expr);
> + TREE_SIDE_EFFECTS (ret) = TREE_SIDE_EFFECTS (expr);
> + TREE_THIS_VOLATILE (ret) = TREE_THIS_VOLATILE (expr);
> + }
> goto out;
I don't particularly like modifying what RET points to in this way, but it's the
same that we're already doing for INDIRECT_REF. So OK.
jeff
