This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH PR80153]Always generate folded type conversion in tree-affine
- From: Richard Biener <richard dot guenther at gmail dot com>
- To: Bin Cheng <Bin dot Cheng at arm dot com>
- Cc: "gcc-patches at gcc dot gnu dot org" <gcc-patches at gcc dot gnu dot org>, nd <nd at arm dot com>
- Date: Tue, 28 Mar 2017 14:34:06 +0200
- Subject: Re: [PATCH PR80153]Always generate folded type conversion in tree-affine
- Authentication-results: sourceware.org; auth=none
- References: <VI1PR0802MB21760C1CA9D2EDD0FF155D5FE7320@VI1PR0802MB2176.eurprd08.prod.outlook.com>
On Tue, Mar 28, 2017 at 2:01 PM, Bin Cheng <Bin.Cheng@arm.com> wrote:
> Hi,
> This patch is to fix PR80153. As analyzed in the PR, root cause is tree_affine lacks
> ability differentiating (unsigned)(ptr + offset) and (unsigned)ptr + (unsigned)offset,
> even worse, it always returns the former expression in aff_combination_tree, which
> is wrong if the original expression has the latter form. The patch resolves the issue
> by always returning the latter form expression, i.e, always trying to generate folded
> expression. Also as analyzed in comment, I think this change won't result in substantial
> code gen difference.
> I also need to adjust get_computation_aff for test case gcc.dg/tree-ssa/reassoc-19.c.
> Well, I think the changed behavior is correct, but for case the original pointer candidate
> is chosen, it should be unnecessary to compute in uutype. Also this adjustment only
> generates (unsigned)(pointer + offset) which is generated by tree-affine.c.
> Bootstrap and test on x86_64 and AArch64. Is it OK?
Hmm. What is the desired goal? To have all elts added have
comb->type as type? Then
the type passed to add_elt_to_tree is redundant with comb->type. It
looks like it
is always passed comb->type now.
ISTR from past work in this area that it was important for pointer
combinations to allow
both pointer and sizetype elts at least.
Your change is incomplete I think, for the scale == -1 and POINTER_TYPE_P case
elt is sizetype now, not of pointer type. As said above, we are
trying to maintain
both pointer and sizetype elts with like:
if (scale == 1)
{
if (!expr)
{
if (POINTER_TYPE_P (TREE_TYPE (elt)))
return elt;
else
return fold_convert (type1, elt);
}
where your earilier fold to type would result in not all cases handled the same
(depending whether scale was -1 for example).
Thus - shouldn't we simply drop the type argument (or rather the comb one?
that wide_int_ext_for_comb looks weird given we get a widest_int as input
and all the other wide_int_ext_for_comb calls around).
And unconditionally convert to type, simplifying the rest of the code?
Richard.
> 2017-03-27 Bin Cheng <bin.cheng@arm.com>
>
> PR tree-optimization/80153
> * tree-affine.c (add_elt_to_tree): Convert to type as required
> by function's parameter.
> * tree-ssa-loop-ivopts.c (alloc_iv): Pass in consistent types.
> (get_computation_aff): Use utype directly for original candidate.
>
> gcc/testsuite/ChangeLog
> 2017-03-27 Bin Cheng <bin.cheng@arm.com>
>
> PR tree-optimization/80153
> * gcc.c-torture/execute/pr80153.c: New.