Reduce multiple_of_p uses
Commit Message
There are a few cases where we know we're dealing with (poly-)integer
constants, so remove the use of multiple_of_p in those cases to make
the PR100499 fix less impactful.
Bootstrapped and tested on x86_64-unknown-linux-gnu.
OK?
Thanks,
Richard.
2022-01-24 Richard Biener <rguenther@suse.de>
PR tree-optimization/100499
* tree-cfg.cc (verify_gimple_assign_ternary): Use multiple_p
on poly-ints instead of multiple_of_p.
* tree-ssa.cc (maybe_rewrite_mem_ref_base): Likewise.
(non_rewritable_mem_ref_base): Likewise.
(non_rewritable_lvalue_p): Likewise.
(execute_update_addresses_taken): Likewise.
---
gcc/tree-cfg.cc | 3 ++-
gcc/tree-ssa.cc | 18 +++++++++---------
2 files changed, 11 insertions(+), 10 deletions(-)
Comments
On 1/24/2022 7:45 AM, Richard Biener via Gcc-patches wrote:
> There are a few cases where we know we're dealing with (poly-)integer
> constants, so remove the use of multiple_of_p in those cases to make
> the PR100499 fix less impactful.
>
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
>
> OK?
>
> Thanks,
> Richard.
>
> 2022-01-24 Richard Biener <rguenther@suse.de>
>
> PR tree-optimization/100499
> * tree-cfg.cc (verify_gimple_assign_ternary): Use multiple_p
> on poly-ints instead of multiple_of_p.
> * tree-ssa.cc (maybe_rewrite_mem_ref_base): Likewise.
> (non_rewritable_mem_ref_base): Likewise.
> (non_rewritable_lvalue_p): Likewise.
> (execute_update_addresses_taken): Likewise.
So it's not a full fix for this class of problems, but removes some of
the cases where we could potentially overflow and give the wrong
result. I'd be happier if we had a concrete testcase for these
instances you're fixing, but I think it's a move in the right direction
and I obviously trust your judgment on whether or not to install it now
or wait for gcc-13.
So your call. Both on whether or not to try and construct testcases to
trigger these instances and whether or not to install now or wait for
gcc-13.
jeff
On Fri, 28 Jan 2022, Jeff Law wrote:
>
>
> On 1/24/2022 7:45 AM, Richard Biener via Gcc-patches wrote:
> > There are a few cases where we know we're dealing with (poly-)integer
> > constants, so remove the use of multiple_of_p in those cases to make
> > the PR100499 fix less impactful.
> >
> > Bootstrapped and tested on x86_64-unknown-linux-gnu.
> >
> > OK?
> >
> > Thanks,
> > Richard.
> >
> > 2022-01-24 Richard Biener <rguenther@suse.de>
> >
> > PR tree-optimization/100499
> > * tree-cfg.cc (verify_gimple_assign_ternary): Use multiple_p
> > on poly-ints instead of multiple_of_p.
> > * tree-ssa.cc (maybe_rewrite_mem_ref_base): Likewise.
> > (non_rewritable_mem_ref_base): Likewise.
> > (non_rewritable_lvalue_p): Likewise.
> > (execute_update_addresses_taken): Likewise.
> So it's not a full fix for this class of problems, but removes some of the
> cases where we could potentially overflow and give the wrong result. I'd be
> happier if we had a concrete testcase for these instances you're fixing, but I
> think it's a move in the right direction and I obviously trust your judgment
> on whether or not to install it now or wait for gcc-13.
No, multiple_of_p has no issues with these cases since we only ever pass
it [POLY_]INT_CSTs, it's just that there's no need to use multiple_of_p
since we have a better tool for these cases (and thus later do not have
to decide whether to use the no-wrap or wrap variant).
> So your call. Both on whether or not to try and construct testcases to
> trigger these instances and whether or not to install now or wait for gcc-13.
I've pushed it now.
Richard.
@@ -4410,7 +4410,8 @@ verify_gimple_assign_ternary (gassign *stmt)
TREE_TYPE (rhs2_type))
&& multiple_p (TYPE_VECTOR_SUBPARTS (rhs1_type),
TYPE_VECTOR_SUBPARTS (rhs2_type))
- && multiple_of_p (bitsizetype, rhs3, TYPE_SIZE (rhs2_type)))))
+ && multiple_p (wi::to_poly_offset (rhs3),
+ wi::to_poly_offset (TYPE_SIZE (rhs2_type))))))
{
error ("not allowed type combination in %qs", code_name);
debug_generic_expr (rhs1_type);
@@ -1417,8 +1417,8 @@ maybe_rewrite_mem_ref_base (tree *tp, bitmap suitable_for_renaming)
if (TREE_CODE (TREE_TYPE (sym)) == VECTOR_TYPE
&& useless_type_conversion_p (TREE_TYPE (*tp),
TREE_TYPE (TREE_TYPE (sym)))
- && multiple_of_p (sizetype, TREE_OPERAND (*tp, 1),
- TYPE_SIZE_UNIT (TREE_TYPE (*tp))))
+ && multiple_p (mem_ref_offset (*tp),
+ wi::to_poly_offset (TYPE_SIZE_UNIT (TREE_TYPE (*tp)))))
{
*tp = build3 (BIT_FIELD_REF, TREE_TYPE (*tp), sym,
TYPE_SIZE (TREE_TYPE (*tp)),
@@ -1504,8 +1504,8 @@ non_rewritable_mem_ref_base (tree ref)
&& known_ge (mem_ref_offset (base), 0)
&& known_gt (wi::to_poly_offset (TYPE_SIZE_UNIT (TREE_TYPE (decl))),
mem_ref_offset (base))
- && multiple_of_p (sizetype, TREE_OPERAND (base, 1),
- TYPE_SIZE_UNIT (TREE_TYPE (base))))
+ && multiple_p (mem_ref_offset (base),
+ wi::to_poly_offset (TYPE_SIZE_UNIT (TREE_TYPE (base)))))
return NULL_TREE;
/* For same sizes and zero offset we can use a VIEW_CONVERT_EXPR. */
if (integer_zerop (TREE_OPERAND (base, 1))
@@ -1596,8 +1596,8 @@ non_rewritable_lvalue_p (tree lhs)
&& known_ge (mem_ref_offset (lhs), 0)
&& known_gt (wi::to_poly_offset (TYPE_SIZE_UNIT (TREE_TYPE (decl))),
mem_ref_offset (lhs))
- && multiple_of_p (sizetype, TREE_OPERAND (lhs, 1),
- TYPE_SIZE_UNIT (TREE_TYPE (lhs)))
+ && multiple_p (mem_ref_offset (lhs),
+ wi::to_poly_offset (TYPE_SIZE_UNIT (TREE_TYPE (lhs))))
&& known_ge (wi::to_poly_offset (TYPE_SIZE (TREE_TYPE (decl))),
wi::to_poly_offset (TYPE_SIZE (TREE_TYPE (lhs)))))
{
@@ -1969,9 +1969,9 @@ execute_update_addresses_taken (void)
&& known_gt (wi::to_poly_offset
(TYPE_SIZE_UNIT (TREE_TYPE (sym))),
mem_ref_offset (lhs))
- && multiple_of_p (sizetype,
- TREE_OPERAND (lhs, 1),
- TYPE_SIZE_UNIT (TREE_TYPE (lhs))))
+ && multiple_p (mem_ref_offset (lhs),
+ wi::to_poly_offset
+ (TYPE_SIZE_UNIT (TREE_TYPE (lhs)))))
{
tree val = gimple_assign_rhs1 (stmt);
if (! types_compatible_p (TREE_TYPE (val),