tree-optimization/106904 - bogus -Wstringopt-overflow with vectors
Commit Message
The following avoids CSE of &ps->wp to &ps->wp.hwnd confusing
-Wstringopt-overflow by making sure to produce addresses to the
biggest container from vectorization. For this I introduce
strip_zero_offset_components which turns &ps->wp.hwnd into
&(*ps) and use that to base the vector data references on.
That will also work for addresses with variable components,
alternatively emitting pointer arithmetic via calling
get_inner_reference and gimplifying that would be possible
but likely more intrusive.
This is by no means a complete fix for all of those issues
(avoiding ADDR_EXPRs in favor of pointer arithmetic might be).
Other passes will have similar issues.
In theory that might now cause false negatives.
Bootstrap and regtest running on x86_64-unknown-linux-gnu.
Any opinion?
Thanks,
Richard.
PR tree-optimization/106904
* tree.h (strip_zero_offset_components): Declare.
* tree.cc (strip_zero_offset_components): Define.
* tree-vect-data-refs.cc (vect_create_addr_base_for_vector_ref):
Strip zero offset components before building the address.
* gcc.dg/Wstringop-overflow-pr106904.c: New testcase.
---
.../gcc.dg/Wstringop-overflow-pr106904.c | 30 +++++++++++++++++++
gcc/tree-vect-data-refs.cc | 12 ++++----
gcc/tree.cc | 12 ++++++++
gcc/tree.h | 1 +
4 files changed, 50 insertions(+), 5 deletions(-)
create mode 100644 gcc/testsuite/gcc.dg/Wstringop-overflow-pr106904.c
Comments
On 12/7/22 06:54, Richard Biener via Gcc-patches wrote:
> The following avoids CSE of &ps->wp to &ps->wp.hwnd confusing
> -Wstringopt-overflow by making sure to produce addresses to the
> biggest container from vectorization. For this I introduce
> strip_zero_offset_components which turns &ps->wp.hwnd into
> &(*ps) and use that to base the vector data references on.
> That will also work for addresses with variable components,
> alternatively emitting pointer arithmetic via calling
> get_inner_reference and gimplifying that would be possible
> but likely more intrusive.
>
> This is by no means a complete fix for all of those issues
> (avoiding ADDR_EXPRs in favor of pointer arithmetic might be).
> Other passes will have similar issues.
>
> In theory that might now cause false negatives.
>
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
>
> Any opinion?
>
> Thanks,
> Richard.
>
> PR tree-optimization/106904
> * tree.h (strip_zero_offset_components): Declare.
> * tree.cc (strip_zero_offset_components): Define.
> * tree-vect-data-refs.cc (vect_create_addr_base_for_vector_ref):
> Strip zero offset components before building the address.
>
> * gcc.dg/Wstringop-overflow-pr106904.c: New testcase.
So you're just canonicalizing to the widest container for a zero offset
access. While it may not fix everything, that seems like a good thing
in general when we can do so. I wouldn't be surprised if other passes
could do the same thing to fix some of the missed CSE opportunities.
jeff
new file mode 100644
@@ -0,0 +1,30 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -Wstringop-overflow -fno-vect-cost-model" } */
+
+struct windowpos
+{
+ int hwnd;
+ int hwnd2;
+};
+
+struct packed_windowpos
+{
+ int hwnd;
+ int pad1;
+ int hwnd2;
+ int pad2;
+};
+
+struct packed_structs
+{
+ struct packed_windowpos wp;
+};
+
+void func(struct packed_structs *ps)
+{
+ struct windowpos wp;
+
+ wp.hwnd = ps->wp.hwnd;
+ wp.hwnd2 = ps->wp.hwnd2;
+ __builtin_memcpy(&ps->wp, &wp, sizeof(wp)); /* { dg-bogus "into a region" } */
+}
@@ -4845,11 +4845,13 @@ vect_create_addr_base_for_vector_ref (vec_info *vinfo, stmt_vec_info stmt_info,
if (loop_vinfo)
addr_base = fold_build_pointer_plus (data_ref_base, base_offset);
else
- {
- addr_base = build1 (ADDR_EXPR,
- build_pointer_type (TREE_TYPE (DR_REF (dr))),
- unshare_expr (DR_REF (dr)));
- }
+ addr_base = build1 (ADDR_EXPR,
+ build_pointer_type (TREE_TYPE (DR_REF (dr))),
+ /* Strip zero offset components since we don't need
+ them and they can confuse late diagnostics if
+ we CSE them wrongly. See PR106904 for example. */
+ unshare_expr (strip_zero_offset_components
+ (DR_REF (dr))));
vect_ptr_type = build_pointer_type (TREE_TYPE (DR_REF (dr)));
dest = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, base_name);
@@ -12014,6 +12014,18 @@ strip_invariant_refs (const_tree op)
return op;
}
+/* Strip handled components with zero offset from OP. */
+
+tree
+strip_zero_offset_components (tree op)
+{
+ while (TREE_CODE (op) == COMPONENT_REF
+ && integer_zerop (DECL_FIELD_OFFSET (TREE_OPERAND (op, 1)))
+ && integer_zerop (DECL_FIELD_BIT_OFFSET (TREE_OPERAND (op, 1))))
+ op = TREE_OPERAND (op, 0);
+ return op;
+}
+
static GTY(()) tree gcc_eh_personality_decl;
/* Return the GCC personality function decl. */
@@ -5373,6 +5373,7 @@ extern bool tree_nop_conversion_p (const_tree, const_tree);
extern tree tree_strip_nop_conversions (tree);
extern tree tree_strip_sign_nop_conversions (tree);
extern const_tree strip_invariant_refs (const_tree);
+extern tree strip_zero_offset_components (tree);
extern tree lhd_gcc_personality (void);
extern void assign_assembler_name_if_needed (tree);
extern bool warn_deprecated_use (tree, tree);