]> git.ipfire.org Git - people/ms/gcc.git/commit
tree-optimization/106904 - bogus -Wstringopt-overflow with vectors
authorRichard Biener <rguenther@suse.de>
Wed, 7 Dec 2022 13:42:24 +0000 (14:42 +0100)
committerRichard Biener <rguenther@suse.de>
Sun, 11 Dec 2022 13:34:44 +0000 (14:34 +0100)
commitf8d136e50e6f82cba793483d910a2b2643108508
tree3339f7ecb0d9cd8b0b2bb00b2f8990e99dcc0a89
parent045592f665bcb67b75dc6b86badbe2fd44aed3e6
tree-optimization/106904 - bogus -Wstringopt-overflow with vectors

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.

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/testsuite/gcc.dg/Wstringop-overflow-pr106904.c [new file with mode: 0644]
gcc/tree-vect-data-refs.cc
gcc/tree.cc
gcc/tree.h