https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119085
Bug ID: 119085
Summary: tree-sra generates wrong code for unions containing
structs with holes
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: pipcet at protonmail dot com
Target Milestone: ---
Created attachment 60633
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=60633&action=edit
test case
The attached test case asserts when compiled with trunk gcc:
$ gcc -O3 -o gcc-001 gcc-001.c -fdump-tree-sra
$ ./gcc-001
gcc-001: gcc-001.c:32: main: Assertion `u.without_hole.y == -1' failed.
Aborted
I believe it should succeed, as it does when I specify -fno-tree-sra.
Po Lu identified the problem to most likely reside in
sort_and_splice_var_accesses, which chooses as the representative access for an
access group one which uses the "with_hole" structure, which has a hole, which
results in an incomplete copy being made.
This is similar to PR 58416, which introduced types_risk_mangled_binary_repr_p,
except that it is aggregate types which aren't "compatible" rather than integer
types.
A quick fix would be to replace types_risk_mangled_binary_repr_p by
!types_compatible_p, but that uses the language backend's definition of
compatibility, whereas we're only interested in the types having equivalent
"bit masks", meaning that the same bits are copied.
My apologies if I messed up in reducing the test case, but I'm convinced the
bug is real.
