On 8/31/26 12:45 PM, Klimkin, Alexey wrote:
_cpp_find_header_unit closes the source descriptor after resolving a
header unit, but records zero instead of the established -1 sentinel.
A later textual include through an alternate pathname consequently
skips reopening the source and reads stdin. This gives a misleading
"file is shorter than expected" warning followed by redefinition
errors. The existing fd > 0 check also misses a real header descriptor
when open returns zero because stdin was closed.
Close every valid descriptor and restore -1 after closing it.
The regression imports a header unit and then includes the same
physical file through a sys/../ spelling. It fails on unpatched trunk
and compiles, links, and runs after the fix.
Bootstrapped and regression-tested on x86_64-pc-linux-gnu, configured with
--disable-multilib --enable-languages=c,c++,fortran,objc,obj-c++,lto.
Compared against an unpatched build of the same parent commit: identical
FAIL/XPASS sets, and 15 new PASSes from the added test.
Per https://gcc.gnu.org/ai-policy.html: this change was LLM-assisted.
I have reviewed and understand the change and can
answer questions about it.
libcpp/
PR c++/127097
* files.cc (_cpp_find_header_unit): Close descriptor zero and set
the closed descriptor sentinel to -1.
gcc/testsuite/
PR c++/127097
* g++.dg/modules/pragma-once-hu-1_a.H: New test.
* g++.dg/modules/pragma-once-hu-1_b.C: New test.
Pushed, thanks!
Jason
