https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124111
Sam James <sjames at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Resolution|WORKSFORME |---
Known to fail| |15.2.1, 16.0
Target Milestone|16.0 |15.3
Status|RESOLVED |UNCONFIRMED
Known to work| |14.3.1
Summary|[16 regression] |[15/16 regression]
|lapack-3.12.1 miscompiled |lapack-3.12.1 miscompiled
| |since r15-2820
CC| |tsamismanolis at gmail dot com
--- Comment #3 from Sam James <sjames at gcc dot gnu.org> ---
OK, it actually started with 15's r15-2820-gab18785840d7b8.
The miscompiled object is BLAS/SRC/CMakeFiles/blas_obj.dir/cgemmtr.f.o, which
gets built with:
```
/usr/bin/gfortran -O2 -march=x86-64-v3 -fno-vect-cost-model -ggdb3 -O2
-march=x86-64-v3 -fno-vect-cost-model -ggdb3 -fPIC -frecursive -c
/home/sam/bugs/lapack/lapack-3.12.1/BLAS/SRC/cgemmtr.f -o
CMakeFiles/blas_obj.dir/cgemmtr.f.o
```
Rebuilding that with -fno-tree-vectorize and make -C build to get libblas
relinked makes the test pass w/
```
LD_LIBRARY_PATH=/tmp/gcc-pfx/lib64
/home/sam/bugs/lapack/lapack-3.12.1/build/bin/xblat3c <
/home/sam/bugs/lapack/lapack-3.12.1/BLAS/TESTING/cblat3.in
```
