https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99218
anlauf at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
Known to fail| |8.4.1
--- Comment #3 from anlauf at gcc dot gnu.org ---
Playing with the reduced testcase:
fTmp = matmul (transpose (G), lambda(::2))
is clean while
fTmp = matmul (transpose (G), lambda(::1))
is not.
It seems we run into a highly optimized code for rank-2 times rank-2.
The following patch fixes the testcase:
diff --git a/libgfortran/m4/matmul_internal.m4
b/libgfortran/m4/matmul_internal.m4
index 13fd7696238..0e96207a0fc 100644
--- a/libgfortran/m4/matmul_internal.m4
+++ b/libgfortran/m4/matmul_internal.m4
@@ -192,7 +192,8 @@ sinclude(`matmul_asm_'rtype_code`.m4')dnl
}
}
- if (rxstride == 1 && axstride == 1 && bxstride == 1)
+ if (rxstride == 1 && axstride == 1 && bxstride == 1
+ && GFC_DESCRIPTOR_RANK (b) != 1)
{
/* This block of code implements a tuned matmul, derived from
Superscalar GEMM-based level 3 BLAS, Beta version 0.1
