Posted a bit hastily, since a few minutes later I found this: BLAS calls not assigned to callers in profiler · Issue #33605 · JuliaLang/julia · GitHub
Seems that what’s happening is that the BLAS calls just don’t get assigned to the proper callers. Indeed, when I look at Profile.print(; C=true), at the bottom, outside the usual call tree are my missing BLAS counts (from different run than the above, so numbers don’t match):
4316╎4316 ...-1.4.0/bin/../lib/julia/libopenblas64_.so:?; dgemm_kernel_HASWELL
These don’t appear without C=true, which is why I at first didn’t spot them at all.