David Nadlinger

Can't reproduce on https://github.com/m-labs/artiq/commit/3d25092cbd7db5d4532df10d8d6494a490423f36 on Kasli `tester` either; thousands of `artiq_run` iterations proceed fine.

@fanmingyu212 Would you mind preparing an ELF file with the dummy variable workaround as well? This does seem likely to be related to the toolchain, and I don't have a...

@fanmingyu212: Thanks for the files, this it helpful. I see you are using LLD 15.0 – where are you getting that from? Is that a default package from upstream Conda?...

@fanmingyu212 Okay, many thanks for the tests, that confirms my suspicion (and explains why it doesn't occur on Nix), which gives me a handle on reproducing the issue. Will look...

Occurs on LLD 14 and 15, but not LLD 13; also on Linux/Nix. Now to actually debug what's going on here…

When building the runtime using LLD 14 (maybe even 13), also get: ``` artiq.coredevice.comm_kernel.LoadError: cannot load kernel: symbol lookup error: __artiq_personality ``` Looking into what's going on with the linker...

Okay, this has been a wild goose chase. My current theory (at least for the kernel-side issue, i.e. when building the firmware with LLD 11) is that LLD 14 orders...

In fact, my simulation is so much limited by the sparse-matrix-dense-vector multiplication inside the `expmv_fun` loop that I hacked together a patch to use the `SuiteSparse`/`CHOLMOD` implementation of that for...

Going to submit the `iamax` optimization to `Base`. Regarding the 1-norm estimation: This would definitely be useful for me in the future, but since I can cache the `select_taylor_degree` call...

D'oh, forgot that the notion of the absolute value of a complex number in BLAS is `abs(re) + abs(im)` instead of the mathematical definition. The `iamax` optimization is thus not...