[BUG]: Internal compiler error with MSVC 2022 when building `nb_func.h`

[jhale]

Problem description

We have recently started building our library Basix on Windows using MSVC 2022 using GitHub Actions. We are using scikit-build-core to orchestrate the CMake build and we build nanobind statically. We use vcpkg to bring in LAPACK and BLAS.

After switching to nanobind 2.0.0 we are consistently triggering an internal compiler error when building nanobind statically. The error points to L42 in nb_func.cpp, but that line looks relatively innocent.

https://github.com/FEniCS/basix/actions/runs/9404421938/job/25903298628#step:5:763
https://github.com/wjakob/nanobind/blob/master/src/nb_func.cpp#L42

We don't have Windows machines locally so we haven't tried reproducing this locally. Your CI seems to be running green on MSVC 2022. Any ideas?

Reproducible example code

No response

[miklhh]

We have been getting the same internal compiler error pointing to L42 in nb_func.cpp since about a day ago.

[wjakob]

That's quite mysterious since it doesn't seem to happen in the nanobind CI. Two requests:

1. Could you please open an issue with Microsoft? The true source of these issues is never fixed unless escalated to the compiler vendor.
2. The failure is either happening because of a change in the compiler (e.g. update) or because of a change in nanobind. Do you know which one it is? If the latter, can you bisect it to a specific commit? (That said, even in that case, the issue should be reported to Microsoft)

[miklhh]

Regarding your second point @wjakob

We are using the 2.0.0 tag point release from wjacob/nanobind. We started noticing this error roughly from yesterday. I think it's safe to say that something upstream, either in the MSVC compiler or in the GitHub Windows Server runner image, has changed.

Possibly related issue: actions/runner-images#10004

[miklhh]

@wjakob, is it possible for you to trigger a windows-latest build job for your nanobind Tests job? From looking around a bit over at actions/runner-images, there seems to be a bit of talk going on that something has happened to that image within the last 24 hours. I'm curious if your test build can trigger this internal compiler error.

[nicholasjng]

To throw another datapoint into the mix:

I maintain nanobind-bazel, a Bazel ruleset for Python bindings builds with nanobind, and my CI (which consists of building @wjakob 's nanobind_example across multiple Python versions with the respective installed C++ toolchain on every OS) is still entirely green on 20240603.1.0: https://github.com/nicholasjng/nanobind-bazel/actions/runs/9416682418/job/25940329652#step:1:9

EDIT: I should also say that this is built with nanobind v2.0.0.

[jhale]

Tried on a fork https://github.com/jhale/nanobind, doesn't trigger the issue on either a slightly tweaked v2.0.0 or a current main.

[jhale]

Could you tell us more about your project @miklhh ?

[metab0t]

https://developercommunity.visualstudio.com/t/Internal-compiler-error-w-nanobind-v19/10662883?sort=newest

@jhale The issue has been reported and fixed, but they haven't released the patch.

The bug is introduced in MSVC 19.40.xxxxx

[jhale]

Thanks @metab0t - Even though this is not an issue with nanobind itself, I propose we leave this issue open until the fix is rolled out @wjakob?

[metab0t]

@jhale I found out that windows-2019 image can be used instead of windows-latest as a temporary solution to compile nanobind.

[henryiii]

Also see pypa/cibuildwheel#1875 for another.

[Yikai-Liao]

It's quite a strange bug. Since I could build my lib successfully on my win10 laptop with vs2022. But on github action, it just failed. I don't quite understand what's the difference.