ci: try running coverage for all OSes

[davidhewitt]

According to https://docs.codecov.com/docs/merging-reports if we upload multiple coverage reports they all get automatically merged. So I'm opening this PR to see what happens if I try that!

[davidhewitt]

Hmmm @taiki-e have you seen errors like these from running cargo-llvm-cov on windows before?

libpyo3-8bd03e2eb0b9a94d.rlib(pyo3-8bd03e2eb0b9a94d.pyo3.e32d66c5-cgu.2.rcgu.o) : fatal error LNK1227: conflicting weak extern definition for '_RNvNtCs9ZwDYBEfEGf_4pyo311type_object12get_tp_allocB3_'.  New default '.weak._RNvNtCs9ZwDYBEfEGf_4pyo311type_object12get_tp_allocB3_.default._RINvMNtCsk1Jcy3OMEV4_5alloc5sliceSh9to_vec_inNtNtB5_5alloc6GlobalECs9ZwDYBEfEGf_4pyo3' conflicts with old default '.weak._RNvNtCs9ZwDYBEfEGf_4pyo311type_object12get_tp_allocB3_.default._RINvMNtCsjJUIXhzXbX_4core3stre13get_uncheckedINtNtNtB5_3ops5range5RangejEECs9ZwDYBEfEGf_4pyo3' in libpyo3-8bd03e2eb0b9a94d.rlib(pyo3-8bd03e2eb0b9a94d.pyo3.e32d66c5-cgu.9.rcgu.o)

I can reproduce locally with latest nightly if I run cargo llvm-cov --package pyo3 pyo3-build-config pyo3-macros-backend pyo3-macros --no-report on my Windows machine.

[taiki-e]

@davidhewitt That is a rustc bug: rust-lang/rust#85461

I'll implement a workaround on the cargo-llvm-cov side.

[davidhewitt]

Aha brilliant, thank you!

[taiki-e]

@davidhewitt I've released cargo-llvm-cov 0.1.11, which implements a workaround for that issue. Could you try that version?

[taiki-e]

Btw, cargo-llvm-cov has both .tar.gz and .zip distribution on windows, so you can unify the installation script on both windows and unix (with shell: bash). For example:

host=$(rustc -Vv | grep host | sed 's/host: //')
curl -fsSL https://github.com/taiki-e/cargo-llvm-cov/releases/download/v${CARGO_LLVM_COV_VERSION}/cargo-llvm-cov-"$host".tar.gz | tar xzf - -C ~/.cargo/bin

[davidhewitt]

Thanks @taiki-e, the windows job working great now!

I think this looks like codecov is merging reports? Not super clear, I'm going to go ahead and merge this, might play around more another day.

[taiki-e]

I think this looks like codecov is merging reports? Not super clear, I'm going to go ahead and merge this, might play around more another day.

This may be due to the different path between windows and unix. (I feel it's a bit odd that codecov can't handle it, but we may need to do what cargo-llvm-cov's test is doing)

[davidhewitt]

Hmm interesting. We're not generating JSON coverage here though; we're using --lcov. I don't know much about lcov format and how easy it would be to do that replace.

I have since noticed that codecov does update the coverage number reported in the PR as the separate jobs finish, so it's quite possible it is actually doing the right thing already.