cranelift: fma_f32 tests fail on GitHub's `windows-latest` image

[alexcrichton]

I was testing out using the windows-latest image with Nick's stack walking PR because I think that should fix a longstanding blocker for updating the CI image.

In doing so though I get:

FAIL filetests\filetests\runtests\fma.clif: interpret
Caused by:
    Failed test: run: %fma_f32(0x1.060000p7, 0x1.340488p7, 0x1.310e42p4) == 0x1.3b88e6p14, actual: 0x1.3b88e8p14

I'm not sure what that corresponds to in the test itself but I suspect that there's just a difference between the old fma symbol on Windows and the new.

cc @afonso360

[afonso360]

This is interesting.

I tried to make a CI run of fma.clif alone and it did pass the tests. (This is based on the main branch from yesterday).

I briefly read the stack walking changes, but I don't see any compile flags or other build changes that would explain this difference.

I pulled your fast-stack-walking2 branch and ran it again on CI. It fails only on windows-latest x86_64-pc-windows-gnu but not on windows-latest. But I don't understand why it works on main, and not with fast-stack-walking2 applied.

Edit: This is not true, I accidentally deleted the CARGO_BUILD_TARGET env var in the original run. So the only difference is that this fails in windows-latest x86_64-pc-windows-gnu instead of windows-latest. Both on main and fast-stack-walking2

---

Note:

I also had some issues building the cmake examples on windows-latest
Which don't seem to trigger on the stack walking branch (possibly due to #4369 ?) I haven't investigated this too far.

I can file an issue for this if you think it's relevant.

[afonso360]

I suspect this is related to fma/fmaf bugs on MinGW-w64.

See: https://sourceforge.net/p/mingw-w64/bugs/848/

I compiled the test case in that bug report, with the additional inputs that we are seeing in the interpreter and it also failed. (Called Testcase[6] on the output)

It does not fail that test case on windows-2019, however it fails some others that we don't have in our test suite.

It looks like fma/fmaf is buggy in MinGW. That bug report has been open for 2 years. What should we do here? Grab a polyfill from some other library?

Edit: We can trigger this on the current CI, see: #4517

[TheGreatRambler]

I also had some issues building the cmake examples on windows-latest

That is because cmake --target for tests is not consistent between compilers. --target test works for gcc, mingw, macos. --target RUN_TESTS works for msvc, really dumb platform difference. It looks like you already know updating should fix it.

[alexcrichton]

This has since been resolved so I think this is good to close.