test failures on x86 32bit (i686) #1044
Comments
|
Interesting, I don't see the same behavior when I run locally (using |
|
This is on Fedora rawhide (development branch). |
|
CXXFLAGS used: |
|
Interestingly, |
|
Yes, #1046 only canonicalized NaNs. It looks like |
|
Testing w/ gcc in Compiler Explorer seems to produce the correct value too (negative zero). |
|
I don't know how faithful Compiler Explorer is and I'm no compiler expert. Two things may explain the difference:
I'm only guessing though. The issue is 100% reproducible on Fedora 29. |
|
I did some more builds and it looks like it's reproducible on Fedora 30 with gcc-9.0.1, too. |
|
Thanks for the info -- I'll get an image to test against. |
|
Meanwhile, I can no longer reproduce this on Fedora 30+. Fedora 29 is nearing end of life, so I don't think it's worth spending time on. |
|
Thanks for the update @rathann. |
|
The following 5 tests are still failing with 1.0.15 on i686 on Fedora 31 (gcc-9.3.1): and I've got 5 new failures: |
|
Still reproducible with 1.0.23 on Fedora rawhide. Same 10 failures are occurring. |
|
New failures on Fedora rawhide (36) with gcc-12 pre-release: Log extract: I'm not sure what's going on with spec/local_tee.txt, because @binji any clues? |
|
Some more failures in 1.0.30: |
|
Two new test failures in 1.0.33 test suite on i686: and |
|
I'm seeing similar errors in https://github.com/turbolent/w2c2. Has anyone investigated why these tests fail? |
|
I think the root cause might be something like, "The SSE2 floating-point instructions largely conform to IEEE 754 (and the Wasm semantics wrt canonical/arithmetic NaNs), but x87 doesn't. So if somebody cares about i686 (which is a 32-bit x86 architecture without SSE2, so the compiler is generating x87 instructions for floating-point), the declarations file would need a lot more TLC to make the floating-point operations match the IEEE and Wasm semantics." If we just care about 32-bit x86 (but not necessarily the base i686) and are willing to restrict to chips that have SSE2 (e.g. Pentium-M, Pentium 4, and later), then compiling with If somebody cares about conforming to the spec exactly on pre-SSE2 32-bit x86, it would be helpful to have commit-by-commit CI on this architecture so we can avoid screwing it up in the future. And somebody will need access to a test machine and will have to do the work of adapting the declarations file to make it conform on x87. |
|
I wonder if we could #error out with a useful error message when building for x86 without SSE2? Perhaps with a link it this bug |
FWIW, Fedora has required SSE2 for years and we don't even build x86 32-bit kernels anymore. |
|
It's cool that Fedora's building WABT with those CFLAGS, but to get the wasm2c tests to build their output with those CFLAGS, I think they have to be added to the If that doesn't do the trick, then bigger picture, I think for this to get fixed, we'd probably need access to a test machine (and ideally commit-by-commit CI) with this architecture to figure out what's really going on and apply the necessary TLC to the declarations... :-/ |
Good point! I'll try that, thanks.
You can do this on any x86_64 with multiarch gcc and i686 flags. This could be set up in GitHub Actions to run in a 32bit Debian container, for example. Fedora x86_64 gcc is capable of building 32-bit binaries, too. |
|
New failure with 1.0.36: |
FWIW, running the tests with |
|
Just got the same issue on debian wabt package, and |
Running the 1.0.10 testsuite (
test/run-tests.py) on x86 32bit (i686) yields the following failed tests:I suspect all but the first are due to lower float precision.
The text was updated successfully, but these errors were encountered: