wasm2c/spec/simd_address.txt fails on non-x86

[rathann]

When running the test suite of wabt-1.0.33 on aarch64 on Fedora 38, I get this (new) test failure:

- test/wasm2c/spec/simd_address.txt
  expected error code 0, got 1.
  STDERR MISMATCH:
  --- expected
  +++ actual
  @@ -0,0 +1,11 @@
  +Traceback (most recent call last):
  +  File "/builddir/build/BUILD/wabt-1.0.33/test/run-spec-wasm2c.py", line 648, in <module>
  +    sys.exit(main(sys.argv[1:]))
  +             ^^^^^^^^^^^^^^^^^^
  +  File "/builddir/build/BUILD/wabt-1.0.33/test/run-spec-wasm2c.py", line 642, in main
  +    utils.Executable(main_exe, forward_stdout=True).RunWithArgs()
  +  File "/builddir/build/BUILD/wabt-1.0.33/test/utils.py", line 95, in RunWithArgs
  +    raise error
  +utils.Error: Error running "out/test/wasm2c/spec/simd_address/simd_address" (1):
  +None
  +out/test/wasm2c/spec/simd_address/simd_address-main.c:499: assertion failed: expected w2c_simd__address__1__wasm_v1280x2Eload_offset_65521(&simd__address__1__wasm_instance) to trap.
  STDOUT MISMATCH:
  --- expected
  +++ actual
  @@ -1 +1 @@
  -42/42 tests passed.
  +41/42 tests passed.

[keithw]

Thanks as always for these reports! I think this particular one is expected and "disclosed" here: https://github.com/WebAssembly/wabt/blob/main/src/template/wasm2c_simd.declarations.c#L6

The root issue here is that:

• out-of-bounds accesses on Wasm must trap deterministically, but
• wasm2c tries to use the MMU to detect out-of-bounds accesses and raise a signal, but this kind of side-effect isn't something the compiler cares to preserve if the load can otherwise be optizmed out, so
• wasm2c needs to force the C compiler not to optimize out a load, which is done with an empty inline asm statement that marks the result of the load as "read,"
• but (unlike with a 32-bit or 64-bit register), there's no architecture-independent way to describe a 128-bit SIMD register, so at present we only have this for SIMD on x86.

Would happily take a contribution of the appropriate asm for aarch64 (possibly "w" per https://gcc.gnu.org/onlinedocs/gcc/Machine-Constraints.html) if somebody is able to test and say it works... I do wish we had CI tests on aarch64 given the interest in this architecture.

[sbc100]

Should we update the TODO such that it will #warning or #error when building on unsupported arch?

Also, I guess we should skip these tests on those platforms too..

[sbc100]

@rathann do you actually care about using the SIMD support in wasm2c on arm64 or are you just interested in getting the tests to pass?

[rathann]

Personally, it's the latter. We do have a growing interest in WebAssembly in Fedora, so others might actually be interested. I'll ask the other SIG members.

[rathann]

I see this on ppc64le and s390x, too.

[rathann]

Looks like this is fixed on s390x, now that I'm running the tests using cmake --target run-tests.

[SoniEx2]

indeed, s390x doesn't use guard pages and falls back to range checking, while this issue only occurs with guard pages. #2545 would fix it everywhere.