Fix singlepass for Aarch64

[ptitSeb]

Description

Following debug of ticket #3343 , some issues in the singlepass were found.
In particular, some ADD/SUB opcode involving XSp with large value were not using the correct opcode, and sub xsp, xsp, x17 were encoded as neg xzr, x17 (a.k.a. sub xzr, xzr, x17). By forcing the dummy UXTX modifier, the correct opcode is selected.
Changed both SUB and ADD emiters.
Some alignement function specific to macOS ABI were also adjusted, and Local variable storage/retreival optimized a bit.

[syrusakbary]

Can we add a test verifying the fix? (triggering what was failing)

Otherwise it looks good

[ptitSeb]

Can we add a test verifying the fix? (triggering what was failing)

Otherwise it looks good

I can try using the wasi test suite facility, that seems the more sane way to be sure to reproduce the issue (needs a .rs compile, so comparing between native / wasm as the wasi testdoes seems the best way).

[theduke]

As Syrus said, a test would be great.
We should try to get into the habit of adding regression tests for each bug we fix.

Adding the Rust reproduction sounds good.

Would be nice to also have a minimal .wat test case, but probably not worth spending extra time on that if it's not trivial to recreate.

In the future we might also want to have some (singlepass) codegen snapshot tests.
As in: test that a specific input compiles to the expected Assembly.

[ptitSeb]

As Syrus said, a test would be great. We should try to get into the habit of adding regression tests for each bug we fix.

Adding the Rust reproduction sounds good.

Would be nice to also have a minimal .wat test case, but probably not worth spending extra time on that if it's not trivial to recreate.

In the future we might also want to have some (singlepass) codegen snapshot tests. As in: test that a specific input compiles to the expected Assembly.

Yeah, well.
I have already spent 6h trying to get a "working" minimal test. A test for an architecture (Aarch64) that is NOT tested on the CI.
I'll add the test, but I'm not convinced this is good investment of time.

[theduke]

👍

I actually also wanted to mention above: we really need to get M1 CI up and running.

[ptitSeb]

bors r+

[bors]

Build succeeded:

• Code lint