Stream: git-wasmtime

Topic: wasmtime / Issue #2267 Fix AArch64 ABI to respect half-ca...


view this post on Zulip Wasmtime GitHub notifications bot (Oct 06 2020 at 20:25):

cfallin commented on Issue #2267:

Thanks! Updated based on comments. I definitely like the renaming in terms of explicit actions ("clobbered by call", "saved in prologue") in the ABI impls. The only open question is the naming of is_included_in_clobbers() in the MachInst interface; I'm curious if you have any other good names in mind. Open to whatever makes sense!

view this post on Zulip Wasmtime GitHub notifications bot (Oct 06 2020 at 21:29):

akirilov-arm commented on Issue #2267:

I also had a quick look at the changes and they are good, but now I have realized that there is a gap in the tests - there is nothing that covers the callee side. However, that is definitely a job for another PR, in particular one that is going to fix function prologues and epilogues, so that they deal only with the lower 64 bits of the SIMD & FP registers. At the very least we should have a function with a chain of operations such as v1 = v0 + v0, v2 = v1 + v1, ..., f(v0, v1, v2, ...) to force the compiler to use the callee-saved registers, possibly mixing data types (f32, f64 and i8x16).

view this post on Zulip Wasmtime GitHub notifications bot (Oct 06 2020 at 21:45):

cfallin commented on Issue #2267:

@akirilov-arm good point; thanks! I just updated the PR to include a test (prologue.clif) that does exactly that, and checks that all of the callee-saves are saved (lower halves of v8-v15, covered by saving all of v8-v15 currently). Let me know if this covers what you had in mind.

view this post on Zulip Wasmtime GitHub notifications bot (Oct 06 2020 at 22:11):

akirilov-arm commented on Issue #2267:

@cfallin Yes, and it is simpler than what I had in mind, which is even better. I think that together with the other tests it will be a good exercise for an optimal with respect to the AAPCS64 implementation (trying to be a little bit forward-thinking here), and yet it demonstrates the current issues that are simpler, namely handling the full registers and the lack of paired loads and stores.


Last updated: Nov 22 2024 at 16:03 UTC