jit: make sure RSP is 16 byte aligned when we call into rust code

[alessandrod]

This aligns $rsp in emit_rust_call() so that if the rust code uses SIMD instructions that require alignment we don't crash

[codecov-commenter]

Codecov Report

Merging #515 (ccfd894) into main (e7aaea2) will increase coverage by 0.06%.
The diff coverage is 80.00%.

❗ Current head ccfd894 differs from pull request most recent head 147f5f1. Consider uploading reports for the commit 147f5f1 to get more accurate results

@@            Coverage Diff             @@
##             main     #515      +/-   ##
==========================================
+ Coverage   89.63%   89.70%   +0.06%     
==========================================
  Files          23       23              
  Lines       10301    10311      +10     
==========================================
+ Hits         9233     9249      +16     
+ Misses       1068     1062       -6     

Files Changed	Coverage Δ
src/jit.rs	92.18% <80.00%> (-0.18%)	⬇️

... and 1 file with indirect coverage changes

[Lichtso]

I would like to emit an assertion before each rust call to find the offending call sites first and then investigate further. Working on another PR for that.

[alessandrod]

I would like to emit an assertion before each rust call to find the offending call sites first and then investigate further. Working on another PR for that.

I'm working on unit tests for this, so please leave it to me 😊 But the answer is "all the call sites".

We don't try to maintain alignment anywhere. Before doing this patch I actually started changing that, but it seemed like too big a change for a backport. Also since we don't have an explicit stack allocator but we manipulate the stack as we go, it seems like it would be fragile.

[Lichtso]

We don't try to maintain alignment anywhere.

That is not true, while we do lack explicit tests, I have implemented stack padding in various places to reach 16 byte alignment. Some are marked as such in comments, e.g.

rbpf/src/jit.rs

Line 1407 in e7aaea2

self.emit_ins(X86Instruction::push_immediate(OperandSize::S64, -1)); // Used as PC value in error case, acts as stack padding otherwise

[ryleung-solana]

Does this fix need to be merged into main as well?

[Lichtso]

Yes, but we are working on a more detailed analysis of the places which cause misalignment first.