AOT_code_discussion_07_23_2026.txt

gistfile1.txt

Notes from code discussion:

- Make AOT loading infallible.

    - Baseline interpreter must always be present. Other artifacts
      (ICs/ Baseline functions) can still fail to load since corpus is
      finite. More MOZ_ASSERTS.

- Remove aot blob YAML field for "kind", becomes internally generated. 

- ImmGCPtr overloads are a misnomer, we only need them for 5 AtomsZone
  pointers. Change from slot list traversal to whitelist.

- AOT preamble trampoline sets the AOT table base into a pass register
  for every transition point into baseline. Explore a "C++ way". Can
  we use C++ calling convention to setup the transition for us somehow?
    
    - Recall the issue here is that AOT artifacts are all position
      independent and can not embed _any_ runtime pointers. The AOT
      indirection table is filled at JitRuntime init. For any AOT
      blob, it is the key to attain all other pointers via the
      AOT_SLOTS contract.

      We need some way to get the AOT table base into BaselineFrame
      and StubFrame. Currently the frame builders _expect_ a specifc
      register to contain the table base, which the preamble trampoline
      is required to setup.

      The other way we tried was TLS (recall from maybe 2 months ago or
      so). TLS in AOT does not work, offsets are sensitive to dynamic
      library loading order and are brittle in offline settings. The
      preamble trampoline works for now, but is not overly elegant.

- YAML descriptions of Baseline artifact meta-data causes duplicate
  source of truth with corresponding C++ classes. If someone adds a
  new blinterp offset field for example, AOT serde breaks. Explore
  some options.
    
 1. Generate the C++ class describing baseline metadata from YAML.
    This consolidates definitions, but is relatively invasive.
    Definitely makes sense for this fork, but for upstream, maybe not.
 
 2. Other (TODO)