Port RPython simplify_graph parity and track remaining graph/codewriter parity gaps

[youknowone]

Summary

Static parity review found that Pyre still lacks several RPython/PyPy graph/codewriter normalization properties that are assumed before flatten_graph runs upstream.

This issue tracks the newly identified work as one umbrella issue. Existing issues cover the broader tracer/codewriter architecture (#50, #73, #97) and tuple specialization (#59); this issue is specifically for the remaining graph/codewriter parity gaps found during the review.

Background

RPython's normal pipeline is:

build_flow(func)
  -> simplify_graph(graph)
  -> perform_register_allocation(graph)
  -> flatten_graph(graph, regallocs)
  -> compute_liveness(ssarepr)
  -> assemble(ssarepr)

flatten_graph assumes it receives a graph already normalized by rpython/translator/simplify.py::simplify_graph. Pyre currently has a split path where the Python-bytecode walker emits SSARepr inline while graph dual-writes are later used for regalloc/canonical checks. That split makes missing graph normalization visible as label/renaming/liveness repair logic that PyPy does not need.

The recent dead-forwarder work is one example: eliminate_empty_blocks graph collapse is PyPy-orthodox, but Pyre also needs a pyre-only TLabel retarget bridge because the walker may already have emitted labels before the graph reaches the final simplified shape.

Scope

1. Port missing simplify_graph passes

PyPy/RPython runs the pass list in rpython/translator/simplify.py before codewriter flattening:

• transform_dead_op_vars
• eliminate_empty_blocks
• remove_assertion_errors
• remove_identical_vars_SSA
• constfold_exitswitch
• remove_trivial_links
• SSA_to_SSI
• coalesce_bool
• transform_ovfcheck
• simplify_exceptions
• transform_xxxitem
• remove_dead_exceptions

Pyre should port the passes that are relevant to its graph/codewriter pipeline, with explicit structural-adaptation notes where a 1:1 port is impossible.

Initial priority candidates:

• remove_trivial_links
• remove_identical_vars_SSA
• SSA_to_SSI
• simplify_exceptions
• constfold_exitswitch

Acceptance criteria for each ported pass:

• document the corresponding RPython source function and any intentional adaptation;
• add focused graph-shape tests where feasible;
• ensure downstream regalloc/flatten sees the normalized shape;
• avoid adding walker-only byte-stream repairs when a graph-level normalization is the PyPy source of truth.

2. Match PyPy regalloc coalesce semantics

Static review found an existing parity gap in CFG coalescing.

PyPy regalloc.py coalesces after dependency/interference analysis and rejects a candidate when the two variables interfere. Pyre currently pre-merges CFG link.args <-> target.inputargs pairs alongside walker pin pairs in a way that can bypass the equivalent interference check.

Work needed:

• compare Pyre coalescing against PyPy regalloc.py _try_coalesce semantics;
• avoid unconditional pre-merge for CFG pairs that PyPy would reject;
• preserve or replace performance-sensitive walker pinning with an orthodox equivalent;
• measure whether extra ref_copy emission comes from correct PyPy parity or from a remaining walker/regalloc mismatch.

This is related to #50 and #73, but it can be investigated independently at the graph/regalloc layer.

3. Investigate unmarked JitCode label panic

A pre-existing panic has been observed in the assembler patching path:

majit/majit-metainterp/src/jitcode/assembler.rs:4121
jitcode label was never marked

This should be tracked here as a graph/codewriter/label parity gap until proven otherwise.

Work needed:

• trace every TLabel / patch allocation path to a corresponding emitted Label / mark_label path;
• classify whether the missing label originates in graph shape, flatten emission order, switch/default handling, trampoline/forwarder handling, or assembler bookkeeping;
• add an invariant or diagnostic that reports the originating opname/label name/graph before the final patch_labels panic;
• determine whether any missing simplify_graph pass makes the problematic label shape reachable.

If investigation proves this is purely backend assembler state unrelated to graph/codewriter parity, split it out into a narrower follow-up issue.

4. Keep structural adaptations explicit

Some Pyre/PyPy differences are structural and should not be forced into a literal port:

• CPython bytecode version and opcode differences;
• Python 3.11/3.14 differences;
• Rust vs RPython implementation differences;
• GIL/free-threading model differences;
• Pyre's current inline walker / Python-PC coordinate system, until Eliminate pc_map and per-PC labels by switching the tracer to interpret JitCode bytecode #73 retires it.

Each pass or repair should say whether it is:

• direct PyPy parity;
• a temporary bridge caused by Pyre's current pipeline;
• a structural adaptation that should remain;
• a workaround that should disappear after majit-translate auto-emit SSA per opcode handler #50/Eliminate pc_map and per-PC labels by switching the tracer to interpret JitCode bytecode #73/Replace syn-AST front-end with Charon-extracted MIR for the JIT lowering pipeline #97.

Related existing issues / PRs

• majit-translate auto-emit SSA per opcode handler #50: opcode-handler SSA emission is not fully represented in the codewriter path.
• Eliminate pc_map and per-PC labels by switching the tracer to interpret JitCode bytecode #73: tracer/codewriter coordinate mismatch creates pc_map, per-PC labels, and label-side repair pressure.
• Replace syn-AST front-end with Charon-extracted MIR for the JIT lowering pipeline #97: MIR-based front-end may eventually provide a more orthodox explicit CFG input.
• Optimize tuple unpacking and specialization in JIT trace #59: tuple specialization/performance should remain tracked in that PR, not here, except where trace shape depends on graph/codewriter normalization.

Non-goals

• Do not make this issue responsible for completing majit-translate auto-emit SSA per opcode handler #50 or Eliminate pc_map and per-PC labels by switching the tracer to interpret JitCode bytecode #73.
• Do not hide semantic mismatches as structural adaptations without citing the PyPy source and the Pyre constraint.
• Do not add new byte-stream surgery when an upstream graph simplification pass is the right layer.

Suggested execution order

1. Add a small parity test harness for graph simplification outputs.
2. Port remove_trivial_links and remove_identical_vars_SSA first.
3. Port or classify SSA_to_SSI and simplify_exceptions.
4. Re-audit regalloc coalescing after graph normalization changes.
5. Add label-origin diagnostics for the unmarked-label panic.
6. Split any clearly independent backend-only finding into a separate issue.