Externalize size cache to remove __buffa_cached_size struct field#22
Draft
Externalize size cache to remove __buffa_cached_size struct field#22
Conversation
Generated structs previously carried a hidden __buffa_cached_size: CachedSize
(AtomicU32) field that compute_size wrote and write_to read. This made
exhaustive destructuring impossible — let Msg { a, b } = m; forced a ..
that then absorbed any new proto fields too, defeating the use case of
having the compiler flag every destructure site when the schema grows.
Replace the in-struct cache with an external SizeCache (Vec<u32> + cursor)
threaded through the trait methods. compute_size reserves slots in pre-order
(reserve before recursing into a length-delimited nested message, set after),
write_to consumes in the same pre-order. Both passes are generated from the
same field list, both take &self, both use identical presence guards — the
traversal order is structurally guaranteed to match. Groups thread the cache
without reserving (no length prefix).
The provided encode* methods construct and discard a fresh cache internally,
so the common case is unchanged at call sites. A new encoded_len() provided
method replaces direct compute_size() calls for users who just want the size.
Side fix: map_write_to_stmt previously recomputed v.compute_size() inside
the entry_size formula during write — idempotent with the old in-struct
cache, but would corrupt a push-based one. Now reads the value size from
the cache once and derives entry_size from it. Also a minor win: map message
values are no longer walked three times (compute, entry_size-recompute,
write) but twice.
Generated structs now contain only proto fields plus __buffa_unknown_fields.
No interior mutability remaining — Sync comes for free instead of via
AtomicU32-with-Relaxed contortion.
Closes #14.
…essage Review follow-up: - map_compute_size_stmt uses .values() when key size is constant, while map_write_to_stmt always uses for (k, v). The cache-slot order between these must match. For HashMap (both std and hashbrown) it does — both walk the table in slot order — but this is not obvious from a cold read. Added a comment at the optimization site. - next_size() previously panicked with a generic index-out-of-bounds message. Now reports cursor position vs slot count, with #[track_caller] so the panic location points at the caller. Helps manual Message implementers diagnose traversal-order mismatches. - Dropped model: opus from the rust-code-reviewer agent definition. That model was emitting text-based pseudo-tool-call syntax the harness does not parse (tool_uses: 0 on every run), so all findings were confabulated from nonexistent file content. Inherits the main loop model now.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #14.
Problem
Generated structs carried a hidden
__buffa_cached_size: CachedSize(AtomicU32) field. This made exhaustive destructuring impossible —let Msg { a, b } = m;forced a..that then absorbed any new proto fields too, defeating the use case of having the compiler flag every destructure site when the schema grows.Approach
Replace the in-struct cache with an external
SizeCache(aVec<u32>+ cursor) threaded through the trait methods:compute_size(&self, cache: &mut SizeCache)reserves slots in pre-order (reserve before recursing into a length-delimited nested message, set after)write_to(&self, cache: &mut SizeCache, buf)consumes in the same pre-order viacache.next_size()Both passes are generated from the same field list, both take
&self, both use identical presence guards — traversal order is structurally guaranteed. The providedencode*methods construct and discard a fresh cache internally; a newencoded_len()provided method covers the "just want the size" case.Side fix:
map_write_to_stmtpreviously recomputedv.compute_size()inside the entry_size formula during write — idempotent with the old in-struct cache, but would corrupt a push-based one. Now reads the value size from the cache once. Minor win: map message values are walked twice instead of three times.Generated structs now contain only proto fields plus
__buffa_unknown_fields. No interior mutability remaining —Synccomes for free instead of via AtomicU32-with-Relaxed.Verification
task lintcleantask test— 221 pass, 0 failtask conformance— all 6 suites PASSED, 0 unexpected failures