[jasperpotts]

Closes #1029. Stacked on #1028 (1027-synthetic-block-bench).

Summary

The flyweight writers (#1023) paid most of their time per nested message: opening a child, checking every field, and patching its length when it ended. This PR cuts that bookkeeping, and for small messages removes it entirely. Everything here is in the writers, so it only applies with pbj.writers on (tier B); models, codecs and the default configuration are unchanged.

Performance (AMD EPYC 9275F): on the synthetic HAPI block, writing through the writers is 18.9% faster (1,381 → 1,121 ns per transaction) and still allocates 0 B; block writes with SHA-384 per item are 6.5% faster. Tier B, opt-in with pbj.writers.

• The writer's state is one guard field (531cbf59, 698ba8e6, a96d0161):
  • Sentinels mean "not open" (Integer.MAX_VALUE), "a child is open" (MAX_VALUE - 1) and "written by writeFrom" (MAX_VALUE - 2), so checking a field stays one compare, and ending an open child moves onto that compare's slow path.
  • The separate state and offset fields are gone, and closing a nested message stores less.
  • endChild() is small enough for the JIT to inline into the field checks.
  • Two- and three-byte varints are written inline.
• Small messages are written in one call (5493cdac, 56f153f3, 987f5f07): a message with up to six number, bool or enum fields, and no repeated fields, maps or other messages, gets a setter on its parent's writer that takes the fields as arguments. It computes the length and writes the tag, length and fields directly, with no child writer and nothing to close:

  amount.accountID(0, 0, accountNum);              // was: amount.accountID().accountNum(accountNum) + an implicit close
  result.consensusTimestamp(seconds, nanos);
  associate.addTokens(0, 0, tokenNum);             // elements of repeated small messages too

  Enums are passed as their proto ordinal. This covers ids, timestamps, durations, fractions, key ids and token relations; 439 such setters across HAPI.
• Messages under 128 bytes encoded are opened without a guess (openSmallChild): their length is always one byte, whatever the buffer's length reservation.

Tried and reverted (recorded in docs/fast-codegen-plan.md §9)

• The buffer tracking the innermost open writer, so field checks needn't look for an open child: 12% slower, an extra load and compare per field.
• Four- to eight-byte varints inlined (a 316-byte copy of the bulk writer, which is over the JIT's 325-byte inlining limit): 2% slower here, and 5-9% slower on a second, external block workload.
• Nine-byte varints in one store: neutral on both workloads.

EPYC results

AMD EPYC 9275F, JDK 25.0.4, one CCD (taskset -c 0-2,24-26) of an idle machine, 3 forks × 5 iterations of 2 s, -prof gc. ns per transaction of the synthetic HAPI block (#1028), every option on.

Step	produceWriters	produceWritersBlockHashed	Allocation
Base (#1028)	1,381 ± 22	3,246 ± 14	0 B
Inlinable auto-close, 2-3 byte varints	1,383 ± 10 (+0.1%)	3,276 ± 63 (+0.9%)	0 B
State folded into the guard	1,362 ± 35 (−1.4%)	3,253 ± 19 (+0.2%)	0 B
Smaller close	1,348 ± 20 (−2.4%)	3,233 ± 29 (−0.4%)	0 B
One-call small messages (head)	1,121 ± 6 (−18.9%)	3,033 ± 51 (−6.5%)	0 B

The one-call setters are the whole win; the state changes before them are neutral, and kept because the writer is simpler with them. In the full benchmark run at each PR head, produceWriters went from 1,359 to 1,146 ns (−15.7%), which is 1.52× the hand-written reference encoder (756 ns in the same run), down from 1.79×. produceWritersBlockHashed includes SHA-384 of every item.

API

The one-call setters add public methods to generated writers (439 in HAPI). They are overloads next to the existing child accessors, so nothing is removed or renamed, but it is new surface to review.

Tests

• InlineSetterTest: ids (with their oneof's one number alternative), timestamps and repeated elements written in one call give the same bytes as the child writer and as the model, for zero, small, large and negative values, and the field order is still checked.
• WriterGenerationTest: which messages get one-call setters, enums included: six fields qualify, while seven, a string field or two number alternatives of one oneof don't; and small children are opened with openSmallChild.
• The whole integration suite with every option on writes every test object through its writer and compares it with the model's encoding.

🤖 Generated with Claude Code

[github-actions]

JUnit Test Report

   559 files     559 suites   1m 18s ⏱️
 2 112 tests  2 108 ✅  4 💤 0 ❌
11 644 runs  11 616 ✅ 28 💤 0 ❌

Results for commit de84353.

♻️ This comment has been updated with latest results.

[github-actions]

Integration Test Report

    660 files      660 suites   44m 41s ⏱️
164 377 tests 164 377 ✅ 0 💤 0 ❌
652 340 runs  652 340 ✅ 0 💤 0 ❌

Results for commit de84353.

♻️ This comment has been updated with latest results.