Repository navigation
Conversation
Builds compiled everything from scratch in every job (macOS build steps grew from 47 to 166 job-minutes between Oct 3 and Oct 11) and MSVC built serially with NMake. Jobs now restore a per-job ccache saved by master pushes. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SiUW6zog3Gpn13eX59fE1c
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SiUW6zog3Gpn13eX59fE1c
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SiUW6zog3Gpn13eX59fE1c
ComposedRenderer::Render calls raw text and {{ ... }} children directly,
switching on the node's kind byte, instead of one shared virtual call that
mispredicts as a body alternates the two. A plain name renders through a
direct ValueRefExpression::Render. Measured in docs/tasks/0141.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc
Inline in the header, its larger body changed GCC's inlining in the parser files that include it: Load +0.06..+1.22% in instructions on CI. In statements.cpp Load is back to -0.04..+0.08%. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc
rmorozov
pushed a commit
that referenced
this pull request
Oct 11, 2026
Switch dispatch for body children and plain names (closed PR #468) traded +0.1..+1.35% Render instructions for fewer mispredicts in callgrind's last-target branch model only; the perf track rejected it. The write-up, the rejected wide variant, the header-inlining trap and the condition for reopening stay in the task file. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc
rmorozov
added a commit
that referenced
this pull request
Oct 11, 2026
Switch dispatch for body children and plain names (closed PR #468) traded +0.1..+1.35% Render instructions for fewer mispredicts in callgrind's last-target branch model only; the perf track rejected it. The write-up, the rejected wide variant, the header-inlining trap and the condition for reopening stay in the task file. Claude-Session: https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc Co-authored-by: Claude <noreply@anthropic.com>
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
2 participants
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.
Requested by Ruslan · project thread
Before: every child of a template body rendered through one shared virtual call in
ComposedRenderer::Render. A body alternates raw text and{{ ... }}, so that one indirect call site changed target on almost every child: in callgrind's branch simulation it was 95% of all indirect mispredicts onmitsuhiko_table(105,000 of 110,265).After: the body loop switches on the child's kind byte and calls raw text and
{{ ... }}directly (inlined); every other kind still goes through the vptr. A{{ name }}renders through a directValueRefExpression::Render. Simulated mispredicts drop 3-71% (mitsuhiko_table-71%,macros-24%,for_loop_vars-27%), instructions move -0.19..+1.35%.How:
src/renderer.h(ComposedRenderer::Renderdefined after the classes it dispatches to),RenderOfinsrc/expression_evaluator.h,ValueRefExpression::Renderoverride (andSelfRefExpressionkeeps the base's). The full numbers, and the rejected wider variant (the same switch at everyEvaluate/EvaluateRefcall site: Ir up to +5.5%, mispredicts up onexpressions), are indocs/tasks/0141-switch-dispatch.md.For the perf verdict: instruction counts rise on most Render cases (
for_loop_vars+1.35%,expressions+1.08%,many_tags+0.76%, others +0.1..+0.7%) in exchange for the mispredict drop. The cloud VM has no hardware counters and its wall time is too noisy (medians -2..-11%, minimums -3..+4%) to settle it; a quiet machine should.Includes #464 (CI ccache branch) as a merge commit, at Ruslan's request so this PR's CI runs use it; its files drop out of this diff once #464 merges.
Roles: main session, verifier (1 round, ready: tests, ASan+UBSan, differential vs base and Python Jinja2 across 5 modes). Red pushes: none so far.
🤖 Generated with Claude Code
https://claude.ai/code/session_018j47FUYwQ9s69rR6Jesrwc
Generated by Claude Code