perf(codegen): size-optimize oversized IR units at -Os instead of -O0#5674
Merged
Conversation
Oversized modules (IR past PERRY_LL_O0_THRESHOLD_BYTES, 6 MB) were compiled at clang -O0 to dodge the #4880 pathology: a unit dominated by ONE enormous generated function (a multi-thousand-element data literal lowering to a single ~800k-line function) makes LLVM's -O1+ pipeline super-linear and effectively never finishes. But -O0 is the wrong default for the *common* oversized case. A large minified CLI bundle is tens of thousands of ordinary functions, none individually huge, and -O0 emits register-pressure-blind, un-folded code: ~30-50% more __text than -Os for no benefit. That generated code is the overwhelming majority of such a binary's __text (measured: 253 MB of a 266 MB binary is codegen output, not the runtime/stdlib), so the opt level on these units dominates final binary size. Pick the level by average IR bytes-per-function, which cleanly separates the two regimes (>100x gap): a pathological monolith is megabytes-per-function (the #4880 400k-element literal is ~9 MB/fn), a real bundle is ~20 KB/fn. Oversized units at or below PERRY_LL_SIZE_OPT_MAX_FN_BYTES (256 KB) size-optimize at -Os; denser units keep -O0. A giant-literal unit always has very few, very large functions and so never reaches the -Os pipeline. Measured (whole real CLI-bundle IR unit, 15440 functions, ~20 KB/fn): -O0 __text 63 MB -> -Os __text 32 MB (-49%), clang 47s -> 93s. End-to-end (synthetic 9000-function program): -Os and -O0 produce byte-identical program output; binary __text 50.7 MB -> 40.7 MB. Small modules (<6 MB IR) are untouched (still -O3). New escape hatches: PERRY_LL_SIZE_OPT=0/1 forces the old -O0 / always -Os, PERRY_LL_SIZE_OPT_MAX_FN_BYTES tunes the density cap.
📝 WalkthroughWalkthroughAdds function-density-aware optimization selection for oversized LLVM IR units. ChangesOversized LLVM IR optimization selection
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~20 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@crates/perry-codegen/src/linker.rs`:
- Around line 123-125: The function counting logic in count_ll_functions only
matches occurrences preceded by a newline, so it misses a function when the IR
starts with define on the first line. Update count_ll_functions in linker.rs to
count a leading define as well as newline-prefixed ones, so the
bytes-per-function heuristic used for optimization level selection stays
accurate and does not incorrectly fall back to -O0 for borderline units.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: f4266e4f-7218-4c72-81a8-d329f33eb1bf
📒 Files selected for processing (1)
crates/perry-codegen/src/linker.rs
Comment on lines
+123
to
+125
| fn count_ll_functions(ll_text: &str) -> usize { | ||
| ll_text.match_indices("\ndefine ").count() | ||
| } |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
Make function counting robust to a first-line define.
Line 124 misses a function if IR starts with define on line 1, which can skew bytes-per-function and select -O0 unnecessarily for borderline oversized units.
Suggested patch
fn count_ll_functions(ll_text: &str) -> usize {
- ll_text.match_indices("\ndefine ").count()
+ ll_text
+ .lines()
+ .filter(|line| line.starts_with("define "))
+ .count()
}📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| fn count_ll_functions(ll_text: &str) -> usize { | |
| ll_text.match_indices("\ndefine ").count() | |
| } | |
| fn count_ll_functions(ll_text: &str) -> usize { | |
| ll_text | |
| .lines() | |
| .filter(|line| line.starts_with("define ")) | |
| .count() | |
| } |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@crates/perry-codegen/src/linker.rs` around lines 123 - 125, The function
counting logic in count_ll_functions only matches occurrences preceded by a
newline, so it misses a function when the IR starts with define on the first
line. Update count_ll_functions in linker.rs to count a leading define as well
as newline-prefixed ones, so the bytes-per-function heuristic used for
optimization level selection stays accurate and does not incorrectly fall back
to -O0 for borderline units.
This was referenced Jun 25, 2026
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.
Problem
A large minified CLI bundle compiles to a ~320 MB native binary. Attribution shows the bloat is perry's own generated code, not the Rust runtime/stdlib/ext libs:
__textcli_ts.o(perry codegen output)So ~95% of the binary's code is the generated
.o. The runtime + stdlib are a small fraction; opt-level/panic/feature-gating levers on them can't move the needle. The lever is how perry compiles its generated IR.Root cause
Oversized modules (IR past
PERRY_LL_O0_THRESHOLD_BYTES, 6 MB) are compiled at clang-O0. That exists to dodge #4880: a unit dominated by ONE enormous generated function (a multi-thousand-element data literal → a single ~800k-line function) makes LLVM's-O1+pipeline super-linear and effectively never finishes (verified: clang times out >10 min at any optimizing level on such a function).But
-O0is the wrong default for the common oversized case. A large bundle is tens of thousands of ordinary functions, none individually huge, and-O0emits register-pressure-blind, un-folded code — ~30-50% more__textthan-Osfor no benefit.Fix
Pick the opt level by average IR bytes-per-function, which cleanly separates the two regimes (>100x gap):
-O0-OsOversized units at or below
PERRY_LL_SIZE_OPT_MAX_FN_BYTES(256 KB/fn) size-optimize at-Os; denser units keep-O0. A giant-literal unit always has very few, very large functions, so it never reaches the-Ospipeline. Modules under 6 MB IR are untouched (still-O3).Measurements
-O0__text63 MB ->-Os32 MB (-49%); clang 47s -> 93s.-Osand-O0produce byte-identical program output; binary__text50.7 MB -> 40.7 MB.-O3, output matchesnode --experimental-strip-types.-Ozmatches-Oson size but is 2-3x slower, so-Osis the sweet spot.Safety / escape hatches
PERRY_LL_SIZE_OPT=0-> force old-O0;=1-> force-Osregardless of density.PERRY_LL_SIZE_OPT_MAX_FN_BYTEStunes the density cap.Tests
New unit tests cover both branches. Full
perry-codegenlib suite green (129 tests).Summary by CodeRabbit
New Features
Bug Fixes