Bypass interpreter bridge for blueprint-backed class construction

[frostney]

Summary

• Move the native blueprint instantiation path before PushVMFramesToGoccia so standalone classes and blueprint-inherited classes never cross the interpreter bridge, eliminating O(VM_depth) frame iteration and the try..except wrapper per construction
• Cache delegate chain setup so subsequent constructions of the same class skip the O(depth) no-op traversal
• Hoist SetLength(VMArgs, 1) outside the field initializer loop to eliminate redundant dynamic array allocations

Test plan

• All 3,422 interpreter tests pass
• All 3,463 bytecode tests pass
• CI benchmark comparison for class-heavy workloads

Made with Cursor

Summary by CodeRabbit

• Refactor
  • Introduced dual-path blueprint handling to support cached and on-demand construction flows, improving consistency of instance creation.
  • Deferred bridge creation and strengthened delegate/prototype initialization so delegation chains are reliably established during construction.
  • Guarded and reordered field initialization and allocation to avoid redundant allocations and ensure field inits run in the correct order before constructor invocation.

[coderabbitai]

📝 Walkthrough

Walkthrough

Handle TSouffleBlueprint constructors early with a dual-path: cached-bridge lookup versus on-demand bridge creation. Allocate the instance record earlier, set Prototype/Method Delegates during prototype walking, guard and run field initializers in reverse, and separate blueprint vs non-blueprint constructor flows.

Changes

Cohort / File(s)	Summary
Blueprint Construction Refactor units/Goccia.Runtime.Operations.pas	Reworked Construct to detect TSouffleBlueprint early and split into cached-bridge vs on-demand bridge paths; allocate Rec earlier (once) for blueprint path; ensure Prototype.Delegate and Methods.Delegate are established during proto walk; guard FieldInit (only if >0) and run in reverse; defer bridge creation when using delegation-walker; update bridge-cache on creation.

Sequence Diagram(s)

sequenceDiagram
participant Caller
participant Runtime
participant Blueprint
participant BridgeCache
participant Bridge
participant VM

Note right of Runtime: Construct start
Caller->>Runtime: Construct(...)
Runtime->>Blueprint: is TSouffleBlueprint?
alt Blueprint path
    Runtime->>Blueprint: ConvertBlueprintToClassValue(...)
    Runtime->>BridgeCache: lookup bridge for blueprint
    alt cached
        BridgeCache-->>Runtime: cached bridge
        Runtime->>Bridge: invoke cached bridge with Rec
    else not cached
        Runtime->>Bridge: create bridge from blueprint
        Bridge-->>BridgeCache: write cache entry
        Runtime->>Bridge: invoke bridge with Rec
    end
else Non-blueprint path
    Runtime->>VM: Push frames / prepare context
    Runtime->>Bridge: build or locate bridge
    Runtime->>Bridge: invoke bridge with Rec
end
Bridge-->>Runtime: constructor result
Runtime-->>Caller: return instance

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Poem

🐇 I hopped through blueprints, bridges in tow,

Delegates sorted in a tidy row,
Fields set backward, one-by-one,
Cache or craft — constructors run,
A rabbit cheers: the runtime's aglow!

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title 'Bypass interpreter bridge for blueprint-backed class construction' accurately captures the main optimization objective of the PR—avoiding the interpreter bridge for blueprint-based class instantiation.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch opt/native-class-construct

📝 Coding Plan

• Generate coding plan for human review comments

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🧹 Nitpick comments (1)

units/Goccia.Runtime.Operations.pas (1)

3342-3342: Use PROP_CONSTRUCTOR instead of a string literal.

This keeps constructor lookups aligned with the rest of the runtime constants.

♻️ Suggested change

-        if WalkBp.Methods.Get('constructor', CtorMethod) then
+        if WalkBp.Methods.Get(PROP_CONSTRUCTOR, CtorMethod) then

As per coding guidelines, "Use runtime constant units instead of hardcoded string literals: Goccia.Constants.PropertyNames for property names, Goccia.Constants.TypeNames for type names, Goccia.Constants.ErrorNames for error names, Goccia.Constants.ConstructorNames for constructor names, Goccia.Constants.SymbolNames for symbol names, Goccia.Constants for literal value strings"

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@units/Goccia.Runtime.Operations.pas` at line 3342, Replace the hardcoded
string 'constructor' with the runtime constant PROP_CONSTRUCTOR so constructor
lookups use the shared constant; specifically update the call to
WalkBp.Methods.Get('constructor', CtorMethod) in
units/Goccia.Runtime.Operations.pas (the WalkBp.Methods.Get invocation that
fills CtorMethod) to use PROP_CONSTRUCTOR from the constants unit
(Goccia.Constants.PropertyNames or the module where PROP_CONSTRUCTOR is
defined).

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@units/Goccia.Runtime.Operations.pas`:
- Around line 3287-3288: The branch currently uses Assigned(Bp.SuperBlueprint)
which incorrectly treats derived blueprints as roots; change the gate to test
for the root (i.e., not Assigned(Bp.SuperBlueprint)) so the native fast-path
only runs for root blueprints and not for leaves. Update the condition "if
Assigned(Bp.SuperBlueprint) or not FBlueprintSuperValues.ContainsKey(Bp) then"
to "if not Assigned(Bp.SuperBlueprint) or not
FBlueprintSuperValues.ContainsKey(Bp) then" (and make the same change in the
other occurrences around the blocks referencing Bp, FBlueprintSuperValues,
__fields__, and delegate wiring to FVM.RecordDelegate at the noted ranges).
- Around line 3373-3377: The FBlueprintBridgeCache is being reused for two
different bridge conversions (BlueprintToClassValue vs
ConvertBlueprintToClassValue), causing lost behavior when
ConvertBlueprintToClassValue is required; in the branch that currently reads
FBlueprintBridgeCache for Bp (where CachedBridge is assigned for use by
ConvertBlueprintToClassValue), change the logic to avoid reusing entries
produced by the other converter — either bypass the cache and always call
ConvertBlueprintToClassValue for this path, or key into a separate cache (e.g.,
a new FBlueprintClassBridgeCache) so entries created by
UnwrapToGocciaValue/BlueprintToClassValue are not returned here; update the code
paths referencing UnwrapToGocciaValue, ConvertBlueprintToClassValue,
BlueprintToClassValue, FBlueprintBridgeCache, and CachedBridge accordingly so
the bridge used here always includes getters/setters, static fields, private
members and field initializers.

---

Nitpick comments:
In `@units/Goccia.Runtime.Operations.pas`:
- Line 3342: Replace the hardcoded string 'constructor' with the runtime
constant PROP_CONSTRUCTOR so constructor lookups use the shared constant;
specifically update the call to WalkBp.Methods.Get('constructor', CtorMethod) in
units/Goccia.Runtime.Operations.pas (the WalkBp.Methods.Get invocation that
fills CtorMethod) to use PROP_CONSTRUCTOR from the constants unit
(Goccia.Constants.PropertyNames or the module where PROP_CONSTRUCTOR is
defined).

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: c53d0e42-ca2a-49e0-9ec1-b22fad110d7b

📥 Commits

Reviewing files that changed from the base of the PR and between 6feb97a and cb97506.

📒 Files selected for processing (1)

• units/Goccia.Runtime.Operations.pas

[github-actions]

Suite Timing

Suite	Metric	Interpreted	Bytecode
Tests	Total	3497	3497
Tests	Passed	3456 ✅	3497 ✅
Tests	Skipped	41	0
Tests	Execution	151.5ms	159.0ms
Tests	Engine	309.3ms	669.3ms
Benchmarks	Total	254	254
Benchmarks	Duration	6.74min	11.49min

_{Measured on ubuntu-latest x64.}

[github-actions]

Benchmark Results

254 benchmarks

Interpreted: 🟢 3 improved · 🔴 2 regressed · 249 unchanged · avg -1.5%
Bytecode: 🟢 9 improved · 🔴 15 regressed · 230 unchanged · avg -0.2%

arraybuffer.js — Interp: 14 unch. · avg -0.5% · Bytecode: 14 unch. · avg -0.6%

Benchmark	Interpreted	Δ	Bytecode	Δ
create ArrayBuffer(0)	447,306 → 443,982	-0.7%	147,311 → 137,205	-6.9%
create ArrayBuffer(64)	438,474 → 437,448	-0.2%	144,538 → 142,585	-1.4%
create ArrayBuffer(1024)	332,856 → 331,325	-0.5%	131,065 → 127,770	-2.5%
create ArrayBuffer(8192)	142,324 → 136,290	-4.2%	81,299 → 80,506	-1.0%
slice full buffer (64 bytes)	527,266 → 522,705	-0.9%	391,336 → 383,665	-2.0%
slice half buffer (512 of 1024 bytes)	445,849 → 439,641	-1.4%	342,123 → 340,064	-0.6%
slice with negative indices	454,883 → 448,048	-1.5%	361,501 → 367,377	+1.6%
slice empty range	513,460 → 510,320	-0.6%	386,354 → 386,114	-0.1%
byteLength access	1,572,618 → 1,574,259	+0.1%	1,257,193 → 1,319,679	+5.0%
Symbol.toStringTag access	1,175,305 → 1,178,174	+0.2%	621,350 → 640,038	+3.0%
ArrayBuffer.isView	749,533 → 791,091	+5.5%	503,196 → 497,593	-1.1%
clone ArrayBuffer(64)	389,251 → 388,538	-0.2%	332,320 → 332,152	-0.1%
clone ArrayBuffer(1024)	305,496 → 301,421	-1.3%	253,827 → 254,113	+0.1%
clone ArrayBuffer inside object	269,889 → 266,383	-1.3%	158,202 → 155,127	-1.9%

arrays.js — Interp: 19 unch. · avg -1.4% · Bytecode: 🟢 1, 18 unch. · avg +1.5%

Benchmark	Interpreted	Δ	Bytecode	Δ
Array.from length 100	14,023 → 13,906	-0.8%	14,922 → 15,294	+2.5%
Array.from 10 elements	229,745 → 229,583	-0.1%	160,993 → 159,218	-1.1%
Array.of 10 elements	318,933 → 312,949	-1.9%	230,520 → 228,407	-0.9%
spread into new array	345,772 → 346,630	+0.2%	911,425 → 901,807	-1.1%
map over 50 elements	28,307 → 27,413	-3.2%	24,479 → 26,230	🟢 +7.2%
filter over 50 elements	23,959 → 23,480	-2.0%	25,118 → 24,317	-3.2%
reduce sum 50 elements	27,608 → 26,625	-3.6%	21,730 → 21,834	+0.5%
forEach over 50 elements	23,229 → 23,485	+1.1%	27,233 → 27,589	+1.3%
find in 50 elements	36,052 → 35,510	-1.5%	32,704 → 32,978	+0.8%
sort 20 elements	12,183 → 12,140	-0.4%	12,848 → 13,278	+3.3%
flat nested array	119,485 → 119,455	-0.0%	481,835 → 488,192	+1.3%
flatMap	74,373 → 73,936	-0.6%	316,753 → 323,033	+2.0%
map inside map (5x5)	21,930 → 21,241	-3.1%	106,299 → 107,799	+1.4%
filter inside map (5x10)	16,200 → 15,564	-3.9%	14,990 → 14,957	-0.2%
reduce inside map (5x10)	19,798 → 19,074	-3.7%	14,976 → 15,447	+3.1%
forEach inside forEach (5x10)	16,763 → 16,952	+1.1%	16,888 → 17,594	+4.2%
find inside some (10x10)	14,029 → 13,851	-1.3%	11,959 → 12,181	+1.9%
map+filter chain nested (5x20)	5,532 → 5,304	-4.1%	4,937 → 5,103	+3.4%
reduce flatten (10x5)	38,781 → 39,034	+0.7%	6,332 → 6,455	+1.9%

async-await.js — Interp: 6 unch. · avg -1.3% · Bytecode: 6 unch. · avg -1.6%

Benchmark	Interpreted	Δ	Bytecode	Δ
single await	385,555 → 380,467	-1.3%	289,542 → 281,507	-2.8%
multiple awaits	170,985 → 170,252	-0.4%	121,319 → 122,581	+1.0%
await non-Promise value	885,738 → 870,190	-1.8%	1,046,947 → 1,011,408	-3.4%
await with try/catch	378,032 → 370,394	-2.0%	280,169 → 281,478	+0.5%
await Promise.all	52,819 → 52,745	-0.1%	43,593 → 42,024	-3.6%
nested async function call	192,811 → 188,318	-2.3%	219,213 → 216,837	-1.1%

classes.js — Interp: 31 unch. · avg -1.1% · Bytecode: 🔴 1, 30 unch. · avg +0.0%

Benchmark	Interpreted	Δ	Bytecode	Δ
simple class new	116,942 → 116,617	-0.3%	397,491 → 414,321	+4.2%
class with defaults	94,296 → 93,861	-0.5%	279,469 → 291,907	+4.5%
50 instances via Array.from	5,551 → 5,515	-0.7%	6,874 → 7,167	+4.3%
instance method call	59,842 → 59,701	-0.2%	184,815 → 191,728	+3.7%
static method call	92,293 → 92,225	-0.1%	421,755 → 441,358	+4.6%
single-level inheritance	46,621 → 46,322	-0.6%	181,087 → 185,903	+2.7%
two-level inheritance	39,095 → 39,093	-0.0%	153,416 → 158,051	+3.0%
private field access	58,898 → 58,282	-1.0%	196,849 → 204,636	+4.0%
private methods	64,377 → 63,673	-1.1%	264,030 → 275,888	+4.5%
getter/setter access	66,679 → 66,157	-0.8%	195,356 → 198,238	+1.5%
class decorator (identity)	82,709 → 81,489	-1.5%	54,737 → 51,860	-5.3%
class decorator (wrapping)	47,503 → 46,955	-1.2%	38,388 → 38,310	-0.2%
identity method decorator	58,802 → 58,066	-1.3%	46,150 → 45,373	-1.7%
wrapping method decorator	48,462 → 47,761	-1.4%	38,803 → 41,096	+5.9%
stacked method decorators (x3)	34,744 → 34,349	-1.1%	28,794 → 25,838	🔴 -10.3%
identity field decorator	66,917 → 65,235	-2.5%	47,230 → 46,458	-1.6%
field initializer decorator	56,435 → 55,913	-0.9%	42,446 → 41,604	-2.0%
getter decorator (identity)	57,201 → 55,380	-3.2%	41,727 → 40,974	-1.8%
setter decorator (identity)	48,242 → 47,581	-1.4%	36,189 → 35,967	-0.6%
static method decorator	62,847 → 62,004	-1.3%	65,636 → 64,371	-1.9%
static field decorator	72,446 → 72,081	-0.5%	68,717 → 67,120	-2.3%
private method decorator	47,974 → 47,557	-0.9%	38,541 → 38,070	-1.2%
private field decorator	54,452 → 52,169	-4.2%	39,971 → 39,021	-2.4%
plain auto-accessor (no decorator)	90,281 → 89,903	-0.4%	55,352 → 54,966	-0.7%
auto-accessor with decorator	52,590 → 52,009	-1.1%	37,968 → 37,324	-1.7%
decorator writing metadata	43,367 → 42,720	-1.5%	43,419 → 42,752	-1.5%
static getter read	105,374 → 104,581	-0.8%	453,598 → 446,249	-1.6%
static getter/setter pair	79,978 → 79,230	-0.9%	241,432 → 235,261	-2.6%
inherited static getter	58,859 → 58,916	+0.1%	306,062 → 301,374	-1.5%
inherited static setter	64,511 → 63,544	-1.5%	237,311 → 233,185	-1.7%
inherited static getter with this binding	53,996 → 53,719	-0.5%	181,360 → 181,863	+0.3%

closures.js — Interp: 11 unch. · avg -2.0% · Bytecode: 11 unch. · avg -0.7%

Benchmark	Interpreted	Δ	Bytecode	Δ
closure over single variable	132,203 → 131,557	-0.5%	798,508 → 796,288	-0.3%
closure over multiple variables	122,054 → 118,952	-2.5%	489,326 → 485,256	-0.8%
nested closures	126,264 → 126,717	+0.4%	707,950 → 707,042	-0.1%
function as argument	96,456 → 95,194	-1.3%	705,524 → 704,813	-0.1%
function returning function	120,424 → 119,413	-0.8%	760,121 → 773,129	+1.7%
compose two functions	72,835 → 72,109	-1.0%	468,619 → 474,049	+1.2%
fn.call	151,906 → 147,014	-3.2%	144,931 → 143,913	-0.7%
fn.apply	115,121 → 111,410	-3.2%	99,602 → 96,845	-2.8%
fn.bind	140,277 → 134,149	-4.4%	151,652 → 143,035	-5.7%
recursive sum to 50	11,986 → 11,669	-2.6%	53,100 → 52,439	-1.2%
recursive tree traversal	20,016 → 19,397	-3.1%	75,466 → 76,343	+1.2%

collections.js — Interp: 🟢 3, 9 unch. · avg +1.4% · Bytecode: 🟢 1, 🔴 1, 10 unch. · avg +1.1%

Benchmark	Interpreted	Δ	Bytecode	Δ
add 50 elements	7,228 → 7,419	+2.6%	5,792 → 6,029	+4.1%
has lookup (50 elements)	89,207 → 87,461	-2.0%	88,367 → 90,105	+2.0%
delete elements	47,673 → 48,797	+2.4%	36,902 → 37,290	+1.1%
forEach iteration	16,368 → 16,101	-1.6%	17,093 → 17,032	-0.4%
spread to array	33,550 → 33,158	-1.2%	153,847 → 155,977	+1.4%
deduplicate array	43,035 → 42,054	-2.3%	47,803 → 48,390	+1.2%
set 50 entries	5,267 → 5,303	+0.7%	6,117 → 6,143	+0.4%
get lookup (50 entries)	80,954 → 88,177	🟢 +8.9%	91,477 → 101,707	🟢 +11.2%
has check	122,887 → 133,292	🟢 +8.5%	151,309 → 156,941	+3.7%
delete entries	43,155 → 46,425	🟢 +7.6%	35,357 → 34,813	-1.5%
forEach iteration	16,307 → 16,246	-0.4%	17,448 → 17,069	-2.2%
keys/values/entries	9,027 → 8,488	-6.0%	22,828 → 21,063	🔴 -7.7%

destructuring.js — Interp: 22 unch. · avg -0.9% · Bytecode: 🟢 1, 🔴 2, 19 unch. · avg -0.7%

Benchmark	Interpreted	Δ	Bytecode	Δ
simple array destructuring	439,362 → 421,595	-4.0%	1,076,709 → 982,360	🔴 -8.8%
with rest element	288,832 → 297,594	+3.0%	779,532 → 716,917	🔴 -8.0%
with defaults	452,776 → 426,042	-5.9%	993,040 → 974,380	-1.9%
skip elements	449,817 → 463,079	+2.9%	1,232,955 → 1,148,440	-6.9%
nested array destructuring	190,239 → 184,735	-2.9%	524,469 → 519,064	-1.0%
swap variables	544,467 → 535,993	-1.6%	1,397,374 → 1,418,860	+1.5%
simple object destructuring	306,518 → 292,695	-4.5%	623,641 → 627,390	+0.6%
with defaults	367,528 → 373,477	+1.6%	330,146 → 329,573	-0.2%
with renaming	314,400 → 324,421	+3.2%	732,212 → 738,235	+0.8%
nested object destructuring	152,868 → 153,537	+0.4%	295,149 → 292,538	-0.9%
rest properties	188,722 → 188,546	-0.1%	245,407 → 240,608	-2.0%
object parameter	92,326 → 92,369	+0.0%	207,831 → 212,390	+2.2%
array parameter	126,167 → 124,202	-1.6%	436,353 → 442,617	+1.4%
mixed destructuring in map	35,161 → 36,290	+3.2%	38,387 → 39,201	+2.1%
forEach with array destructuring	69,013 → 66,949	-3.0%	186,323 → 203,353	🟢 +9.1%
map with array destructuring	70,241 → 68,957	-1.8%	256,424 → 260,243	+1.5%
filter with array destructuring	73,112 → 71,189	-2.6%	301,203 → 306,895	+1.9%
reduce with array destructuring	79,417 → 75,453	-5.0%	287,791 → 286,679	-0.4%
map with object destructuring	80,294 → 81,074	+1.0%	80,389 → 81,596	+1.5%
map with nested destructuring	66,249 → 66,645	+0.6%	69,669 → 69,518	-0.2%
map with rest in destructuring	39,755 → 38,198	-3.9%	24,371 → 23,606	-3.1%
map with defaults in destructuring	60,029 → 60,829	+1.3%	37,692 → 36,137	-4.1%

fibonacci.js — Interp: 8 unch. · avg -1.5% · Bytecode: 8 unch. · avg -2.0%

Benchmark	Interpreted	Δ	Bytecode	Δ
recursive fib(15)	326 → 321	-1.2%	1,479 → 1,453	-1.7%
recursive fib(20)	29 → 29	+0.5%	134 → 132	-1.4%
recursive fib(15) typed	330 → 323	-1.9%	1,907 → 1,972	+3.4%
recursive fib(20) typed	29 → 29	+1.3%	174 → 178	+2.7%
iterative fib(20) via reduce	12,648 → 12,239	-3.2%	9,070 → 8,660	-4.5%
iterator fib(20)	9,763 → 9,654	-1.1%	15,926 → 15,101	-5.2%
iterator fib(20) via Iterator.from + take	15,084 → 14,827	-1.7%	17,547 → 16,420	-6.4%
iterator fib(20) last value via reduce	11,659 → 11,149	-4.4%	13,020 → 12,657	-2.8%

for-of.js — Interp: 7 unch. · avg -1.9% · Bytecode: 🔴 1, 6 unch. · avg -4.2%

Benchmark	Interpreted	Δ	Bytecode	Δ
for...of with 10-element array	49,232 → 48,393	-1.7%	165,514 → 164,614	-0.5%
for...of with 100-element array	5,660 → 5,526	-2.4%	24,940 → 24,716	-0.9%
for...of with string (10 chars)	35,793 → 35,366	-1.2%	131,257 → 132,090	+0.6%
for...of with Set (10 elements)	49,787 → 48,771	-2.0%	174,658 → 177,302	+1.5%
for...of with Map entries (10 entries)	31,614 → 31,070	-1.7%	56,530 → 45,903	🔴 -18.8%
for...of with destructuring	42,642 → 41,490	-2.7%	82,965 → 79,164	-4.6%
for-await-of with sync array	47,168 → 46,530	-1.4%	149,255 → 139,327	-6.7%

iterators.js — Interp: 20 unch. · avg -2.1% · Bytecode: 🔴 3, 17 unch. · avg -3.8%

Benchmark	Interpreted	Δ	Bytecode	Δ
Iterator.from({next}).toArray() — 20 elements	15,045 → 14,858	-1.2%	18,613 → 16,440	🔴 -11.7%
Iterator.from({next}).toArray() — 50 elements	6,585 → 6,563	-0.3%	8,432 → 7,721	🔴 -8.4%
spread pre-wrapped iterator — 20 elements	11,823 → 11,682	-1.2%	16,886 → 16,485	-2.4%
Iterator.from({next}).forEach — 50 elements	4,688 → 4,524	-3.5%	5,974 → 5,578	-6.6%
Iterator.from({next}).reduce — 50 elements	4,740 → 4,586	-3.3%	5,622 → 5,265	-6.3%
wrap array iterator	180,285 → 178,322	-1.1%	110,385 → 106,882	-3.2%
wrap plain {next()} object	10,586 → 10,410	-1.7%	12,093 → 11,637	-3.8%
map + toArray (50 elements)	4,831 → 4,654	-3.7%	5,526 → 5,494	-0.6%
filter + toArray (50 elements)	4,687 → 4,561	-2.7%	5,550 → 5,224	-5.9%
take(10) + toArray (50 element source)	27,737 → 27,056	-2.5%	29,407 → 27,684	-5.9%
drop(40) + toArray (50 element source)	6,613 → 6,538	-1.1%	9,052 → 8,314	🔴 -8.1%
chained map + filter + take (100 element source)	8,748 → 8,489	-3.0%	10,063 → 9,582	-4.8%
some + every (50 elements)	2,727 → 2,668	-2.2%	3,537 → 3,360	-5.0%
find (50 elements)	5,951 → 5,819	-2.2%	7,495 → 7,441	-0.7%
array.values().map().filter().toArray()	9,264 → 8,827	-4.7%	10,249 → 10,394	+1.4%
array.values().take(5).toArray()	218,477 → 218,568	+0.0%	149,742 → 147,177	-1.7%
array.values().drop(45).toArray()	204,508 → 208,183	+1.8%	146,228 → 143,747	-1.7%
map.entries() chained helpers	11,271 → 10,890	-3.4%	5,554 → 5,510	-0.8%
set.values() chained helpers	19,188 → 18,540	-3.4%	22,202 → 22,200	-0.0%
string iterator map + toArray	14,674 → 14,280	-2.7%	22,770 → 22,645	-0.6%

json.js — Interp: 20 unch. · avg -2.8% · Bytecode: 🟢 1, 🔴 1, 18 unch. · avg -0.3%

Benchmark	Interpreted	Δ	Bytecode	Δ
parse simple object	172,764 → 171,131	-0.9%	138,358 → 144,245	+4.3%
parse nested object	109,431 → 106,263	-2.9%	93,608 → 95,650	+2.2%
parse array of objects	57,580 → 56,114	-2.5%	50,583 → 48,727	-3.7%
parse large flat object	51,046 → 48,946	-4.1%	45,236 → 45,168	-0.2%
parse mixed types	73,062 → 70,397	-3.6%	65,443 → 60,695	🔴 -7.3%
stringify simple object	201,662 → 194,291	-3.7%	150,355 → 149,696	-0.4%
stringify nested object	113,996 → 110,977	-2.6%	86,398 → 85,189	-1.4%
stringify array of objects	65,300 → 62,174	-4.8%	48,643 → 51,869	+6.6%
stringify mixed types	87,525 → 87,544	+0.0%	71,001 → 71,093	+0.1%
reviver doubles numbers	45,946 → 44,720	-2.7%	41,156 → 45,751	🟢 +11.2%
reviver filters properties	39,445 → 38,654	-2.0%	46,511 → 46,717	+0.4%
reviver on nested object	52,374 → 51,193	-2.3%	52,580 → 51,826	-1.4%
reviver on array	29,617 → 29,361	-0.9%	29,986 → 29,958	-0.1%
replacer function doubles numbers	49,078 → 48,279	-1.6%	51,299 → 52,542	+2.4%
replacer function excludes properties	62,299 → 61,469	-1.3%	63,032 → 61,441	-2.5%
array replacer (allowlist)	118,165 → 112,755	-4.6%	98,854 → 96,744	-2.1%
stringify with 2-space indent	100,950 → 98,445	-2.5%	80,899 → 78,367	-3.1%
stringify with tab indent	102,163 → 97,215	-4.8%	82,317 → 78,629	-4.5%
parse then stringify	55,977 → 53,420	-4.6%	51,693 → 50,128	-3.0%
stringify then parse	33,410 → 32,303	-3.3%	31,277 → 30,106	-3.7%

jsx.jsx — Interp: 21 unch. · avg -2.8% · Bytecode: 21 unch. · avg +0.5%

Benchmark	Interpreted	Δ	Bytecode	Δ
simple element	204,451 → 202,300	-1.1%	790,592 → 800,067	+1.2%
self-closing element	216,509 → 209,427	-3.3%	805,492 → 820,255	+1.8%
element with string attribute	175,040 → 171,640	-1.9%	567,486 → 569,428	+0.3%
element with multiple attributes	151,806 → 145,326	-4.3%	478,647 → 483,586	+1.0%
element with expression attribute	166,544 → 162,834	-2.2%	555,915 → 553,142	-0.5%
text child	208,803 → 200,386	-4.0%	793,208 → 794,060	+0.1%
expression child	204,794 → 196,840	-3.9%	784,096 → 792,923	+1.1%
mixed text and expression	190,859 → 192,227	+0.7%	707,515 → 718,556	+1.6%
nested elements (3 levels)	76,108 → 75,075	-1.4%	314,636 → 320,397	+1.8%
sibling children	57,321 → 56,663	-1.1%	229,447 → 234,764	+2.3%
component element	146,986 → 145,393	-1.1%	540,796 → 550,654	+1.8%
component with children	90,240 → 89,173	-1.2%	343,791 → 356,715	+3.8%
dotted component	123,482 → 121,243	-1.8%	408,237 → 404,140	-1.0%
empty fragment	218,679 → 210,597	-3.7%	808,046 → 826,951	+2.3%
fragment with children	56,974 → 56,118	-1.5%	233,598 → 237,290	+1.6%
spread attributes	113,691 → 106,316	-6.5%	116,270 → 113,421	-2.4%
spread with overrides	98,247 → 93,348	-5.0%	86,320 → 83,226	-3.6%
shorthand props	160,162 → 151,518	-5.4%	530,137 → 521,568	-1.6%
nav bar structure	26,987 → 26,637	-1.3%	100,155 → 101,938	+1.8%
card component tree	33,281 → 30,955	-7.0%	110,630 → 107,167	-3.1%
10 list items via Array.from	14,621 → 14,415	-1.4%	25,714 → 25,533	-0.7%

numbers.js — Interp: 🔴 1, 10 unch. · avg -0.5% · Bytecode: 11 unch. · avg +2.7%

Benchmark	Interpreted	Δ	Bytecode	Δ
integer arithmetic	593,341 → 543,542	🔴 -8.4%	2,009,068 → 2,143,676	+6.7%
floating point arithmetic	575,973 → 591,755	+2.7%	2,251,833 → 2,408,187	+6.9%
number coercion	186,058 → 185,388	-0.4%	140,817 → 138,011	-2.0%
toFixed	104,056 → 102,516	-1.5%	224,655 → 225,448	+0.4%
toString	162,600 → 159,007	-2.2%	835,934 → 892,593	+6.8%
valueOf	235,600 → 231,437	-1.8%	1,302,369 → 1,376,599	+5.7%
toPrecision	149,809 → 146,783	-2.0%	458,383 → 462,944	+1.0%
Number.isNaN	296,962 → 312,821	+5.3%	175,018 → 179,497	+2.6%
Number.isFinite	290,785 → 299,621	+3.0%	171,624 → 171,059	-0.3%
Number.isInteger	287,854 → 295,517	+2.7%	187,695 → 191,042	+1.8%
Number.parseInt and parseFloat	245,237 → 237,641	-3.1%	153,610 → 153,296	-0.2%

objects.js — Interp: 7 unch. · avg -1.9% · Bytecode: 🔴 2, 5 unch. · avg -3.5%

Benchmark	Interpreted	Δ	Bytecode	Δ
create simple object	483,500 → 460,408	-4.8%	1,005,558 → 993,735	-1.2%
create nested object	222,166 → 221,387	-0.4%	446,117 → 434,535	-2.6%
create 50 objects via Array.from	9,156 → 9,130	-0.3%	8,786 → 8,820	+0.4%
property read	626,843 → 618,341	-1.4%	937,386 → 919,271	-1.9%
Object.keys	299,860 → 291,793	-2.7%	188,584 → 182,581	-3.2%
Object.entries	108,949 → 108,860	-0.1%	57,956 → 53,314	🔴 -8.0%
spread operator	195,785 → 188,730	-3.6%	213,388 → 196,047	🔴 -8.1%

promises.js — Interp: 12 unch. · avg -0.9% · Bytecode: 🔴 3, 9 unch. · avg -4.2%

Benchmark	Interpreted	Δ	Bytecode	Δ
Promise.resolve(value)	555,821 → 574,426	+3.3%	348,397 → 340,063	-2.4%
new Promise(resolve => resolve(value))	199,301 → 196,811	-1.2%	150,559 → 141,798	-5.8%
Promise.reject(reason)	563,519 → 551,819	-2.1%	341,904 → 309,791	🔴 -9.4%
resolve + then (1 handler)	177,125 → 177,788	+0.4%	155,720 → 144,693	🔴 -7.1%
resolve + then chain (3 deep)	68,679 → 66,520	-3.1%	64,852 → 60,648	-6.5%
resolve + then chain (10 deep)	22,133 → 21,610	-2.4%	22,031 → 21,761	-1.2%
reject + catch + then	99,590 → 98,607	-1.0%	89,964 → 94,165	+4.7%
resolve + finally + then	86,372 → 86,092	-0.3%	77,761 → 72,349	-7.0%
Promise.all (5 resolved)	34,099 → 34,420	+0.9%	25,556 → 25,401	-0.6%
Promise.race (5 resolved)	35,937 → 35,339	-1.7%	27,459 → 27,646	+0.7%
Promise.allSettled (5 mixed)	28,421 → 27,843	-2.0%	24,341 → 22,016	🔴 -9.6%
Promise.any (5 mixed)	33,387 → 32,747	-1.9%	28,708 → 26,939	-6.2%

strings.js — Interp: 🔴 1, 10 unch. · avg -3.5% · Bytecode: 🟢 1, 10 unch. · avg +3.2%

Benchmark	Interpreted	Δ	Bytecode	Δ
string concatenation	421,276 → 413,284	-1.9%	413,494 → 428,009	+3.5%
template literal	728,613 → 633,650	🔴 -13.0%	754,498 → 785,941	+4.2%
string repeat	421,667 → 413,159	-2.0%	1,143,090 → 1,139,257	-0.3%
split and join	140,038 → 134,408	-4.0%	341,991 → 339,379	-0.8%
indexOf and includes	174,623 → 171,834	-1.6%	762,389 → 787,822	+3.3%
toUpperCase and toLowerCase	263,009 → 258,247	-1.8%	781,113 → 844,556	🟢 +8.1%
slice and substring	164,598 → 158,268	-3.8%	860,232 → 888,409	+3.3%
trim operations	195,535 → 192,717	-1.4%	709,380 → 744,216	+4.9%
replace and replaceAll	216,338 → 208,178	-3.8%	658,906 → 662,723	+0.6%
startsWith and endsWith	141,238 → 138,016	-2.3%	641,482 → 680,667	+6.1%
padStart and padEnd	205,090 → 198,950	-3.0%	696,763 → 714,477	+2.5%

typed-arrays.js — Interp: 22 unch. · avg -1.8% · Bytecode: 🟢 4, 🔴 1, 17 unch. · avg +3.3%

Benchmark	Interpreted	Δ	Bytecode	Δ
new Int32Array(0)	339,724 → 332,043	-2.3%	129,856 → 125,748	-3.2%
new Int32Array(100)	314,576 → 306,395	-2.6%	125,476 → 120,502	-4.0%
new Int32Array(1000)	174,872 → 172,752	-1.2%	78,446 → 72,617	🔴 -7.4%
new Float64Array(100)	278,927 → 272,591	-2.3%	112,275 → 104,884	-6.6%
Int32Array.from([...])	184,204 → 186,447	+1.2%	157,486 → 159,417	+1.2%
Int32Array.of(1, 2, 3, 4, 5)	327,980 → 319,247	-2.7%	239,182 → 251,661	+5.2%
sequential write 100 elements	3,708 → 3,636	-2.0%	13,841 → 15,329	🟢 +10.8%
sequential read 100 elements	3,802 → 3,758	-1.1%	10,932 → 11,694	+7.0%
Float64Array write 100 elements	3,444 → 3,419	-0.7%	13,460 → 14,343	+6.6%
fill(42)	45,037 → 47,888	+6.3%	43,657 → 46,344	+6.2%
slice()	212,706 → 208,389	-2.0%	187,983 → 185,397	-1.4%
map(x => x * 2)	8,352 → 8,115	-2.8%	7,099 → 8,262	🟢 +16.4%
filter(x => x > 50)	8,873 → 8,368	-5.7%	7,966 → 8,484	+6.5%
reduce (sum)	8,364 → 8,031	-4.0%	6,670 → 7,389	🟢 +10.8%
sort()	178,019 → 167,898	-5.7%	150,707 → 156,728	+4.0%
indexOf()	458,980 → 451,742	-1.6%	363,634 → 374,590	+3.0%
reverse()	341,487 → 336,536	-1.4%	278,503 → 278,879	+0.1%
create view over existing buffer	413,640 → 404,009	-2.3%	138,484 → 133,683	-3.5%
subarray()	472,425 → 448,770	-5.0%	349,946 → 366,419	+4.7%
set() from array	601,673 → 603,108	+0.2%	233,331 → 263,004	🟢 +12.7%
for-of loop	5,355 → 5,298	-1.0%	20,123 → 21,084	+4.8%
spread into array	18,955 → 18,934	-0.1%	56,405 → 56,183	-0.4%

_{Measured on ubuntu-latest x64. Changes within ±7% are considered insignificant.}

[coderabbitai]

Actionable comments posted: 1

♻️ Duplicate comments (2)

units/Goccia.Runtime.Operations.pas (2)

3381-3385: ⚠️ Potential issue | 🔴 Critical

Avoid reusing FBlueprintBridgeCache for two different blueprint converters.

At Line 3381, this cache can already contain entries created by BlueprintToClassValue (from UnwrapToGocciaValue), but this branch requires ConvertBlueprintToClassValue. Reusing the same entry can drop constructor-path behavior.

🔧 Minimal safe fix

-      if not FBlueprintBridgeCache.TryGetValue(Bp, CachedBridge) then
-      begin
-        CachedBridge := ConvertBlueprintToClassValue(Bp, Self);
-        FBlueprintBridgeCache.Add(Bp, CachedBridge);
-      end;
+      CachedBridge := ConvertBlueprintToClassValue(Bp, Self);

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@units/Goccia.Runtime.Operations.pas` around lines 3381 - 3385,
FBlueprintBridgeCache is being reused for entries created by two different
conversion paths (ConvertBlueprintToClassValue vs BlueprintToClassValue called
from UnwrapToGocciaValue), which can cause incorrect constructor behavior;
change the caching so entries are distinguished by converter or conversion path
— e.g., introduce a separate cache (like FBlueprintConvertBridgeCache) or
include the converter id/type in the cache key, and update the lookup/insert in
the code around ConvertBlueprintToClassValue and UnwrapToGocciaValue to use the
appropriate cache (or key) so ConvertBlueprintToClassValue never reuses entries
created by BlueprintToClassValue.

---

3295-3296: ⚠️ Potential issue | 🔴 Critical

Gate native construction by the root blueprint mapping, not the leaf.

At Line 3295, this condition still routes derived blueprints into the native path even when their root has a wrapped super mapping, which can bypass wrapped super constructor semantics.

🔧 Minimal safe fix

-    if Assigned(Bp.SuperBlueprint) or
-       not FBlueprintSuperValues.ContainsKey(Bp) then
+    WalkBp := Bp;
+    while Assigned(WalkBp.SuperBlueprint) do
+      WalkBp := WalkBp.SuperBlueprint;
+    if not FBlueprintSuperValues.ContainsKey(WalkBp) then

Based on learnings: “Classes extending built-in constructors must be compiled natively via FBlueprintSuperValues … and TGocciaSuperCallHelper.”

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@units/Goccia.Runtime.Operations.pas` around lines 3295 - 3296, The condition
currently gates native construction using the leaf blueprint (Bp) which allows
derived blueprints to take the native path even when their root has a wrapped
super mapping; change the check to determine the root-most blueprint first
(e.g., walk SuperBlueprint pointers until no further SuperBlueprint exists or
call a GetRootBlueprint helper) and then use
FBlueprintSuperValues.ContainsKey(rootBlueprint) (and keep the
TGocciaSuperCallHelper semantics for classes extending built-ins) so that the
decision to use native construction is based on the root blueprint mapping
rather than the leaf Bp.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@units/Goccia.Runtime.Operations.pas`:
- Around line 3408-3410: The branch that calls SyncArraysBack(Self,
Context.Scope) then clears FArrayBridgeCache must also reset the corresponding
dirty flag: after FArrayBridgeCache.Clear add logic to set FArrayBridgeDirty to
False (or 0, matching how the sibling branch resets it) so the state is
consistent; update the code near SyncArraysBack, Context.Scope,
FArrayBridgeCache.Clear to mirror the sibling branch's handling of
FArrayBridgeDirty.

---

Duplicate comments:
In `@units/Goccia.Runtime.Operations.pas`:
- Around line 3381-3385: FBlueprintBridgeCache is being reused for entries
created by two different conversion paths (ConvertBlueprintToClassValue vs
BlueprintToClassValue called from UnwrapToGocciaValue), which can cause
incorrect constructor behavior; change the caching so entries are distinguished
by converter or conversion path — e.g., introduce a separate cache (like
FBlueprintConvertBridgeCache) or include the converter id/type in the cache key,
and update the lookup/insert in the code around ConvertBlueprintToClassValue and
UnwrapToGocciaValue to use the appropriate cache (or key) so
ConvertBlueprintToClassValue never reuses entries created by
BlueprintToClassValue.
- Around line 3295-3296: The condition currently gates native construction using
the leaf blueprint (Bp) which allows derived blueprints to take the native path
even when their root has a wrapped super mapping; change the check to determine
the root-most blueprint first (e.g., walk SuperBlueprint pointers until no
further SuperBlueprint exists or call a GetRootBlueprint helper) and then use
FBlueprintSuperValues.ContainsKey(rootBlueprint) (and keep the
TGocciaSuperCallHelper semantics for classes extending built-ins) so that the
decision to use native construction is based on the root blueprint mapping
rather than the leaf Bp.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 8c42a987-72bf-4421-ad90-414a7dafd5b4

📥 Commits

Reviewing files that changed from the base of the PR and between cb97506 and 4260c58.

📒 Files selected for processing (1)

• units/Goccia.Runtime.Operations.pas

[frostney]

Re: comment 4047936711

Acknowledged. Bytecode benchmarks show avg +1.3% improvement (16 improved, 5 regressed). Interpreted side shows avg -1.6% which is within noise. The typed-array and class construction paths show the expected gains from bypassing the bridge.

[frostney]

Re: comment 4047936460

Acknowledged. All 3497 tests pass in both modes.

[coderabbitai]

♻️ Duplicate comments (1)

units/Goccia.Runtime.Operations.pas (1)

3383-3387: ⚠️ Potential issue | 🔴 Critical

Don't share FBlueprintBridgeCache between BlueprintToClassValue and ConvertBlueprintToClassValue.

At Line 3383 this can reuse a BlueprintToClassValue entry inserted by UnwrapToGocciaValue, but this branch needs the richer ConvertBlueprintToClassValue bridge. If a wrapped-super blueprint is unwrapped before new, InstantiateClass at Line 3407 can silently lose accessors, static members, private members, and __fields__ initializers.

🔧 Minimal safe fix

-      if not FBlueprintBridgeCache.TryGetValue(Bp, CachedBridge) then
-      begin
-        CachedBridge := ConvertBlueprintToClassValue(Bp, Self);
-        FBlueprintBridgeCache.Add(Bp, CachedBridge);
-      end;
+      CachedBridge := ConvertBlueprintToClassValue(Bp, Self);

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@units/Goccia.Runtime.Operations.pas` around lines 3383 - 3387, The bug is
that FBlueprintBridgeCache can return a lightweight BlueprintToClassValue
created by UnwrapToGocciaValue which lacks accessors/static/private members and
__fields__ initializers required by ConvertBlueprintToClassValue and
InstantiateClass; to fix it, change the lookup/insert logic around
FBlueprintBridgeCache in the block with ConvertBlueprintToClassValue so that you
either use a separate cache for converted bridges or detect the existing entry's
kind (e.g., test if CachedBridge is a BlueprintToClassValue vs a full
ConvertBlueprintToClassValue result) and if it's the lightweight form, call
ConvertBlueprintToClassValue to produce the richer bridge and replace the cache
entry before returning; update the code that uses FBlueprintBridgeCache (the
retrieval before calling ConvertBlueprintToClassValue and the add after
conversion) so InstantiateClass always receives the full converted bridge.

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Duplicate comments:
In `@units/Goccia.Runtime.Operations.pas`:
- Around line 3383-3387: The bug is that FBlueprintBridgeCache can return a
lightweight BlueprintToClassValue created by UnwrapToGocciaValue which lacks
accessors/static/private members and __fields__ initializers required by
ConvertBlueprintToClassValue and InstantiateClass; to fix it, change the
lookup/insert logic around FBlueprintBridgeCache in the block with
ConvertBlueprintToClassValue so that you either use a separate cache for
converted bridges or detect the existing entry's kind (e.g., test if
CachedBridge is a BlueprintToClassValue vs a full ConvertBlueprintToClassValue
result) and if it's the lightweight form, call ConvertBlueprintToClassValue to
produce the richer bridge and replace the cache entry before returning; update
the code that uses FBlueprintBridgeCache (the retrieval before calling
ConvertBlueprintToClassValue and the add after conversion) so InstantiateClass
always receives the full converted bridge.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 0e7355fe-c649-45bf-95f6-6767e8863f98

📥 Commits

Reviewing files that changed from the base of the PR and between 4260c58 and c463328.

📒 Files selected for processing (1)

• units/Goccia.Runtime.Operations.pas