ClojureWasm Optimizations Completed optimizations and future opportunities, ordered by introduction.
Environment : Apple M4 Pro, 48 GB RAM, Darwin 25.2.0, ReleaseSafe
Measurement : hyperfine (bench/history.yaml)
Result : CW wins cold 14/19 vs Babashka, NaN boxing (D85) live, zwasm (D92) integrated
1. Completed Optimizations
#
Optimization
Phase
Key Impact
1
Arithmetic intrinsics
3
Direct opcodes for +,-,*,/,<,>,= etc.
2
BuiltinFn pointer dispatch
3
No var resolution for core builtins
3
Unified callFnVal dispatch (D36)
10
Single dispatch point for all fn types
4
VM heap allocation (D71)
22c
Prevents C stack overflow in nested calls
5
Mark-Sweep GC (D69, D70)
23
Three-allocator architecture
Phase 24A: VM Core (Pre-24 → 24A.9)
#
Task
Optimization
Biggest Improvement
6
24A.1
Switch dispatch + batched GC
fib_loop 56→31ms (1.8x)
7
24A.2
Stack arg buffer (TreeWalk)
Avoids GC allocation per call
8
24A.3
Fused reduce (lazy-seq collapse)
sieve 2152→40ms (54x)
9
24A.4
Arithmetic fast-path inlining
fib_recursive 502→41ms (12x)
10
24A.5
Monomorphic inline cache
protocol_dispatch 30→27ms
11
24A.9
@branchHint annotations
fib_recursive 41→28ms (1.5x)
Phase 24B: Data Structures
#
Task
Optimization
Biggest Improvement
12
24B.2
HAMT persistent hash map
map_ops 26→14ms (1.9x)
13
24B.4
GC tuning + meta tracing
Fixed F97 sieve crash
Phase 24C: Babashka Parity (24A.9 → 24C.10)
#
Task
Optimization
Biggest Improvement
14
24C.1
Fix fused reduce (__zig-lazy-map)
lazy_chain 6655→17ms (391x)
15
24C.2
Multimethod 2-level cache
multimethod 2053→14ms (147x)
16
24C.3
String stack buffer fast path
string_ops 398→28ms (14x)
17
24C.4
Vector geometric COW + Cons cells
vector_ops 180→14ms (13x)
18
24C.5
GC free-pool recycling
gc_stress 324→46ms (7x)
19
24C.5b
Two-phase bootstrap (D73)
transduce 2134→15ms (142x)
20
24C.7
Filter chain collapsing (D74)
sieve 1645→16ms (103x)
21
24C.9
Zig builtins for update-in etc.
nested_update 39→23ms (1.7x)
22
24C.10
Collection constructor intrinsics
gc_stress 55→35ms (1.6x)
Phase 35X: NaN Boxing (D85)
#
Optimization
Impact
23
4-heap-tag NaN boxing
Value 48→8 bytes, 6x cache
28 heap types, 48-bit address
VM stack 1.5MB→256KB
Phase 36.7: Wasm Interpreter (D86)
#
Optimization
Impact
24
VM reuse (stack cache)
wasm_call 931→118ms (7.9x)
25
Sidetable (branch table)
wasm_fib 11046→7663ms (1.44x)
cached_memory + @memset
DECIDED-AGAINST (ROI < 1%)
Phase 37.2-37.3: VM Superinstructions + Branch Fusion
#
Optimization
Impact
26
Superinstructions (10 fused ops)
arith_loop 53→40ms (1.33x)
27
Compare-and-branch fusion (7 ops)
arith_loop 40→31ms (1.29x)
28
Recur-loop fusion
Dispatch: 6→4 per loop iteration
Cumulative (37.1 base → 37.3)
arith_loop 53→31ms (1.71x)
Phase 37.4: JIT PoC — ARM64 Hot Loop Native Code (D87)
#
Optimization
Impact
29
ARM64 JIT (hot integer loops)
arith_loop 31→3ms (10.3x)
used_slots bitset (skip fn_val)
Avoids deopt on closure self-ref
THEN path skip in analyzeLoop
Handles real compiler bytecode
Cumulative (37.1 base → 37.4)
arith_loop 53→3ms (17.7x)
End-to-End Progression (Pre-24 → Post-zwasm)
Benchmark
Pre-24
24C.10
Post-zwasm
Speedup
fib_recursive
542
24
18
30x
map_filter_reduce
4,013
17
6
669x
sieve
2,152
16
6
359x
lazy_chain
21,375
16
9
2,375x
transduce
8,409
16
6
1,402x
multimethod_dispatch
2,373
15
6
396x
real_workload
1,286
22
9
143x
All times in ms (warm), ReleaseSafe, Apple M4 Pro. Post-zwasm = latest entry.
Full table: bench/history.yaml.
Cross-Language Summary (Cold, Phase 24C.10) CW vs Babashka (Cold): CW wins 14/19, BB wins 5/19, 1 skip.
CW vs Ruby: CW wins 20/20. CW vs Java: CW wins 18/20. CW vs Python: CW wins 11/20.
Startup: C 3.9 / Zig 5.8 / BB 8.0 / Py 11.1 / CW 14.2 / Java 21.2 / Ruby 30.1 ms.
Re-run with bash bench/compare_langs.sh --both for latest numbers.
CW-side (can implement now)
ID
Technique
Expected Impact
Effort
Notes
F102
map/filter chunked processing
lazy-seq alloc
MEDIUM
CW range is eager; deferred
F103
Escape analysis
GC overhead
HIGH
Compiler detects local-only
F104
Profile-guided IC extension
2x polymorphic
MEDIUM
Beyond monomorphic IC
—
Generational GC
2-5x allocation
HIGH
Write barriers required
—
SmallVector (inline 2-3 elts)
Small vec alloc
MEDIUM
NaN boxing extension
—
Closure stack allocation
1-2 capture
MEDIUM
Avoid heap for small closures
Wasm-side (in zwasm repository)
Technique
Expected Impact
Notes
Register-based IR
1.5-3x all
Major rewrite of zwasm interpreter
ARM64 JIT for Wasm
5-20x
Reuse CW JIT PoC (D87) patterns
Constant folding / DCE
5-10%
Low ROI for current benchmarks
ID
Technique
Notes
F99
Iterative lazy-seq realize
D74 partial fix, general case left
F4
HAMT persistent vectors
Current ArrayMap sufficient
F120
Native SIMD (@Vector)
Profile first before investing
Technique
Reason
wasmtime-as-library
+20MB binary, Rust dep. Keep zwasm (small, controlled)
Tail-call dispatch
0% improvement on Apple M4 (45.3 measured)
RRB-Tree vectors
Vectors rarely sliced in practice
cached_memory (Wasm)
ROI < 1% in benchmarks
SmallString widening
asString() returns []const u8 — lifetime problem
String interning
string_ops bottleneck is alloc, not comparison
COMPLETED (Phase 36.11+37+45)
Technique
Phase
Result
F101 into() transient
36.11
core.clj transient/persistent!
F105 JIT compilation (PoC)
37.4
ARM64 hot loops, arith_loop 17.7x
Predecoded IR (Wasm)
45.2
Fixed-width 8-byte instrs, 1.7-2.5x
Superinstructions (Wasm)
45.4
11 fused opcodes, fib 1.3x
Cached memory pointer (Wasm)
45.5
Marginal (~3% on sieve)
4. Wasm Performance (Post-zwasm v0.1.0, Register IR + ARM64 JIT)
Benchmark
CW warm (ms)
wasmtime (ms)
Ratio
vs Phase 45
fib(20)x10K
641
211
3.0x
6.8x
tak(18,12,6)x10K
2,786
1,174
2.4x
5.1x
arith(1M)x10
0.8
0.1
8.0x
—
sieve(64K)x100
21
4.8
4.4x
9.4x
fib_loop(25)x1M
22
2.2
10.0x
8.0x
gcd(1M,700K)x1M
54
41
1.3x
5.8x
CW startup (4.1ms) < wasmtime startup (5.5ms).
zwasm Register IR + ARM64 JIT brings most benchmarks within 3-10x of wasmtime.
gcd achieves near-parity (1.3x). Call-heavy workloads (fib, tak) at 2.4-3.0x.
Full history: bench/wasm_history.yaml.
Topic
Location
Benchmark history
bench/history.yaml
Wasm bench history
bench/wasm_history.yaml
Cross-language script
bench/compare_langs.sh
Cross-language results
bench/cross-lang-results.yaml
D85 NaN boxing
.dev/decisions.md
D87 JIT PoC
.dev/decisions.md
D92 zwasm integration
.dev/decisions.md
Checklist items
.dev/checklist.md
zwasm repository
../zwasm/ or GitHub