47 lines
3.0 KiB
Markdown
47 lines
3.0 KiB
Markdown
# Fast Ops Optimizations Plan (Draft)
|
|
|
|
Baseline
|
|
- See `notes/nested_range_baseline.md`
|
|
|
|
Candidates (not started)
|
|
1) Primitive comparisons (done)
|
|
- Emit fast CMP variants for known ObjString/ObjInt/ObjReal using temp/stable slots.
|
|
- MixedCompareBenchmarkTest: 374 ms -> 347 ms.
|
|
2) Mixed numeric ops (done)
|
|
- Allow INT+REAL arithmetic to use primitive REAL ops (no obj fallback).
|
|
- MixedCompareBenchmarkTest: 347 ms -> 275 ms.
|
|
3) Boolean conversion (done; do not revert without review)
|
|
- Skip redundant OBJ_TO_BOOL in logical AND/OR when compiler already emits BOOL.
|
|
- MixedCompareBenchmarkTest: 275 ms -> 249 ms.
|
|
4) Range/loop hot path (done)
|
|
- Reuse a cached ObjVoid slot for if-statements in statement context (avoids per-iteration CONST_OBJ).
|
|
- MixedCompareBenchmarkTest: 249 ms -> 247 ms.
|
|
5) String ops (done)
|
|
- Mark GET_INDEX results as stable only for closed ObjString elements to enable fast compares.
|
|
- MixedCompareBenchmarkTest: 247 ms -> 240 ms.
|
|
6) Box/unbox audit (done)
|
|
- Unbox ObjInt/ObjReal in assign-op when target is INT/REAL to avoid boxing + obj ops.
|
|
- MixedCompareBenchmarkTest: 240 ms -> 234 ms.
|
|
7) Primitive list fill with capacity (done)
|
|
- Extended the compiler/runtime fast path from `List.fill(size) { intExpr }` to `List.fill(size, capacity) { intExpr }`.
|
|
- Added `LIST_NEW_INT_CAP` and `LIST_FILL_INT_CAP` so the 3-arg form keeps primitive-int storage instead of falling back to generic stdlib code.
|
|
- `OptTest.testAddToArray2`: `List.fill(n, n + 10) { ... }` dropped from the prior anomaly (~10x slower than 2-arg fill) to the same range as `List.fill(n) { ... }`, roughly `56-67 ms` vs `46-75 ms` after warmup.
|
|
8) Primitive list append preservation (done)
|
|
- Fixed `ObjList.add(...)` to append through the primitive-aware fast path instead of forcing `.list` and boxing the backing storage.
|
|
- `OptTest.testAddToArray2`: appending to the pre-extended list dropped from the prior anomaly (~10x slower) to sub-millisecond / low-millisecond timings (`~0.05-0.16 ms` for the extended list path, `~1.6-4.3 ms` for the baseline path, depending on warmup).
|
|
9) Mixed compare coverage
|
|
- Emit CMP_*_REAL when one operand is known ObjReal in more expression forms (not just assign-op).
|
|
- Verify with disassembly that fast cmp opcodes are emitted.
|
|
10) Range-loop invariant hoist
|
|
- Cache range end/step into temps once per loop; avoid repeated slot reads/boxing in body.
|
|
- Confirm no extra CONST_OBJ in hot path.
|
|
11) Boxing elision pass
|
|
- Remove redundant BOX_OBJ when value feeds only primitive ops afterward (local liveness).
|
|
- Ensure no impact on closures/escaping values.
|
|
12) Closed-type fast paths expansion
|
|
- Apply closed-type trust for ObjBool/ObjInt/ObjReal/ObjString in ternaries and conditional chains.
|
|
- Guard with exact non-null temp/slot checks only.
|
|
13) VM hot op micro-optimizations
|
|
- Reduce frame reads/writes in ADD_INT, MUL_REAL, CMP_*_INT/REAL when operands are temps.
|
|
- Compare against baseline; revert if regression after 10-run median.
|