SergeychWorks/lyng
1
0
Fork 0
Code Issues 3 Pull Requests Actions Packages Projects 1 Releases 19 Wiki Activity

Expand fast ops optimization plan

Browse Source
This commit is contained in:
Sergey Chernov 2026-02-16 18:09:58 +03:00
parent bfa8a59df3
commit 7f2f99524f
1 changed files with 15 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
notes/fast_ops_optimizations_plan.md
View File

@ -22,3 +22,18 @@ Candidates (not started)
6) Box/unbox audit (done) 6) Box/unbox audit (done)
- Unbox ObjInt/ObjReal in assign-op when target is INT/REAL to avoid boxing + obj ops. - Unbox ObjInt/ObjReal in assign-op when target is INT/REAL to avoid boxing + obj ops.
- MixedCompareBenchmarkTest: 240 ms -> 234 ms. - MixedCompareBenchmarkTest: 240 ms -> 234 ms.
7) 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.
8) 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.
9) Boxing elision pass
- Remove redundant BOX_OBJ when value feeds only primitive ops afterward (local liveness).
- Ensure no impact on closures/escaping values.
10) 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.
11) 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.