Decode the opcode header once per instruction via RuntimeOp.seek

[NullVoxPopuli]

RuntimeOpImpl now decodes type, size, and isMachine in a single seek(offset) call instead of re-reading the heap in three getters.

The RuntimeOp interface fields become readonly and seek is the only way to move the cursor.

[NullVoxPopuli]

Benchmark: main (bdb2580) vs nvp/perf/runtime-op-seek (a922570)

Ran via pnpm bench (the krausest js-framework-benchmark scenario in smoke-tests/benchmark-app, tracerbench, 20 samples each, 8x CPU throttle). Negative delta means the PR is faster. Only rows tracerbench marks significant (p < 0.05) are meaningful.

Phase	main median	PR median	Δ (estimator)	95% CI	Significant
Total duration	9,497ms	9,519ms	8ms (+0.1%)	-444ms .. 410ms
render	143ms	138ms	-3ms (-2.3%)	-9ms .. 3ms
render1000Items1	348ms	329ms	-16ms (-4.7%)	-34ms .. -2ms	✅
clearItems1	37ms	38ms	0ms (-1.2%)	-2ms .. 2ms
render1000Items2	292ms	294ms	1ms (+0.5%)	-16ms .. 16ms
clearItems2	135ms	135ms	0ms (-0.3%)	-2ms .. 2ms
render10000Items1	2,742ms	2,718ms	-40ms (-1.5%)	-183ms .. 43ms
clearManyItems1	369ms	370ms	1ms (+0.3%)	-6ms .. 8ms
render10000Items2	2,732ms	2,730ms	1ms (+0.1%)	-126ms .. 129ms
clearManyItems2	264ms	261ms	-3ms (-1.1%)	-11ms .. 3ms
render1000Items3	260ms	258ms	-7ms (-2.7%)	-25ms .. 10ms
append1000Items1	346ms	364ms	14ms (+4.0%)	-23ms .. 53ms
append1000Items2	320ms	325ms	4ms (+1.4%)	-15ms .. 33ms
updateEvery10thItem1	286ms	290ms	4ms (+1.5%)	-16ms .. 23ms
updateEvery10thItem2	305ms	307ms	-2ms (-0.6%)	-30ms .. 19ms
selectFirstRow1	75ms	79ms	2ms (+2.5%)	-6ms .. 10ms
selectSecondRow1	59ms	60ms	0ms (+0.2%)	-9ms .. 8ms
removeFirstRow1	194ms	193ms	5ms (+2.4%)	-7ms .. 15ms
removeSecondRow1	183ms	191ms	12ms (+6.3%)	-1ms .. 28ms
swapRows1	131ms	127ms	-1ms (-0.9%)	-12ms .. 13ms
swapRows2	124ms	127ms	3ms (+2.2%)	-8ms .. 15ms
clearItems4	104ms	106ms	2ms (+2.0%)	-1ms .. 5ms
paint	9ms	11ms	1ms (+7.4%)	-3ms .. 4ms

[johanrd]

@NullVoxPopuli FWIW: Claude found flaws (but for VM_LOCAL_DEV only) Recorded as two failing test here: johanrd#41

[NullVoxPopuli]

@johanrd can we trust that those very private tests are valid?

afaik, the test setup could be wrong?

[johanrd]

@NullVoxPopuli true, though, the tests were created to prove this failure:

1. debug.ts ends the walk with opcode.seek(-_size)
2. _size comes from size = ((header & OPERAND_LEN_MASK) >> ARG_SHIFT) + 1, so it is never 0 → the argument is always <= -1
3. console.log(new Int32Array(8)[-1]) // undefined

[NullVoxPopuli]

right, but would an app ever do run in to that?

[johanrd]

no, it wont affect real apps, it only affectsVM_LOCAL_DEV (trace logging is perhaps already not working?)

Maybe just delete the line, not port it? (the op seems to be discarded immediately after, nothing reads it? – not very familiar with opcodes, sorry.)

[johanrd]

@NullVoxPopuli else, very cool that you have found these improvements to the vm (also the other prs!). Thanks!