[folbricht]

I reran all three benchmark tables on main at 860bcd2 for the v1.1.0 release. The machine was quiet, with the usual powersave governor, and I used the commands in the method section. The README's summary table is updated to match.

Scripts across implementations: glua averages 1.20x C Lua, down from 1.23x.

• format 1.10x → 0.73x (Format string.format's common conversions without fmt or throwaway strings #45)
• pairs 1.48x → 1.25x (Skip the hash lookup in pairs when the last node still holds the key #42)
• parse 1.57x → 1.49x (Build pattern captures in the result buffer instead of a new slice per match #46)
• binarytrees 0.98x → 0.93x

The other engines came out within 0.02 of their published averages.

glua's two loops:

	before	now
Fast loop	1.22x	1.15x
Checked loop	1.40x	1.31x
Checked loop while the GC is marking	3.01x	2.82x

Calls between Go and Lua: I traced the two rows that moved before publishing them.

• The rule rows are about 15% slower on the fast loop (140 → 162 ns). I bisected that to Build pattern captures in the result buffer instead of a new slice per match #46, which isn't on that path. perf stat shows the same number of instructions (26.0 vs 26.1 billion) but 28 times more branch mispredictions in the dispatch of the interpreter loop. Build pattern captures in the result buffer instead of a new slice per match #46 moved the bound-call wrapper by 128 bytes, which changes the branch history the CPU predicts the dispatch from. So this is code layout, not a code change, and the method section now says that layout can move single rows like this.
• A call into a private VM is about 5 ns slower (77 → 82 ns). This one is real: Long-entry fixes: name registry and stack shrink #40, the long-entry fixes, added 47 instructions to each call (reset, call, VM.enter, and the stack-shrink check in VM.leave), and Measure what an entry or a VM keeps with LiveBytes #37, live bytes, another 7. That's the cost of those features. It may be partly recoverable later.

The table rows are pasted from the tools' output. The raw Go results are in the gitignored bench/results/.