[folbricht]

A SETFIELD that adds a key misses the loop's site hint, because the key is not in the table yet. That is every field written into a table just built, as in setmetatable({x = x, y = y}, Point), or self.z = z in a constructor. The miss went to newindex, which:

• looked the key up twice (Get, then findNode in set),
• hashed it at run time,
• measured the table before and after to charge memory,
• then looked it up a third time to learn the hint.

constSet handles these writes in the slow path when the table has no metatable, is not frozen or guarded, and the key does not start with __. It probes with the constant's hash computed at compile time, as constField does for reads. It sets the key if it is there, or adds it with insert when no resize is needed, then learns the hint. Every other case still goes to newindex. With no resize, the table's size doesn't change, so nothing needs to be charged. insert now returns the node index.

The insert is not done inside the loop. Doing it there makes fast larger than the inliner's big-function threshold. Small helpers like setInt and setBool then become calls, and every opcode spills; asmgate catches this.

Effect, fast loop unless noted:

	before	after
oop, instructions per run	603M	544M (-9.7%)
oop, ms (minimum of 6 interleaved runs)	85.3	77.1 (-9.6%)
oop, checked loop, ms	89.6	81.4 (-9.1%)
binarytrees, json, parse, instructions		within ±0.5%

A new case in testdata/meta.lua, with its expected output from C Lua, covers:

• adds past the constructor's size,
• setting a deleted key again,
• __newindex on absent and present keys,
• a __ key added to a table that is in use as a metatable.