On the Wikimedia cluster, 1.6% of MediaWiki wall-clock time is burnt on
calls from Lua into Scribunto_LuaSandboxCallback::frameExists()[1]. We
can optimize away many of these calls by not calling into PHP to check
if 'empty' or 'current' exist: the engine always reports that the
'empty' frame exists, and 'current' is guaranteed to have been set up
(in LuaEngine::setupCurrentFrames) prior to calling into Lua.
To help validate this, I added debug logging to the current production
branch of Scribunto[2] to see if there are any cases where
Scribunto_LuaSandboxCallback::frameExists('current') is false. As I
write this commit message, the logging code has been active for 24H and
there have not been any occurrences.
[1]: https://performance.wikimedia.org/arclamp/svgs/daily/2021-03-16.excimer-wall.all.reversed.svgz
[2]: https://gerrit.wikimedia.org/r/c/mediawiki/extensions/Scribunto/+/672836
Change-Id: I1902b711c9a442a5a42745a582a6a9ff988a355f
This is getting close to the point of "don't do that, just wrap the
built-in". But since it's a regression in a recent patch, let's restore
the old behavior here.
Bug: T236092
Change-Id: Ieddc23d942bc91fd0246ae14d8a4af7719e3834f
When an #invoke is passed as an argument to another #invoke,
mw.getCurrentFrame() at module scope will return the wrong frame.
On the PHP side, we need to always reset the frame when processing
an #invoke, not just when there's no frame already. I don't remember why
I82dde43e wasn't done that way, but changing it doesn't make any tests
fail and Scribunto tends to have good tests.
On the Lua side, we need to do the same. The logic wih mw.getCurrentFrame()
using a global that gets stored, modified, and reset in several places
was getting confusing, so this patch reworks the logic to inject a
globalless mw.getCurrentFrame() into each #invoke's cloned environment
instead.
Bug: T234368
Change-Id: I8cb5bc4dc14c9b448c9f267e0539daa75e72af4c
