Bug 83678 – Global variable predictions should not be coalesced unnecessarily

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[Filip Pizlo]

The DFG still coalesces all predictions for all global variables even though it isn't strictly necessary.  This is a holdover from when we would sometimes compile code without having complete predictions for the entire code block, and would have no way of recompiling later - so we needed to play strange games to "fill in" predictions for those parts of the code block that we wanted to optimize but for which we had no profiling data.  But these days we either manager to get all of the profiling we need before compiling things (because of our well now mature heuristics) or we trigger recompilation upon repeated attempts to execute code that did not have profiling data.  Either way, in steady state, all running DFG code will have ample profiling information.

[Filip Pizlo]

Created attachment 136639 [details]
the patch

[Geoffrey Garen]

Comment on attachment 136639 [details]
the patch

r=me

[Benjamin Poulain]

Could this have caused the regression in Bindings/event-target-wrapper?

The regression: http://webkit-perf.appspot.com/graph.html#tests=[[8062,2001,32196]]&sel=1334183542158.416,1334203959676.6643,90.90909090909076,1386.3636363636363&displayrange=7&datatype=running
The delta: http://trac.webkit.org/log/?rev=113906&stop_rev=113898&verbose=on

This is the only non-trivial commit in this delta.

[Eric Seidel (no email)]

Attachment 136639 [details] was posted by a committer and has review+, assigning to Filip Pizlo for commit.

[Filip Pizlo]

This was fixed ages ago.