I've been able to get substantial improvements with this technique using GCC. See bug 361343 for Windows data.

This data shows improvements in almost all benchmarks. The exception is regex on Linux (very slight regression) and two benchmarks on Mac. I checked the Mac ones in Shark, and they both spend more than 50% of their time under JS_[Get/Set/Find]Property. That tells me these numbers are going to change substantially if we improve property access, so I'll block on a few of those bugs.

Comment on attachment 295585 [details] PGO data for Linux and Mac The Mac uses different settings than Linux because -fvpt profiling proved buggy on Darwin's gcc.

The blocking bugs are fixed. I assume we'll give this a shot after 1.9.1?

This has been done long ago. I can't find the actual bug for it, but eg bug 467271 talks about some followup fixes.