Bugzilla

Based on the log I've attached, it looks like typescript has the most missing pretenuring. If I run a local comparison report on that benchmark, I do see minor GC as the most important bucket, with a 40x difference in time. WSL also seems to have quite a bit, and a comparison report there again shows that minor GC is the top bucket with a 20x difference.

We probably won't close the entire gap by pretenuring, but that's still enough evidence to mark this as perf-next.

**js-perf-next**: Add alloc sites to scripted constructor calls in baseline, and use this information to pretenure them. Bug 1898270 is a recent example of adding alloc sites for lambdas. As described above, the tricky part with this patch is that the thisObject is allocated by [https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/BaselineCacheIRCompiler.cpp#3666-3670), which is inside shared code used by a variety of different calls. (In particular, both CallScriptedFunction and CallBoundScriptedFunction can reach this point.) Right now, information about the `this` object is stored in a preceding `MetaScriptedThisShape` op that [generates no code](https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/CacheIRCompiler.cpp#9780-9781), but [is used in the transpiler](https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/WarpCacheIRTranspiler.cpp#6751-6776) to allocate `this`.

I think there are two reasonable options for making this work. One, as described above, is to add a new CallKnownScriptedConstructor op that would replace the MetaScriptedThisShape/CallScriptedFunction pair. (This might also need a CallKnownBoundScriptedConstructor op to handle the bound function case?) The second option, easier but more of a hack, would be to add an alloc site to MetaScriptedThisShape (and maybe rename it), rewrite CacheIRCompiler::emitMetaScriptedThis to store an allocSite offset internally in the CacheIRCompiler, and then use that allocsite offset if it exists in BaselineCacheIRCompiler::createThis to load the alloc site out of the stub fields and use it.

Based on the log I've attached, it looks like typescript has the most missing pretenuring. If I run a local comparison report on that benchmark, I do see minor GC as the most important bucket, with a 40x difference in time. WSL also seems to have quite a bit, and a comparison report there again shows that minor GC is the top bucket with a 20x difference.

We probably won't close the entire gap by pretenuring, but that's still enough evidence to mark this as perf-next.

**js-perf-next**: Add alloc sites to scripted constructor calls in baseline, and use this information to pretenure them. Bug 1898270 is a recent example of adding alloc sites for lambdas. As described above, the tricky part with this patch is that the thisObject is allocated by [this code](https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/BaselineCacheIRCompiler.cpp#3666-3670), which is inside shared code used by a variety of different calls. (In particular, both CallScriptedFunction and CallBoundScriptedFunction can reach this point.) Right now, information about the `this` object is stored in a preceding `MetaScriptedThisShape` op that [generates no code](https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/CacheIRCompiler.cpp#9780-9781), but [is used in the transpiler](https://searchfox.org/mozilla-central/rev/771bc161e016e2bd1f7982a88d387916fdf350dd/js/src/jit/WarpCacheIRTranspiler.cpp#6751-6776) to allocate `this`.

I think there are two reasonable options for making this work. One, as described above, is to add a new CallKnownScriptedConstructor op that would replace the MetaScriptedThisShape/CallScriptedFunction pair. (This might also need a CallKnownBoundScriptedConstructor op to handle the bound function case?) The second option, easier but more of a hack, would be to add an alloc site to MetaScriptedThisShape (and maybe rename it), rewrite CacheIRCompiler::emitMetaScriptedThis to store an allocSite offset internally in the CacheIRCompiler, and then use that allocsite offset if it exists in BaselineCacheIRCompiler::createThis to load the alloc site out of the stub fields and use it.