================
@@ -1159,12 +1256,29 @@ class Sema;
return reinterpret_cast<T *>(FreeSpaceStart);
}
+ template <typename T> T *allocateDeferredCandidate() {
+ T *C = slabAllocate<T>(1);
----------------
erichkeane wrote:
Seems funny to me to `slab` allocate in increments of 1. I have a 'silly'
idea... Since 'deferred candidates' are likely to only be a handful per
lookup... I wonder if we could:
1- Add a few-bits to `DeferredTemplateOverloadCandidate` for `leftInSlab`.
2- on `!FirstDeferredCandidate`, allocate (say)4 items. Set the 1st item to
'3' left.
3- Then, find the last candidate the same way. if `leftInSlab` != 0,
emplacement-new/etc into the 'next' location, set the `leftInSlab` to be the
previous `-1`. If `leftInSlab == 0`, allocate.
I realize it is more work here, and can be done in a followup, but there might
be a mild perf improvement for cache locality. I suspect if we set this to
something like 4, which requires 2 bits in the structure, high-90%s of lookups
will take only 1 allocation.
WDYT?
https://github.com/llvm/llvm-project/pull/133426
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
