Allocations are redirected by overriding `operator new` and `operator
delete` on the IRNode class. This allows us to use our existing
`unique_ptr` and `make_unique` calls as-is. The Pool class is simple;
it holds a fixed number of nodes and recycles them as they are returned.
A fixed pool size of 2000 nodes was chosen. That is large enough to hold
the contents of `sksl_large` during compilation, but it can be
overflowed by very large shaders, or if multiple programs are converted
at the same time. Exhausting the pool is not a problem; if this happens,
additional nodes will be allocated via the system allocator as usual.
More elaborate schemes are possible but might not add a lot of value.
Thread safety is accomplished by placing the pool in a `thread_local`
static during a Program's creation and destruction; the pool is freed
when the program is destroyed. One important consequence of this
strategy is that a program must free every node that it allocated during
its creation, or else the node will be leaked. In debug, leaking a node
will be detected and causes a DEBUGFAIL. In release, the pool will be
freed despite having a live node in it, and if that node is later freed,
that pointer will be passed to the system `free` (which is likely to
cause a crash).
In this CL, iOS does not support pooling, since support for
`thread_local` was only added on iOS 9. This is fixed in the followup
CL, http://review.skia.org/328837, which uses pthread keys on iOS.
Nanobench shows ~15% improvement:
(last week) http://screen/5CNBhTaZApcDA8h
(today) http://screen/8ti5Rymvf6LUs8i
Change-Id: I559de73606ee1be54e5eae7f82129dc928a63e3c
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/326876
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
e16eca95f5