On Tue, 17 Nov 2020 14:55:07 GMT, Maurizio Cimadamore <mcimadam...@openjdk.org> wrote:
> The current memory segment implementation defines a hierarchy with 3 concrete > classes: one for heap segments, one for native segments and one for mapped > segments. > > Since there can be many kinds of heap segments (e.g. created from a byte[] or > from a float[]) the current implementation is prone to type profile pollution > problems: if enough heap segments are created (of different kinds), the JIT > compiler will give up on speculating on the heap segment kind, which will > then result in poor performances. > > This issue can be reproduced in one of the existing benchmark, by adding some > initialization code which is enough to pollute the types profiles. When that > happens, performance numbers look like the following: > > Benchmark (polluteProfile) Mode Cnt Score > Error Units > LoopOverNonConstantHeap.segment_loop false avgt 10 0.285 ± > 0.003 ms/op > LoopOverNonConstantHeap.segment_loop true avgt 10 5.540 ± > 0.143 ms/op > > (Thanks to Vlad for coming up for the exact incantation which leads to > profile pollution :-) ) > > The solution is to create a sharp subclass for each heap segment case. With > this, C2 has always a sharp Unsafe *base* to work with, and performances are > stable regardless of profile pollution attempts. > > This patch also tweaks the benchmark for heap segments so that it checks it > with and without profile pollution. Marked as reviewed by jvernee (Committer). ------------- PR: https://git.openjdk.java.net/jdk/pull/1259
