RFR: 8278518: String(byte[], int, int, Charset) constructor and String.translateEscapes() miss bounds check elimination [v3]
Roland Westrelin
roland at openjdk.java.net
Thu Jan 13 10:24:15 UTC 2022
> In the micro benchmark for this bug, the bytecode has a single loop
> head with 2 backedges. One branch is taken a lot more frequently than
> the other (for instance, profiling reports 3820938 for the most
> common, 134068 for the least common). When ciTypeFlow builds its loop
> tree, the least common loop ends up as the inner loop (ciTypeFlow uses
> the tail block number to order loops with a shared header). Then
> ciTypeFlow clones the inner loop head.
>
> The results I get then is:
>
> StringBenchmark.safeDecoding avgt 5 141.965 ± 53.511 ns/op
> StringBenchmark.unsafeDecoding avgt 5 123.051 ± 77.855 ns/op
>
> The unsafe version uses unsafe instead of standard array accesses. The
> reason for the unsafe version is to demonstrate that a missed range
> check elimination in the common path hurts performance (the range
> check is not performed on the loop iv but another phi so can't be
> eliminated).
>
> The patch I propose takes back branch count (from profile data) into
> account when building the ciTypeFlow tree. As a consequence, the inner
> loop (the one for which the loop head is cloned), is the most most
> frequent one. The range check in the common path is also hoisted in
> that case (it's now performed on the iv). Running the micro benchmark
> again, I get:
>
> StringBenchmark.safeDecoding avgt 5 53.368 ± 2.232 ns/op
> StringBenchmark.unsafeDecoding avgt 5 55.565 ± 5.124 ns/op
>
> Both are improved (2-3x) and the unsafe version doesn't perform better.
Roland Westrelin has updated the pull request incrementally with four additional commits since the last revision:
- whitespaces
- more review
- simple benchmark
- benchmark
-------------
Changes:
- all: https://git.openjdk.java.net/jdk/pull/7034/files
- new: https://git.openjdk.java.net/jdk/pull/7034/files/ceac8620..8b61e82a
Webrevs:
- full: https://webrevs.openjdk.java.net/?repo=jdk&pr=7034&range=02
- incr: https://webrevs.openjdk.java.net/?repo=jdk&pr=7034&range=01-02
Stats: 127 lines in 3 files changed: 126 ins; 0 del; 1 mod
Patch: https://git.openjdk.java.net/jdk/pull/7034.diff
Fetch: git fetch https://git.openjdk.java.net/jdk pull/7034/head:pull/7034
PR: https://git.openjdk.java.net/jdk/pull/7034
More information about the hotspot-compiler-dev
mailing list