RFR: 8176188: jdk/internal/misc/JavaLangAccess/NewUnsafeString.java failing since 9-b93 (original) (raw)

Claes Redestad claes.redestad at oracle.com
Mon Dec 4 21:47:20 UTC 2017

Hi Martin,

On 2017-12-04 22:06, Martin Buchholz wrote:

I'm rather sad about what happened to our non-copying String constructions for trusted code.  This is a performance regression with the change in String representation that should have blocked that change IMO.  I think we should have a plan for moving in the opposite direction.  I don't think we can implement something as ambitious as Rust's ownership tracking, so have to restrict ourselves to trusted code.  The use case that keeps coming up is constructing zip entry names, which are much more likely to be pure ASCII than their file contents.

I don't have a good design for how one could do that, and who the trusted set of callers is (at least java.base, not java.lang), but I think we should set a direction.

as I alluded to in a footnote there exists a non-copying String(byte[] value, byte coder) constructor - the problem is that it's somewhat cumbersome to use:

Some of the more performance sensitive uses outside of java.lang was addressed by the Compact Strings update, for example the implementation backing java.util.UUID was somewhat surprisingly moved into java.lang.Long::fastUUID[1]. Something similar is doable for the java.sql types, but further complicated by those classes being in a different module, and ultimately questionable since their implementations in JDK 9 are quite a bit more performant than in any previous release (thus not technically a regression).

That leaves StringJoiner as the one case that stands out. And the fact that existing uses of String(byte[], byte) are a bit of an eye-sore[1!!1!].

One idea I'm tinkering with here is to have a trusted, package-private SharedStringBuilder added to java.lang and exposed via SharedSecrets. It'd more or less mimic StringBuilder (including deal with inflating the byte[] when necessary, encapsulate the awkward String.COMPACT_STRINGS checks etc) but enable calling String(byte[], byte) in the toString() call.  To be effective it'll only have a single constructor taking the capacity, and should probably throw IOOBE rather than resize the internal buffer. Some cases like Long::fastUUID could probably be much simplified by using such a builder instead (for a very minimal overhead). Does that sound reasonable? At any rate I think of this as a possible follow-up RFE, and not an alternative to the cleanup/"bugfix" at hand.

Thanks!

/Claes

[1] http://hg.openjdk.java.net/jdk/jdk/file/532cdc178e42/src/java.base/share/classes/java/lang/Long.java#l427

More information about the core-libs-dev mailing list