Thanks, Claes. I think we're in agreement! I did that shared String optimization for StringJoiner. It's a lot harder to justify in the new String world because we have to handle non-ASCII, and the non-ASCII case is actually fairly likely. Does it make sense to keep COMPACT_STRINGS as an option in openjdk? If essentially every user runs with that flag turned on, bit rot is likely to set in, and even if not, we have to remain ever vigilant to not break anyone turning it off. The code in Long.fastUUID is indeed ugly. I've never heard of UUID creation being a bottleneck. At Google it sometimes seems all our java performance problems are with zip file manipulation. I might have avoided the code duplication in Long.fastUUID by: - build the all-ASCII byte[] unconditionally - if COMPACT_STRINGS, then use the secret LATIN1 constructor, else use the public constructor(byte[], ISO_8859_1) One idea for a public performant API is to add something like Charset.toString(byte[]) or Charset.toString(byte[], int start, int length) via default method on Charset. Then; override that in ISO_8859_1 to do only a single copy (and can we cheat somehow to do no copies?) On Mon, Dec 4, 2017 at 1:47 PM, Claes Redestad <claes.redestad@oracle.com> wrote:

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:

- first off, the caller needs to be aware about the value of String.COMPACT_STRINGS: if false, all strings needs to be UTF-16 encoded and the coder byte always set to String.UTF16 - secondly, the caller needs to know if the byte[] you're constructing needs to be LATIN-1 or UTF-16 encoded up front and act accordingly

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/jav a.base/share/classes/java/lang/Long.java#l427

On 2017-12-05 00:46, Martin Buchholz wrote:

Thanks, Claes.

I think we're in agreement!

Great!

I did that shared String optimization for StringJoiner. It's a lot harder to justify in the new String world because we have to handle non-ASCII, and the non-ASCII case is actually fairly likely.

Let's sleep on it, at least. :-)

Does it make sense to keep COMPACT_STRINGS as an option in openjdk?  If essentially every user runs with that flag turned on, bit rot is likely to set in, and even if not, we have to remain ever vigilant to not break anyone turning it off.

I think it was only kept as a safeguard in case anyone would run into serious trouble with the implementation (unlikely), so it might be reasonable to deprecate the flag in an upcoming release. Maybe survey if anyone ever had to opt-out of it first.

The code in Long.fastUUID is indeed ugly.  I've never heard of UUID creation being a bottleneck.  At Google it sometimes seems all our java performance problems are with zip file manipulation.

I see Steven already said Hi. I helped him get this optimization in, and it did look a bit nicer originally. :-)

I might have avoided the code duplication in Long.fastUUID by: - build the all-ASCII byte[] unconditionally - if COMPACT_STRINGS, then use the secret LATIN1 constructor, else use the public constructor(byte[], ISO_8859_1)

I think the implementors felt it would be unfair to penalize those who had to opt-out of it for some unlikely reason.

One idea for a public performant API is to add something like Charset.toString(byte[]) or Charset.toString(byte[], int start, int length) via default method on Charset.  Then; override that in ISO_8859_1 to do only a single copy (and can we cheat somehow to do no copies?)

I think these would need to clone the byte[] most of the time, so I think it'd not be that much different from new String(byte[], ISO_8859_1). The variant with Charset.toString(byte[], int, int) could optimize away one copy when start > 0 or length < bytes.length, but this seems like a corner case gain that doesn't really motivate adding two new methods to every Charset class. I guess a SharedStringBuilder (InlineStringBuilder?) could be made public if we used $suitable_concurrency_trick to ensure the toString() is callable at most once and that once it's been called the byte[] can't be written to again. Maybe a ThreadLocal-based implementation could suffice if we disallow nested InlineStringBuilder use (shouldn't be too harsh of a limitation). /Claes

On Mon, Dec 4, 2017 at 1:47 PM, Claes Redestad <claes.redestad@oracle.com <mailto:claes.redestad@oracle.com>> wrote:

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:

- first off, the caller needs to be aware about the value of String.COMPACT_STRINGS: if false, all strings needs to be UTF-16 encoded and the coder byte always set to String.UTF16 - secondly, the caller needs to know if the byte[] you're constructing needs to be LATIN-1 or UTF-16 encoded up front and act accordingly

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/cla... <http://hg.openjdk.java.net/jdk/jdk/file/532cdc178e42/src/java.base/share/classes/java/lang/Long.java#l427>