RFR: 8316426: Optimization for HexFormat.formatHex

[温绍锦]

On Fri, 15 Sep 2023 18:20:32 GMT, Claes Redestad <redestad at openjdk.org> wrote:

> What numbers do you get with this?
> 
> In my experiments the improvement was either negligible or small enough to hardly matter. I also tried flipping bits similar to what you're doing for uppercasing but saw a net regression from that.
> 
> I opted for simplicity in #15991 - a simple cleanup that removed a few tiny lookup-tables and still led to a decent improvement. Going the other direction seems like the wrong move.

The byte[] length used by HexFormatBench is 512. In this scenario, the performance improvement of using table lookup is significant.

> I don't think `appendHex` would make the cut as a public API. I think we would need something with much more general utility for that.
> 
> A more high-level idea would be to improve said `String` templates JEP to be able to format more generally into `Appendable`/`StringBuilder`, so that hex, octal - any custom format - can be efficiently appended to a `StringBuilder`. 
> 
> BTW, here's a patch on top of #15768 that simplifies `formatHex(Appendable..)` to simply append the result of `formatHex(byte[]...)`:
> 
> 
> Benchmark                                              (size)  Mode  Cnt     Score     Error   Units
> HexFormatBench.appenderLowerCached                        512  avgt   15     0,202 ±   0,001   us/op
> HexFormatBench.formatLowerCached                          512  avgt   15     0,189 ±   0,016   us/op
> 
> 
> Same speed. Drawback is that the `appender` becomes allocating but short-lived allocations is one of my least concerns:
> 
> ```java 
> diff --git a/src/java.base/share/classes/java/util/HexFormat.java b/src/java.base/share/classes/java/util/HexFormat.java
> index e1d88e3ceb6..68a15307fe0 100644
> --- a/src/java.base/share/classes/java/util/HexFormat.java
> +++ b/src/java.base/share/classes/java/util/HexFormat.java
> @@ -407,38 +407,7 @@ public <A extends Appendable> A formatHex(A out, byte[] bytes, int fromIndex, in
>          int length = toIndex - fromIndex;
>          if (length > 0) {
>              try {
> -                boolean prefixEmpty = prefix.isEmpty();
> -                boolean suffixEmpty = suffix.isEmpty();
> -                boolean delimiterEmpty = delimiter.isEmpty();
> -                if (!prefixEmpty) {
> -                    out.append(prefix);
> -                }
> -                if (suffixEmpty && delimiterEmpty && prefixEmpty) {
> -                    if (out instanceof StringBuilder sb) {
> -                        jla.appendHex(sb, digitCase == Case.UPPERCASE, bytes, fromIndex, toIndex);
> -                    } else {
> -                        for (int i = 0; i < length; i++) {
> -                            toHexDigits(out, bytes[fromIndex + i]);
> -                        }
> -                    }
> -                } else {
> -                    toHexDigits(out, bytes[fromIndex]);
> -                    for (int i = 1; i < length; i++) {
> -                        if (!suffixEmpty) {
> -                            out.append(suffix);
> -                        }
> -                        if (!delimiterEmpty) {
> -                            out.append(delimiter);
> -                 ...

@cl4es Can you help me create an issue for this PR?

-------------

PR Comment: https://git.openjdk.org/jdk/pull/15768#issuecomment-1721726615
PR Comment: https://git.openjdk.org/jdk/pull/15768#issuecomment-1722356329