RFR: 8315585: Optimization for new BigDecimal(String)

Shaojin Wen duke at openjdk.org
Mon Mar 11 11:15:14 UTC 2024


On Sun, 10 Mar 2024 16:11:12 GMT, Shaojin Wen <duke at openjdk.org> wrote:

> The current BigDecimal(String) constructor calls String#toCharArray, which has a memory allocation.
> 
> 
> public BigDecimal(String val) {
>     this(val.toCharArray(), 0, val.length()); // allocate char[]
> }
> 
> 
> When the length is greater than 18, create a char[]
> 
> 
> boolean isCompact = (len <= MAX_COMPACT_DIGITS); // 18
> if (!isCompact) {
>     // ...
> } else {
>     char[] coeff = new char[len]; // allocate char[]
>     // ...
> }
> 
> 
> This PR eliminates the two memory allocations mentioned above, resulting in an approximate 60% increase in performance..

The performance numbers under MacBookPro M1 Max are as follows:

-Benchmark                              Mode  Cnt   Score   Error  Units (master)
-BigDecimals.testConstructorWithString  avgt   15  66.005 ? 0.733  ns/op

+Benchmark                              Mode  Cnt   Score   Error  Units (current version #aaf04f6)
+BigDecimals.testConstructorWithString  avgt   15  39.720 ? 0.266  ns/op +66.17%

Compared to directly using a char[], using CharBuffer.wrap leads to a 27% decrease in performance.

The performance numbers under MacBookPro M1 Max are as follows:

Benchmark                                 Mode  Cnt   Score   Error  Units (char[] direct)
BigDecimals.testConstructorWithCharArray  avgt   15  41.998 ? 0.259  ns/op

Benchmark                                 Mode  Cnt   Score   Error  Units (CharBuffer.wrap)
BigDecimals.testConstructorWithCharArray  avgt   15  53.533 ? 1.297  ns/op -27.46%

Now, new BigDecimal(char[]) and new BigDecimal(String) are using the same algorithm. 

They are two separate paths, and both have seen performance improvements.

The performance numbers under MacBookPro M1 Max are as follows:

-Benchmark                                 Mode  Cnt    Score   Error  Units (master)
-BigDecimals.testConstructorWithCharArray  avgt   15   48.448 ? 1.879  ns/op
-BigDecimals.testConstructorWithString     avgt   15   65.203 ? 3.871  ns/op

+Benchmark                                 Mode  Cnt    Score    Error  Units (current #)
+BigDecimals.testConstructorWithCharArray  avgt   15   41.166 ?  0.940  ns/op +17.68%
+BigDecimals.testConstructorWithString     avgt   15   39.289 ?  0.159  ns/op +65.95%

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

PR Comment: https://git.openjdk.org/jdk/pull/18177#issuecomment-1987283721
PR Comment: https://git.openjdk.org/jdk/pull/18177#issuecomment-1987601447
PR Comment: https://git.openjdk.org/jdk/pull/18177#issuecomment-1988185886


More information about the core-libs-dev mailing list