Repository navigation
Description
What happens
Formatting allocates more than a thousand times the size of its input, and the native binary spends most of its run collecting it. 865 files from java.base of JDK 21, 15 MB of source, --replace with 21.0.2 on 10 cores:
| Wall | Of it in GC pauses | Peak RSS | |
|---|---|---|---|
| Native binary | 6.9 s | 5.0 s: 17 full collections for 3.05 s, 23 incremental for 1.94 s | 2.4 GB |
java -cp ... Main, JDK 21 |
6.2 s | 4.3 GB |
The native run's GC log adds up to at least 16.6 GB allocated, about 1,100 bytes for each byte of source. The image is built without --gc, so its serial collector stops all 20 threads for each collection, and they format on about 2 of the 10 cores (15 s of CPU in 6.9 s).
Input and output are not where it goes: reading the 865 files takes 0.04 s, and Main's reading and writing are 0.5% of the allocation, so memory-mapped files or other I/O changes cannot help.
Where it goes
Flight Recorder on the JVM run, with -XX:FlightRecorderOptions:stackdepth=2048 so that no stack was cut, 1,611 allocation samples and 469 CPU samples:
Step of FormatFileCallable.formatFile |
Allocation | CPU |
|---|---|---|
Formatter.formatSource, the layout |
65.5% | 56.1% |
StringWrapper.wrap |
16.8% | 23.5% |
RemoveUnusedImports |
12.9% | 19.8% |
ImportOrderer |
4.0% | 0.6% |
Main, reading and writing |
0.5% |
- The file is parsed several times. 42% of the CPU is in javac's parser and scanner: 16% for the layout, 16% for the parse of its own that
RemoveUnusedImportsmakes, 10% inStringWrapper, whose rounds parse again and parse before and after to compare the trees. - Each parse sets javac up from scratch: 18.1% of the allocation.
Formatter.parseJcCompilationUnit,JavacTokens.getTokensandJavaInput.buildTokseach create a newContext, and with it aSharedNameTable(NameImpl[0x8000]andbyte[0x20000]), the writers ofLog, and the buffers ofClassWriter(throughModules),ClassReaderand evenGen, which parsing does not use. javac keeps disposed name tables for reuse (SharedNameTable.create), but nothing here callsNames.dispose(). - The top sites otherwise:
OpsBuilder.build11.5%, the memoizing supplier ofHasUniqueId.ord4.6% plus 2.3% for itsSuppliersobjects,Doc.<init>4.3%,InputOutput.makeKToIJ3.1%,JavaOutput.append2.7%. By class:byte[]23.4%,char[]8.3%,Object[]8.1%, boxedInteger7.2%, GuavaRange2.7%.
Proposal
In order of what each is likely to give for its work, each measured on the run above:
- The native image's collector: a larger young generation, another collector where GraalVM offers one for the platform, fewer threads under the serial one (Read, format and write each file on a worker thread, and keep only what the output needs #228).
- javac set up once per thread, or reused: call
Names.dispose()after each parse so that javac hands the name table out again, and find what makes a parse createModules,ClassWriter,ClassReaderandGen, and stop it. - One parse per file:
RemoveUnusedImportsandStringWrapperwork from the compilation unit and tokens of the formatting pass where the text did not change, as google-java-format'sCleanupsreuse its parse (google/google-java-format@633ce6be, Fix the imports before the layout or after it? #224). - Fewer small objects in the layout: token positions as
int, not boxedIntegerinRanges, and the unique ids ofDocs without a memoizing supplier for each.
Activity