[ru551n]

Adds --parallel-compile, which compiles files in parallel, up to -p files at a time. A file is compiled as soon as the files it depends on are compiled. Without the flag, files are compiled one at a time as before, whatever -p is, so python run.py --clean -p 0 still compiles serially.

What was added

• Scheduler: SimulatorInterface.compile_source_files compiles each file once its dependencies are compiled, using a thread pool. As before, the dependents of a failed file are skipped, and --keep-compiling keeps going. Without it, no new files start after a failure.

• Library rule: GHDL and Riviera-PRO update a library's index without locking it.

  • Two compiles into the same library lose units. Riviera-PRO even reports success for every compile.
  • A compile reading a library while GHDL rewrites its index fails with "cannot find resource library", because GHDL deletes the index before renaming the new one into place.

  Unless a simulator sets supports_concurrent_compiles_into_library, a file is therefore not compiled while its own library is compiled into or read, or while a library it reads is compiled into. The libraries a file reads are those of its dependencies. NVC sets the flag, because it holds a reader/writer lock on its libraries (see NVC performance with Vunit #1036).

• GHDL: creates every library with ghdl -i when setting up the library mapping. A file with a library clause can otherwise be compiled before anything has been compiled into that library, which GHDL rejects. With gcc coverage, GHDL compiles one file at a time, because the .gcno files land in the working directory.

• Output: with -p > 1, each file's output is written once its compile finishes, so outputs don't mix.

Measured

A clean compile (--compile --clean --parallel-compile) of hdl-modules, 330 files, wall time including VUnit's start-up, median of 3 runs:

-p	NVC	GHDL	Riviera-PRO*
master	5.88 s	7.51 s	11.85 s
1	5.95 s	7.51 s	11.90 s
4	2.51 s	5.62 s	7.39 s
16	1.65 s	5.63 s	7.24 s

* With --keep-compiling, because hdl-modules does not compile completely with Riviera-PRO on master either ('latency attribute in sine_calculator.vhd). Every run compiled the same 327 files.

GHDL and Riviera-PRO level off early because almost every file reads vunit_lib, so the library rule serialises most of the compile. The follow-up #1236, built on this PR, addresses that with --batch-compile, which compiles the files of a library that are ready at the same time in one call, the automatic form of the soft file groups discussed in #1088. With both flags, the clean compile of hdl-modules takes 1.42 s with NVC, 1.64 s with GHDL and 2.66 s with Riviera-PRO.

Tested

• New unit tests for compiling different libraries concurrently, one file per library at a time, the read rule, concurrent readers of a library, waiting for dependencies, skipping dependents and writing each file's output in one piece. They fail on master.
• UI tests: without --parallel-compile, files are compiled with one thread for -p 4 and -p 0; with it, with -p threads.
• The unit and lint tests pass.
• GHDL stress test: 10 clean hdl-modules compiles each at -p 4, 16 and 32. All 29 runs that reached the compile compiled all 330 files. The 30th failed earlier, on a duplicate configuration name in hdl-modules' own randomised test setup.

A news fragment is included.

🤖 Generated with Claude Code