This fixes some edge cases the previous version didn't handle properly
by simplifying the logic of determining directly driven wires and
representatives to use as buffer inputs.
This hasn't been an issue when using -l to redirect or when stdout is
line buffered, explaining how we didn't notice this earlier, but for
`yosys ... > log` that extra flush is required to ensure all messages
preceding the fatal error are flushed.
This is a complete rewrite of the RTLIL-kernel-side bufnorm code. This
is done to support inout ports and undirected connections as well as to
allow removal of cells while in bufnorm mode.
This doesn't yet update the (experimental) `bufnorm` pass, so to
manually test the new kernel functionality, it is important to only use
`bufnorm -update` and `bufnorm -reset` which rely entirely on the kernel
functionality. Other modes of the `bufnorm` pass may still fail in the
presence of inout ports or undirected connections.
Now that we only call `bitvectorize()` in non-const methods, we can move the casting-away-const to only happen
in `bitvectorize()`, which is deprecated so only some plugins (maybe) are using it.
This means `const` `Const` methods don't change the underlying data, which means
they'll be safe to use from multiple threads if/when we want to do that.
Doing ABC runs in parallel can actually make things slower when every ABC run requires
spawning an ABC subprocess --- especially when using popen(), which on glibc does not
use vfork(). What seems to happen is that constant fork()ing keeps making the main
process data pages copy-on-write, so the main process code that is setting up each ABC
call takes a lot of minor page-faults, slowing it down.
The solution is pretty straightforward although a little tricky to implement.
We just reuse ABC subprocesses. Instead of passing the ABC script name on the command
line, we spawn an ABC REPL and pipe a command into it to source the script. When that's
done we echo an `ABC_DONE` token instead of exiting. Yosys then puts the ABC process
onto a stack which we can pull from the next time we do an ABC run.
For one of our large designs, this is an additional 5x speedup of the primary AbcPass.
It does 5155 ABC runs, all very small; runtime of the AbcPass goes from 760s to 149s
(not very scientific benchmarking but the effect size is large).
Large circuits can run hundreds or thousands of ABCs in a single AbcPass.
For some circuits, some of those ABC runs can run for hundreds of seconds.
Running ABCs in parallel with each other and in parallel with main-thread
processing (reading and writing BLIF files, copying ABC BLIF output into
the design) can give large speedups.
Robert O'Callahan
Emil J
Jannis Harder
Akash Levy
Mohamed Gaber
George Rennie
rhanqtl