Add splitfanout pass
#4741
Add splitfanout pass
#4741
Conversation
| Module *module; | ||
| SigMap sigmap; | ||
| dict<SigBit, tuple<IdString,IdString,int>> bit_drivers_db; | ||
| dict<SigBit, pool<tuple<IdString,IdString,int>>> bit_users_db; |
There was a problem hiding this comment.
This construction is copied from splitcells.cc, it would be best to leave a comment that marks it as such for the sake of a future rework touching both instances of this pattern. I think some rework is possible to reduce the usage of these indices and make them more readable, but for the most part, if the runtime and memory consumption of the current implementation of splitfanout is sufficient for your use case then there isn't any immediate need for that.
Some of the possible improvements:
- In buffer-normalized mode (see
bufnorm.cc),bit_drivers_dbwouldn't be needed - The construction of
toposort, construction of the indices, and final cell duplication could be done with fewer passes over the cells. Though this would also make the pass harder to modify, read, maintain
| SplitfanoutWorker(Module *module) : module(module), sigmap(module) | ||
| { | ||
| // Add nodes to topological sorter for all selected cells | ||
| log("Making toposort nodes for module %s...\n", log_id(module)); |
There was a problem hiding this comment.
Generally, most of the log statements should be demoted to log_debug. For example, opt_clean only logs how many cells and wires it removed. Listing every cell that's ineligible for splitting is very verbose
| int foi = 0; | ||
| cell->unsetPort(outport); | ||
| int num_new_cells = GetSize(bit_users)-1; | ||
| int bit_user_i = num_new_cells; |
There was a problem hiding this comment.
What's the motivation for counting down here?
| } | ||
| void execute(std::vector<std::string> args, RTLIL::Design *design) override | ||
| { | ||
| int limit = -1; |
There was a problem hiding this comment.
Maybe consider std::optional<int>
What are the reasons/motivation for this change?
This PR adds a
splitfanoutpass that splits cells with fanout > 1 into copies, each with fanout 1. It is effectively the opposite ofopt_mergeand is useful for (1) enabling optimizations that cannot work when there is fanout, (2) splitting high fanout cells/nets that might cause problems in PD.The pass can operate on selections and has a
-limitflag to selectively operate on cells with fanout < specified limit. It does not operate on "bit-split" cells (word-level cells where different bits of the word have different/non-contiguous destinations), sosplitcellsshould be run first if fanout splitting is desired for such cells.Explain how this is achieved.
The pass works by building a database of drivers and loads, then operating in reverse topological order on selected cells. It checks that each cell is not "bit-split", then makes a copy to drive each load. The original cell is used to drive the final load.
passes/cmds/splitfanout.cc:splitfanoutpasspasses/cmds/Makefile.inc: Add the new sourceIf applicable, please suggest to reviewers how they can test the change.
equiv_optor FOSSeqy