opt_expr: limit effort to reduce runtime on anomalous designs

[widlarizer]

The first case of the ROM section of tests/arch/ice40/memories.ys takes 326 seconds to run with majority of this runtime spent sorting cells in opt_expr. This PR allows sorting to only fail a limited number of times on a given module before reaching fixpoint without trying to sort the cells. This cuts down the test case runtime to 14 seconds. The limit can be changed with the new -effort argument.

5.3.12. Executing OPT_EXPR pass (perform const folding).
Optimizing module sync_rom.
Couldn't topologically sort cells, optimizing module sync_rom may take a longer time.
Couldn't topologically sort cells, optimizing module sync_rom may take a longer time.
Couldn't topologically sort cells, optimizing module sync_rom may take a longer time.
Couldn't topologically sort cells, optimizing module sync_rom may take a longer time.
Couldn't topologically sort cells, optimizing module sync_rom may take a longer time.
Effort of 5 exceeded, no longer attempting toposort on module sync_rom.
<suppressed ~6151 debug messages>

[povik]

Isn't this empty if you don't run cells.sort()? And if that's the case, then the true cause of the speed-up isn't not bothering to sort anymore but giving up entirely

[widlarizer]

Hm, there's a log_assert(GetSize(sorted) == GetSize(nodes)); right before the return from sort() so I guess it isn't but I'll verify it later

EDIT: yeah if you don't even run sort() then it's going to be empty, yikes

[widlarizer]

Fixed by iterating over a list of references to a snapshot of module->cells(). This increases the runtime to 83 seconds, which I still prefer to 326 seconds. I should say, all runtimes I'm mentioning in this PR are measured with ENABLE_DEBUG=1 for no particular reason other having that around in Makefile.conf

[povik]

We could cache the sorting result provided we remove cells we removed

[widlarizer]

I can't really foresee the consequences of doing that, though - for how many cycles would we cache etc

[povik]

For the lifetime of one opt_expr invocation. It should be ok as long as we zero out the position in the sorted array corresponding to any removed cells.

I am not sure if there's a case where opt_expr replaces a cell with two or more cells, that's a slight complication