Although the two passes started out very similar, they diverged over
time and now have little in common. Moreover, `techmap` is extremely
complex while `flatten` does not have to be, and this complexity
interferes with improving `flatten`.
The $div and $mod cells use truncating division semantics (rounding
towards 0), as defined by e.g. Verilog. Another rounding mode, flooring
(rounding towards negative infinity), can be used in e.g. VHDL. The
new $divfloor cell provides this flooring division.
This commit also fixes the handling of $div in opt_expr, which was
previously optimized as if it was $divfloor.
The $div and $mod cells use truncating division semantics (rounding
towards 0), as defined by e.g. Verilog. Another rounding mode, flooring
(rounding towards negative infinity), can be used in e.g. VHDL. The
new $modfloor cell provides this flooring modulo (also known as "remainder"
in several languages, but this name is ambiguous).
This commit also fixes the handling of $mod in opt_expr, which was
previously optimized as if it was $modfloor.
The chosen value shouldn't have any effect. I considered something
clearly wrong like -1, but there's no checking inside the generated
lexer, and I suspect this will cause even weirder bugs if triggered
than just setting it to INITIAL.
The existing code does a search to figure out whether id is in the
dict (with the call to count()), and then looks it up again to get the
result (with the call to at()). This version calls find() instead,
avoiding the double lookup.
Code size increases slightly (6kb). I think this is because the
contents of find() are getting inlined, and then inlined into lots of
the callsites for cell() and wire().
Looking at the compiled code before this patch, you just get
a (non-inlined) call to count() followed by a call to at(). After the
patch, the contents of find() have been inlined (so you see do_hash,
then do_lookup). The result for each function is about 30 bytes / 40%
bigger, which presumably also enlarges call-sites that inline it.
There was a handwritten copy constructor, which I'm not sure was
actually legal C++ (it unconditionally read from the 'data' member of
a union, which wouldn't have been written if wire was true). It was
also a bit less efficient than the constructor you get from the
compiler by default (which is allowed to just copy the memory).
This gives a marginal (~0.25%) decrease in code size when compiled
with GCC 9.3.