drw/drf suboptimality

[Nic30]

Description:
For larger nets some trivial cases do not optimize. with drw/drf while they do with rw/rf

• using latest master 474e7fb, gcc 15.2.0, Ubuntu 25.10

---

Input formulas:

a = n1
b = n2
c = n3
n17 =  (a ^ ((b & a) & a)) ^ 1
n23 = (a & b) | (c & ~b)
n26 = (c | a) & ~(c & b)

---

script:

 resyn2; compress2;  resyn2; compress2;
 # with replaced:  
 #   rf -> drf,-l
 #   rw -> drw,-l

replaced as suggested by YosysHQ/yosys#4039

---

Expected output (output with the original rf, rw in opt. script):

n17 = ~a | b

---

Actual output for n17:

n17 = ~(a & ~(a & b))

---

• Doing more optimization rounds does not help.
• Removing n23 or n26 will "fix" the issue.
• DRefactor fExtend=true seems to fix this (but by adding more formulas it can be broken again)

Is this an expected behavior?

Is this a limitation of drf/drw or I am doing something incorrectly, or is there some parameter like nCutsMax which needs tweaking to achieve same behavior as rf/rw?

[Nic30]

After reading original paper https://people.eecs.berkeley.edu/~alanmi/publications/2006/dac06_rwr.pdf
I think that this is an expected behavior as DAR selects 4 input cuts and this one is 3 input one.
However, this does not explain why case with just n17 works.

Can someone confirm that this is the case?
Do you think that writing an additional rewriting function for 3-input cuts is a valid solution there?

[alanminko]

Thank you for bringing it up. AIG-based optimization commands minimize the number of AIG node. In this case, the number is the same (5). It may be possible to add a tie-breaker to optimize the number of levels, which in this case would have preferred the 1-node solution for n17.