fraig_store: restore the PI order when the name comparison fails

[marcelwa]

Follow-up to #537, which I filed before I had the mechanism right. It is not PI invention in Abc_NtkAppendToCone() — that branch is never reached here. It is PI reordering, one level up, and it has a two-line fix.

What happens

Abc_NtkCompareSignals() calls Abc_NtkOrderObjsByName() on both networks, which qsorts vPis, vPos and vBoxes in place before the comparison. That is deliberate: it is what lets fraig_store accept two networks that use the same port names in a different order.

When the names genuinely differ, the comparison fails and Abc_NtkFraigStore() takes this path:

if ( !Abc_NtkCompareSignals( pNtk, (Abc_Ntk_t *)Vec_PtrEntry(vStore, 0), 1, 1 ) )
{
    printf( "Trying to store the network with different primary inputs.\n" );
    printf( "The previously stored networks are deleted and this one is added.\n" );
    Abc_NtkFraigStoreClean();
}
Vec_PtrPush( vStore, pNtk );

The store is reset and pNtk is kept — but pNtk has already been sorted by the comparison that failed. What goes into the store is a permutation of the network the caller read in. The two printed lines say the store was reset; nothing says the interface changed.

Sorting is by name as a string, so numeric port names are the worst case: 1, 2, …, 10 sorts to 1, 10, 2, 3, ….

Reproducing

EPFL cavlc and its published reference netlist, whose ports are named 1…10:

read_aiger cavlc.aig;        strash; fraig_store
read_blif  cavlc_size.blif;  strash; fraig_store
fraig_restore; write_blif out.blif

$ abc -q "read_blif out.blif; print_io"
Primary inputs (10):  0=1 1=10 2=2 3=3 4=4 5=5 6=6 7=7 8=8 9=9
                            ^^^^ should be 1=2

$ abc -q "cec -n out.blif cavlc.aig"
Output po00: Value in Network1 = 0. Value in Network2 = 1.

Right number of inputs and outputs, passes Abc_NtkCheck(), wrong function. The reference netlist on its own is equivalent — only going through fraig_store breaks it, and only because the store then discards it and keeps the sorted copy.

I ran into this building choice networks out of differently-named implementations of the same circuit; 9 of 20 merges came back non-equivalent with plausible-looking area and depth. The workaround is to round-trip everything through write_aiger/read_aiger first so names are regenerated positionally, but the sort surviving a failed comparison seemed worth fixing rather than working around.

The change

Save vPis/vPos/vBoxes before the comparison, put them back if it fails, and rebuild vCis/vCos with Abc_NtkOrderCisCos().

Deliberately narrow:

• The success path is untouched — when the names do line up, both networks stay sorted into the common order, which is the point.
• Networks whose names already agree never reach Abc_NtkCompareSignals() at all, because Abc_NodeCompareCiCo() has already returned 1.
• The warning still prints, and the store is still reset.

Checks

Built clean. Against the EPFL suite:

• cavlc with the mismatched reference: PI order now 0=1 1=2 2=3 …, and cec -n reports equivalent.
• ctrl, arbiter, max — names already agree — merge and verify exactly as before.
• Same-interface merge, and a plain single-network fraig_store/fraig_restore, both unchanged.
• The warning is still emitted on the mismatch path.

Happy to add the name-mismatch case to a regression test if you have a preferred place for it.