Prefer last indices

Is (B A)_{c a} a valid answer for (B)_{b c} (A)_{a b}? It isn't if "a"
and "c" are indices on the same footing as "b". However, they could be
pure labels which would turn the above into a commutative dot product.
To avoid this, we should assume that indices go outside the bracket
while labels go inside. It is therefore not enough to begin the
contraction on B's last participating index. We should begin with a last
index that is also participating. Sometimes this does not exist so we
keep the old logic in place. Garbage in / garbage out is better than a
crash.

Signed-off-by: Connor Behan <[email protected]>

Author: cbehan

core/algorithms/combine.cc

@@ -36,7 +36,7 @@ Algorithm::result_t combine::apply(iterator& it)
 
 	while(ind_dummy.begin()!=ind_dummy.end()) {
 		bool found=false;
-		index_map_t::iterator start=ind_dummy.begin();
+		index_map_t::iterator start=ind_dummy.begin(), backup;
 		while(!found && start!=ind_dummy.end()) {
 			iterator parent=tr.parent(start->second);
 			sibling_iterator ch=tr.begin(parent), last_part;
@@ -46,11 +46,17 @@ Algorithm::result_t combine::apply(iterator& it)
 				++ch;
 				}
 			if(last_part==start->second) {
-				// We are on a rightmost contracted index
-				found=true;
+				++last_part;
+				if(last_part==tr.end(parent)) {
+					// Dummy index with nothing to the right is preferred
+					found=true;
+					}
+				else backup=start;
 				}
-			else ++start;
+			if(!found) ++start;
 			}
+		// As a backup, we use a dummy index with only non-dummies to the right
+		if(!found) start=backup;
 		bool paired=true;
 		while(paired && start!=ind_dummy.end()) {
 			iterator parent=tr.parent(start->second);

tests/implicit.cdb

@@ -97,3 +97,14 @@ def test08():
 
 test08()
 
+def test09():
+    __cdbkernel__=create_scope()
+    {a,b,c}::Indices(vector);
+    ex:=(B)_{b c} (A)_{a b};
+    combine(_);
+    tst:= (A B)_{a c} - @(ex);
+    assert(tst==0)
+    print("Test 09 passed")
+
+test09()
+