diff --git a/R/data.table.R b/R/data.table.R
index 801482147e..15e270753c 100644
--- a/R/data.table.R
+++ b/R/data.table.R
@@ -527,12 +527,15 @@ replace_dot_alias = function(e) {
           # Really, `anyDuplicated` in base is AWESOME!
           # allow.cartesian shouldn't error if a) not-join, b) 'i' has no duplicates
           if (verbose) {last.started.at=proc.time();catf("Constructing irows for '!byjoin || nqbyjoin' ... ");flush.console()}
-          irows = if (allLen1) f__ else vecseq(f__,len__,
-            if (allow.cartesian ||
-                notjoin || # #698. When notjoin=TRUE, ignore allow.cartesian. Rows in answer will never be > nrow(x).
-                !anyDuplicated(f__, incomparables = c(0L, NA_integer_))) {
-              NULL # #742. If 'i' has no duplicates, ignore
-            } else as.double(nrow(x)+nrow(i))) # rows in i might not match to x so old max(nrow(x),nrow(i)) wasn't enough. But this limit now only applies when there are duplicates present so the reason now for nrow(x)+nrow(i) is just to nail it down and be bigger than max(nrow(x),nrow(i)).
+          if (allLen1)
+            irows = f__
+          else {
+            # notjoin condition for #698. When notjoin=TRUE, ignore allow.cartesian. Rows in answer will never be > nrow(x).
+            # anyDuplicated condition for # #742. If 'i' has no duplicates, no clamp needed.
+            # rows in i might not match to x so old max(nrow(x),nrow(i)) wasn't enough. But this limit now only applies when there are duplicates present so the reason now for nrow(x)+nrow(i) is just to nail it down and be bigger than max(nrow(x),nrow(i)).
+            clamp = if (!allow.cartesian && !notjoin && anyDuplicated(f__, incomparables = c(0L, NA_integer_))) as.double(nrow(x) + nrow(i))
+            irows = vecseq(f__, len__, clamp)
+          }
           if (verbose) {cat(timetaken(last.started.at),"\n"); flush.console()}
           # Fix for #1092 and #1074
           # TODO: implement better version of "any"/"all"/"which" to avoid