Optimize temporary insertion for single bit replicates in DFG (#8110)

A '{N{bit}}' mask expands to a bit vector where every word is the same
value, obtained by negating the replicated bit. Recomputing that at each
use is no more expensive than loading it

src/V3Dfg.h

@@ -829,6 +829,13 @@ bool DfgVertex::isCheaperThanLoad() const {
         const uint32_t msb = lsb + selp->width() - 1;
         return VL_BITWORD_E(msb) == VL_BITWORD_E(lsb);
     }
+    // Replication of a single cheap bit. Each word of the result is the same
+    // mask computed by negating that bit, so recomputing it at each use costs
+    // no more than the load it replaces.
+    if (const DfgRep* const repp = cast<DfgRep>()) {
+        const DfgVertex* const srcp = repp->srcp();
+        return srcp->width() == 1 && srcp->isCheaperThanLoad();
+    }
     // Zero extend of a cheap vertex - Extend(_) was converted to Concat(0, _)
     if (const DfgConcat* const catp = cast<DfgConcat>()) {
         if (catp->width() > VL_QUADSIZE) return false;