Speed up Verilator itself by prefetching pointer-chase

src/V3Ast.cpp

@@ -680,11 +680,38 @@ void AstNode::operator delete(void* objp, size_t size) {
 // Iterators
 
 void AstNode::iterateChildren(AstNVisitor& v, AstNUser* vup) {
+    // This is a very hot function
     if (!this) return;
-    m_op1p->iterateAndNext(v, vup);
+    ASTNODE_PREFETCH(m_op1p);
-    m_op2p->iterateAndNext(v, vup);
+    ASTNODE_PREFETCH(m_op2p);
-    m_op3p->iterateAndNext(v, vup);
+    ASTNODE_PREFETCH(m_op3p);
-    m_op4p->iterateAndNext(v, vup);
+    ASTNODE_PREFETCH(m_op4p);
+    // if () not needed since iterateAndNext accepts null this, but faster with it.
+    if (m_op1p) m_op1p->iterateAndNext(v, vup);
+    if (m_op2p) m_op2p->iterateAndNext(v, vup);
+    if (m_op3p) m_op3p->iterateAndNext(v, vup);
+    if (m_op4p) m_op4p->iterateAndNext(v, vup);
+}
+
+void AstNode::iterateAndNext(AstNVisitor& v, AstNUser* vup) {
+    // This is a very hot function
+    // IMPORTANT: If you replace a node that's the target of this iterator,
+    // then the NEW node will be iterated on next, it isn't skipped!
+    // if (!this) return;  // Part of for()
+    for (AstNode* nodep=this; nodep;) {
+	AstNode* niterp = nodep;
+	ASTNODE_PREFETCH(nodep->m_nextp);
+	niterp->m_iterpp = &niterp;
+	niterp->accept(v, vup);
+	// accept may do a replaceNode and change niterp on us...
+	if (!niterp) return;
+	niterp->m_iterpp = NULL;
+	if (niterp!=nodep) { // Edited it
+	    nodep = niterp;
+	} else {  // Same node, just loop
+	    nodep = niterp->m_nextp;
+	}
+    }
 }
 
 void AstNode::iterateListBackwards(AstNVisitor& v, AstNUser* vup) {
@@ -707,25 +734,6 @@ void AstNode::iterateChildrenBackwards(AstNVisitor& v, AstNUser* vup) {
     this->op4p()->iterateListBackwards(v,vup);
 }
 
-void AstNode::iterateAndNext(AstNVisitor& v, AstNUser* vup) {
-    // IMPORTANT: If you replace a node that's the target of this iterator,
-    // then the NEW node will be iterated on next, it isn't skipped!
-    // if (!this) return;  // Part of for()
-    for (AstNode* nodep=this; nodep;) {
-	AstNode* niterp = nodep;
-	niterp->m_iterpp = &niterp;
-	niterp->accept(v, vup);
-	// accept may do a replaceNode and change niterp on us...
-	if (!niterp) return;
-	niterp->m_iterpp = NULL;
-	if (niterp!=nodep) { // Edited it
-	    nodep = niterp;
-	} else {  // Same node, just loop
-	    nodep = niterp->m_nextp;
-	}
-    }
-}
-
 void AstNode::iterateAndNextIgnoreEdit(AstNVisitor& v, AstNUser* vup) {
     // Keep following the current list even if edits change it
     if (!this) return;
@@ -796,6 +804,11 @@ void AstNode::checkTreeIter(AstNode* backp) {
     if (backp != this->backp()) {
 	this->v3fatalSrc("Back node inconsistent");
     }
+    if (castNodeTermop()) {
+	// Termops have a short-circuited iterateChildren, so check usage
+	if (op1p()||op2p()||op3p()||op4p())
+	    this->v3fatalSrc("Terminal operation with non-terminals");
+    }
     if (op1p()) op1p()->checkTreeIterList(this);
     if (op2p()) op2p()->checkTreeIterList(this);
     if (op3p()) op3p()->checkTreeIterList(this);

src/V3Ast.h

@@ -541,16 +541,23 @@ ostream& operator<<(ostream& os, V3Hash rhs);
 //######################################################################
 // AstNode -- Base type of all Ast types
 
+// Prefetch a node.
+// The if() makes it faster, even though prefetch won't fault on null pointers
+#define ASTNODE_PREFETCH(nodep) \
+    { if (nodep) { VL_PREFETCH_RD(&(nodep->m_nextp)); VL_PREFETCH_RD(&(nodep->m_iterpp)); }}
+
 class AstNode {
-private:
+    // v ASTNODE_PREFETCH depends on below ordering of members
     AstNode*	m_nextp;	// Next peer in the parent's list
     AstNode*	m_backp;	// Node that points to this one (via next/op1/op2/...)
-    AstNode*	m_headtailp;	// When at begin/end of list, the opposite end of the list
     AstNode*	m_op1p;		// Generic pointer 1
     AstNode*	m_op2p;		// Generic pointer 2
     AstNode*	m_op3p;		// Generic pointer 3
     AstNode*	m_op4p;		// Generic pointer 4
     AstNode**	m_iterpp;	// Pointer to node iterating on, change it if we replace this node.
+    // ^ ASTNODE_PREFETCH depends on above ordering of members
+
+    AstNode*	m_headtailp;	// When at begin/end of list, the opposite end of the list
 
     AstNode*	m_clonep;	// Pointer to clone of/ source of this module (for *LAST* cloneTree() ONLY)
     int		m_cloneCnt;	// Mark of when userp was set

src/V3Gate.cpp

@@ -421,7 +421,7 @@ void GateVisitor::optimizeSignals(bool allowMultiIn) {
 		if (!vvertexp->isTop()		// Ok if top inputs are driverless
 		    && !vvertexp->varScp()->varp()->initp()
 		    && !vvertexp->varScp()->varp()->isSigPublic()) {
-		    UINFO(1, "No drivers "<<vvertexp->varScp()<<endl);
+		    UINFO(4, "No drivers "<<vvertexp->varScp()<<endl);
 		    if (0) {
 			// If we warned here after constant propagation, what the user considered
 			// reasonable logic may have disappeared.  Issuing a warning would