Merge 1ac45cb7e7 into 3afb7e1a21

src/V3AstNodeOther.h

@@ -1889,6 +1889,26 @@ public:
     bool isPure() override { return false; }
     bool isOutputter() override { return true; }
 };
+class AstTaskLocalVar final : public AstNode {
+    // A task activation-local variable before loop unrolling is complete. The local variable is
+    // deliberately unscoped so cloning this node cannot invalidate an external AstVarScope. The
+    // temporary scoped variable carries references until V3Task materializes the local variable.
+    // @astgen op1 := varp : AstVar
+    // @astgen ptr := m_templateVscp : AstVarScope
+public:
+    AstTaskLocalVar(FileLine* fl, AstVar* varp, AstVarScope* templateVscp)
+        : ASTGEN_SUPER_TaskLocalVar(fl)
+        , m_templateVscp{templateVscp} {
+        if (varp) this->varp(varp);
+    }
+    ASTGEN_MEMBERS_AstTaskLocalVar;
+    AstVarScope* templateVscp() const { return m_templateVscp; }
+    void dump(std::ostream& str) const override;
+    void dumpJson(std::ostream& str) const override;
+    bool sameNode(const AstNode* samep) const override {
+        return templateVscp() == VN_DBG_AS(samep, TaskLocalVar)->templateVscp();
+    }
+};
 class AstText final : public AstNode {
     // Represents a piece of text to be emitted into the output
     //

src/V3AstNodes.cpp

@@ -3077,6 +3077,16 @@ void AstSystemCSection::dumpJson(std::ostream& str) const {
     dumpJsonStr(str, "sectionType", sectionType().ascii());
     dumpJsonGen(str);
 }
+void AstTaskLocalVar::dump(std::ostream& str) const {
+    this->AstNode::dump(str);
+    str << " -> ";
+    if (templateVscp()) {
+        templateVscp()->dump(str);
+    } else {
+        str << "%E:UNLINKED";
+    }
+}
+void AstTaskLocalVar::dumpJson(std::ostream& str) const { dumpJsonGen(str); }
 void AstTypeTable::dump(std::ostream& str) const {
     this->AstNode::dump(str);
     for (int i = 0; i < static_cast<int>(VBasicDTypeKwd::_ENUM_MAX); ++i) {

src/V3Sched.cpp

@@ -50,6 +50,14 @@ VL_DEFINE_DEBUG_FUNCTIONS;
 
 namespace V3Sched {
 
+LogicByScope LogicByScope::clone() const {
+    LogicByScope result;
+    for (const auto& pair : *this) {
+        result.emplace_back(pair.first, VN_AS(util::cloneLogic(pair.second), Active));
+    }
+    return result;
+}
+
 namespace {
 
 //============================================================================
@@ -928,6 +936,28 @@ cloneMapWithNewTriggerReferences(const std::unordered_map<const AstSenTree*, Ast
     return std::unordered_map<const AstSenTree*, AstSenTree*>{pairs.begin(), pairs.end()};
 }
 
+void removeVarScopesForDeletedLogic(AstNetlist* netlistp) {
+    // Scheduler ordering deletes logic replicas after moving their statements. Remove only
+    // unreferenced scopes whose cloned local declaration was deleted with such a replica.
+    std::unordered_set<const AstVar*> liveVarps;
+    std::unordered_set<const AstVarScope*> referencedVscps;
+    std::vector<AstVarScope*> staleVscps;
+
+    netlistp->foreach([&](const AstVar* varp) { liveVarps.emplace(varp); });
+    netlistp->foreach([&](const AstNodeVarRef* refp) {
+        if (refp->varScopep()) referencedVscps.emplace(refp->varScopep());
+    });
+    netlistp->foreach([&](AstVarScope* vscp) {
+        if (!liveVarps.count(vscp->varp())) staleVscps.push_back(vscp);
+    });
+
+    for (AstVarScope* const vscp : staleVscps) {
+        UASSERT_OBJ(!referencedVscps.count(vscp), vscp,
+                    "Variable scope still referenced after its declaration was deleted");
+        VL_DO_DANGLING(vscp->unlinkFrBack()->deleteTree(), vscp);
+    }
+}
+
 //============================================================================
 // Top level entry-point to scheduling
 
@@ -1204,6 +1234,7 @@ void schedule(AstNetlist* netlistp) {
 
     // Step 16: Clean up
     netlistp->clearStlFirstIterationp();
+    removeVarScopesForDeletedLogic(netlistp);
 
     // Haven't split static initializer yet
     util::splitCheck(staticp);

src/V3Sched.h

@@ -37,6 +37,8 @@ namespace V3Sched {
 namespace util {
 using VarScopeSet = std::unordered_set<const AstVarScope*>;
 
+AstNode* cloneLogic(AstNode* logicp);
+
 inline bool isVlForceVec(const AstVarScope* vscp) {
     const AstCDType* const dtypep = VN_CAST(vscp->dtypep()->skipRefp(), CDType);
     return dtypep && dtypep->name() == "VlForceVec";
@@ -92,13 +94,7 @@ struct LogicByScope final : public std::vector<std::pair<AstScope*, AstActive*>>
     };
 
     // Create copy, with the AstActives cloned
-    LogicByScope clone() const {
+    LogicByScope clone() const;
-        LogicByScope result;
-        for (const auto& pair : *this) {
-            result.emplace_back(pair.first, pair.second->cloneTree(false));
-        }
-        return result;
-    }
 
     // Delete actives (they should all be empty)
     void deleteActives() {

src/V3SchedReplicate.cpp

@@ -274,7 +274,7 @@ LogicReplicas replicate(Graph* graphp) {
     for (V3GraphVertex& vtx : graphp->vertices()) {
         if (SchedReplicateLogicVertex* const lvtxp = vtx.cast<SchedReplicateLogicVertex>()) {
             const auto replicateTo = [&](LogicByScope& lbs) {
-                lbs.add(lvtxp->scopep(), lvtxp->senTreep(), lvtxp->logicp()->cloneTree(false));
+                lbs.add(lvtxp->scopep(), lvtxp->senTreep(), util::cloneLogic(lvtxp->logicp()));
             };
             const uint8_t targetRegions = lvtxp->drivingRegions() & ~lvtxp->assignedRegion();
             UASSERT(!lvtxp->senTreep()->hasClocked() || targetRegions == 0,

src/V3SchedUtil.cpp

@@ -33,6 +33,32 @@ VL_DEFINE_DEBUG_FUNCTIONS;
 namespace V3Sched {
 namespace util {
 
+AstNode* cloneLogic(AstNode* logicp) {
+    AstNode* const clonep = logicp->cloneTree(false);
+    std::map<AstVarScope*, AstVarScope*> clonedVscps;
+    clonep->foreach([&](AstNodeVarRef* refp) {
+        AstVarScope* const oldVscp = refp->varScopep();
+        // cloneTree relinks an in-tree local AstVar, but its out-of-tree AstVarScope still points
+        // at the original declaration. Give each cloned declaration an independent scope.
+        if (!oldVscp || oldVscp->varp() == refp->varp()) return;
+
+        const auto pair = clonedVscps.emplace(oldVscp, nullptr);
+        if (pair.second) {
+            AstVarScope* const newVscp
+                = new AstVarScope{refp->fileline(), oldVscp->scopep(), refp->varp()};
+            newVscp->trace(oldVscp->isTrace());
+            newVscp->optimizeLifePost(oldVscp->optimizeLifePost());
+            oldVscp->scopep()->addVarsp(newVscp);
+            pair.first->second = newVscp;
+        } else {
+            UASSERT_OBJ(pair.first->second->varp() == refp->varp(), refp,
+                        "Cloned variable scope maps to multiple variables");
+        }
+        refp->varScopep(pair.first->second);
+    });
+    return clonep;
+}
+
 AstCFunc* makeSubFunction(AstNetlist* netlistp, const string& name, bool slow) {
     AstScope* const scopeTopp = netlistp->topScopep()->scopep();
     AstCFunc* const funcp = new AstCFunc{netlistp->fileline(), name, scopeTopp, ""};

src/V3Split.cpp

@@ -522,9 +522,8 @@ public:
             m_addAfter[color] = placeholderp;
             m_newBlocksp->push_back(alwaysp);
         }
-        // Scan the body of the always. We'll handle if/else
+        // Scan the body of the always. We'll handle if/else specially; every other leaf belongs to
-        // specially, everything else is a leaf node that we can
+        // one color and can move into the corresponding split always block.
-        // just clone into one of the split always blocks.
         iterateAndNextNull(m_origAlwaysp->stmtsp());
     }
 
@@ -547,12 +546,13 @@ protected:
         // Each leaf must have a user3p
         UASSERT_OBJ(nodep->user3p(), nodep, "null user3p in V3Split leaf");
 
-        // Clone the leaf into its new always block
+        // A leaf belongs to exactly one color, so move it into its new always block. Moving also
+        // preserves lexical declarations whose AstVarScopes are owned outside the leaf subtree.
         const SplitLogicVertex* const vxp = reinterpret_cast<SplitLogicVertex*>(nodep->user3p());
         const uint32_t color = vxp->color();
-        AstNode* const clonedp = nodep->cloneTree(false);
+        nodep->unlinkFrBack();
-        m_addAfter[color]->addNextHere(clonedp);
+        m_addAfter[color]->addNextHere(nodep);
-        m_addAfter[color] = clonedp;
+        m_addAfter[color] = nodep;
     }
 
     void visit(AstNodeIf* nodep) override {

src/V3Task.cpp

@@ -34,6 +34,7 @@
 #include "V3Stats.h"
 #include "V3UniqueNames.h"
 
+#include <set>
 #include <tuple>
 
 VL_DEFINE_DEBUG_FUNCTIONS;
@@ -483,6 +484,45 @@ class TaskVisitor final : public VNVisitor {
             return newvscp;
         }
     }
+    AstVarScope* createTaskLocalVarScope(AstVar* invarp, const string& name,
+                                         AstTaskLocalVar*& localDeclpr) {
+        AstVarScope* const templateVscp = createVarScope(invarp, name);
+        AstVar* const localVarp
+            = new AstVar{invarp->fileline(), VVarType::BLOCKTEMP, name, invarp};
+        localVarp->funcLocal(true);
+        localVarp->propagateAttrFrom(invarp);
+        localVarp->isInternal(true);
+        localDeclpr = new AstTaskLocalVar{localVarp->fileline(), localVarp, templateVscp};
+        return templateVscp;
+    }
+    AstVarScope* createPortVarScope(AstVar* invarp, const string& name, bool activationLocalPorts,
+                                    AstTaskLocalVar*& localDeclpr) {
+        localDeclpr = nullptr;
+        if (activationLocalPorts) return createTaskLocalVarScope(invarp, name, localDeclpr);
+        return createVarScope(invarp, name);
+    }
+
+    void addToFront(AstNode* const beginp, AstNode* const newp) const {
+        if (!newp) return;
+        if (AstNode* const afterp = beginp->nextp()) {
+            afterp->unlinkFrBackWithNext();
+            AstNode::addNext<AstNode, AstNode>(newp, afterp);
+        }
+        beginp->addNext(newp);
+    }
+
+    static bool isTaskActivationLocal(AstVar* const varp, const bool activationLocalPorts) {
+        return activationLocalPorts && !varp->direction().isAny() && varp->lifetime().isAutomatic()
+               && (varp->isFuncLocal()
+                   || (varp->varType() == VVarType::BLOCKTEMP
+                       && VN_IS(varp->dtypep()->skipRefp(), ClassRefDType)));
+    }
+    static bool bodyNeedsActivationLocals(const AstNode* const bodyp) {
+        return bodyp && bodyp->existsAndNext([](const AstNode* const nodep) {
+            // A join_none child can outlive the task without itself containing a timing control.
+            return nodep->isTimingControl() || VN_IS(nodep, Fork);
+        });
+    }
 
     // Replace varrefs with new var pointer
     void relink(AstNode* nodep) {
@@ -552,8 +592,9 @@ class TaskVisitor final : public VNVisitor {
         }
     }
 
-    void connectPort(AstVar* portp, AstArg* argp, const string& namePrefix, AstNode* beginp,
+    bool connectPort(AstVar* portp, AstArg* argp, const string& namePrefix, AstNode* beginp,
-                     bool inlineTask) {
+                     bool inlineTask, bool activationLocalPorts) {
+        bool needsCLocalScope = false;
         AstNodeExpr* pinp = argp->exprp();
         if (inlineTask) {
             portp->unlinkFrBack();
@@ -619,8 +660,10 @@ class TaskVisitor final : public VNVisitor {
             } else if (portp->isInout()) {
                 // UINFOTREE(9, pinp, "", "pinrsize-");
 
+                AstTaskLocalVar* localDeclp = nullptr;
                 AstVarScope* const newvscp
-                    = createVarScope(portp, namePrefix + "__" + portp->shortName());
+                    = createPortVarScope(portp, namePrefix + "__" + portp->shortName(),
+                                         activationLocalPorts, localDeclp);
                 portp->user2p(newvscp);
                 if (!inlineTask) {
                     pinp->replaceWith(
@@ -629,42 +672,47 @@ class TaskVisitor final : public VNVisitor {
                 }
                 // Put input assignment in FRONT of all other statements
                 AstAssign* const preassp = connectPortMakeInAssign(pinp, newvscp, true);
-                if (AstNode* const afterp = beginp->nextp()) {
+                AstNode* const frontp
-                    afterp->unlinkFrBackWithNext();
+                    = AstNode::addNextNull<AstNode, AstNode>(localDeclp, preassp);
-                    AstNode::addNext<AstNode, AstNode>(preassp, afterp);
+                addToFront(beginp, frontp);
-                }
-                beginp->addNext(preassp);
 
                 AstAssign* const postassp = connectPortMakeOutAssign(portp, pinp, newvscp, true);
                 beginp->addNext(postassp);
+                needsCLocalScope = localDeclp != nullptr;
                 // if (debug() >= 9) beginp->dumpTreeAndNext(cout, "-pinrsize-out- ");
             } else if (portp->isWritable()) {
                 // Even if it's referencing a varref, we still make a temporary
                 // Else task(x,x,x) might produce incorrect results
+                AstTaskLocalVar* localDeclp = nullptr;
                 AstVarScope* const newvscp
-                    = createVarScope(portp, namePrefix + "__" + portp->shortName());
+                    = createPortVarScope(portp, namePrefix + "__" + portp->shortName(),
+                                         activationLocalPorts, localDeclp);
                 portp->user2p(newvscp);
                 if (!inlineTask) {
                     pinp->replaceWith(new AstVarRef{newvscp->fileline(), newvscp, VAccess::WRITE});
                     pushDeletep(pinp);  // Cloned by connectPortMakeOutAssign
                 }
+                addToFront(beginp, localDeclp);
                 AstAssign* const postassp = connectPortMakeOutAssign(portp, pinp, newvscp, false);
                 // Put assignment BEHIND of all other statements
                 beginp->addNext(postassp);
+                needsCLocalScope = localDeclp != nullptr;
             } else if (inlineTask && portp->isNonOutput()) {
                 // Make input variable
+                AstTaskLocalVar* localDeclp = nullptr;
                 AstVarScope* const newvscp
-                    = createVarScope(portp, namePrefix + "__" + portp->shortName());
+                    = createPortVarScope(portp, namePrefix + "__" + portp->shortName(),
+                                         activationLocalPorts, localDeclp);
                 portp->user2p(newvscp);
                 AstAssign* const preassp = connectPortMakeInAssign(pinp, newvscp, false);
                 // Put assignment in FRONT of all other statements
-                if (AstNode* const afterp = beginp->nextp()) {
+                AstNode* const frontp
-                    afterp->unlinkFrBackWithNext();
+                    = AstNode::addNextNull<AstNode, AstNode>(localDeclp, preassp);
-                    AstNode::addNext<AstNode, AstNode>(preassp, afterp);
+                addToFront(beginp, frontp);
-                }
+                needsCLocalScope = localDeclp != nullptr;
-                beginp->addNext(preassp);
             }
         }
+        return needsCLocalScope;
     }
 
     bool hasRefArgument(AstNodeFTask* nodep) {
@@ -688,13 +736,20 @@ class TaskVisitor final : public VNVisitor {
         if (debug() >= 9) beginp->dumpTreeAndNext(cout, "-  newbegi: ");
         //
         // Create input variables
+        // Task bodies that can suspend or fork need activation-local input temporaries. Inlined
+        // functions cannot suspend; keeping those temporaries as normal scoped temps avoids
+        // exposing AstCLocalScope to pre-timing constant/table simulation.
+        const bool activationLocalPorts
+            = !refp->taskp()->verilogFunction() && bodyNeedsActivationLocals(newbodysp);
+        bool needsCLocalScope = false;
         AstNode::user2ClearTree();
         {
             const V3TaskConnects tconnects = V3Task::taskConnects(refp, beginp);
             for (const auto& itr : tconnects) {
                 AstVar* const portp = itr.first;
                 AstArg* const argp = itr.second;
-                connectPort(portp, argp, namePrefix, beginp, true);
+                needsCLocalScope
+                    |= connectPort(portp, argp, namePrefix, beginp, true, activationLocalPorts);
             }
         }
         UASSERT_OBJ(!refp->argsp(), refp, "Arg wasn't removed by above loop");
@@ -705,6 +760,29 @@ class TaskVisitor final : public VNVisitor {
                 if (AstVar* const portp = VN_CAST(stmtp, Var)) {
                     // Any I/O variables that fell out of above loop were already linked
                     if (!portp->user2p()) {
+                        // Preserve automatic locals introduced by earlier passes inside the
+                        // inlined task body so each activation keeps its own state.
+                        if (isTaskActivationLocal(portp, activationLocalPorts)) {
+                            const string localName = namePrefix + "__" + portp->shortName();
+                            portp->name(localName);
+                            portp->funcLocal(true);
+                            AstVarScope* const templateVscp = createVarScope(portp, localName);
+                            portp->user2p(templateVscp);
+                            AstTaskLocalVar* const localDeclp
+                                = new AstTaskLocalVar{portp->fileline(), nullptr, templateVscp};
+                            portp->replaceWith(localDeclp);
+                            localDeclp->varp(portp);
+                            if (portp->needsCReset() && !portp->valuep()) {
+                                // Reset automatic var to its default on each invocation of task.
+                                AstNode* const crstp = new AstAssign{
+                                    portp->fileline(),
+                                    new AstVarRef{portp->fileline(), portp, VAccess::WRITE},
+                                    new AstCReset{portp->fileline(), portp, false}};
+                                localDeclp->addNextHere(crstp);
+                            }
+                            needsCLocalScope = true;
+                            continue;
+                        }
                         // Move it to a new localized variable
                         AstVarScope* const localVscp
                             = createVarScope(portp, namePrefix + "__" + portp->shortName());
@@ -734,7 +812,7 @@ class TaskVisitor final : public VNVisitor {
         relink(beginp);
         //
         if (debug() >= 9) beginp->dumpTreeAndNext(cout, "-  iotask: ");
-        return beginp;
+        return needsCLocalScope ? new AstCLocalScope{refp->fileline(), beginp} : beginp;
     }
 
     AstNode* createNonInlinedFTask(AstNodeFTaskRef* refp, const string& namePrefix,
@@ -769,13 +847,20 @@ class TaskVisitor final : public VNVisitor {
             ccallp->superReference(taskRefp->superReference());
         }
 
-        // Convert complicated outputs to temp signals
+        // Convert complicated outputs to temp signals. Task calls that can suspend or fork need
+        // staging temps in the caller activation. Non-inlined functions cannot suspend; keep their
+        // temps as normal scoped temps to avoid exposing AstCLocalScope to earlier function/table
+        // simulation.
+        bool needsCLocalScope = false;
+        const bool activationLocalPorts
+            = !refp->taskp()->verilogFunction() && bodyNeedsActivationLocals(cfuncp->stmtsp());
         {
             const V3TaskConnects tconnects = V3Task::taskConnects(refp, refp->taskp()->stmtsp());
             for (const auto& itr : tconnects) {
                 AstVar* const portp = itr.first;
                 AstArg* const argp = itr.second;
-                connectPort(portp, argp, namePrefix, beginp, false);
+                needsCLocalScope
+                    |= connectPort(portp, argp, namePrefix, beginp, false, activationLocalPorts);
             }
         }
         // First argument is symbol table, then output if a function
@@ -805,7 +890,7 @@ class TaskVisitor final : public VNVisitor {
         if (outvscp) ccallp->addArgsp(new AstVarRef{refp->fileline(), outvscp, VAccess::WRITE});
 
         if (debug() >= 9) beginp->dumpTreeAndNext(cout, "-  nitask: ");
-        return beginp;
+        return needsCLocalScope ? new AstCLocalScope{refp->fileline(), beginp} : beginp;
     }
 
     string dpiSignature(AstNodeFTask* nodep, AstVar* rtnvarp) const {
@@ -1879,6 +1964,99 @@ public:
     }
 };
 
+//######################################################################
+// Materialize task activation-local variables after loop unrolling
+
+using TaskLocalVarScopeMap = std::map<AstVarScope*, AstVarScope*>;
+
+class TaskLocalDeclVisitor final : public VNVisitor {
+    std::vector<AstTaskLocalVar*> m_declps;
+
+    void visit(AstCLocalScope*) override {
+        // A nested C-local scope is a separate task activation.
+    }
+    void visit(AstTaskLocalVar* nodep) override { m_declps.push_back(nodep); }
+    void visit(AstNode* nodep) override { iterateChildren(nodep); }
+
+public:
+    explicit TaskLocalDeclVisitor(AstNode* nodep) { iterateAndNextNull(nodep); }
+    const std::vector<AstTaskLocalVar*>& declps() const { return m_declps; }
+};
+
+class TaskLocalRefVisitor final : public VNVisitor {
+    const TaskLocalVarScopeMap& m_replacements;
+
+    void visit(AstCLocalScope* nodep) override {
+        // A nested activation can capture variables from this outer activation. Its own locals
+        // have different template scopes, so only references to this replacement map are changed.
+        iterateChildren(nodep);
+    }
+    void visit(AstVarRef* nodep) override {
+        const auto it = m_replacements.find(nodep->varScopep());
+        if (it == m_replacements.end()) return;
+        nodep->varScopep(it->second);
+        nodep->varp(it->second->varp());
+    }
+    void visit(AstNode* nodep) override { iterateChildren(nodep); }
+
+public:
+    TaskLocalRefVisitor(AstNode* nodep, const TaskLocalVarScopeMap& replacements)
+        : m_replacements{replacements} {
+        iterateAndNextNull(nodep);
+    }
+};
+
+class TaskLocalizeVisitor final : public VNVisitor {
+    std::set<AstVarScope*> m_templateVscps;
+    AstCLocalScope* m_localScopep = nullptr;
+
+    void visit(AstNetlist* nodep) override {
+        iterateChildren(nodep);
+
+        std::set<AstVar*> templateVarps;
+        for (AstVarScope* const vscp : m_templateVscps) {
+            templateVarps.insert(vscp->varp());
+            VL_DO_DANGLING(pushDeletep(vscp->unlinkFrBack()), vscp);
+        }
+        for (AstVar* const varp : templateVarps) {
+            VL_DO_DANGLING(pushDeletep(varp->unlinkFrBack()), varp);
+        }
+    }
+    void visit(AstCLocalScope* nodep) override {
+        VL_RESTORER(m_localScopep);
+        m_localScopep = nodep;
+
+        // Process nested task activations before replacing declarations in this one.
+        iterateChildren(nodep);
+
+        const TaskLocalDeclVisitor declVisitor{nodep->stmtsp()};
+        TaskLocalVarScopeMap replacements;
+        for (AstTaskLocalVar* const declp : declVisitor.declps()) {
+            AstVarScope* const templateVscp = declp->templateVscp();
+            UASSERT_OBJ(templateVscp, declp, "Task-local declaration has no template scope");
+            AstVar* const localVarp = declp->varp();
+            localVarp->unlinkFrBack();
+            UASSERT_OBJ(localVarp->isFuncLocal(), localVarp, "Task-local variable is not local");
+            AstVarScope* const localVscp
+                = new AstVarScope{localVarp->fileline(), templateVscp->scopep(), localVarp};
+            templateVscp->scopep()->addVarsp(localVscp);
+            UASSERT_OBJ(replacements.emplace(templateVscp, localVscp).second, declp,
+                        "Duplicate task-local template in one activation");
+            m_templateVscps.insert(templateVscp);
+            declp->replaceWith(localVarp);
+            VL_DO_DANGLING(pushDeletep(declp), declp);
+        }
+        if (!replacements.empty()) TaskLocalRefVisitor{nodep->stmtsp(), replacements};
+    }
+    void visit(AstTaskLocalVar* nodep) override {
+        UASSERT_OBJ(m_localScopep, nodep, "Task-local declaration is outside a C-local scope");
+    }
+    void visit(AstNode* nodep) override { iterateChildren(nodep); }
+
+public:
+    explicit TaskLocalizeVisitor(AstNetlist* nodep) { iterate(nodep); }
+};
+
 //######################################################################
 // Task class functions
 
@@ -2300,3 +2478,9 @@ void V3Task::taskAll(AstNetlist* nodep) {
     }  // Destruct before checking
     V3Global::dumpCheckGlobalTree("task", 0, dumpTreeEitherLevel() >= 3);
 }
+
+void V3Task::taskLocalizeAll(AstNetlist* nodep) {
+    UINFO(2, __FUNCTION__ << ":");
+    { const TaskLocalizeVisitor visitor{nodep}; }
+    V3Global::dumpCheckGlobalTree("tasklocal", 0, dumpTreeEitherLevel() >= 3);
+}

src/V3Task.h

@@ -53,6 +53,7 @@ class V3Task final {
 
 public:
     static void taskAll(AstNetlist* nodep) VL_MT_DISABLED;
+    static void taskLocalizeAll(AstNetlist* nodep) VL_MT_DISABLED;
     /// Return vector of [port, pin-connects-to]  (SLOW)
     static V3TaskConnects taskConnects(AstNodeFTaskRef* nodep, AstNode* taskStmtsp,
                                        V3TaskConnectState* statep = nullptr,

src/V3Timing.cpp

@@ -478,6 +478,7 @@ class TimingControlVisitor final : public VNVisitor {
     AstScope* m_scopep = nullptr;  // Current scope
     AstActive* m_activep = nullptr;  // Current active
     AstNode* m_procp = nullptr;  // NodeProcedure/CFunc/Begin we're under
+    AstNode* m_varInsertp = nullptr;  // Nearest node that owns function-local variables
     bool m_hasProcess = false;  // True if current scope has a VlProcess handle available
     int m_forkCnt = 0;  // Number of forks inside a module
     bool m_underJumpBlock = false;  // True if we are inside of a jump-block
@@ -806,8 +807,8 @@ class TimingControlVisitor final : public VNVisitor {
         forkp->addNextHere(new AstCAwait{flp, joinp});
     }
 
-    // `procp` shall be a NodeProcedure/CFunc/Begin and within it vars from `varsp` will be placed.
+    // `procp` shall be a NodeProcedure/CFunc/Begin/CLocalScope and within it vars from `varsp`
-    // `varsp` vector of vars which shall be localized.
+    // will be placed. `varsp` vector of vars which shall be localized.
     static void localizeVars(AstNode* const procp, const std::vector<AstVar*>& varsp) {
         UASSERT(procp, "procp is nullptr");
         AstNode* firstStmtp;
@@ -824,15 +825,19 @@ class TimingControlVisitor final : public VNVisitor {
             firstStmtp = cfuncp->varsp();
         } else if (AstBegin* const beginp = VN_CAST(procp, Begin)) {
             firstStmtp = beginp->stmtsp();
+        } else if (AstCLocalScope* const localScopep = VN_CAST(procp, CLocalScope)) {
+            firstStmtp = localScopep->stmtsp();
         } else {
             procp->v3fatalSrc(
                 procp->prettyNameQ()
-                << " is not of an expected type NodeProcedure/CFunc/Begin instead it is: "
+                << " is not of an expected type NodeProcedure/CFunc/Begin/CLocalScope instead it "
+                   "is: "
                 << procp->prettyTypeName());
         }
         UASSERT_OBJ(firstStmtp, procp,
-                    procp->prettyNameQ() << " has no non-var statement. 'localizeVars()' is ment "
+                    procp->prettyNameQ()
-                                            "to be called on non-empty NodeProcedure/CFunc/Begin");
+                        << " has no non-var statement. 'localizeVars()' is meant to be called on "
+                           "non-empty NodeProcedure/CFunc/Begin/CLocalScope");
         for (AstVar* const varp : varsp) {
             varp->funcLocal(true);
             firstStmtp->addHereThisAsNext(varp->unlinkFrBack());
@@ -889,8 +894,10 @@ class TimingControlVisitor final : public VNVisitor {
     }
     void visit(AstNodeProcedure* nodep) override {
         VL_RESTORER(m_procp);
+        VL_RESTORER(m_varInsertp);
         VL_RESTORER(m_hasProcess);
         m_procp = nodep;
+        m_varInsertp = nodep;
         m_hasProcess = hasFlags(nodep, T_HAS_PROC);
         VL_RESTORER(m_underProcedure);
         m_underProcedure = true;
@@ -913,8 +920,10 @@ class TimingControlVisitor final : public VNVisitor {
     void visit(AstAlways* nodep) override {
         if (nodep->user1SetOnce()) return;
         VL_RESTORER(m_procp);
+        VL_RESTORER(m_varInsertp);
         VL_RESTORER(m_hasProcess);
         m_procp = nodep;
+        m_varInsertp = nodep;
         m_hasProcess = hasFlags(nodep, T_HAS_PROC);
         VL_RESTORER(m_underProcedure);
         m_underProcedure = true;
@@ -947,8 +956,10 @@ class TimingControlVisitor final : public VNVisitor {
     }
     void visit(AstCFunc* nodep) override {
         VL_RESTORER(m_procp);
+        VL_RESTORER(m_varInsertp);
         VL_RESTORER(m_hasProcess);
         m_procp = nodep;
+        m_varInsertp = nodep;
         m_hasProcess = hasFlags(nodep, T_HAS_PROC);
         iterateChildren(nodep);
         if (hasFlags(nodep, T_HAS_PROC)) nodep->setNeedProcess();
@@ -1101,7 +1112,7 @@ class TimingControlVisitor final : public VNVisitor {
                                                 m_senExprBuilderp->build(sentreep).first};
             // Get the SenExprBuilder results
             const SenExprBuilder::Results senResults = m_senExprBuilderp->getAndClearResults();
-            localizeVars(m_procp, senResults.m_vars);
+            localizeVars(m_varInsertp, senResults.m_vars);
             // If post updates are destructive (e.g. clearFired on events), perform a
             // conservative pre-clear once before entering the wait loop so stale state from a
             // previous wait does not cause an immediate false-positive trigger.
@@ -1419,12 +1430,19 @@ class TimingControlVisitor final : public VNVisitor {
     }
     void visit(AstBegin* nodep) override {
         VL_RESTORER(m_procp);
+        VL_RESTORER(m_varInsertp);
         VL_RESTORER(m_hasProcess);
         m_hasProcess |= hasFlags(nodep, T_HAS_PROC);
         m_procp = nodep;
+        m_varInsertp = nodep;
         if (m_hasProcess) nodep->setNeedProcess();
         iterateChildren(nodep);
     }
+    void visit(AstCLocalScope* nodep) override {
+        VL_RESTORER(m_varInsertp);
+        m_varInsertp = nodep;
+        iterateChildren(nodep);
+    }
     void visit(AstFork* nodep) override {
         if (nodep->user1SetOnce()) return;
         UINFO(9, "control-visit " << nodep);

src/Verilator.cpp

@@ -397,6 +397,9 @@ static void process() {
             // Loop unrolling
             V3Unroll::unrollAll(v3Global.rootp());
 
+            // Materialize task activation locals after statement-cloning loop transforms.
+            V3Task::taskLocalizeAll(v3Global.rootp());
+
             // Expand slices of arrays
             V3Slice::sliceAll(v3Global.rootp());
 

test_regress/t/t_always_split.py

@@ -14,7 +14,7 @@ test.scenarios('simulator')
 test.compile(verilator_flags2=["--stats"])
 
 if test.vlt_all:
-    test.file_grep(test.stats, r'Optimizations, Split always\s+(\d+)', 4)
+    test.file_grep(test.stats, r'Optimizations, Split always\s+(\d+)', 6)
 
 test.execute()
 

test_regress/t/t_always_split.v

@@ -55,6 +55,24 @@ module t (
     l_split_1 <= l_split_2 | m_din;
   end
 
+  task automatic copy_task(input logic [15:0] value, output logic [15:0] result);
+    result = value;
+  endtask
+
+  reg [15:0] t_split_1, t_split_2;
+  always @(  /*AS*/ m_din) begin
+    copy_task(m_din, t_split_1);
+    t_split_2 = ~m_din;
+  end
+
+  reg [15:0] n_split_1, n_split_2;
+  always @(  /*AS*/ m_din) begin
+    for (int i = 0; i < m_din[15]; ++i) begin
+      copy_task(m_din, n_split_1);
+    end
+    n_split_2 = ~m_din;
+  end
+
   // (The checker block is an exception, it won't split.)
   always @(posedge clk) begin
     if (cyc != 0) begin
@@ -68,18 +86,24 @@ module t (
         m_din <= 16'he11e;
         //$write(" A %x %x\n", a_split_1, a_split_2);
         if (!(a_split_1 == 16'hfeed && a_split_2 == 16'hfeed)) $stop;
+        if (!(t_split_1 == 16'hfeed && t_split_2 == 16'h0112)) $stop;
+        if (!(n_split_1 == 16'hfeed && n_split_2 == 16'h0112)) $stop;
         if (!(d_split_1 == 16'h0112 && d_split_2 == 16'h0112)) $stop;
         if (!(h_split_1 == 16'hfeed && h_split_2 == 16'h0112)) $stop;
       end
       if (cyc == 5) begin
         m_din <= 16'he22e;
         if (!(a_split_1 == 16'he11e && a_split_2 == 16'he11e)) $stop;
+        if (!(t_split_1 == 16'he11e && t_split_2 == 16'h1ee1)) $stop;
+        if (!(n_split_1 == 16'he11e && n_split_2 == 16'h1ee1)) $stop;
         if (!(d_split_1 == 16'h0112 && d_split_2 == 16'h0112)) $stop;
         if (!(h_split_1 == 16'hfeed && h_split_2 == 16'h0112)) $stop;
       end
       if (cyc == 6) begin
         m_din <= 16'he33e;
         if (!(a_split_1 == 16'he22e && a_split_2 == 16'he22e)) $stop;
+        if (!(t_split_1 == 16'he22e && t_split_2 == 16'h1dd1)) $stop;
+        if (!(n_split_1 == 16'he22e && n_split_2 == 16'h1dd1)) $stop;
         if (!(d_split_1 == 16'h1ee1 && d_split_2 == 16'h0112)) $stop;
         if (!(h_split_1 == 16'he11e && h_split_2 == 16'h1ee1)) $stop;
       end

test_regress/t/t_debug_emitv.out

@@ -114,7 +114,8 @@ module Vt_debug_emitv_t;
         logic [11:10]  nn3;
     } [5:4] nibblearray[3:2];
     task t;
-        $display("stmt");
+        input int signed value;
+        $display("stmt %d", value);
     endtask
     function f;
         input int signed v;
@@ -137,7 +138,7 @@ module Vt_debug_emitv_t;
                             other = f(i);
                             $display("stmt %d %d", 
                                      i, other);
-                            t();
+                            t(i);
                         end
                         i = (i + 'h1);
                     end

test_regress/t/t_debug_emitv.v

@@ -129,8 +129,8 @@ module t (/*AUTOARG*/
   } nibble_t;
   nibble_t [5:4] nibblearray[3:2];
 
-  task t;
+  task t(input int value);
-    $display("stmt");
+    $display("stmt %d", value);
   endtask
   function int f(input int v);
     $display("stmt");
@@ -146,7 +146,7 @@ module t (/*AUTOARG*/
       for (int i = 0; i < 3; ++i) begin
         other = f(i);
         $display("stmt %d %d", i, other);
-        t();
+        t(i);
       end
     end
     begin : named

test_regress/t/t_fork_join_none_capture.v

@@ -4,9 +4,162 @@
 // SPDX-FileCopyrightText: 2026 by Antmicro Ltd.
 // SPDX-License-Identifier: CC0-1.0
 
+// verilog_format: off
+`define stop $stop;
+`define checkd(gotv, expv) do if ((gotv) !== (expv)) begin $write("%%Error: %s:%0d:  got=%0d exp=%0d\n", `__FILE__, `__LINE__, (gotv), (expv)); `stop; end while (0);
+// verilog_format: on
+
 module test;
   mailbox #(int) mbox;
+  int seen_10 = 0;
+  int seen_30 = 0;
+
+  task automatic check_mbox_item(input int item);
+    #0;
+    if (item == 10) begin
+      seen_10++;
+    end
+    else if (item == 30) begin
+      seen_30++;
+    end
+    else begin
+      $write("%%Error: unexpected item '%0d'\n", item);
+      `stop;
+    end
+  endtask
+
+  class driver;
+    string names[string];
+    int seen_a = 0;
+    int seen_b = 0;
+
+    task automatic check_name(input string name);
+      #0;
+      if (name == "a") begin
+        seen_a++;
+      end
+      else if (name == "b") begin
+        seen_b++;
+      end
+      else begin
+        $write("%%Error: unexpected name '%0s'\n", name);
+        `stop;
+      end
+    endtask
+
+    task automatic make_name(input bit pick_b, output string name);
+      name = pick_b ? "b" : "a";
+      #0;
+    endtask
+
+    task automatic append_name(inout string name);
+      // verilator no_inline_task
+      #0;
+      name = {name, "_done"};
+    endtask
+
+    task automatic check_arg_capture();
+      names["a"] = "alpha";
+      names["b"] = "bravo";
+
+      foreach (names[name]) begin
+        fork
+          check_name(name);
+        join_none
+      end
+
+      wait fork;
+
+      `checkd(seen_a, 1);
+      `checkd(seen_b, 1);
+    endtask
+
+    task automatic check_local_lifetime();
+      seen_a = 0;
+      seen_b = 0;
+
+      for (int i = 0; i < 2; ++i) begin
+        automatic bit pick_b = i[0];
+        fork
+          begin
+            string name;
+            name = pick_b ? "b" : "a";
+            #0;
+            check_name(name);
+          end
+        join_none
+      end
+
+      wait fork;
+
+      `checkd(seen_a, 1);
+      `checkd(seen_b, 1);
+    endtask
+
+    task automatic check_output_capture();
+      seen_a = 0;
+      seen_b = 0;
+
+      for (int i = 0; i < 2; ++i) begin
+        automatic bit pick_b = i[0];
+        fork
+          begin
+            string name;
+            make_name(pick_b, name);
+            #0;
+            check_name(name);
+          end
+        join_none
+      end
+
+      wait fork;
+
+      `checkd(seen_a, 1);
+      `checkd(seen_b, 1);
+    endtask
+
+    task automatic check_inout_capture();
+      seen_a = 0;
+      seen_b = 0;
+
+      for (int i = 0; i < 2; ++i) begin
+        automatic bit pick_b = i[0];
+        fork
+          begin
+            string name = pick_b ? "b" : "a";
+            append_name(name);
+            #0;
+            if (name == "a_done") begin
+              seen_a++;
+            end
+            else if (name == "b_done") begin
+              seen_b++;
+            end
+            else begin
+              $write("%%Error: unexpected inout name '%0s'\n", name);
+              `stop;
+            end
+          end
+        join_none
+      end
+
+      wait fork;
+
+      `checkd(seen_a, 1);
+      `checkd(seen_b, 1);
+    endtask
+
+    task automatic run();
+      check_arg_capture();
+      check_local_lifetime();
+      check_output_capture();
+      check_inout_capture();
+    endtask
+  endclass
+
   initial begin
+    driver drv;
+
     mbox = new();
     mbox.put(10);
     mbox.put(30);
@@ -15,14 +168,19 @@ module test;
       automatic int item;
       mbox.get(item);
       fork
-        begin
+        check_mbox_item(item);
-          $display("got", item);
-          if (item == 10) $finish;
-        end
       join_none
     end
 
-    #0;
+    wait fork;
-    $stop;
+
+    `checkd(seen_10, 1);
+    `checkd(seen_30, 1);
+
+    drv = new;
+    drv.run();
+
+    $write("*-* All Finished *-*\n");
+    $finish;
   end
 endmodule

test_regress/t/t_fork_task_local_replicate.py

@@ -0,0 +1,16 @@
+#!/usr/bin/env python3
+# DESCRIPTION: Verilator: Task-local variables survive scheduler logic replication
+#
+# This program is free software; you can redistribute it and/or modify it
+# under the terms of either the GNU Lesser General Public License Version 3
+# or the Perl Artistic License Version 2.0.
+# SPDX-FileCopyrightText: 2026 Wilson Snyder
+# SPDX-License-Identifier: LGPL-3.0-only OR Artistic-2.0
+
+import vltest_bootstrap
+
+test.scenarios('simulator')
+
+test.lint(verilator_flags2=["--timing"])
+
+test.passes()

test_regress/t/t_fork_task_local_replicate.v

@@ -0,0 +1,20 @@
+// DESCRIPTION: Verilator: Task-local variables survive scheduler logic replication
+//
+// This file ONLY is placed under the Creative Commons Public Domain.
+// SPDX-FileCopyrightText: 2026 by Wilson Snyder
+// SPDX-License-Identifier: CC0-1.0
+
+module t (
+    input logic a,
+    output logic unused
+);
+
+  task automatic spawn(input logic value);
+    fork
+      unused = value;
+    join_none
+  endtask
+
+  always @* spawn(a);
+
+endmodule

test_regress/t/t_unroll_stmt.py

@@ -26,7 +26,7 @@ test.file_grep(test.stats,
 test.file_grep(test.stats,
                r'Optimizations, Loop unrolling, Failed - unknown loop condition\s+(\d+)', 0)
 test.file_grep(test.stats, r'Optimizations, Loop unrolling, Pragma unroll_disable\s+(\d+)', 0)
-test.file_grep(test.stats, r'Optimizations, Loop unrolling, Unrolled loops\s+(\d+)', 13)
+test.file_grep(test.stats, r'Optimizations, Loop unrolling, Unrolled loops\s+(\d+)', 14)
-test.file_grep(test.stats, r'Optimizations, Loop unrolling, Unrolled iterations\s+(\d+)', 77)
+test.file_grep(test.stats, r'Optimizations, Loop unrolling, Unrolled iterations\s+(\d+)', 80)
 
 test.passes()

test_regress/t/t_unroll_stmt.v

@@ -14,11 +14,21 @@ module t;
 
   int nonConst3 = $c("3");
 
+  task automatic copy_task(input int value, output int result);
+    result = value;
+  endtask
+
   initial begin
     // Basic loop
     for (int i = 0; i < 3; ++i) begin : loop_0
       $display("loop_0 %0d", i);
     end
+    // Loop containing activation-local task staging
+    for (int i = 0; i < 2; ++i) begin : loop_task
+      int result;
+      copy_task(i, result);
+      if (result != i) $stop;
+    end
     // Loop with 2 init/step
     for (int i = 0, j = 5; i < j; i += 2, j += 1) begin : loop_1
       $display("loop_1 %0d %0d", i, j);