AVX2

commit 5e6f06cf59
Author: Aaron Gutierrez <gutierrez.aaron.m@gmail.com>
Date:   Thu May 11 15:42:11 2017 -0400

    Fixed issue with aliasing local variables

    ISPC++ now produces valid code, or an appropriate error message, for all
    of my test cases.

commit bfe723e1b7
Author: Aaron Gutierrez <gutierrez.aaron.m@gmail.com>
Date:   Thu May 11 03:09:38 2017 -0400

    Actually copy the AST.

    Type replacement works except for function parameters.

commit f65b3e6300
Author: Aaron Gutierrez <gutierrez.aaron.m@gmail.com>
Date:   Thu May 11 01:19:50 2017 -0400

    [WIP] Remove cases for ForeachStmt and SymbolExpr

commit 2e28640860
Merge: 6a91c5d d020107
Author: Aaron Gutierrez <gutierrez.aaron.m@gmail.com>
Date:   Wed May 10 23:13:40 2017 -0400

    Merge branch 'master' into copy_ast

commit 6a91c5d5ac
Author: Aaron Gutierrez <gutierrez.aaron.m@gmail.com>
Date:   Wed May 10 11:11:39 2017 -0400

    Attempt to replicate AST when expanding polytypes

ast.cpp

@@ -86,7 +86,7 @@ AST::GenerateIR() {
 
 ASTNode *
 WalkAST(ASTNode *node, ASTPreCallBackFunc preFunc, ASTPostCallBackFunc postFunc,
-        void *data) {
+        void *data, ASTPostCallBackFunc preUpdate) {
     if (node == NULL)
         return node;
 
@@ -97,6 +97,10 @@ WalkAST(ASTNode *node, ASTPreCallBackFunc preFunc, ASTPostCallBackFunc postFunc,
             return node;
     }
 
+    if (preUpdate != NULL) {
+        node = preUpdate(node, data);
+    }
+
     ////////////////////////////////////////////////////////////////////////////
     // Handle Statements
     if (llvm::dyn_cast<Stmt>(node) != NULL) {
@@ -120,54 +124,54 @@ WalkAST(ASTNode *node, ASTPreCallBackFunc preFunc, ASTPostCallBackFunc postFunc,
         UnmaskedStmt *ums;
 
         if ((es = llvm::dyn_cast<ExprStmt>(node)) != NULL)
-            es->expr = (Expr *)WalkAST(es->expr, preFunc, postFunc, data);
+            es->expr = (Expr *)WalkAST(es->expr, preFunc, postFunc, data, preUpdate);
         else if ((ds = llvm::dyn_cast<DeclStmt>(node)) != NULL) {
             for (unsigned int i = 0; i < ds->vars.size(); ++i)
                 ds->vars[i].init = (Expr *)WalkAST(ds->vars[i].init, preFunc,
-                                                   postFunc, data);
+                                                   postFunc, data, preUpdate);
         }
         else if ((is = llvm::dyn_cast<IfStmt>(node)) != NULL) {
-            is->test = (Expr *)WalkAST(is->test, preFunc, postFunc, data);
+            is->test = (Expr *)WalkAST(is->test, preFunc, postFunc, data, preUpdate);
             is->trueStmts = (Stmt *)WalkAST(is->trueStmts, preFunc,
-                                            postFunc, data);
+                                            postFunc, data, preUpdate);
             is->falseStmts = (Stmt *)WalkAST(is->falseStmts, preFunc,
-                                             postFunc, data);
+                                             postFunc, data, preUpdate);
         }
         else if ((dos = llvm::dyn_cast<DoStmt>(node)) != NULL) {
             dos->testExpr = (Expr *)WalkAST(dos->testExpr, preFunc,
-                                            postFunc, data);
+                                            postFunc, data, preUpdate);
             dos->bodyStmts = (Stmt *)WalkAST(dos->bodyStmts, preFunc,
-                                             postFunc, data);
+                                             postFunc, data, preUpdate);
         }
         else if ((fs = llvm::dyn_cast<ForStmt>(node)) != NULL) {
-            fs->init = (Stmt *)WalkAST(fs->init, preFunc, postFunc, data);
-            fs->test = (Expr *)WalkAST(fs->test, preFunc, postFunc, data);
-            fs->step = (Stmt *)WalkAST(fs->step, preFunc, postFunc, data);
-            fs->stmts = (Stmt *)WalkAST(fs->stmts, preFunc, postFunc, data);
+            fs->init = (Stmt *)WalkAST(fs->init, preFunc, postFunc, data, preUpdate);
+            fs->test = (Expr *)WalkAST(fs->test, preFunc, postFunc, data, preUpdate);
+            fs->step = (Stmt *)WalkAST(fs->step, preFunc, postFunc, data, preUpdate);
+            fs->stmts = (Stmt *)WalkAST(fs->stmts, preFunc, postFunc, data, preUpdate);
         }
         else if ((fes = llvm::dyn_cast<ForeachStmt>(node)) != NULL) {
             for (unsigned int i = 0; i < fes->startExprs.size(); ++i)
                 fes->startExprs[i] = (Expr *)WalkAST(fes->startExprs[i], preFunc,
-                                                     postFunc, data);
+                                                     postFunc, data, preUpdate);
             for (unsigned int i = 0; i < fes->endExprs.size(); ++i)
                 fes->endExprs[i] = (Expr *)WalkAST(fes->endExprs[i], preFunc,
-                                                   postFunc, data);
-            fes->stmts = (Stmt *)WalkAST(fes->stmts, preFunc, postFunc, data);
+                                                   postFunc, data, preUpdate);
+            fes->stmts = (Stmt *)WalkAST(fes->stmts, preFunc, postFunc, data, preUpdate);
         }
         else if ((fas = llvm::dyn_cast<ForeachActiveStmt>(node)) != NULL) {
-            fas->stmts = (Stmt *)WalkAST(fas->stmts, preFunc, postFunc, data);
+            fas->stmts = (Stmt *)WalkAST(fas->stmts, preFunc, postFunc, data, preUpdate);
         }
         else if ((fus = llvm::dyn_cast<ForeachUniqueStmt>(node)) != NULL) {
-            fus->expr = (Expr *)WalkAST(fus->expr, preFunc, postFunc, data);
-            fus->stmts = (Stmt *)WalkAST(fus->stmts, preFunc, postFunc, data);
+            fus->expr = (Expr *)WalkAST(fus->expr, preFunc, postFunc, data, preUpdate);
+            fus->stmts = (Stmt *)WalkAST(fus->stmts, preFunc, postFunc, data, preUpdate);
         }
         else if ((cs = llvm::dyn_cast<CaseStmt>(node)) != NULL)
-            cs->stmts = (Stmt *)WalkAST(cs->stmts, preFunc, postFunc, data);
+            cs->stmts = (Stmt *)WalkAST(cs->stmts, preFunc, postFunc, data, preUpdate);
         else if ((defs = llvm::dyn_cast<DefaultStmt>(node)) != NULL)
-            defs->stmts = (Stmt *)WalkAST(defs->stmts, preFunc, postFunc, data);
+            defs->stmts = (Stmt *)WalkAST(defs->stmts, preFunc, postFunc, data, preUpdate);
         else if ((ss = llvm::dyn_cast<SwitchStmt>(node)) != NULL) {
-            ss->expr = (Expr *)WalkAST(ss->expr, preFunc, postFunc, data);
-            ss->stmts = (Stmt *)WalkAST(ss->stmts, preFunc, postFunc, data);
+            ss->expr = (Expr *)WalkAST(ss->expr, preFunc, postFunc, data, preUpdate);
+            ss->stmts = (Stmt *)WalkAST(ss->stmts, preFunc, postFunc, data, preUpdate);
         }
         else if (llvm::dyn_cast<BreakStmt>(node) != NULL ||
                  llvm::dyn_cast<ContinueStmt>(node) != NULL ||
@@ -175,22 +179,22 @@ WalkAST(ASTNode *node, ASTPreCallBackFunc preFunc, ASTPostCallBackFunc postFunc,
             // nothing
         }
         else if ((ls = llvm::dyn_cast<LabeledStmt>(node)) != NULL)
-            ls->stmt = (Stmt *)WalkAST(ls->stmt, preFunc, postFunc, data);
+            ls->stmt = (Stmt *)WalkAST(ls->stmt, preFunc, postFunc, data, preUpdate);
         else if ((rs = llvm::dyn_cast<ReturnStmt>(node)) != NULL)
-            rs->expr = (Expr *)WalkAST(rs->expr, preFunc, postFunc, data);
+            rs->expr = (Expr *)WalkAST(rs->expr, preFunc, postFunc, data, preUpdate);
         else if ((sl = llvm::dyn_cast<StmtList>(node)) != NULL) {
             std::vector<Stmt *> &sls = sl->stmts;
             for (unsigned int i = 0; i < sls.size(); ++i)
-                sls[i] = (Stmt *)WalkAST(sls[i], preFunc, postFunc, data);
+                sls[i] = (Stmt *)WalkAST(sls[i], preFunc, postFunc, data, preUpdate);
         }
         else if ((ps = llvm::dyn_cast<PrintStmt>(node)) != NULL)
-            ps->values = (Expr *)WalkAST(ps->values, preFunc, postFunc, data);
+            ps->values = (Expr *)WalkAST(ps->values, preFunc, postFunc, data, preUpdate);
         else if ((as = llvm::dyn_cast<AssertStmt>(node)) != NULL)
-            as->expr = (Expr *)WalkAST(as->expr, preFunc, postFunc, data);
+            as->expr = (Expr *)WalkAST(as->expr, preFunc, postFunc, data, preUpdate);
         else if ((dels = llvm::dyn_cast<DeleteStmt>(node)) != NULL)
-            dels->expr = (Expr *)WalkAST(dels->expr, preFunc, postFunc, data);
+            dels->expr = (Expr *)WalkAST(dels->expr, preFunc, postFunc, data, preUpdate);
         else if ((ums = llvm::dyn_cast<UnmaskedStmt>(node)) != NULL)
-            ums->stmts = (Stmt *)WalkAST(ums->stmts, preFunc, postFunc, data);
+            ums->stmts = (Stmt *)WalkAST(ums->stmts, preFunc, postFunc, data, preUpdate);
         else
             FATAL("Unhandled statement type in WalkAST()");
     }
@@ -215,57 +219,57 @@ WalkAST(ASTNode *node, ASTPreCallBackFunc preFunc, ASTPostCallBackFunc postFunc,
         NewExpr *newe;
 
         if ((ue = llvm::dyn_cast<UnaryExpr>(node)) != NULL)
-            ue->expr = (Expr *)WalkAST(ue->expr, preFunc, postFunc, data);
+            ue->expr = (Expr *)WalkAST(ue->expr, preFunc, postFunc, data, preUpdate);
         else if ((be = llvm::dyn_cast<BinaryExpr>(node)) != NULL) {
-            be->arg0 = (Expr *)WalkAST(be->arg0, preFunc, postFunc, data);
-            be->arg1 = (Expr *)WalkAST(be->arg1, preFunc, postFunc, data);
+            be->arg0 = (Expr *)WalkAST(be->arg0, preFunc, postFunc, data, preUpdate);
+            be->arg1 = (Expr *)WalkAST(be->arg1, preFunc, postFunc, data, preUpdate);
         }
         else if ((ae = llvm::dyn_cast<AssignExpr>(node)) != NULL) {
-            ae->lvalue = (Expr *)WalkAST(ae->lvalue, preFunc, postFunc, data);
-            ae->rvalue = (Expr *)WalkAST(ae->rvalue, preFunc, postFunc, data);
+            ae->lvalue = (Expr *)WalkAST(ae->lvalue, preFunc, postFunc, data, preUpdate);
+            ae->rvalue = (Expr *)WalkAST(ae->rvalue, preFunc, postFunc, data, preUpdate);
         }
         else if ((se = llvm::dyn_cast<SelectExpr>(node)) != NULL) {
-            se->test = (Expr *)WalkAST(se->test, preFunc, postFunc, data);
-            se->expr1 = (Expr *)WalkAST(se->expr1, preFunc, postFunc, data);
-            se->expr2 = (Expr *)WalkAST(se->expr2, preFunc, postFunc, data);
+            se->test = (Expr *)WalkAST(se->test, preFunc, postFunc, data, preUpdate);
+            se->expr1 = (Expr *)WalkAST(se->expr1, preFunc, postFunc, data, preUpdate);
+            se->expr2 = (Expr *)WalkAST(se->expr2, preFunc, postFunc, data, preUpdate);
         }
         else if ((el = llvm::dyn_cast<ExprList>(node)) != NULL) {
             for (unsigned int i = 0; i < el->exprs.size(); ++i)
                 el->exprs[i] = (Expr *)WalkAST(el->exprs[i], preFunc,
-                                               postFunc, data);
+                                               postFunc, data, preUpdate);
         }
         else if ((fce = llvm::dyn_cast<FunctionCallExpr>(node)) != NULL) {
-            fce->func = (Expr *)WalkAST(fce->func, preFunc, postFunc, data);
-            fce->args = (ExprList *)WalkAST(fce->args, preFunc, postFunc, data);
+            fce->func = (Expr *)WalkAST(fce->func, preFunc, postFunc, data, preUpdate);
+            fce->args = (ExprList *)WalkAST(fce->args, preFunc, postFunc, data, preUpdate);
             for (int k = 0; k < 3; k++)
               fce->launchCountExpr[0] = (Expr *)WalkAST(fce->launchCountExpr[0], preFunc,
-                                                   postFunc, data);
+                                                   postFunc, data, preUpdate);
         }
         else if ((ie = llvm::dyn_cast<IndexExpr>(node)) != NULL) {
-            ie->baseExpr = (Expr *)WalkAST(ie->baseExpr, preFunc, postFunc, data);
-            ie->index = (Expr *)WalkAST(ie->index, preFunc, postFunc, data);
+            ie->baseExpr = (Expr *)WalkAST(ie->baseExpr, preFunc, postFunc, data, preUpdate);
+            ie->index = (Expr *)WalkAST(ie->index, preFunc, postFunc, data, preUpdate);
         }
         else if ((me = llvm::dyn_cast<MemberExpr>(node)) != NULL)
-            me->expr = (Expr *)WalkAST(me->expr, preFunc, postFunc, data);
+            me->expr = (Expr *)WalkAST(me->expr, preFunc, postFunc, data, preUpdate);
         else if ((tce = llvm::dyn_cast<TypeCastExpr>(node)) != NULL)
-            tce->expr = (Expr *)WalkAST(tce->expr, preFunc, postFunc, data);
+            tce->expr = (Expr *)WalkAST(tce->expr, preFunc, postFunc, data, preUpdate);
         else if ((re = llvm::dyn_cast<ReferenceExpr>(node)) != NULL)
-            re->expr = (Expr *)WalkAST(re->expr, preFunc, postFunc, data);
+            re->expr = (Expr *)WalkAST(re->expr, preFunc, postFunc, data, preUpdate);
         else if ((ptrderef = llvm::dyn_cast<PtrDerefExpr>(node)) != NULL)
             ptrderef->expr = (Expr *)WalkAST(ptrderef->expr, preFunc, postFunc,
-                                             data);
+                                             data, preUpdate);
         else if ((refderef = llvm::dyn_cast<RefDerefExpr>(node)) != NULL)
             refderef->expr = (Expr *)WalkAST(refderef->expr, preFunc, postFunc,
-                                             data);
+                                             data, preUpdate);
         else if ((soe = llvm::dyn_cast<SizeOfExpr>(node)) != NULL)
-            soe->expr = (Expr *)WalkAST(soe->expr, preFunc, postFunc, data);
+            soe->expr = (Expr *)WalkAST(soe->expr, preFunc, postFunc, data, preUpdate);
         else if ((aoe = llvm::dyn_cast<AddressOfExpr>(node)) != NULL)
-            aoe->expr = (Expr *)WalkAST(aoe->expr, preFunc, postFunc, data);
+            aoe->expr = (Expr *)WalkAST(aoe->expr, preFunc, postFunc, data, preUpdate);
         else if ((newe = llvm::dyn_cast<NewExpr>(node)) != NULL) {
             newe->countExpr = (Expr *)WalkAST(newe->countExpr, preFunc,
-                                              postFunc, data);
+                                              postFunc, data, preUpdate);
             newe->initExpr = (Expr *)WalkAST(newe->initExpr, preFunc,
-                                             postFunc, data);
+                                             postFunc, data, preUpdate);
         }
         else if (llvm::dyn_cast<SymbolExpr>(node) != NULL ||
                  llvm::dyn_cast<ConstExpr>(node) != NULL ||
@@ -536,11 +540,21 @@ lTranslatePolyNode(ASTNode *node, void *d) {
     return node->ReplacePolyType(data->polyType, data->replacement);
 }
 
+static ASTNode *
+lCopyNode(ASTNode *node, void *) {
+    return node->Copy();
+}
+
 ASTNode *
 TranslatePoly(ASTNode *root, const PolyType *polyType, const Type *replacement) {
     struct PolyData data;
     data.polyType = polyType;
     data.replacement = replacement;
 
-    return WalkAST(root, NULL, lTranslatePolyNode, &data);
+    return WalkAST(root, NULL, lTranslatePolyNode, &data, lCopyNode);
+}
+
+ASTNode *
+CopyAST(ASTNode *root) {
+    return WalkAST(root, NULL, NULL, NULL, lCopyNode);
 }

ast.h

@@ -68,6 +68,8 @@ public:
         pointer in place of the original ASTNode *. */
     virtual ASTNode *TypeCheck() = 0;
 
+    virtual ASTNode *Copy() = 0;
+
     virtual ASTNode *ReplacePolyType(const PolyType *, const Type *) = 0;
 
     /** Estimate the execution cost of the node (not including the cost of
@@ -177,7 +179,8 @@ typedef ASTNode * (* ASTPostCallBackFunc)(ASTNode *node, void *data);
     doing so, calls postFunc, at the node.  The return value from the
     postFunc call is ignored. */
 extern ASTNode *WalkAST(ASTNode *root, ASTPreCallBackFunc preFunc,
-                        ASTPostCallBackFunc postFunc, void *data);
+                        ASTPostCallBackFunc postFunc, void *data,
+                        ASTPostCallBackFunc preUpdate = NULL);
 
 /** Perform simple optimizations on the AST or portion thereof passed to
     this function, returning the resulting AST. */
@@ -209,6 +212,8 @@ extern int EstimateCost(ASTNode *root);
 
 extern ASTNode * TranslatePoly(ASTNode *root, const PolyType *polyType, const Type *replacement);
 
+extern ASTNode * CopyAST(ASTNode *root);
+
 /** Returns true if it would be safe to run the given code with an "all
     off" mask. */
 extern bool SafeToRunWithMaskAllOff(ASTNode *root);

expr.cpp

@@ -113,7 +113,81 @@ Expr::GetBaseSymbol() const {
 }
 
 Expr *
-Expr::ReplacePolyType(const PolyType *polyType, const Type *replacement) {
+Expr::Copy() {
+    Expr *copy;
+    switch (getValueID()) {
+        case AddressOfExprID:
+            copy = (Expr*)new AddressOfExpr(*(AddressOfExpr*)this);
+            break;
+        case AssignExprID:
+            copy = (Expr*)new AssignExpr(*(AssignExpr*)this);
+            break;
+        case BinaryExprID:
+            copy = (Expr*)new BinaryExpr(*(BinaryExpr*)this);
+            break;
+        case ConstExprID:
+            copy = (Expr*)new ConstExpr(*(ConstExpr*)this);
+            break;
+        case PtrDerefExprID:
+            copy = (Expr*)new PtrDerefExpr(*(PtrDerefExpr*)this);
+            break;
+        case RefDerefExprID:
+            copy = (Expr*)new RefDerefExpr(*(RefDerefExpr*)this);
+            break;
+        case ExprListID:
+            copy = (Expr*)new ExprList(*(ExprList*)this);
+            break;
+        case FunctionCallExprID:
+            copy = (Expr*)new FunctionCallExpr(*(FunctionCallExpr*)this);
+            break;
+        case FunctionSymbolExprID:
+            copy = (Expr*)new FunctionSymbolExpr(*(FunctionSymbolExpr*)this);
+            break;
+        case IndexExprID:
+            copy = (Expr*)new IndexExpr(*(IndexExpr*)this);
+            break;
+        case StructMemberExprID:
+            copy = (Expr*)new StructMemberExpr(*(StructMemberExpr*)this);
+            break;
+        case VectorMemberExprID:
+            copy = (Expr*)new VectorMemberExpr(*(VectorMemberExpr*)this);
+            break;
+        case NewExprID:
+            copy = (Expr*)new NewExpr(*(NewExpr*)this);
+            break;
+        case NullPointerExprID:
+            copy = (Expr*)new NullPointerExpr(*(NullPointerExpr*)this);
+            break;
+        case ReferenceExprID:
+            copy = (Expr*)new ReferenceExpr(*(ReferenceExpr*)this);
+            break;
+        case SelectExprID:
+            copy = (Expr*)new SelectExpr(*(SelectExpr*)this);
+            break;
+        case SizeOfExprID:
+            copy = (Expr*)new SizeOfExpr(*(SizeOfExpr*)this);
+            break;
+        case SymbolExprID:
+            copy = (Expr*)new SymbolExpr(*(SymbolExpr*)this);
+            break;
+        case SyncExprID:
+            copy = (Expr*)new SyncExpr(*(SyncExpr*)this);
+            break;
+        case TypeCastExprID:
+            copy = (Expr*)new TypeCastExpr(*(TypeCastExpr*)this);
+            break;
+        case UnaryExprID:
+            copy = (Expr*)new UnaryExpr(*(UnaryExpr*)this);
+            break;
+        default:
+            FATAL("Unmatched case in Expr::Copy");
+            copy = this; // just to silence the compiler
+    }
+    return copy;
+}
+
+Expr *
+Expr::ReplacePolyType(const PolyType *, const Type *) {
     return this;
 }
 
@@ -4680,11 +4754,11 @@ IndexExpr::ReplacePolyType(const PolyType *from, const Type *to) {
     if (index == NULL || baseExpr == NULL)
         return NULL;
 
-    if (Type::EqualForReplacement(this->GetType()->GetBaseType(), from)) {
+    if (Type::EqualForReplacement(GetType()->GetBaseType(), from)) {
         type = PolyType::ReplaceType(type, to);
     }
 
-    if (Type::EqualForReplacement(this->GetLValueType()->GetBaseType(), from)) {
+    if (Type::EqualForReplacement(GetLValueType()->GetBaseType(), from)) {
         lvalueType = new PointerType(to, lvalueType->GetVariability(),
                                      lvalueType->IsConstType());
     }
@@ -4757,27 +4831,6 @@ lIdentifierToVectorElement(char id) {
 //////////////////////////////////////////////////
 // StructMemberExpr
 
-class StructMemberExpr : public MemberExpr
-{
-public:
-    StructMemberExpr(Expr *e, const char *id, SourcePos p,
-                     SourcePos idpos, bool derefLValue);
-
-    static inline bool classof(StructMemberExpr const*) { return true; }
-    static inline bool classof(ASTNode const* N) {
-        return N->getValueID() == StructMemberExprID;
-    }
-
-    const Type *GetType() const;
-    const Type *GetLValueType() const;
-    int getElementNumber() const;
-    const Type *getElementType() const;
-
-private:
-    const StructType *getStructType() const;
-};
-
-
 StructMemberExpr::StructMemberExpr(Expr *e, const char *id, SourcePos p,
                                    SourcePos idpos, bool derefLValue)
     : MemberExpr(e, id, p, idpos, derefLValue, StructMemberExprID) {
@@ -4929,31 +4982,6 @@ StructMemberExpr::getStructType() const {
 //////////////////////////////////////////////////
 // VectorMemberExpr
 
-class VectorMemberExpr : public MemberExpr
-{
-public:
-    VectorMemberExpr(Expr *e, const char *id, SourcePos p,
-                     SourcePos idpos, bool derefLValue);
-
-    static inline bool classof(VectorMemberExpr const*) { return true; }
-    static inline bool classof(ASTNode const* N) {
-        return N->getValueID() == VectorMemberExprID;
-    }
-
-    llvm::Value *GetValue(FunctionEmitContext* ctx) const;
-    llvm::Value *GetLValue(FunctionEmitContext* ctx) const;
-    const Type *GetType() const;
-    const Type *GetLValueType() const;
-
-    int getElementNumber() const;
-    const Type *getElementType() const;
-
-private:
-    const VectorType *exprVectorType;
-    const VectorType *memberType;
-};
-
-
 VectorMemberExpr::VectorMemberExpr(Expr *e, const char *id, SourcePos p,
                                    SourcePos idpos, bool derefLValue)
     : MemberExpr(e, id, p, idpos, derefLValue, VectorMemberExprID) {
@@ -5339,14 +5367,10 @@ MemberExpr::ReplacePolyType(const PolyType *from, const Type *to) {
     if (expr == NULL)
         return NULL;
 
-    if (Type::EqualForReplacement(this->GetType()->GetBaseType(), from)) {
+    if (Type::EqualForReplacement(GetType()->GetBaseType(), from)) {
         type = PolyType::ReplaceType(type, to);
     }
 
-    if (Type::EqualForReplacement(this->GetLValueType()->GetBaseType(), from)) {
-        lvalueType =  PolyType::ReplaceType(lvalueType, lvalueType);
-    }
-
     return this;
 }
 
@@ -8075,6 +8099,12 @@ SymbolExpr::ReplacePolyType(const PolyType *from, const Type *to) {
     if (!symbol)
         return NULL;
 
+    Symbol *tmp = m->symbolTable->LookupVariable(symbol->name.c_str());
+    if (tmp) {
+        tmp->parentFunction = symbol->parentFunction;
+        symbol = tmp;
+    }
+
     if (Type::EqualForReplacement(symbol->type->GetBaseType(), from)) {
         symbol->type = PolyType::ReplaceType(symbol->type, to);
     }
@@ -8151,6 +8181,14 @@ FunctionSymbolExpr::Optimize() {
     return this;
 }
 
+Expr *
+FunctionSymbolExpr::ReplacePolyType(const PolyType *from, const Type *to) {
+    // force re-evaluation of overloaded type
+    this->triedToResolve = false;
+
+    return this;
+}
+
 
 int
 FunctionSymbolExpr::EstimateCost() const {
@@ -8397,6 +8435,16 @@ FunctionSymbolExpr::computeOverloadCost(const FunctionType *ftype,
                 cost[i] += 8 * costScale;
                 continue;
             }
+            if (callTypeNC->IsPolymorphicType()) {
+                const PolyType *callTypeP =
+                    CastType<PolyType>(callTypeNC->GetBaseType());
+                if (callTypeP->CanBeType(fargTypeNC->GetBaseType()) &&
+                    callTypeNC->IsArrayType() == fargTypeNC->IsArrayType() &&
+                    callTypeNC->IsPointerType() == fargTypeNC->IsPointerType()){
+                    cost[i] += 8 * costScale;
+                    continue;
+                }
+            }
             if (fargType->IsVaryingType() && callType->IsUniformType()) {
                 // Here we deal with brodcasting uniform to varying.
                 // callType - varying and fargType - uniform is forbidden.
@@ -8523,6 +8571,12 @@ FunctionSymbolExpr::ResolveOverloads(SourcePos argPos,
         return true;
     }
     else if (matches.size() > 1) {
+        for (size_t i=0; i<argTypes.size(); i++) {
+            if (argTypes[i]->IsPolymorphicType()) {
+                matchingFunc = matches[0];
+                return true;
+            }
+        }
         // Multiple matches: ambiguous
         std::string candidateMessage =
             lGetOverloadCandidateMessage(matches, argTypes, argCouldBeNULL);

expr.h

@@ -96,6 +96,7 @@ public:
         encountered, NULL should be returned. */
     virtual Expr *TypeCheck() = 0;
 
+    Expr *Copy();
 
     /** This method replaces a polymorphic type with a specific atomic type */
     Expr *ReplacePolyType(const PolyType *polyType, const Type *replacement);
@@ -385,6 +386,51 @@ protected:
     mutable const Type *type, *lvalueType;
 };
 
+class StructMemberExpr : public MemberExpr
+{
+public:
+    StructMemberExpr(Expr *e, const char *id, SourcePos p,
+                     SourcePos idpos, bool derefLValue);
+
+    static inline bool classof(StructMemberExpr const*) { return true; }
+    static inline bool classof(ASTNode const* N) {
+        return N->getValueID() == StructMemberExprID;
+    }
+
+    const Type *GetType() const;
+    const Type *GetLValueType() const;
+    int getElementNumber() const;
+    const Type *getElementType() const;
+
+private:
+    const StructType *getStructType() const;
+};
+
+
+class VectorMemberExpr : public MemberExpr
+{
+public:
+    VectorMemberExpr(Expr *e, const char *id, SourcePos p,
+                     SourcePos idpos, bool derefLValue);
+
+    static inline bool classof(VectorMemberExpr const*) { return true; }
+    static inline bool classof(ASTNode const* N) {
+        return N->getValueID() == VectorMemberExprID;
+    }
+
+    llvm::Value *GetValue(FunctionEmitContext* ctx) const;
+    llvm::Value *GetLValue(FunctionEmitContext* ctx) const;
+    const Type *GetType() const;
+    const Type *GetLValueType() const;
+
+    int getElementNumber() const;
+    const Type *getElementType() const;
+
+private:
+    const VectorType *exprVectorType;
+    const VectorType *memberType;
+};
+
 
 /** @brief Expression representing a compile-time constant value.
 
@@ -715,6 +761,7 @@ public:
     Symbol *GetBaseSymbol() const;
     Expr *TypeCheck();
     Expr *Optimize();
+    Expr *ReplacePolyType(const PolyType *from, const Type *to);
     void Print() const;
     int EstimateCost() const;
     llvm::Constant *GetConstant(const Type *type) const;

func.cpp

@@ -90,7 +90,7 @@
 #endif
 #include <llvm/Support/ToolOutputFile.h>
 
-Function::Function(Symbol *s, Stmt *c) {
+Function::Function(Symbol *s, Stmt *c, bool typecheck) {
     sym = s;
     code = c;
 
@@ -98,13 +98,15 @@ Function::Function(Symbol *s, Stmt *c) {
     Assert(maskSymbol != NULL);
 
     if (code != NULL) {
-        code = TypeCheck(code);
+        if (typecheck) {
+            code = TypeCheck(code);
 
-        if (code != NULL && g->debugPrint) {
-            printf("After typechecking function \"%s\":\n",
-                    sym->name.c_str());
-            code->Print(0);
-            printf("---------------------\n");
+            if (code != NULL && g->debugPrint) {
+                printf("After typechecking function \"%s\":\n",
+                        sym->name.c_str());
+                code->Print(0);
+                printf("---------------------\n");
+            }
         }
 
         if (code != NULL) {
@@ -135,6 +137,7 @@ Function::Function(Symbol *s, Stmt *c) {
         args.push_back(sym);
 
         const Type *t = type->GetParameterType(i);
+
         if (sym != NULL && CastType<ReferenceType>(t) == NULL)
             sym->parentFunction = this;
     }
@@ -632,7 +635,7 @@ Function::GenerateIR() {
 const bool
 Function::IsPolyFunction() const {
     for (size_t i = 0; i < args.size(); i++) {
-        if (args[i]->type->IsPolymorphicType()) {
+        if (args[i] && args[i]->type->IsPolymorphicType()) {
             return true;
         }
     }
@@ -640,7 +643,6 @@ Function::IsPolyFunction() const {
     return false;
 }
 
-
 std::vector<Function *> *
 Function::ExpandPolyArguments(SymbolTable *symbolTable) const {
     Assert(symbolTable != NULL);
@@ -651,37 +653,64 @@ Function::ExpandPolyArguments(SymbolTable *symbolTable) const {
 
     const FunctionType *func = CastType<FunctionType>(sym->type);
 
-    if (g->debugPrint) {
-        printf("%s before replacing anything:\n", sym->name.c_str());
-        code->Print(0);
-    }
-
     for (size_t i=0; i<versions.size(); i++) {
+        if (g->debugPrint) {
+            printf("%s before replacing anything:\n", sym->name.c_str());
+            code->Print(0);
+        }
         const FunctionType *ft = CastType<FunctionType>(versions[i]->type);
-        Stmt *ncode = code;
+
+        symbolTable->PushScope();
+
+        Symbol *s = symbolTable->LookupFunction(versions[i]->name.c_str(), ft);
+        Stmt *ncode = (Stmt*)CopyAST(code);
+
+        Function *f = new Function(s, ncode, false);
+
+        for (size_t j=0; j<args.size(); j++) {
+            f->args[j] = new Symbol(*args[j]);
+            symbolTable->AddVariable(f->args[j], false);
+        }
 
         for (int j=0; j<ft->GetNumParameters(); j++) {
             if (func->GetParameterType(j)->IsPolymorphicType()) {
                 const PolyType *from = CastType<PolyType>(
                         func->GetParameterType(j)->GetBaseType());
 
-                ncode = (Stmt*)TranslatePoly(ncode, from,
+                f->code = (Stmt*)TranslatePoly(f->code, from,
                         ft->GetParameterType(j)->GetBaseType());
                 if (g->debugPrint) {
                     printf("%s after replacing %s with %s:\n\n",
                             sym->name.c_str(), from->GetString().c_str(),
                             ft->GetParameterType(j)->GetBaseType()->GetString().c_str());
 
-                    ncode->Print(0);
+                    f->code->Print(0);
 
                     printf("------------------------------------------\n\n");
                 }
             }
         }
 
-        Symbol *s = symbolTable->LookupFunction(versions[i]->name.c_str(), ft);
+        // we didn't typecheck before, now we can
+        f->code = TypeCheck(f->code);
 
-        expanded->push_back(new Function(s, ncode));
+        f->code = Optimize(f->code);
+
+        if (g->debugPrint) {
+            printf("After optimizing expanded function \"%s\":\n",
+                    f->sym->name.c_str());
+            f->code->Print(0);
+            printf("---------------------\n");
+        }
+
+
+
+        symbolTable->PopScope();
+
+
+
+
+        expanded->push_back(f);
     }
 
     return expanded;

func.h

@@ -44,7 +44,7 @@
 
 class Function {
 public:
-    Function(Symbol *sym, Stmt *code);
+    Function(Symbol *sym, Stmt *code, bool typecheck=true);
 
     const Type *GetReturnType() const;
     const FunctionType *GetType() const;

module.cpp

@@ -1061,8 +1061,6 @@ Module::AddFunctionDeclaration(const std::string &name,
 
                 const Type *ret = eft->GetReturnType();
                 if (Type::EqualForReplacement(ret, pt)) {
-                    printf("Replaced return type %s\n",
-                           ret->GetString().c_str());
                     ret = PolyType::ReplaceType(ret, *te);
                 }
 
@@ -1998,11 +1996,13 @@ lPrintPolyFunctionWrappers(FILE *file, const std::vector<std::string> &funcs) {
         for (size_t j=0; j<poly.size(); j++) {
             const FunctionType *ftype = CastType<FunctionType>(poly[j]->type);
             Assert(ftype);
-            std::string decl = ftype->GetCDeclaration(funcs[i]);
-            fprintf(file, "    %s {\n", decl.c_str());
+            if (ftype->isExported || ftype->isExternC) {
+                std::string decl = ftype->GetCDeclaration(funcs[i]);
+                fprintf(file, "    %s {\n", decl.c_str());
 
-            std::string call = ftype->GetCCall(poly[j]->name);
-            fprintf(file, "        return %s;\n    }\n", call.c_str());
+                std::string call = ftype->GetCCall(poly[j]->name);
+                fprintf(file, "        return %s;\n    }\n", call.c_str());
+            }
         }
     }
 

parse.yy

@@ -37,7 +37,7 @@
 /* one for 'if', one for 'cif' */
 %expect 2
 
-%error-verbose
+%define parse.error verbose
 
 %code requires {
 

stmt.cpp

@@ -35,6 +35,7 @@
     @brief File with definitions classes related to statements in the language
 */
 
+#include "ast.h"
 #include "stmt.h"
 #include "ctx.h"
 #include "util.h"
@@ -78,7 +79,81 @@ Stmt::Optimize() {
 }
 
 Stmt *
-Stmt::ReplacePolyType(const PolyType *polyType, const Type *replacement) {
+Stmt::Copy() {
+    Stmt *copy;
+    switch (getValueID()) {
+        case AssertStmtID:
+            copy = (Stmt*)new AssertStmt(*(AssertStmt*)this);
+            break;
+        case BreakStmtID:
+            copy = (Stmt*)new BreakStmt(*(BreakStmt*)this);
+            break;
+        case CaseStmtID:
+            copy = (Stmt*)new CaseStmt(*(CaseStmt*)this);
+            break;
+        case ContinueStmtID:
+            copy = (Stmt*)new ContinueStmt(*(ContinueStmt*)this);
+            break;
+        case DeclStmtID:
+            copy = (Stmt*)new DeclStmt(*(DeclStmt*)this);
+            break;
+        case DefaultStmtID:
+            copy = (Stmt*)new DefaultStmt(*(DefaultStmt*)this);
+            break;
+        case DeleteStmtID:
+            copy = (Stmt*)new DeleteStmt(*(DeleteStmt*)this);
+            break;
+        case DoStmtID:
+            copy = (Stmt*)new DoStmt(*(DoStmt*)this);
+            break;
+        case ExprStmtID:
+            copy = (Stmt*)new ExprStmt(*(ExprStmt*)this);
+            break;
+        case ForeachActiveStmtID:
+            copy = (Stmt*)new ForeachActiveStmt(*(ForeachActiveStmt*)this);
+            break;
+        case ForeachStmtID:
+            copy = (Stmt*)new ForeachStmt(*(ForeachStmt*)this);
+            break;
+        case ForeachUniqueStmtID:
+            copy = (Stmt*)new ForeachUniqueStmt(*(ForeachUniqueStmt*)this);
+            break;
+        case ForStmtID:
+            copy = (Stmt*)new ForStmt(*(ForStmt*)this);
+            break;
+        case GotoStmtID:
+            copy = (Stmt*)new GotoStmt(*(GotoStmt*)this);
+            break;
+        case IfStmtID:
+            copy = (Stmt*)new IfStmt(*(IfStmt*)this);
+            break;
+        case LabeledStmtID:
+            copy = (Stmt*)new LabeledStmt(*(LabeledStmt*)this);
+            break;
+        case PrintStmtID:
+            copy = (Stmt*)new PrintStmt(*(PrintStmt*)this);
+            break;
+        case ReturnStmtID:
+            copy = (Stmt*)new ReturnStmt(*(ReturnStmt*)this);
+            break;
+        case StmtListID:
+            copy = (Stmt*)new StmtList(*(StmtList*)this);
+            break;
+        case SwitchStmtID:
+            copy = (Stmt*)new SwitchStmt(*(SwitchStmt*)this);
+            break;
+        case UnmaskedStmtID:
+            copy = (Stmt*)new UnmaskedStmt(*(UnmaskedStmt*)this);
+            break;
+        default:
+            FATAL("Unmatched case in Stmt::Copy");
+            copy = this; // just to silence the compiler
+    }
+    return copy;
+}
+
+Stmt *
+Stmt::ReplacePolyType(const PolyType *, const Type *) {
     return this;
 }
 
@@ -484,7 +559,8 @@ DeclStmt::TypeCheck() {
         // an int as the constValue later...
         const Type *type = vars[i].sym->type;
         if (CastType<AtomicType>(type) != NULL ||
-            CastType<EnumType>(type) != NULL) {
+            CastType<EnumType>(type) != NULL ||
+            CastType<PolyType>(type) != NULL) {
             // If it's an expr list with an atomic type, we'll later issue
             // an error.  Need to leave vars[i].init as is in that case so
             // it is in fact caught later, though.
@@ -502,9 +578,15 @@ DeclStmt::TypeCheck() {
 Stmt *
 DeclStmt::ReplacePolyType(const PolyType *from, const Type *to) {
     for (size_t i = 0; i < vars.size(); i++) {
+        vars[i].sym = new Symbol(*vars[i].sym);
+        m->symbolTable->AddVariable(vars[i].sym, false);
         Symbol *s = vars[i].sym;
         if (Type::EqualForReplacement(s->type->GetBaseType(), from)) {
             s->type = PolyType::ReplaceType(s->type, to);
+
+            // this typecast *should* be valid after typechecking
+            vars[i].init = TypeConvertExpr(vars[i].init, s->type,
+                                           "initializer");
         }
     }
 
@@ -1487,6 +1569,17 @@ ForeachStmt::ForeachStmt(const std::vector<Symbol *> &lvs,
       stmts(s) {
 }
 
+/*
+ForeachStmt::ForeachStmt(ForeachStmt *base)
+    : Stmt(base->pos, ForeachStmtID) {
+    dimVariables = base->dimVariables;
+    startExprs = base->startExprs;
+    endExprs = base->endExprs;
+    isTiled = base->isTiled;
+    stmts = base->stmts;
+}
+*/
+
 
 /* Given a uniform counter value in the memory location pointed to by
    uniformCounterPtr, compute the corresponding set of varying counter
@@ -1730,8 +1823,10 @@ ForeachStmt::EmitCode(FunctionEmitContext *ctx) const {
         // Start and end value for this loop dimension
         llvm::Value *sv = startExprs[i]->GetValue(ctx);
         llvm::Value *ev = endExprs[i]->GetValue(ctx);
-        if (sv == NULL || ev == NULL)
+        if (sv == NULL || ev == NULL) {
+            fprintf(stderr, "ev is NULL again :(\n");
             return;
+        }
         startVals.push_back(sv);
         endVals.push_back(ev);
 
@@ -2191,21 +2286,6 @@ ForeachStmt::TypeCheck() {
     return anyErrors ? NULL : this;
 }
 
-Stmt *
-ForeachStmt::ReplacePolyType(const PolyType *from, const Type *to) {
-    if (!stmts)
-        return NULL;
-
-    for (size_t i=0; i<dimVariables.size(); i++) {
-        const Type *t = dimVariables[i]->type;
-        if (Type::EqualForReplacement(t->GetBaseType(), from)) {
-            t = PolyType::ReplaceType(t, to);
-        }
-    }
-
-    return this;
-}
-
 
 int
 ForeachStmt::EstimateCost() const {

stmt.h

@@ -70,6 +70,7 @@ public:
     // Stmts don't have anything to do here.
     virtual Stmt *Optimize();
     virtual Stmt *TypeCheck() = 0;
+    Stmt *Copy();
     Stmt *ReplacePolyType(const PolyType *polyType, const Type *replacement);
 };
 
@@ -283,7 +284,6 @@ public:
     void Print(int indent) const;
 
     Stmt *TypeCheck();
-    Stmt *ReplacePolyType(const PolyType *from, const Type *to);
     int EstimateCost() const;
 
     std::vector<Symbol *> dimVariables;

sym.cpp

@@ -95,14 +95,14 @@ SymbolTable::PopScope() {
 
 
 bool
-SymbolTable::AddVariable(Symbol *symbol) {
+SymbolTable::AddVariable(Symbol *symbol, bool issueScopeWarning) {
     Assert(symbol != NULL);
 
     // Check to see if a symbol of the same name has already been declared.
     for (int i = (int)variables.size() - 1; i >= 0; --i) {
         SymbolMapType &sm = *(variables[i]);
         if (sm.find(symbol->name) != sm.end()) {
-            if (i == (int)variables.size()-1) {
+            if (i == (int)variables.size()-1 && issueScopeWarning) {
                 // If a symbol of the same name was declared in the
                 // same scope, it's an error.
                 Error(symbol->pos, "Ignoring redeclaration of symbol \"%s\".",
@@ -112,9 +112,11 @@ SymbolTable::AddVariable(Symbol *symbol) {
             else {
                 // Otherwise it's just shadowing something else, which
                 // is legal but dangerous..
-                Warning(symbol->pos,
-                        "Symbol \"%s\" shadows symbol declared in outer scope.",
-                        symbol->name.c_str());
+                if (issueScopeWarning) {
+                    Warning(symbol->pos,
+                            "Symbol \"%s\" shadows symbol declared in outer scope.",
+                            symbol->name.c_str());
+                }
                 (*variables.back())[symbol->name] = symbol;
                 return true;
             }

sym.h

@@ -141,7 +141,7 @@ public:
         with a symbol defined at the same scope.  (Symbols may shaodow
         symbols in outer scopes; a warning is issued in this case, but this
         method still returns true.) */
-    bool AddVariable(Symbol *symbol);
+    bool AddVariable(Symbol *symbol, bool issueScopeWarning=true);
 
     /** Looks for a variable with the given name in the symbol table.  This
         method searches outward from the innermost scope to the outermost,

tests_ispcpp/CycleTimer.h

@@ -0,0 +1,177 @@
+#ifndef _SYRAH_CYCLE_TIMER_H_
+#define _SYRAH_CYCLE_TIMER_H_
+
+#if defined(__APPLE__)
+  #if defined(__x86_64__)
+    #include <sys/sysctl.h>
+  #else
+    #include <mach/mach.h>
+    #include <mach/mach_time.h>
+  #endif // __x86_64__ or not
+
+  #include <stdio.h>  // fprintf
+  #include <stdlib.h> // exit
+
+#elif _WIN32
+#  include <windows.h>
+#  include <time.h>
+#else
+#  include <stdio.h>
+#  include <stdlib.h>
+#  include <string.h>
+#  include <sys/time.h>
+#endif
+
+
+  // This uses the cycle counter of the processor.  Different
+  // processors in the system will have different values for this.  If
+  // you process moves across processors, then the delta time you
+  // measure will likely be incorrect.  This is mostly for fine
+  // grained measurements where the process is likely to be on the
+  // same processor.  For more global things you should use the
+  // Time interface.
+
+  // Also note that if you processors' speeds change (i.e. processors
+  // scaling) or if you are in a heterogenous environment, you will
+  // likely get spurious results.
+  class CycleTimer {
+  public:
+    typedef unsigned long long SysClock;
+
+    //////////
+    // Return the current CPU time, in terms of clock ticks.
+    // Time zero is at some arbitrary point in the past.
+    static SysClock currentTicks() {
+#if defined(__APPLE__) && !defined(__x86_64__)
+      return mach_absolute_time();
+#elif defined(_WIN32)
+      LARGE_INTEGER qwTime;
+      QueryPerformanceCounter(&qwTime);
+      return qwTime.QuadPart;
+#elif defined(__x86_64__)
+      unsigned int a, d;
+      asm volatile("rdtsc" : "=a" (a), "=d" (d));
+      return static_cast<unsigned long long>(a) |
+        (static_cast<unsigned long long>(d) << 32);
+#elif defined(__ARM_NEON__) && 0 // mrc requires superuser.
+      unsigned int val;
+      asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(val));
+      return val;
+#else
+      timespec spec;
+      clock_gettime(CLOCK_THREAD_CPUTIME_ID, &spec);
+      return CycleTimer::SysClock(static_cast<float>(spec.tv_sec) * 1e9 + static_cast<float>(spec.tv_nsec));
+#endif
+    }
+
+    //////////
+    // Return the current CPU time, in terms of seconds.
+    // This is slower than currentTicks().  Time zero is at
+    // some arbitrary point in the past.
+    static double currentSeconds() {
+      return currentTicks() * secondsPerTick();
+    }
+
+    //////////
+    // Return the conversion from seconds to ticks.
+    static double ticksPerSecond() {
+      return 1.0/secondsPerTick();
+    }
+
+    static const char* tickUnits() {
+#if defined(__APPLE__) && !defined(__x86_64__)
+      return "ns";
+#elif defined(__WIN32__) || defined(__x86_64__)
+      return "cycles";
+#else
+      return "ns"; // clock_gettime
+#endif
+    }
+
+    //////////
+    // Return the conversion from ticks to seconds.
+    static double secondsPerTick() {
+      static bool initialized = false;
+      static double secondsPerTick_val;
+      if (initialized) return secondsPerTick_val;
+#if defined(__APPLE__)
+  #ifdef __x86_64__
+      int args[] = {CTL_HW, HW_CPU_FREQ};
+      unsigned int Hz;
+      size_t len = sizeof(Hz);
+      if (sysctl(args, 2, &Hz, &len, NULL, 0) != 0) {
+         fprintf(stderr, "Failed to initialize secondsPerTick_val!\n");
+         exit(-1);
+      }
+      secondsPerTick_val = 1.0 / (double) Hz;
+  #else
+      mach_timebase_info_data_t time_info;
+      mach_timebase_info(&time_info);
+
+      // Scales to nanoseconds without 1e-9f
+      secondsPerTick_val = (1e-9*static_cast<double>(time_info.numer))/
+        static_cast<double>(time_info.denom);
+  #endif // x86_64 or not
+#elif defined(_WIN32)
+      LARGE_INTEGER qwTicksPerSec;
+      QueryPerformanceFrequency(&qwTicksPerSec);
+      secondsPerTick_val = 1.0/static_cast<double>(qwTicksPerSec.QuadPart);
+#else
+      FILE *fp = fopen("/proc/cpuinfo","r");
+      char input[1024];
+      if (!fp) {
+         fprintf(stderr, "CycleTimer::resetScale failed: couldn't find /proc/cpuinfo.");
+         exit(-1);
+      }
+      // In case we don't find it, e.g. on the N900
+      secondsPerTick_val = 1e-9;
+      while (!feof(fp) && fgets(input, 1024, fp)) {
+        // NOTE(boulos): Because reading cpuinfo depends on dynamic
+        // frequency scaling it's better to read the @ sign first
+        float GHz, MHz;
+        if (strstr(input, "model name")) {
+          char* at_sign = strstr(input, "@");
+          if (at_sign) {
+            char* after_at = at_sign + 1;
+            char* GHz_str = strstr(after_at, "GHz");
+            char* MHz_str = strstr(after_at, "MHz");
+            if (GHz_str) {
+              *GHz_str = '\0';
+              if (1 == sscanf(after_at, "%f", &GHz)) {
+                //printf("GHz = %f\n", GHz);
+                secondsPerTick_val = 1e-9f / GHz;
+                break;
+              }
+            } else if (MHz_str) {
+              *MHz_str = '\0';
+              if (1 == sscanf(after_at, "%f", &MHz)) {
+                //printf("MHz = %f\n", MHz);
+                secondsPerTick_val = 1e-6f / GHz;
+                break;
+              }
+            }
+          }
+        } else if (1 == sscanf(input, "cpu MHz : %f", &MHz)) {
+          //printf("MHz = %f\n", MHz);
+          secondsPerTick_val = 1e-6f / MHz;
+          break;
+        }
+      }
+      fclose(fp);
+#endif
+
+      initialized = true;
+      return secondsPerTick_val;
+    }
+
+    //////////
+    // Return the conversion from ticks to milliseconds.
+    static double msPerTick() {
+      return secondsPerTick() * 1000.0;
+    }
+
+  private:
+    CycleTimer();
+  };
+
+#endif // #ifndef _SYRAH_CYCLE_TIMER_H_

tests_ispcpp/Makefile

@@ -1,12 +1,15 @@
 CXX=g++
-CXXFLAGS=-std=c++11
+CXXFLAGS=-std=c++11 -O3 -lm -lpthread
 
 ISPC=../ispc
-ISPCFLAGS=--target=sse4-x2 -O2 --arch=x86-64
+ISPCFLAGS=--target=avx2 -O3 --arch=x86-64
 
-%.out : %.cpp %.o
+%.out : %.cpp %.o tasksys.o
 	$(CXX) $(CXXFLAGS) -o $@ $^
 
+tasksys.o : ../examples/tasksys.cpp
+	$(CXX) $(CXXFLAGS) -c -o $@ $^
+
 $ : $.o
 
 %.o : %.ispc

tests_ispcpp/array.ispc

@@ -0,0 +1,13 @@
+export void array(uniform int N, uniform integer * uniform X) {
+    integer *A = new integer[N/2];
+
+    foreach (i = 0 ... N/2) {
+        A[i] = X[i] + X[N/2 + i];
+    }
+
+    foreach (i = 0 ... N) {
+        X[i] = A[i/2];
+    }
+
+    delete[] A;
+}

tests_ispcpp/error_5.ispc

@@ -0,0 +1,17 @@
+//@error
+
+floating$0 mult(floating$0 x, floating$1 y) {
+    return x * y;
+}
+
+export void saxpy(uniform int N,
+                  uniform floating$0 scale,
+                  uniform floating$1 X[],
+                  uniform floating$1 Y[],
+                  uniform floating$2 result[])
+{
+    foreach (i = 0 ... N) {
+        floating$ tmp = mult(scale, X[i]) + Y[i];
+        result[i] = tmp;
+    }
+}

tests_ispcpp/error_6.ispc

@@ -0,0 +1,14 @@
+number pow(number b, int a) {
+    number out = b;
+    for (int i = 1; i<a; i++) {
+        out *= b;
+    }
+
+    return out;
+}
+
+export void square(uniform int N, uniform number$-1 b[], uniform number$-1 out[]) {
+    foreach (i = 0 ... N) {
+        out[i] = pow(b[i], 2);
+    }
+}

tests_ispcpp/error_7.ispc

@@ -0,0 +1,13 @@
+floating foo(floating a, floating b) {
+    floating d = a / b;
+    if (d < 0.)
+        return 0.;
+
+    return d;
+}
+
+export void bar(uniform integer * uniform X, uniform int N) {
+    foreach (i = 0 ... N-1) {
+        X[i] = foo(X[i], X[i+1]);
+    }
+}

tests_ispcpp/function.ispc

@@ -1,15 +1,15 @@
 floating saxpy_helper(floating scale,
-                      floating<0> x,
-                      floating<0> y) {
+                      floating$3 x,
+                      floating$3 y) {
     return scale * x + y;
 }
 
 
 export void saxpy(uniform int N,
-                  uniform floating<0> scale,
-                  uniform floating<1> X[],
-                  uniform floating<1> Y[],
-                  uniform floating<2> result[])
+                  uniform floating$0 scale,
+                  uniform floating$1 X[],
+                  uniform floating$1 Y[],
+                  uniform floating$2 result[])
 {
     foreach (i = 0 ... N) {
         result[i] = saxpy_helper(scale, X[i], Y[i]);

tests_ispcpp/hello.cpp

@@ -1,7 +1,7 @@
 #include <stdlib.h>
 #include <stdio.h>
 
-#include "hello.ispc.h"
+#include "hello.h"
 
 int main() {
     float A[100];

tests_ispcpp/sqrt.cpp

@@ -0,0 +1,212 @@
+#include <stdio.h>
+#include <algorithm>
+#include <pthread.h>
+#include <math.h>
+#include <cmath>
+
+#include "CycleTimer.h"
+#include "sqrt.h"
+
+using namespace ispc;
+
+void sqrtSerial(int N,
+                float initialGuess,
+                float values[],
+                float output[])
+{
+
+    static const float kThreshold = 0.00001f;
+
+    for (int i=0; i<N; i++) {
+
+        float x = values[i];
+        float guess = initialGuess;
+
+        float error = fabs(guess * guess * x - 1.f);
+
+        while (error > kThreshold) {
+            guess = (3.f * guess - x * guess * guess * guess) * 0.5f;
+            error = fabs(guess * guess * x - 1.f);
+        }
+
+        output[i] = x * guess;
+    }
+}
+
+void sqrtSerial(int N,
+                double initialGuess,
+                double values[],
+                double output[])
+{
+
+    static const double kThreshold = 0.00001f;
+
+    for (int i=0; i<N; i++) {
+
+        double x = values[i];
+        double guess = initialGuess;
+
+        double error = std::abs(guess * guess * x - 1.);
+
+        while (error > kThreshold) {
+            guess = (3. * guess - x * guess * guess * guess) * 0.5;
+            error = std::abs(guess * guess * x - 1.);
+        }
+
+        output[i] = x * guess;
+    }
+}
+
+static void verifyResult(int N, float* result, float* gold) {
+    for (int i=0; i<N; i++) {
+        if (fabs(result[i] - gold[i]) > 1e-4) {
+            printf("Error: [%d] Got %f expected %f\n", i, result[i], gold[i]);
+        }
+    }
+}
+
+static void verifyResult(int N, double* result, double* gold) {
+    for (int i=0; i<N; i++) {
+        if (std::abs(result[i] - gold[i]) > 1e-4) {
+            printf("Error: [%d] Got %f expected %f\n", i, result[i], gold[i]);
+        }
+    }
+}
+
+int main() {
+
+    const unsigned int N = 20 * 1000 * 1000;
+    const float initialGuess = 1.0f;
+    const double dinitialGuess = 1.0;
+
+    float* values = new float[N];
+    float* output = new float[N];
+    float* gold = new float[N];
+
+    double* dvalues = new double[N];
+    double* doutput = new double[N];
+    double* dgold = new double[N];
+
+    for (unsigned int i=0; i<N; i++)
+    {
+        // random input values
+        values[i] = .001f + 2.998f * static_cast<float>(rand()) / RAND_MAX;
+        dvalues[i] = .001 + 2.998 * static_cast<double>(rand()) / RAND_MAX;
+        output[i] = 0.f;
+        doutput[i] = 0.;
+    }
+
+    // generate a gold version to check results
+    for (unsigned int i=0; i<N; i++) {
+        gold[i] = sqrt(values[i]);
+        dgold[i] = sqrt(dvalues[i]);
+    }
+
+    //
+    // And run the serial implementation 3 times, again reporting the
+    // minimum time.
+    //
+    double minSerial = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        sqrtSerial(N, initialGuess, values, output);
+        double endTime = CycleTimer::currentSeconds();
+        minSerial = std::min(minSerial, endTime - startTime);
+    }
+
+    printf("[sqrt float serial]:\t\t[%.3f] ms\n", minSerial * 1000);
+    verifyResult(N, output, gold);
+
+    double minDSerial = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        sqrtSerial(N, dinitialGuess, dvalues, doutput);
+        double endTime = CycleTimer::currentSeconds();
+        minDSerial = std::min(minDSerial, endTime - startTime);
+    }
+
+    printf("[sqrt double serial]:\t\t[%.3f] ms\n", minDSerial * 1000);
+    verifyResult(N, doutput, dgold);
+
+    // Clear out the buffer
+    for (unsigned int i = 0; i < N; ++i) {
+        output[i] = 0;
+        doutput[i] = 0;
+    }
+
+
+    //
+    // Compute the image using the ispc implementation; report the minimum
+    // time of three runs.
+    //
+    double minISPC = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        ispc::sqrt_ispc(N, initialGuess, values, output);
+        double endTime = CycleTimer::currentSeconds();
+        minISPC = std::min(minISPC, endTime - startTime);
+    }
+
+    printf("[sqrt float ispc]:\t\t[%.3f] ms\n", minISPC * 1000);
+
+    verifyResult(N, output, gold);
+
+    double minDISPC = 1e30;
+    for (int i = 0; i < 5; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        ispc::sqrt_ispc(N, dinitialGuess, dvalues, doutput);
+        double endTime = CycleTimer::currentSeconds();
+        minDISPC = std::min(minDISPC, endTime - startTime);
+    }
+
+    printf("[sqrt double ispc]:\t\t[%.3f] ms\n", minDISPC * 1000);
+
+    verifyResult(N, doutput, dgold);
+
+    // Clear out the buffer
+    for (unsigned int i = 0; i < N; ++i) {
+        output[i] = 0;
+        doutput[i] = 0;
+    }
+
+    //
+    // Tasking version of the ISPC code
+    //
+    double minTaskISPC = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        ispc::sqrt_ispc_withtasks(N, initialGuess, values, output);
+        double endTime = CycleTimer::currentSeconds();
+        minTaskISPC = std::min(minTaskISPC, endTime - startTime);
+    }
+
+    printf("[sqrt float task ispc]:\t\t[%.3f] ms\n", minTaskISPC * 1000);
+
+    verifyResult(N, output, gold);
+
+    double minDTaskISPC = 1e30;
+    for (int i = 0; i < 3; ++i) {
+        double startTime = CycleTimer::currentSeconds();
+        ispc::sqrt_ispc_withtasks(N, dinitialGuess, dvalues, doutput);
+        double endTime = CycleTimer::currentSeconds();
+        minDTaskISPC = std::min(minDTaskISPC, endTime - startTime);
+    }
+
+    printf("[sqrt double task ispc]:\t[%.3f] ms\n", minDTaskISPC * 1000);
+
+    verifyResult(N, output, gold);
+
+    printf("\t\t\t\t(%.2fx speedup from ISPC float)\n", minSerial/minISPC);
+    printf("\t\t\t\t(%.2fx speedup from ISPC double)\n", minDSerial/minDISPC);
+    printf("\t\t\t\t(%.2fx speedup from task ISPC float)\n", minSerial/minTaskISPC);
+    printf("\t\t\t\t(%.2fx speedup from task ISPC double)\n", minDSerial/minDTaskISPC);
+
+    delete[] values;
+    delete[] output;
+    delete[] gold;
+    delete[] dvalues;
+    delete[] doutput;
+    delete[] dgold;
+
+    return 0;
+}

tests_ispcpp/sqrt.ispc

@@ -0,0 +1,62 @@
+
+static const float kThreshold = 0.00001f;
+
+export void sqrt_ispc(uniform int N,
+                      uniform floating initialGuess,
+                      uniform floating values[],
+                      uniform floating output[])
+{
+    foreach (i = 0 ... N) {
+
+        floating x = values[i];
+        floating guess = initialGuess;
+
+        floating pred = abs(guess * guess * x - 1.f);
+
+        while (pred > kThreshold) {
+            guess = (3.f * guess - x * guess * guess * guess) * 0.5f;
+            pred = abs(guess * guess * x - 1.f);
+        }
+
+        output[i] = x * guess;
+
+    }
+}
+
+task void sqrt_ispc_task(uniform int N,
+                         uniform int span,
+                         uniform floating initialGuess,
+                         uniform floating values[],
+                         uniform floating output[])
+{
+
+    uniform int indexStart = taskIndex * span;
+    uniform int indexEnd = min(N, indexStart + span);
+
+    foreach (i = indexStart ... indexEnd) {
+
+        floating x = values[i];
+        floating guess = initialGuess;
+
+        floating pred = abs(guess * guess * x - 1.f);
+
+        while (pred > kThreshold) {
+            guess = (3.f * guess - x * guess * guess * guess) * 0.5f;
+            pred = abs(guess * guess * x - 1.f);
+        }
+
+        output[i] = x * guess;
+
+    }
+}
+
+export void sqrt_ispc_withtasks(uniform int N,
+                                uniform floating initialGuess,
+                                uniform floating values[],
+                                uniform floating output[])
+{
+
+    uniform int span = N / 64;  // 64 tasks
+
+    launch[N/span] sqrt_ispc_task(N, span, initialGuess, values, output);
+}

tests_ispcpp/varying.cpp

@@ -0,0 +1,27 @@
+#include <stdlib.h>
+#include <stdio.h>
+
+#include "varying.h"
+
+int main() {
+    float A[256];
+    double B[256];
+    float outA[256];
+    double outB[256];
+
+
+    for (int i=0; i<256; i++) {
+        A[i] = 1. / (i+1);
+        B[i] = 1. / (i+1);
+    }
+
+    ispc::square(256, (float*)&A, (float*)&outA);
+
+    ispc::square(256, (double*)&B, (double*)&outB);
+
+    for (int i=0; i<256; i++) {
+        printf("float: %.16f\tdouble: %.16f\n", outA[i], outB[i]);
+    }
+
+    return 0;
+}

tests_ispcpp/varying.ispc

@@ -0,0 +1,14 @@
+number pow(number b, int a) {
+    number out = b;
+    for (int i = 1; i<a; i++) {
+        out *= b;
+    }
+
+    return out;
+}
+
+export void square(uniform int N, uniform number b[], uniform number out[]) {
+    foreach (i = 0 ... N) {
+        out[i] = pow(b[i], 2);
+    }
+}

type.cpp

@@ -721,9 +721,11 @@ PolyType::ReplaceType(const Type *from, const Type *to) {
     if (from->IsVaryingType())
         t = t->GetAsVaryingType();
 
-    fprintf(stderr, "Replacing type \"%s\" with \"%s\"\n",
-            from->GetString().c_str(),
-            t->GetString().c_str());
+    if (g->debugPrint) {
+        fprintf(stderr, "Replacing type \"%s\" with \"%s\"\n",
+                from->GetString().c_str(),
+                t->GetString().c_str());
+    }
 
     return t;
 }