C ABI: Do not pass empty structs as parameters at all

This is most visible on x86 (32-bit), where the stack
alignment is off otherwise.

This change is quite messy because many places assumed
that there was always exactly one LLVM parameter per
TypeFunction::parameters entry.

gen/functions.cpp

@@ -164,6 +164,17 @@ llvm::FunctionType* DtoFunctionType(Type* type, IrFuncTy &irFty, Type* thistype,
         }
         else if (!passPointer)
         {
+            if (loweredDType->toBasetype()->ty == Tstruct)
+            {
+                // Do not pass empty structs at all for C++ ABI compatibility.
+                // Tests with clang reveal that more complex "empty" types, for
+                // example a struct containing an empty struct, are not
+                // optimized in the same way.
+                StructDeclaration *sd =
+                    static_cast<TypeStruct*>(loweredDType->toBasetype())->sym;
+                if (sd->fields.empty()) continue;
+            }
+
             if (abi->passByVal(loweredDType))
             {
                 attrBuilder.add(LDC_ATTRIBUTE(ByVal));
@@ -177,6 +188,7 @@ llvm::FunctionType* DtoFunctionType(Type* type, IrFuncTy &irFty, Type* thistype,
             }
         }
         newIrFty.args.push_back(new IrFuncTyArg(loweredDType, passPointer, attrBuilder));
+        newIrFty.args.back()->parametersIdx = i;
         ++nextLLArgIdx;
     }
 
@@ -411,13 +423,11 @@ static void set_param_attrs(TypeFunction* f, llvm::Function* func, FuncDeclarati
 
     #undef ADD_PA
 
-    // set attrs on the rest of the arguments
-    size_t n = Parameter::dim(f->parameters);
+    // Set attributes on the explicit parameters.
+    const size_t n = irFty.args.size();
     for (size_t k = 0; k < n; k++)
     {
-        assert(Parameter::getNth(f->parameters, k));
-
-        unsigned i = idx + (irFty.reverseParams ? n-k-1 : k);
+        const size_t i = idx + (irFty.reverseParams ? (n - k - 1) : k);
         newAttrs.add(i, irFty.args[k]->attrs);
     }
 
@@ -576,32 +586,28 @@ void DtoDeclareFunction(FuncDeclaration* fdecl)
         ++iarg;
     }
 
-    // we never reference parameters of function prototypes
     unsigned int k = 0;
     for (; iarg != func->arg_end(); ++iarg)
     {
-        if (fdecl->parameters && fdecl->parameters->dim > k)
-        {
-            int paramIndex = irFty.reverseParams ? fdecl->parameters->dim-k-1 : k;
-            Dsymbol* argsym = static_cast<Dsymbol*>(fdecl->parameters->data[paramIndex]);
+        size_t llExplicitIdx = irFty.reverseParams ? irFty.args.size() - k - 1 : k;
+        ++k;
+        IrFuncTyArg *arg = irFty.args[llExplicitIdx];
 
-            VarDeclaration* argvd = argsym->isVarDeclaration();
-            assert(argvd);
-            assert(!isIrLocalCreated(argvd));
-            std::string str(argvd->ident->toChars());
-            str.append("_arg");
-            iarg->setName(str);
-
-            IrParameter *irParam = getIrParameter(argvd, true);
-            irParam->value = iarg;
-            irParam->arg = irFty.args[paramIndex];
-
-            k++;
-        }
-        else
+        if (!fdecl->parameters || arg->parametersIdx >= fdecl->parameters->dim)
         {
             iarg->setName("unnamed");
+            continue;
         }
+
+        Dsymbol* const argsym = (*fdecl->parameters)[arg->parametersIdx];
+        VarDeclaration* argvd = argsym->isVarDeclaration();
+        assert(argvd);
+
+        iarg->setName(argvd->ident->toChars() + llvm::Twine("_arg"));
+
+        IrParameter *irParam = getIrParameter(argvd, true);
+        irParam->arg = arg;
+        irParam->value = iarg;
     }
 }
 
@@ -842,24 +848,34 @@ void DtoDefineFunction(FuncDeclaration* fd)
         irFunc->nestArg = val;
     }
 
-    // give arguments storage
-    // and debug info
+    // give arguments storage and debug info
     if (fd->parameters)
     {
-        size_t n = irFty.args.size();
-        assert(n == fd->parameters->dim);
-        for (size_t i=0; i < n; ++i)
+        // Not all arguments are necessarily passed on the LLVM level
+        // (e.g. zero-member structs), so we need to keep track of the
+        // index in the IrFuncTy args array separately.
+        size_t llArgIdx = 0;
+        for (size_t i = 0; i < fd->parameters->dim; ++i)
         {
-            Dsymbol* argsym = static_cast<Dsymbol*>(fd->parameters->data[i]);
-            VarDeclaration* vd = argsym->isVarDeclaration();
+            Dsymbol* const argsym = (*fd->parameters)[i];
+            VarDeclaration* const vd = argsym->isVarDeclaration();
             assert(vd);
+            const bool refout = vd->storage_class & (STCref | STCout);
 
             IrParameter* irparam = getIrParameter(vd);
-            assert(irparam);
-
-            bool refout = vd->storage_class & (STCref | STCout);
-            bool lazy = vd->storage_class & STClazy;
-            bool firstClassVal = !refout && (!irparam->arg->byref || lazy);
+            Type* debugInfoType = vd->type;
+            if (!irparam)
+            {
+                // This is a parameter that is not passed on the LLVM level.
+                // Create the param here and set it to a "dummy" alloca that
+                // we do not store to here.
+                irparam = getIrParameter(vd, true);
+                irparam->value = DtoAlloca(vd->type, vd->ident->toChars());
+            }
+            else
+            {
+                const bool lazy = vd->storage_class & STClazy;
+                const bool firstClassVal = !refout && (!irparam->arg->byref || lazy);
                 if (firstClassVal)
                 {
                     // alloca a stack slot for this first class value arg
@@ -867,14 +883,18 @@ void DtoDefineFunction(FuncDeclaration* fd)
 
                     // let the abi transform the argument back first
                     DImValue arg_dval(vd->type, irparam->value);
-                irFty.getParam(vd->type, i, &arg_dval, mem);
+                    irFty.getParam(vd->type, llArgIdx, &arg_dval, mem);
 
                     // set the arg var value to the alloca
                     irparam->value = mem;
+
+                    debugInfoType = irparam->arg->type;
+                }
+                ++llArgIdx;
             }
 
             if (global.params.symdebug && !(isaArgument(irparam->value) && isaArgument(irparam->value)->hasByValAttr()) && !refout)
-                gIR->DBuilder.EmitLocalVariable(irparam->value, vd, firstClassVal ? irparam->arg->type : 0);
+                gIR->DBuilder.EmitLocalVariable(irparam->value, vd, debugInfoType);
         }
     }
 

gen/nested.cpp

@@ -158,7 +158,7 @@ DValue* DtoNestedVariable(Loc& loc, Type* astype, VarDeclaration* vd, bool byref
         Logger::cout() << "Addr: " << *val << '\n';
         Logger::cout() << "of type: " << *val->getType() << '\n';
     }
-    if (byref || (vd->isParameter() && getIrParameter(vd)->arg->byref)) {
+    if (byref || (vd->isParameter() && getIrParameter(vd)->arg && getIrParameter(vd)->arg->byref)) {
         val = DtoAlignedLoad(val);
         //dwarfOpDeref(dwarfAddr);
         IF_LOG {
@@ -364,9 +364,11 @@ static void DtoCreateNestedContextType(FuncDeclaration* fd) {
             irLocal->nestedIndex = types.size();
             irLocal->nestedDepth = depth;
 
-            if (vd->isParameter()) {
-                // Parameters will have storage associated with them (to handle byref etc.),
-                // so handle those cases specially by storing a pointer instead of a value.
+            if (vd->isParameter() && getIrParameter(vd)->arg) {
+                // Parameters that are part of the LLVM signature will have
+                // storage associated with them (to handle byref etc.), so
+                // handle those cases specially by storing a pointer instead
+                // of a value.
                 const IrParameter* irparam = getIrParameter(vd);
                 const bool refout = vd->storage_class & (STCref | STCout);
                 const bool lazy = vd->storage_class & STClazy;
@@ -479,7 +481,7 @@ void DtoCreateNestedContext(FuncDeclaration* fd) {
                 LOG_SCOPE
                 IrParameter* parm = getIrParameter(vd);
 
-                if (parm->arg->byref)
+                if (parm->arg && parm->arg->byref)
                 {
                     storeVariable(vd, gep);
                 }

gen/tocall.cpp

@@ -117,14 +117,20 @@ LLFunctionType* DtoExtractFunctionType(LLType* type)
 static void addExplicitArguments(std::vector<LLValue*>& args, AttrSet& attrs,
     IrFuncTy& irFty, LLFunctionType* callableTy, const std::vector<DValue*>& argvals, int numFormalParams)
 {
-    const int numImplicitArgs = args.size();
-    const int numExplicitArgs = argvals.size();
+    // Number of arguments added to the LLVM type that are implicit on the
+    // frontend side of things (this, context pointers, etc.)
+    const size_t implicitLLArgCount = args.size();
 
-    args.resize(numImplicitArgs + numExplicitArgs, static_cast<LLValue*>(0));
+    // Number of formal arguments in the LLVM type (i.e. excluding varargs).
+    const size_t formalLLArgCount = irFty.args.size();
+
+    // The number of explicit arguments in the D call expression (including
+    // varargs), not all of which necessarily generate a LLVM argument.
+    const size_t explicitDArgCount = argvals.size();
 
     // construct and initialize an IrFuncTyArg object for each vararg
     std::vector<IrFuncTyArg*> optionalIrArgs;
-    for (int i = numFormalParams; i < numExplicitArgs; i++) {
+    for (size_t i = numFormalParams; i < explicitDArgCount; i++) {
         Type* argType = argvals[i]->getType();
         bool passByVal = gABI->passByVal(argType);
 
@@ -135,64 +141,68 @@ static void addExplicitArguments(std::vector<LLValue*>& args, AttrSet& attrs,
             initialAttrs.add(DtoShouldExtend(argType));
 
         optionalIrArgs.push_back(new IrFuncTyArg(argType, passByVal, initialAttrs));
+        optionalIrArgs.back()->parametersIdx = i;
     }
 
     // let the ABI rewrite the IrFuncTyArg objects
     gABI->rewriteVarargs(irFty, optionalIrArgs);
 
-    for (int i = 0; i < numExplicitArgs; i++)
+    const size_t explicitLLArgCount = formalLLArgCount + optionalIrArgs.size();
+    args.resize(implicitLLArgCount + explicitLLArgCount, static_cast<llvm::Value*>(0));
+
+    // Iterate the explicit arguments from left to right in the D source,
+    // which is the reverse of the LLVM order if irFty.reverseParams is true.
+    for (size_t i = 0; i < explicitLLArgCount; ++i)
     {
-        int j = numImplicitArgs + (irFty.reverseParams ? numExplicitArgs - i - 1 : i);
-
-        DValue* argval = argvals[i];
-        Type* argType = argval->getType();
-
-        const bool isVararg = (i >= numFormalParams);
+        const bool isVararg = (i >= irFty.args.size());
         IrFuncTyArg* irArg = NULL;
-        LLValue* arg = NULL;
-
-        if (!isVararg)
-        {
-            irArg = irFty.args[i];
-            arg = irFty.putParam(argType, i, argval);
-        }
-        else
-        {
+        if (isVararg)
             irArg = optionalIrArgs[i - numFormalParams];
-            arg = irFty.putParam(argType, *irArg, argval);
-        }
+        else
+            irArg = irFty.args[i];
 
-        LLType* callableArgType = (isVararg ? NULL : callableTy->getParamType(j));
+        DValue* const argval = argvals[irArg->parametersIdx];
+        Type* const argType = argval->getType();
+
+        llvm::Value* llVal = NULL;
+        if (isVararg)
+            llVal = irFty.putParam(argType, *irArg, argval);
+        else
+            llVal = irFty.putParam(argType, i, argval);
+
+        const size_t llArgIdx = implicitLLArgCount +
+            (irFty.reverseParams ? explicitLLArgCount - i - 1 : i);
+        llvm::Type* const callableArgType =
+            (isVararg ? NULL : callableTy->getParamType(llArgIdx));
 
         // Hack around LDC assuming structs and static arrays are in memory:
         // If the function wants a struct, and the argument value is a
         // pointer to a struct, load from it before passing it in.
-        if (isaPointer(arg) && DtoIsPassedByRef(argType) &&
-            ( (!isVararg && !isaPointer(callableArgType)) ||
-              (isVararg && !irArg->byref && !irArg->isByVal()) ) )
+        if (isaPointer(llVal) && DtoIsPassedByRef(argType) &&
+            ((!isVararg && !isaPointer(callableArgType)) ||
+             (isVararg && !irArg->byref && !irArg->isByVal())))
         {
             Logger::println("Loading struct type for function argument");
-            arg = DtoLoad(arg);
+            llVal = DtoLoad(llVal);
         }
 
         // parameter type mismatch, this is hard to get rid of
-        if (!isVararg && arg->getType() != callableArgType)
+        if (!isVararg && llVal->getType() != callableArgType)
         {
-#if 1
-            IF_LOG {
-                Logger::cout() << "arg:     " << *arg << '\n';
-                Logger::cout() << "of type: " << *arg->getType() << '\n';
+            IF_LOG
+            {
+                Logger::cout() << "arg:     " << *llVal << '\n';
                 Logger::cout() << "expects: " << *callableArgType << '\n';
             }
-#endif
-            if (isaStruct(arg))
-                arg = DtoAggrPaint(arg, callableArgType);
+            if (isaStruct(llVal))
+                llVal = DtoAggrPaint(llVal, callableArgType);
             else
-                arg = DtoBitCast(arg, callableArgType);
+                llVal = DtoBitCast(llVal, callableArgType);
         }
 
-        args[j] = arg;
-        attrs.add(j + 1, irArg->attrs);
+        args[llArgIdx] = llVal;
+        // +1 as index 0 contains the function attributes.
+        attrs.add(llArgIdx + 1, irArg->attrs);
 
         if (isVararg)
             delete irArg;

ir/irfuncty.cpp

@@ -16,7 +16,7 @@
 #include "gen/tollvm.h"
 
 IrFuncTyArg::IrFuncTyArg(Type* t, bool bref, const AttrBuilder& a)
-    : type(t),
+    : type(t), parametersIdx(0),
       ltype(t != Type::tvoid && bref ? DtoType(t->pointerTo()) : DtoType(t)),
       attrs(a), byref(bref), rewrite(0)
 {

ir/irfuncty.h

@@ -37,14 +37,21 @@ namespace llvm {
     class FunctionType;
 }
 
-// represents a function type argument
-// both explicit and implicit as well as return values
+/// Represents a function type argument (both explicit and implicit as well as
+/// return values).
+///
+/// Instances of this only exist for arguments that are actually lowered to an
+/// LLVM parameter (e.g. not for empty structs).
 struct IrFuncTyArg
 {
     /** This is the original D type as the frontend knows it
      *  May NOT be rewritten!!! */
     Type* const type;
 
+    /// The index of the declaration in the FuncDeclaration::parameters array
+    /// corresponding to this argument.
+    size_t parametersIdx;
+
     /// This is the final LLVM Type used for the parameter/return value type
     llvm::Type* ltype;