Fix for operator assignment issue (#2951)

* WIP handling LValue coercion via LValueImplicitCast

* Need to have the ptr type for the cast.

* Casting conversion working on C++.

* Make the LValue casts record if in or in/out as we can produce better code if we know the difference.

* WIP LValueCast pass

* Fix tests so we don't fail because downstream compilers detect use of uninitialized variable.

* Do conversions through through tmp for l-value scenarios that can't work other ways.

* Fix a typo.

* Change diagnostic implicit-cast-lvalue for a type that still exhibits the issue.

* Add matrix test.

* Added a bit more clarity around LValue casting choices.

* Small comment improvements.
Improvements based on comments on PR.

* Use findOuterGeneric.

22 files changed, 687 insertions, 16 deletions

diff --git a/build/visual-studio/slang/slang.vcxproj b/build/visual-studio/slang/slang.vcxproj
index 3f1f580ed..3bff1e419 100644
--- a/build/visual-studio/slang/slang.vcxproj
+++ b/build/visual-studio/slang/slang.vcxproj
@@ -411,6 +411,7 @@ IF EXIST ..\..\..\external\slang-glslang\bin\windows-aarch64\release\slang-glsla
     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-generic-function.h" />

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-generic-type.h" />

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-generics.h" />

+    <ClInclude Include="..\..\..\source\slang\slang-ir-lower-l-value-cast.h" />

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-optional-type.h" />

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-reinterpret.h" />

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-result-type.h" />

@@ -609,6 +610,7 @@ IF EXIST ..\..\..\external\slang-glslang\bin\windows-aarch64\release\slang-glsla
     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-generic-function.cpp" />

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-generic-type.cpp" />

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-generics.cpp" />

+    <ClCompile Include="..\..\..\source\slang\slang-ir-lower-l-value-cast.cpp" />

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-optional-type.cpp" />

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-reinterpret.cpp" />

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-result-type.cpp" />

diff --git a/build/visual-studio/slang/slang.vcxproj.filters b/build/visual-studio/slang/slang.vcxproj.filters
index 3484c406a..f0b6b5324 100644
--- a/build/visual-studio/slang/slang.vcxproj.filters
+++ b/build/visual-studio/slang/slang.vcxproj.filters
@@ -321,6 +321,9 @@
     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-generics.h">

       <Filter>Header Files</Filter>

     </ClInclude>

+    <ClInclude Include="..\..\..\source\slang\slang-ir-lower-l-value-cast.h">

+      <Filter>Header Files</Filter>

+    </ClInclude>

     <ClInclude Include="..\..\..\source\slang\slang-ir-lower-optional-type.h">

       <Filter>Header Files</Filter>

     </ClInclude>

@@ -911,6 +914,9 @@
     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-generics.cpp">

       <Filter>Source Files</Filter>

     </ClCompile>

+    <ClCompile Include="..\..\..\source\slang\slang-ir-lower-l-value-cast.cpp">

+      <Filter>Source Files</Filter>

+    </ClCompile>

     <ClCompile Include="..\..\..\source\slang\slang-ir-lower-optional-type.cpp">

       <Filter>Source Files</Filter>

     </ClCompile>

diff --git a/source/slang/slang-ast-expr.h b/source/slang/slang-ast-expr.h
index da213e8d4..36d6546de 100644
--- a/source/slang/slang-ast-expr.h
+++ b/source/slang/slang-ast-expr.h
@@ -302,6 +302,32 @@ class ImplicitCastExpr : public TypeCastExpr
     SLANG_AST_CLASS(ImplicitCastExpr)
 };
 
+class LValueImplicitCastExpr : public TypeCastExpr
+{
+    SLANG_AST_CLASS(LValueImplicitCastExpr)
+
+    explicit LValueImplicitCastExpr(const TypeCastExpr& rhs) :Super(rhs) {}
+};
+
+// To work around situations like int += uint
+// where we want to allow an LValue to work with an implicit cast.
+// The argument being cast *must* be an LValue.
+class OutImplicitCastExpr : public LValueImplicitCastExpr
+{
+    SLANG_AST_CLASS(OutImplicitCastExpr)
+
+        /// Allow explict construction from any TypeCastExpr
+    explicit OutImplicitCastExpr(const TypeCastExpr& rhs) :Super(rhs) {}
+};
+
+class InOutImplicitCastExpr : public LValueImplicitCastExpr
+{
+    SLANG_AST_CLASS(InOutImplicitCastExpr)
+
+    /// Allow explict construction from any TypeCastExpr
+    explicit InOutImplicitCastExpr(const TypeCastExpr& rhs) :Super(rhs) {}
+};
+
     /// A cast of a value to a super-type of its type.
     ///
     /// The type being cast to is stored as this expression's `type`.
diff --git a/source/slang/slang-check-expr.cpp b/source/slang/slang-check-expr.cpp
index e0084e08e..00ece3628 100644
--- a/source/slang/slang-check-expr.cpp
+++ b/source/slang/slang-check-expr.cpp
@@ -1952,6 +1952,51 @@ namespace Slang
         return checkedExpr;
     }
 
+    static bool _canLValueCoerceScalarType(Type* a, Type* b)
+    {
+        auto basicTypeA = as<BasicExpressionType>(a);
+        auto basicTypeB = as<BasicExpressionType>(b);
+
+        if (basicTypeA && basicTypeB)
+        {
+            const auto& infoA = BaseTypeInfo::getInfo(basicTypeA->baseType);
+            const auto& infoB = BaseTypeInfo::getInfo(basicTypeB->baseType);
+
+            // TODO(JS): Initially this tries to limit where LValueImplict casts happen.
+            // We could in principal allow different sizes, as long as we converted to a temprorary
+            // and back again. 
+            // 
+            // For now we just stick with the simple case. 
+            // // We only allow on integer types for now. In effect just allowing any size uint/int conversions
+            if (infoA.sizeInBytes == infoB.sizeInBytes && 
+                (infoA.flags & infoB.flags & BaseTypeInfo::Flag::Integer))
+            {
+                return true;
+            }
+
+        }
+        return false;
+    }
+
+    static bool _canLValueCoerce(Type* a, Type* b)
+    {
+        // We can *assume* here that if they are coercable, that dimensions of vectors
+        // and matrices match. We might want to assert to be sure...
+        SLANG_ASSERT(a != b);
+        if (a->astNodeType == b->astNodeType)
+        {
+            if (auto matA = as<MatrixExpressionType>(a))
+            {
+                return _canLValueCoerceScalarType(matA->getElementType(), static_cast<MatrixExpressionType*>(b)->getElementType());   
+            }
+            else if (auto vecA = as<VectorExpressionType>(a))
+            {
+                return  _canLValueCoerceScalarType(vecA->getScalarType(), static_cast<VectorExpressionType*>(b)->getScalarType());   
+            }
+        }
+        return _canLValueCoerceScalarType(a, b);
+    }
+
     Expr* SemanticsVisitor::CheckInvokeExprWithCheckedOperands(InvokeExpr *expr)
     {
         auto rs = ResolveInvoke(expr);
@@ -1985,23 +2030,63 @@ namespace Slang
                         if( pp < expr->arguments.getCount() )
                         {
                             auto argExpr = expr->arguments[pp];
-                            if( !argExpr->type.isLeftValue )
+                            if( !argExpr->type.isLeftValue)
                             {
-                                getSink()->diagnose(
-                                    argExpr,
-                                    Diagnostics::argumentExpectedLValue,
-                                    pp);
+                                auto implicitCastExpr = as<ImplicitCastExpr>(argExpr);
 
-                                if( auto implicitCastExpr = as<ImplicitCastExpr>(argExpr) )
+                                if (implicitCastExpr && _canLValueCoerce(implicitCastExpr->arguments[0]->type, implicitCastExpr->type))
+                                {
+                                    // This is to work around issues like
+                                    //
+                                    // ```
+                                    // int a = 0;
+                                    // uint b = 1;
+                                    // a += b;
+                                    // ```
+                                    // That strictly speaking it's not allowed, but we are going to allow it for now 
+                                    // for situations were the types are uint/int and vector/matrix varieties of those types
+                                    // 
+                                    // Then in lowering we are going to insert code to do something like
+                                    // ```
+                                    // var OutType: tmp = arg;
+                                    // f(... tmp);
+                                    // arg = tmp;
+                                    // ```
+
+                                    TypeCastExpr* lValueImplicitCast;
+
+                                    // We want to record if the cast is being used for `out` or `inout`/`ref` as 
+                                    // if it's just `out` we won't need to convert before passing in.
+                                    if (as<OutType>(paramType))
+                                    {
+                                        lValueImplicitCast = getASTBuilder()->create<OutImplicitCastExpr>(*implicitCastExpr);
+                                    }
+                                    else
+                                    {
+                                        lValueImplicitCast = getASTBuilder()->create<InOutImplicitCastExpr>(*implicitCastExpr);
+                                    }
+
+                                    // Replace the expression. This should make this situation easier to detect.
+                                    expr->arguments[pp] = lValueImplicitCast;
+                                }
+                                else
                                 {
                                     getSink()->diagnose(
                                         argExpr,
-                                        Diagnostics::implicitCastUsedAsLValue,
-                                        implicitCastExpr->arguments[0]->type,
-                                        implicitCastExpr->type);
+                                        Diagnostics::argumentExpectedLValue,
+                                        pp);
+
+                                    if(implicitCastExpr)
+                                    {
+                                        getSink()->diagnose(
+                                            argExpr,
+                                            Diagnostics::implicitCastUsedAsLValue,
+                                            implicitCastExpr->arguments[pp]->type,
+                                            implicitCastExpr->type);
+                                    }
+
+                                    maybeDiagnoseThisNotLValue(argExpr);
                                 }
-
-                                maybeDiagnoseThisNotLValue(argExpr);
                             }
                         }
                         else
diff --git a/source/slang/slang-compiler.h b/source/slang/slang-compiler.h
index 4e396e1f5..7082eccd7 100755
--- a/source/slang/slang-compiler.h
+++ b/source/slang/slang-compiler.h
@@ -1633,6 +1633,12 @@ namespace Slang
         /// Are we generating code for a CUDA API (CUDA / OptiX)?
     bool isCUDATarget(TargetRequest* targetReq);
 
+        /// Given a target request returns which (if any) intermediate source language is required
+        /// to produce it.
+        ///
+        /// If no intermediate source language is required, will return SourceLanguage::Unknown
+    SourceLanguage getIntermediateSourceLanguageForTarget(TargetRequest* req);
+
         /// Are resource types "bindless" (implemented as ordinary data) on the given `target`?
     bool areResourceTypesBindlessOnTarget(TargetRequest* target);
 
diff --git a/source/slang/slang-emit-cpp.cpp b/source/slang/slang-emit-cpp.cpp
index 49edec92b..ef9b7f54e 100644
--- a/source/slang/slang-emit-cpp.cpp
+++ b/source/slang/slang-emit-cpp.cpp
@@ -1200,6 +1200,20 @@ bool CPPSourceEmitter::tryEmitInstExprImpl(IRInst* inst, const EmitOpInfo& inOut
         {
             return false;
         }
+       
+        case kIROp_InOutImplicitCast:
+        case kIROp_OutImplicitCast:
+        {
+            // We'll just the LValue to be the desired type
+            m_writer->emit("reinterpret_cast<");
+            emitType(inst->getDataType());
+            m_writer->emit(">(");
+            
+            emitOperand(inst->getOperand(0), getInfo(EmitOp::General));
+
+            m_writer->emit(")");
+            return true;
+        }
         case kIROp_MakeVector:
         {
             IRType* retType = inst->getFullType();
diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index bd6542e2a..827c69e50 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -35,6 +35,7 @@
 #include "slang-ir-lower-result-type.h"
 #include "slang-ir-lower-optional-type.h"
 #include "slang-ir-lower-bit-cast.h"
+#include "slang-ir-lower-l-value-cast.h"
 #include "slang-ir-lower-reinterpret.h"
 #include "slang-ir-loop-unroll.h"
 #include "slang-ir-metadata.h"
@@ -866,6 +867,9 @@ Result linkAndOptimizeIR(
         legalizeUniformBufferLoad(irModule);
     }
 
+    // Lower all the LValue implict casts (used for out/inout/ref scenarios)
+    lowerLValueCast(targetRequest, irModule);
+
     // Lower all bit_cast operations on complex types into leaf-level
     // bit_cast on basic types.
     lowerBitCast(targetRequest, irModule);
diff --git a/source/slang/slang-ir-inst-defs.h b/source/slang/slang-ir-inst-defs.h
index b33565700..de3735a55 100644
--- a/source/slang/slang-ir-inst-defs.h
+++ b/source/slang/slang-ir-inst-defs.h
@@ -923,6 +923,8 @@ INST(ExtractTaggedUnionPayload,         extractTaggedUnionPayload,  1, 0)
 
 INST(BitCast,                           bitCast,                    1, 0)
 INST(Reinterpret,                       reinterpret,                1, 0)
+INST(OutImplicitCast,                   outImplicitCast,           1, 0)
+INST(InOutImplicitCast,                 inOutImplicitCast,         1, 0)
 INST(IntCast, intCast, 1, 0)
 INST(FloatCast, floatCast, 1, 0)
 INST(CastIntToFloat, castIntToFloat, 1, 0)
diff --git a/source/slang/slang-ir-insts.h b/source/slang/slang-ir-insts.h
index 4c540bfdd..f2c00f406 100644
--- a/source/slang/slang-ir-insts.h
+++ b/source/slang/slang-ir-insts.h
@@ -3302,6 +3302,8 @@ public:
     IRUndefined* emitUndefined(IRType* type);
 
     IRInst* emitReinterpret(IRInst* type, IRInst* value);
+    IRInst* emitOutImplicitCast(IRInst* type, IRInst* value);
+    IRInst* emitInOutImplicitCast(IRInst* type, IRInst* value);
 
     IRFunc* createFunc();
     IRGlobalVar* createGlobalVar(
diff --git a/source/slang/slang-ir-lower-l-value-cast.cpp b/source/slang/slang-ir-lower-l-value-cast.cpp
new file mode 100644
index 000000000..cd03d2bd5
--- /dev/null
+++ b/source/slang/slang-ir-lower-l-value-cast.cpp
@@ -0,0 +1,247 @@
+#include "slang-ir-lower-l-value-cast.h"
+
+#include "slang-ir.h"
+#include "slang-ir-insts.h"
+
+#include "slang-ir-clone.h"
+
+#include "slang-ir-util.h"
+
+namespace Slang
+{
+
+struct LValueCastLoweringContext
+{    
+    void _addToWorkList(IRInst* inst)
+    {
+        if (!findOuterGeneric(inst) && !m_workList.contains(inst))
+        {
+            m_workList.add(inst);
+        }
+    }
+
+    void _processInst(IRInst* inst)
+    {
+        switch (inst->getOp())
+        {
+        case kIROp_InOutImplicitCast:
+        case kIROp_OutImplicitCast:
+            _processLValueCast(inst);
+            break;
+        default:
+            break;
+        }
+    }
+
+    void processModule()
+    {
+        _addToWorkList(m_module->getModuleInst());
+
+        while (m_workList.getCount() != 0)
+        {
+            IRInst* inst = m_workList.getLast();
+            m_workList.removeLast();
+
+            _processInst(inst);
+
+            for (auto child = inst->getLastChild(); child; child = child->getPrevInst())
+            {
+                _addToWorkList(child);
+            }
+        }
+    }
+
+        /// True if the conversion from a to b, can be achieved 
+        /// via a reinterpret cast/bitcast
+        /// Only some targets can allow such conversions
+    bool _canReinterpretCast(IRType* a, IRType* b)
+    {
+        auto ptrA = as<IRPtrType>(a);
+        auto ptrB = as<IRPtrType>(b);
+
+        // They must both be pointers...
+        SLANG_ASSERT(ptrA && ptrB);
+
+        a = ptrA->getValueType();
+        b = ptrB->getValueType();
+
+        if (a->m_op == b->m_op)
+        {
+            if (auto matA = as<IRMatrixType>(a))
+            {
+                auto matB = static_cast<IRMatrixType*>(b);
+
+                if (getIntVal(matA->getColumnCount()) != getIntVal(matB->getColumnCount()))
+                {
+                    return false;
+                }
+
+                a = matA->getElementType();
+                b = matB->getElementType();
+            }
+            else if (auto vecA = as<IRVectorType>(a))
+            {
+                auto vecB = static_cast<IRVectorType*>(b);
+
+                if (getIntVal(vecA->getElementCount()) != getIntVal(vecB->getElementCount()))
+                {
+                    return false;
+                }
+
+                a = vecA->getElementType();
+                b = vecB->getElementType();
+            }
+        }
+
+        auto basicA = as<IRBasicType>(a);
+        auto basicB = as<IRBasicType>(b);
+
+        if (basicA && basicB)
+        {
+            auto baseA = basicA->getBaseType();
+            auto baseB = basicB->getBaseType();
+
+            const auto& infoA = BaseTypeInfo::getInfo(baseA);
+            const auto& infoB = BaseTypeInfo::getInfo(baseB);
+
+            // We allow reinterpret case for int type conversions of the same bit size for now
+            if (infoA.sizeInBytes == infoB.sizeInBytes && 
+                (infoA.flags & infoB.flags & BaseTypeInfo::Flag::Integer))
+            {
+                return true;
+            }
+        }
+
+        return false;
+    }
+
+        /// True if for HLSL the cast can be removed entirely
+    bool _canRemoveCastForHLSL(IRType* a, IRType* b)
+    {
+        // Currently _canReinterpret is exactly the same class of types that we can just ignore the cast totally 
+        // for HLSL
+        // If _canReinterpretCast changes, this will need to be updated
+        return _canReinterpretCast(a, b);
+    }
+
+    void _processLValueCast(IRInst* castInst)
+    {
+        auto castOperand = castInst->getOperand(0);
+        auto fromType = castOperand->getDataType();
+        auto toType = castInst->getDataType();
+
+        switch (m_intermediateSourceLanguage)
+        {
+            case SourceLanguage::HLSL:
+            {
+                // If the conversion can just be ignored for HLSL, just remove it
+                if (_canRemoveCastForHLSL(fromType, toType))
+                {
+                    castInst->replaceUsesWith(castOperand);
+                    castInst->removeAndDeallocate();
+                    return;
+                }
+                break;
+            }
+            case SourceLanguage::C:
+            case SourceLanguage::CPP:
+            case SourceLanguage::CUDA:
+            {
+                // For languages with pointers, out parameter differences can *sometimes* just be sidestepped with 
+                // a reinterpret cast.
+                if (_canReinterpretCast(fromType, toType))
+                {
+                    return;
+                }
+                break;
+            }
+            default: break;
+        }
+
+        // If we can't use the other mechanisms we are going to do a conversion
+        // via a cast into a temporary of the approprite time before the useSite,
+        // then immediately after converting back into the original location.
+        //
+        // With a special case for uses which are just out - where we don't need to
+        // convert in.
+        
+        // Okay we are going to replace the implicit casts with temporaries around call sites/uses.
+        List<IRInst*> useSites;
+        for (auto use = castInst->firstUse; use; use = use->nextUse)
+        {
+            auto useSite = use->getUser();
+
+            if (useSites.indexOf(useSite) < 0)
+            {
+                useSites.add(useSite);
+            }
+        }
+
+        // If there is a name hint on the source, we'll copy it over to the temporaries
+        auto nameHintDecoration = castOperand->findDecoration<IRNameHintDecoration>();
+
+        IRBuilder builder(m_module);
+        
+        IRType* toValueType = as<IRPtrType>(toType)->getValueType();
+        IRType* fromValueType = as<IRPtrType>(fromType)->getValueType();
+
+        for (auto useSite : useSites)
+        {
+            builder.setInsertBefore(useSite);
+            auto tmpVar = builder.emitVar(toValueType);
+
+            if (nameHintDecoration)
+            {
+                cloneDecoration(nameHintDecoration, tmpVar);
+            }
+
+            // If it's inout we convert via cast whats in the castOperand
+            if (castInst->getOp() == kIROp_InOutImplicitCast)
+            {
+                builder.emitStore(tmpVar, builder.emitCast(toValueType, builder.emitLoad(castOperand)));
+            }
+
+            // Convert the temporary back to the original location
+            builder.setInsertAfter(useSite);
+            builder.emitStore(castOperand, builder.emitCast(fromValueType, builder.emitLoad(tmpVar)));
+
+            // Go through all of the operands of the use inst relacing, with the temporary
+            const auto operandCount = Count(useSite->getOperandCount());
+            auto operands = useSite->getOperands();
+
+            for (Index i = 0; i < operandCount; ++i)
+            {
+                auto& callSiteOperand = operands[i];
+                if(callSiteOperand.get() == castInst)
+                {
+                    callSiteOperand.set(tmpVar);
+                }
+            }
+        }
+
+        // When we are done we can destroy the inst
+        castInst->removeAndDeallocate();
+    }
+
+    LValueCastLoweringContext(TargetRequest* targetRequest, IRModule* module):
+        m_targetReq(targetRequest),
+        m_module(module)
+    {
+        m_intermediateSourceLanguage = getIntermediateSourceLanguageForTarget(targetRequest);
+    }
+
+    // The intermediate source language used to produce code for the target.
+    // If no intermediate source language is used will be SourceLanguage::Unknown.
+    SourceLanguage m_intermediateSourceLanguage = SourceLanguage::Unknown;
+    TargetRequest* m_targetReq;
+    IRModule* m_module;
+    OrderedHashSet<IRInst*> m_workList;
+};
+
+void lowerLValueCast(TargetRequest* targetReq, IRModule* module)
+{ 
+    LValueCastLoweringContext context(targetReq, module);
+    context.processModule();
+}
+
+} // namespace Slang
diff --git a/source/slang/slang-ir-lower-l-value-cast.h b/source/slang/slang-ir-lower-l-value-cast.h
new file mode 100644
index 000000000..3d93d1200
--- /dev/null
+++ b/source/slang/slang-ir-lower-l-value-cast.h
@@ -0,0 +1,24 @@
+#ifndef SLANG_IR_LOWER_L_VALUE_CAST_H
+#define SLANG_IR_LOWER_L_VALUE_CAST_H
+
+// This defines an IR pass that lowers LValue implicit casts. These are typically formed 
+// when an in/inout paramter is passed a type that doesn't match. 
+// 
+// Depending on the target this could produce
+// 
+// * Nothing - some kinds of casts are implicit for some targets such as HLSL on out parameters for same sized integer types
+// * A reinterpret cast. On targets with pointers, such as C++/CUDA we can fix the problem by just casting to the appropriate pointer (for some kinds of conversions)
+// * Creating a temporary of the right type and calling the function, and *converting* to the target (say an out parameter)
+// * Creating a temporary, converting the value into the temporary, calling the function, and converting back to the source
+
+namespace Slang
+{
+
+struct IRModule;
+class TargetRequest;
+
+void lowerLValueCast(TargetRequest* targetReq, IRModule* module);
+
+} // namespace Slang
+
+#endif
diff --git a/source/slang/slang-ir.cpp b/source/slang/slang-ir.cpp
index fb121d245..789349b4c 100644
--- a/source/slang/slang-ir.cpp
+++ b/source/slang/slang-ir.cpp
@@ -3130,7 +3130,14 @@ namespace Slang
     {
         return emitIntrinsicInst((IRType*)type, kIROp_Reinterpret, 1, &value);
     }
-
+    IRInst* IRBuilder::emitInOutImplicitCast(IRInst* type, IRInst* value)
+    {
+        return emitIntrinsicInst((IRType*)type, kIROp_InOutImplicitCast, 1, &value);
+    }
+    IRInst* IRBuilder::emitOutImplicitCast(IRInst* type, IRInst* value)
+    {
+        return emitIntrinsicInst((IRType*)type, kIROp_OutImplicitCast, 1, &value);
+    }
     IRLiveRangeStart* IRBuilder::emitLiveRangeStart(IRInst* referenced)
     {
         // This instruction doesn't produce any result, 
diff --git a/source/slang/slang-lower-to-ir.cpp b/source/slang/slang-lower-to-ir.cpp
index 58768f2ad..85b17dafb 100644
--- a/source/slang/slang-lower-to-ir.cpp
+++ b/source/slang/slang-lower-to-ir.cpp
@@ -4677,6 +4677,34 @@ struct LValueExprLoweringVisitor : ExprLoweringVisitorBase<LValueExprLoweringVis
 {
     static bool _isLValueContext() { return true; }
 
+    LoweredValInfo visitLValueImplicitCastExpr(LValueImplicitCastExpr* expr)
+    {
+        auto builder = getBuilder();
+
+        auto irType = lowerType(context, expr->type);
+        auto irPtrType = builder->getPtrType(irType);
+
+        auto loweredArg = lowerLValueExpr(context, expr->arguments[0]);
+
+        // It should be a ptr, because it is a LValue
+        SLANG_ASSERT(loweredArg.flavor == LoweredValInfo::Flavor::Ptr);
+
+        // We have the irValue (which should be a Ptr because it's an LValue)
+        auto irLValue = loweredArg.val;
+
+        IRInst* irCast = nullptr;
+        if (as<OutImplicitCastExpr>(expr))
+        {
+            irCast = builder->emitOutImplicitCast(irPtrType, irLValue);
+        }
+        else
+        {
+            irCast = builder->emitInOutImplicitCast(irPtrType, irLValue);
+        }
+
+        return LoweredValInfo::ptr(irCast);
+    }
+
     // When visiting a swizzle expression in an l-value context,
     // we need to construct a "swizzled l-value."
     LoweredValInfo visitMatrixSwizzleExpr(MatrixSwizzleExpr* expr)
diff --git a/source/slang/slang-type-layout.cpp b/source/slang/slang-type-layout.cpp
index 29cf86f5e..4d52b76eb 100644
--- a/source/slang/slang-type-layout.cpp
+++ b/source/slang/slang-type-layout.cpp
@@ -1689,6 +1689,62 @@ bool isCUDATarget(TargetRequest* targetReq)
     }
 }
 
+SourceLanguage getIntermediateSourceLanguageForTarget(TargetRequest* req)
+{
+    // If we are emitting directly, there is no intermediate source language
+    if (req->shouldEmitSPIRVDirectly())
+    {
+        return SourceLanguage::Unknown;
+    }
+
+    switch (req->getTarget())
+    {
+        case CodeGenTarget::GLSL:
+        case CodeGenTarget::GLSL_Vulkan:
+        case CodeGenTarget::GLSL_Vulkan_OneDesc:
+            // If we aren't emitting directly we are going to output GLSL to feed to GLSLANG
+        case CodeGenTarget::SPIRV:
+        case CodeGenTarget::SPIRVAssembly:
+        {
+            return SourceLanguage::GLSL;
+        }
+        case CodeGenTarget::HLSL:
+        case CodeGenTarget::DXBytecode:
+        case CodeGenTarget::DXBytecodeAssembly:
+        case CodeGenTarget::DXIL:
+        case CodeGenTarget::DXILAssembly:
+        {
+            // Currently DXBytecode and DXIL are generated via HLSL
+            return SourceLanguage::HLSL;
+        }
+        case CodeGenTarget::CSource:
+        {
+            return SourceLanguage::C;
+        }
+        case CodeGenTarget::ShaderSharedLibrary:
+        case CodeGenTarget::ObjectCode:
+        case CodeGenTarget::HostExecutable:
+        case CodeGenTarget::HostHostCallable:
+        case CodeGenTarget::ShaderHostCallable:
+        case CodeGenTarget::CPPSource:
+        case CodeGenTarget::HostCPPSource:
+        case CodeGenTarget::PyTorchCppBinding:
+        {
+            // For CPU based scenarios are generated via C++
+            return SourceLanguage::CPP;
+        }
+        case CodeGenTarget::CUDAObjectCode:
+        case CodeGenTarget::CUDASource:
+        case CodeGenTarget::PTX:
+        {
+            return SourceLanguage::CUDA;
+        }
+        default: break;
+    }
+
+    return SourceLanguage::Unknown;
+}
+
 bool areResourceTypesBindlessOnTarget(TargetRequest* targetReq)
 {
     return isCPUTarget(targetReq) || isCUDATarget(targetReq);
diff --git a/tests/bugs/op-assignment-unify-mat.slang b/tests/bugs/op-assignment-unify-mat.slang
new file mode 100644
index 000000000..13b6ecf48
--- /dev/null
+++ b/tests/bugs/op-assignment-unify-mat.slang
@@ -0,0 +1,47 @@
+// We can't test on VK, as currently we don't support integer matrix types
+//DISABLE_TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -vk
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj 
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -cpu
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0], stride=4):out,name outputBuffer
+RWStructuredBuffer<uint> outputBuffer;
+
+void silly<T : __BuiltinIntegerType, let ROWS : int, let COLS : int>(matrix<T, ROWS, COLS> a, inout matrix<T, ROWS, COLS> b)
+{
+    b *= a;
+}
+
+void silly2<T : __BuiltinIntegerType, let ROWS : int, let COLS : int>(matrix<T, ROWS, COLS> a, out matrix<T, ROWS, COLS> b)
+{
+    b = a;
+}
+
+
+[numthreads(4, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    matrix<int, 2, 2> v = { 1, 2, 3, 4};
+    matrix<uint, 2, 2> u = { 5, 6, 7, 8};
+
+    int idx = int(dispatchThreadID.x);
+    
+    matrix<uint, 2, 2> uidx = { idx, idx * idx, idx + idx, idx + 2 };
+
+    v += uidx;
+    u += idx;
+
+    silly(u, v);
+    silly(v, u);
+
+    // Check that detects issues with undefined variables.
+
+    matrix<int, 2, 2> undef1, undef2;          // undefined
+    // Downstream compilers detect this
+    //silly(u, undef1);
+    silly2(u, undef2);
+    
+    matrix<uint, 2, 2> added = (u + v + undef2);
+    uint2 added2 = added[0] + added[1];
+
+    outputBuffer[idx] = added2.x + added2.y;
+}
diff --git a/tests/bugs/op-assignment-unify-mat.slang.expected.txt b/tests/bugs/op-assignment-unify-mat.slang.expected.txt
new file mode 100644
index 000000000..f5973ac46
--- /dev/null
+++ b/tests/bugs/op-assignment-unify-mat.slang.expected.txt
@@ -0,0 +1,4 @@
+53E
+92C
+FA8
+19C0
diff --git a/tests/bugs/op-assignment-unify-vec.slang b/tests/bugs/op-assignment-unify-vec.slang
new file mode 100644
index 000000000..4cc3154a0
--- /dev/null
+++ b/tests/bugs/op-assignment-unify-vec.slang
@@ -0,0 +1,47 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -vk
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj 
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -cpu
+
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):out,name outputBuffer
+RWStructuredBuffer<uint> outputBuffer;
+
+
+void silly<T : __BuiltinIntegerType, let N : int>(vector<T, N> a, inout vector<T, N> b)
+{
+    b *= a;
+}
+
+void silly2<T : __BuiltinIntegerType, let N : int>(vector<T, N> a, out vector<T, N> b)
+{
+    b = a;
+}
+
+
+[numthreads(4, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    int2 v = int2(0, 0);
+    uint2 u = uint2(0, 0);
+
+    int idx = int(dispatchThreadID.x);
+    
+    uint2 uidx = uint2(idx, idx * idx);
+
+    v += uidx;
+    u += idx;
+
+    silly(u, v);
+    silly(v, u);
+
+    // Check that detects issues with undefined variables.
+
+    int2 undef1, undef2;          // undefined
+    // Downstream compilers detect this
+    //silly(u, undef1);
+    silly2(u, undef2);
+    
+    uint2 added = (u + v + undef2);
+
+    outputBuffer[idx] = added.x + added.y;
+}
diff --git a/tests/bugs/op-assignment-unify-vec.slang.expected.txt b/tests/bugs/op-assignment-unify-vec.slang.expected.txt
new file mode 100644
index 000000000..d92319a24
--- /dev/null
+++ b/tests/bugs/op-assignment-unify-vec.slang.expected.txt
@@ -0,0 +1,6 @@
+0
+6
+3C
+FC
+0
+0
diff --git a/tests/bugs/op-assignment-unify.slang b/tests/bugs/op-assignment-unify.slang
new file mode 100644
index 000000000..738b49677
--- /dev/null
+++ b/tests/bugs/op-assignment-unify.slang
@@ -0,0 +1,44 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -vk
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj 
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -cpu
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):out,name outputBuffer
+RWStructuredBuffer<uint> outputBuffer;
+
+
+void silly<T : __BuiltinIntegerType>(T a, inout T b)
+{
+    b *= a;
+}
+
+void silly2<T : __BuiltinIntegerType>(T a, out T b)
+{
+    b = a;
+}
+
+
+[numthreads(4, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    int v = 0;
+    uint u = 0;
+
+    int idx = int(dispatchThreadID.x);
+    uint uidx = uint(idx);
+
+    v += uidx;
+    u += idx;
+
+    silly(u, v);
+    silly(v, u);
+
+    // Check that detects issues with undefined variables.
+
+    int undef1, undef2;          // undefined
+    
+    // Downstream compilers detect this... 
+    //silly(u, undef1);
+    silly2(u, undef2);
+    
+    outputBuffer[idx] = u + v + undef2;
+}
diff --git a/tests/bugs/op-assignment-unify.slang.expected.txt b/tests/bugs/op-assignment-unify.slang.expected.txt
new file mode 100644
index 000000000..539ab696c
--- /dev/null
+++ b/tests/bugs/op-assignment-unify.slang.expected.txt
@@ -0,0 +1,6 @@
+0
+3
+14
+3F
+0
+0
diff --git a/tests/diagnostics/implicit-cast-lvalue.slang b/tests/diagnostics/implicit-cast-lvalue.slang
index 79b619443..daf196378 100644
--- a/tests/diagnostics/implicit-cast-lvalue.slang
+++ b/tests/diagnostics/implicit-cast-lvalue.slang
@@ -1,15 +1,20 @@
 //DIAGNOSTIC_TEST:SIMPLE:
 
+
+// Generally speaking:
 // Passing an argument for an `out` parameter such
 // that implicit conversion would be required in
 // both directions.
+//
+// But we do have special case code to handle uint/int of the same type and scalar/vector/matrix forms. 
+// So for this to fail we need to use floats or something else
 
-void a(out uint x)
+void a(out float x)
 {
 	x = 0;
 }
 
-void b(int y)
+void b(double y)
 {
 	a(y);
 }
diff --git a/tests/diagnostics/implicit-cast-lvalue.slang.expected b/tests/diagnostics/implicit-cast-lvalue.slang.expected
index 3fdd846a5..920ad17c2 100644
--- a/tests/diagnostics/implicit-cast-lvalue.slang.expected
+++ b/tests/diagnostics/implicit-cast-lvalue.slang.expected
@@ -1,9 +1,12 @@
 result code = -1
 standard error = {
-tests/diagnostics/implicit-cast-lvalue.slang(14): error 30047: argument passed to parameter '0' must be l-value.
+tests/diagnostics/implicit-cast-lvalue.slang(19): warning 30081: implicit conversion from 'double' to 'float' is not recommended
     a(y);
       ^
-tests/diagnostics/implicit-cast-lvalue.slang(14): note 30048: argument was implicitly cast from 'int' to 'uint', and Slang does not support using an implicit cast as an l-value
+tests/diagnostics/implicit-cast-lvalue.slang(19): error 30047: argument passed to parameter '0' must be l-value.
+    a(y);
+      ^
+tests/diagnostics/implicit-cast-lvalue.slang(19): note 30048: argument was implicitly cast from 'double' to 'float', and Slang does not support using an implicit cast as an l-value
 }
 standard output = {
 }