HLSL: add implicit promotions for assignments and function returns.

Test/hlsl.promotions.frag

+
+struct PS_OUTPUT
+{
+    float4 Color : SV_Target0;
+};
+
+uniform int3    i3;
+uniform bool3   b3;
+uniform float3  f3;
+uniform uint3   u3;
+uniform double3 d3;
+
+uniform int    is;
+uniform bool   bs;
+uniform float  fs;
+uniform uint   us;
+uniform double ds;
+
+void Fn_F3(float3 x) { }
+void Fn_I3(int3 x) { }
+void Fn_U3(uint3 x) { }
+void Fn_B3(bool3 x) { }
+void Fn_D3(double3 x) { }
+
+// ----------- Test implicit conversions on function returns -----------
+float3  Fn_R_F3I(out float3 p) { p = i3; return i3; }
+float3  Fn_R_F3U(out float3 p) { p = u3; return u3; }
+float3  Fn_R_F3B(out float3 p) { p = b3; return b3; }
+float3  Fn_R_F3D(out float3 p) { p = d3; return d3; }  // valid, but loss of precision on downconversion.
+
+int3    Fn_R_I3U(out int3 p) { p = u3; return u3; }
+int3    Fn_R_I3B(out int3 p) { p = b3; return b3; }
+int3    Fn_R_I3F(out int3 p) { p = f3; return f3; }
+int3    Fn_R_I3D(out int3 p) { p = d3; return d3; }  // valid, but loss of precision on downconversion.
+
+uint3   Fn_R_U3I(out uint3 p) { p = i3; return i3; }
+uint3   Fn_R_U3F(out uint3 p) { p = f3; return f3; }
+uint3   Fn_R_U3B(out uint3 p) { p = b3; return b3; }
+uint3   Fn_R_U3D(out uint3 p) { p = d3; return d3; }  // valid, but loss of precision on downconversion.
+
+bool3   Fn_R_B3I(out bool3 p) { p = i3; return i3; }
+bool3   Fn_R_B3U(out bool3 p) { p = u3; return u3; }
+bool3   Fn_R_B3F(out bool3 p) { p = f3; return f3; }
+bool3   Fn_R_B3D(out bool3 p) { p = d3; return d3; }
+
+double3 Fn_R_D3I(out double3 p) { p = i3; return i3; }
+double3 Fn_R_D3U(out double3 p) { p = u3; return u3; }
+double3 Fn_R_D3B(out double3 p) { p = b3; return b3; }
+double3 Fn_R_D3F(out double3 p) { p = f3; return f3; }
+
+PS_OUTPUT main()
+{
+    // ----------- assignment conversions -----------
+    float3 r00 = i3;
+    float3 r01 = b3;
+    float3 r02 = u3;
+    float3 r03 = d3;  // valid, but loss of precision on downconversion.
+
+    int3   r10 = b3;
+    int3   r11 = u3;
+    int3   r12 = f3;
+    int3   r13 = d3;  // valid, but loss of precision on downconversion.
+
+    uint3  r20 = b3;
+    uint3  r21 = i3;
+    uint3  r22 = f3;
+    uint3  r23 = d3;  // valid, but loss of precision on downconversion.
+
+    bool3  r30 = i3;
+    bool3  r31 = u3;
+    bool3  r32 = f3;
+    bool3  r33 = d3;
+
+    double3 r40 = i3;
+    double3 r41 = u3;
+    double3 r42 = f3;
+    double3 r43 = b3;
+
+    // ----------- assign ops: vector times vector ----------- 
+    r00 *= i3;
+    r01 *= b3;
+    r02 *= u3;
+    r03 *= d3;  // valid, but loss of precision on downconversion.
+    
+    r10 *= b3;
+    r11 *= u3;
+    r12 *= f3;
+    r13 *= d3;  // valid, but loss of precision on downconversion.
+    
+    r20 *= b3;
+    r21 *= i3;
+    r22 *= f3;
+    r23 *= d3;  // valid, but loss of precision on downconversion.
+
+    // No mul operator for bools
+    
+    r40 *= i3;
+    r41 *= u3;
+    r42 *= f3;
+    r43 *= b3;
+
+    // ----------- assign ops: vector times scalar ----------- 
+    r00 *= is;
+    r01 *= bs;
+    r02 *= us;
+    r03 *= ds;  // valid, but loss of precision on downconversion.
+    
+    r10 *= bs;
+    r11 *= us;
+    r12 *= fs;
+    r13 *= ds;  // valid, but loss of precision on downconversion.
+    
+    r20 *= bs;
+    r21 *= is;
+    r22 *= fs;
+    r23 *= ds;  // valid, but loss of precision on downconversion.
+
+    // No mul operator for bools
+    
+    r40 *= is;
+    r41 *= us;
+    r42 *= fs;
+    r43 *= bs;
+
+
+#define FN_OVERLOADS 0 // change to 1 when overloads under promotions are in place
+
+#if FN_OVERLOADS
+    Fn_F3(i3);
+    Fn_F3(u3);
+    Fn_F3(f3);
+    Fn_F3(b3);
+    Fn_F3(d3);  // valid, but loss of precision on downconversion.
+
+    Fn_I3(i3);
+    Fn_I3(u3);
+    Fn_I3(f3);
+    Fn_I3(b3);
+    Fn_I3(d3);  // valid, but loss of precision on downconversion.
+
+    Fn_U3(i3);
+    Fn_U3(u3);
+    Fn_U3(f3);
+    Fn_U3(b3);
+    Fn_U3(d3);  // valid, but loss of precision on downconversion.
+
+    Fn_B3(i3);
+    Fn_B3(u3);
+    Fn_B3(f3);
+    Fn_B3(b3);
+    Fn_B3(d3);
+
+    Fn_D3(i3);
+    Fn_D3(u3);
+    Fn_D3(f3);
+    Fn_D3(b3);
+    Fn_D3(d3);
+
+    Fn_F3(i3.x);
+    Fn_F3(u3.x);
+    Fn_F3(f3.x);
+    Fn_F3(b3.x);
+    Fn_F3(d3.x);  // valid, but loss of precision on downconversion.
+
+    Fn_I3(i3.x);
+    Fn_I3(u3.x);
+    Fn_I3(f3.x);
+    Fn_I3(b3.x);
+    Fn_I3(d3.x);  // valid, but loss of precision on downconversion.
+
+    Fn_U3(i3.x);
+    Fn_U3(u3.x);
+    Fn_U3(f3.x);
+    Fn_U3(b3.x);
+    Fn_U3(d3.x);  // valid, but loss of precision on downconversion.
+
+    Fn_B3(i3.x);
+    Fn_B3(u3.x);
+    Fn_B3(f3.x);
+    Fn_B3(b3.x);
+    Fn_B3(d3.x);
+
+    Fn_D3(i3.x);
+    Fn_D3(u3.x);
+    Fn_D3(f3.x);
+    Fn_D3(b3.x);
+    Fn_D3(d3.x);
+#endif
+
+    const int   si = 3;
+    const float sf = 1.2;
+
+    int   c1 = si * sf;  // 3.6 (not 3!)
+    int   c2 = sf * si;  // 3.6 (not 3!)
+
+    float4 outval = float4(si * sf, sf*si, c1, c2);
+
+    PS_OUTPUT psout;
+    psout.Color = outval;
+    return psout;
+}

glslang/MachineIndependent/Intermediate.cpp

@@ -524,7 +524,7 @@ TIntermTyped* TIntermediate::addConversion(TOperator op, const TType& type, TInt
         if (type.getBasicType() == node->getType().getBasicType())
             return node;
 
-        if (canImplicitlyPromote(node->getType().getBasicType(), type.getBasicType()))
+        if (canImplicitlyPromote(node->getType().getBasicType(), type.getBasicType(), op))
             promoteTo = type.getBasicType();
         else
             return 0;
@@ -726,11 +726,37 @@ TIntermTyped* TIntermediate::addShapeConversion(TOperator op, const TType& type,
 // See if the 'from' type is allowed to be implicitly converted to the
 // 'to' type.  This is not about vector/array/struct, only about basic type.
 //
-bool TIntermediate::canImplicitlyPromote(TBasicType from, TBasicType to) const
+bool TIntermediate::canImplicitlyPromote(TBasicType from, TBasicType to, TOperator op) const
 {
     if (profile == EEsProfile || version == 110)
         return false;
 
+    // Some languages allow more general (or potentially, more specific) conversions under some conditions.
+    if (source == EShSourceHlsl) {
+        const bool fromConvertable = (from == EbtFloat || from == EbtDouble || from == EbtInt || from == EbtUint || from == EbtBool);
+        const bool toConvertable = (to == EbtFloat || to == EbtDouble || to == EbtInt || to == EbtUint || to == EbtBool);
+
+        if (fromConvertable && toConvertable) {
+            switch (op) {
+            case EOpAndAssign:               // assignments can perform arbitrary conversions
+            case EOpInclusiveOrAssign:       // ... 
+            case EOpExclusiveOrAssign:       // ... 
+            case EOpAssign:                  // ... 
+            case EOpAddAssign:               // ... 
+            case EOpSubAssign:               // ... 
+            case EOpMulAssign:               // ... 
+            case EOpVectorTimesScalarAssign: // ... 
+            case EOpMatrixTimesScalarAssign: // ... 
+            case EOpDivAssign:               // ... 
+            case EOpModAssign:               // ... 
+            case EOpReturn:                  // function returns can also perform arbitrary conversions
+                return true;
+            default:
+                break;
+            }
+        }
+    }
+
     switch (to) {
     case EbtDouble:
         switch (from) {

glslang/MachineIndependent/localintermediate.h

@@ -189,7 +189,7 @@ public:
     TIntermTyped* addIndex(TOperator op, TIntermTyped* base, TIntermTyped* index, TSourceLoc);
     TIntermTyped* addUnaryMath(TOperator, TIntermTyped* child, TSourceLoc);
     TIntermTyped* addBuiltInFunctionCall(const TSourceLoc& line, TOperator, bool unary, TIntermNode*, const TType& returnType);
-    bool canImplicitlyPromote(TBasicType from, TBasicType to) const;
+    bool canImplicitlyPromote(TBasicType from, TBasicType to, TOperator op = EOpNull) const;
     TOperator mapTypeToConstructorOp(const TType&) const;
     TIntermAggregate* growAggregate(TIntermNode* left, TIntermNode* right);
     TIntermAggregate* growAggregate(TIntermNode* left, TIntermNode* right, const TSourceLoc&);

gtests/Hlsl.FromFile.cpp

@@ -121,6 +121,7 @@ INSTANTIATE_TEST_CASE_P(
         {"hlsl.load.offsetarray.dx10.frag", "main"},
         {"hlsl.numericsuffixes.frag", "main"},
         {"hlsl.pp.line.frag", "main"},
+        {"hlsl.promotions.frag", "main"},
         {"hlsl.sample.array.dx10.frag", "main"},
         {"hlsl.sample.basic.dx10.frag", "main"},
         {"hlsl.sample.offset.dx10.frag", "main"},

hlsl/hlslGrammar.cpp

@@ -2446,7 +2446,7 @@ bool HlslGrammar::acceptJumpStatement(TIntermNode*& statement)
         TIntermTyped* node;
         if (acceptExpression(node)) {
             // hook it up
-            statement = intermediate.addBranch(EOpReturn, node, token.loc);
+            statement = parseContext.handleReturnValue(token.loc, node);
         } else
             statement = intermediate.addBranch(EOpReturn, token.loc);
         break;

hlsl/hlslParseHelper.cpp

@@ -796,6 +796,25 @@ TIntermAggregate* HlslParseContext::handleFunctionDefinition(const TSourceLoc& l
     return paramNodes;
 }
 
+// Handle function returns, including type conversions to the function return type
+// if necessary.
+TIntermNode* HlslParseContext::handleReturnValue(const TSourceLoc& loc, TIntermTyped* value)
+{
+    if (currentFunctionType->getBasicType() == EbtVoid) {
+        error(loc, "void function cannot return a value", "return", "");
+        return intermediate.addBranch(EOpReturn, loc);
+    } else if (*currentFunctionType != value->getType()) {
+        TIntermTyped* converted = intermediate.addConversion(EOpReturn, *currentFunctionType, value);
+        if (converted) {
+            return intermediate.addBranch(EOpReturn, converted, loc);
+        } else {
+            error(loc, "type does not match, or is not convertible to, the function's return type", "return", "");
+            return intermediate.addBranch(EOpReturn, value, loc);
+        }
+    } else
+        return intermediate.addBranch(EOpReturn, value, loc);
+}
+
 void HlslParseContext::handleFunctionArgument(TFunction* function, TIntermTyped*& arguments, TIntermTyped* newArg)
 {
     TParameter param = { 0, new TType };

hlsl/hlslParseHelper.h

@@ -85,6 +85,7 @@ public:
     TIntermTyped* handleDotDereference(const TSourceLoc&, TIntermTyped* base, const TString& field);
     TFunction* handleFunctionDeclarator(const TSourceLoc&, TFunction& function, bool prototype);
     TIntermAggregate* handleFunctionDefinition(const TSourceLoc&, TFunction&);
+    TIntermNode* handleReturnValue(const TSourceLoc&, TIntermTyped*);
     void handleFunctionArgument(TFunction*, TIntermTyped*& arguments, TIntermTyped* newArg);
     TIntermTyped* handleFunctionCall(const TSourceLoc&, TFunction*, TIntermNode*);
     void decomposeIntrinsic(const TSourceLoc&, TIntermTyped*& node, TIntermNode* arguments);