[HLSL] Shore up floating point conversions

This PR fixes bugs in HLSL floating conversions. HLSL always has
`half`, `float` and `double` types, which promote in the order:

`half`->`float`->`double`

and convert in the order:

`double`->`float`->`half`

As with other conversions in C++, promotions are preferred over
conversions.

We do have floating conversions documented in the draft language
specification (https://microsoft.github.io/hlsl-specs/specs/hlsl.pdf
[Conv.rank.float]) although the exact language is still in flux
(microsoft/hlsl-specs#206).

Resolves llvm#81047

clang/lib/Sema/SemaOverload.cpp

@@ -2587,7 +2587,8 @@ bool Sema::IsIntegralPromotion(Expr *From, QualType FromType, QualType ToType) {
 
   // In HLSL an rvalue of integral type can be promoted to an rvalue of a larger
   // integral type.
-  if (Context.getLangOpts().HLSL)
+  if (Context.getLangOpts().HLSL && FromType->isIntegerType() &&
+      ToType->isIntegerType())
     return Context.getTypeSize(FromType) < Context.getTypeSize(ToType);
 
   return false;
@@ -2616,6 +2617,13 @@ bool Sema::IsFloatingPointPromotion(QualType FromType, QualType ToType) {
            ToBuiltin->getKind() == BuiltinType::Ibm128))
         return true;
 
+      // In HLSL, `half` promotes to `float` or `double`, regardless of whether
+      // or not native half types are enabled.
+      if (getLangOpts().HLSL && FromBuiltin->getKind() == BuiltinType::Half &&
+          (ToBuiltin->getKind() == BuiltinType::Float ||
+           ToBuiltin->getKind() == BuiltinType::Double))
+        return true;
+
       // Half can be promoted to float.
       if (!getLangOpts().NativeHalfType &&
            FromBuiltin->getKind() == BuiltinType::Half &&
@@ -4393,6 +4401,24 @@ getFixedEnumPromtion(Sema &S, const StandardConversionSequence &SCS) {
   return FixedEnumPromotion::ToPromotedUnderlyingType;
 }
 
+static ImplicitConversionSequence::CompareKind
+HLSLCompareFloatingRank(QualType LHS, QualType RHS) {
+  assert(LHS->isVectorType() == RHS->isVectorType() &&
+         "Either both elements should be vectors or neither should.");
+  if (const auto *VT = LHS->getAs<VectorType>())
+    LHS = VT->getElementType();
+
+  if (const auto *VT = RHS->getAs<VectorType>())
+    RHS = VT->getElementType();
+
+  const auto L = LHS->getAs<BuiltinType>()->getKind();
+  const auto R = RHS->getAs<BuiltinType>()->getKind();
+  if (L == R)
+    return ImplicitConversionSequence::Indistinguishable;
+  return L < R ? ImplicitConversionSequence::Better
+               : ImplicitConversionSequence::Worse;
+}
+
 /// CompareStandardConversionSequences - Compare two standard
 /// conversion sequences to determine whether one is better than the
 /// other or if they are indistinguishable (C++ 13.3.3.2p3).
@@ -4634,6 +4660,21 @@ CompareStandardConversionSequences(Sema &S, SourceLocation Loc,
                  : ImplicitConversionSequence::Worse;
   }
 
+  if (S.getLangOpts().HLSL) {
+    // On a promotion we prefer the lower rank to disambiguate.
+    if ((SCS1.Second == ICK_Floating_Promotion &&
+         SCS2.Second == ICK_Floating_Promotion) ||
+        (SCS1.Element == ICK_Floating_Promotion &&
+         SCS2.Element == ICK_Floating_Promotion))
+      return HLSLCompareFloatingRank(SCS1.getToType(2), SCS2.getToType(2));
+    // On a conversion we prefer the higher rank to disambiguate.
+    if ((SCS1.Second == ICK_Floating_Conversion &&
+         SCS2.Second == ICK_Floating_Conversion) ||
+        (SCS1.Element == ICK_Floating_Conversion &&
+         SCS2.Element == ICK_Floating_Conversion))
+      return HLSLCompareFloatingRank(SCS2.getToType(2), SCS1.getToType(2));
+  }
+
   return ImplicitConversionSequence::Indistinguishable;
 }
 

clang/test/SemaHLSL/ScalarOverloadResolution.hlsl

@@ -0,0 +1,229 @@
+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -fnative-half-type -finclude-default-header -Wconversion -verify -o - %s
+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -fnative-half-type -finclude-default-header -ast-dump %s | FileCheck %s
+
+// This test verifies floating point type implicit conversion ranks for overload
+// resolution. In HLSL the built-in type ranks are half < float < double. This
+// applies to both scalar and vector types.
+
+// HLSL allows implicit truncation fo types, so it differentiates between
+// promotions (converting to larger types) and conversions (converting to
+// smaller types). Promotions are preferred over conversions. Promotions prefer
+// promoting to the next lowest type in the ranking order. Conversions prefer
+// converting to the next highest type in the ranking order.
+
+void HalfFloatDouble(double D);
+void HalfFloatDouble(float F);
+void HalfFloatDouble(half H);
+
+// CHECK: FunctionDecl {{.*}} used HalfFloatDouble 'void (double)'
+// CHECK: FunctionDecl {{.*}} used HalfFloatDouble 'void (float)'
+// CHECK: FunctionDecl {{.*}} used HalfFloatDouble 'void (half)'
+
+void FloatDouble(double D);
+void FloatDouble(float F);
+
+// CHECK: FunctionDecl {{.*}} used FloatDouble 'void (double)'
+// CHECK: FunctionDecl {{.*}} used FloatDouble 'void (float)'
+
+void HalfDouble(double D);
+void HalfDouble(half H);
+
+// CHECK: FunctionDecl {{.*}} used HalfDouble 'void (double)'
+// CHECK: FunctionDecl {{.*}} used HalfDouble 'void (half)'
+
+void HalfFloat(float F);
+void HalfFloat(half H);
+
+// CHECK: FunctionDecl {{.*}} used HalfFloat 'void (float)'
+// CHECK: FunctionDecl {{.*}} used HalfFloat 'void (half)'
+
+void Double(double D);
+void Float(float F);
+void Half(half H);
+
+// CHECK: FunctionDecl {{.*}} used Double 'void (double)'
+// CHECK: FunctionDecl {{.*}} used Float 'void (float)'
+// CHECK: FunctionDecl {{.*}} used Half 'void (half)'
+
+
+// Case 1: A function declared with overloads for half float and double types.
+//   (a) When called with half, it will resolve to half because half is an exact
+//   match.
+//   (b) When called with float it will resolve to float because float is an
+//   exact match.
+//   (c) When called with double it will resolve to double because it is an
+//   exact match.
+
+// CHECK: FunctionDecl {{.*}} Case1 'void (half, float, double)'
+void Case1(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'HalfFloatDouble' 'void (half)'
+  HalfFloatDouble(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'HalfFloatDouble' 'void (float)'
+  HalfFloatDouble(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'HalfFloatDouble' 'void (double)'
+  HalfFloatDouble(D);
+}
+
+// Case 2: A function declared with double and float overlaods.
+//   (a) When called with half, it will resolve to float because float is lower
+//   ranked than double.
+//   (b) When called with float it will resolve to float because float is an
+//   exact match.
+//   (c) When called with double it will resolve to double because it is an
+//   exact match.
+
+// CHECK: FunctionDecl {{.*}} Case2 'void (half, float, double)'
+void Case2(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'FloatDouble' 'void (float)'
+  FloatDouble(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'FloatDouble' 'void (float)'
+  FloatDouble(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'FloatDouble' 'void (double)'
+  FloatDouble(D);
+}
+
+// Case 3: A function declared with half and double overloads
+//   (a) When called with half, it will resolve to half because it is an exact
+//   match.
+//   (b) When called with flaot, it will resolve to double because double is a
+//   valid promotion.
+//   (c) When called with double, it will resolve to double because it is an
+//   exact match.
+
+// CHECK: FunctionDecl {{.*}} Case3 'void (half, float, double)'
+void Case3(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'HalfDouble' 'void (half)'
+  HalfDouble(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'HalfDouble' 'void (double)'
+  HalfDouble(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'HalfDouble' 'void (double)'
+  HalfDouble(D);
+}
+
+// Case 4: A function declared with half and float overloads.
+//   (a) When called with half, it will resolve to half because half is an exact
+//   match.
+//   (b) When called with float it will resolve to float because float is an
+//   exact match.
+//   (c) When called with double it will resolve to float because it is the
+//   float is higher rank than half.
+
+// CHECK: FunctionDecl {{.*}} Case4 'void (half, float, double)'
+void Case4(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'HalfFloat' 'void (half)'
+  HalfFloat(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'HalfFloat' 'void (float)'
+  HalfFloat(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'HalfFloat' 'void (float)'
+  HalfFloat(D); // expected-warning{{implicit conversion loses floating-point precision: 'double' to 'float'}}
+}
+
+// Case 5: A function declared with only a double overload.
+//   (a) When called with half, it will resolve to double because double is a
+//   valid promotion.
+//   (b) When called with float it will resolve to double because double is a
+//   valid promotion.
+//   (c) When called with double it will resolve to double because it is an
+//   exact match.
+
+// CHECK: FunctionDecl {{.*}} Case5 'void (half, float, double)'
+void Case5(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'Double' 'void (double)'
+  Double(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'Double' 'void (double)'
+  Double(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(double)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (double)' lvalue Function {{.*}} 'Double' 'void (double)'
+  Double(D);
+}
+
+// Case 6: A function declared with only a float overload.
+//   (a) When called with half, it will resolve to float because float is a
+//   valid promotion.
+//   (b) When called with float it will resolve to float because float is an
+//   exact match.
+//   (c) When called with double it will resolve to float because it is a
+//   valid conversion.
+
+// CHECK: FunctionDecl {{.*}} Case6 'void (half, float, double)'
+void Case6(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'Float' 'void (float)'
+  Float(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'Float' 'void (float)'
+  Float(F);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(float)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (float)' lvalue Function {{.*}} 'Float' 'void (float)'
+  Float(D); // expected-warning{{implicit conversion loses floating-point precision: 'double' to 'float'}}
+}
+
+// Case 7: A function declared with only a half overload.
+//   (a) When called with half, it will resolve to half because half is an
+//   exact match
+//   (b) When called with float it will resolve to half because half is a
+//   valid conversion.
+//   (c) When called with double it will resolve to float because it is a
+//   valid conversion.
+
+// CHECK: FunctionDecl {{.*}} Case7 'void (half, float, double)'
+void Case7(half H, float F, double D) {
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'Half' 'void (half)'
+  Half(H);
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'Half' 'void (half)'
+  Half(F); // expected-warning{{implicit conversion loses floating-point precision: 'float' to 'half'}}
+
+  // CHECK: CallExpr {{.*}} 'void'
+  // CHECK-NEXT: ImplicitCastExpr {{.*}} 'void (*)(half)' <FunctionToPointerDecay>
+  // CHECK-NEXT: DeclRefExpr {{.*}} 'void (half)' lvalue Function {{.*}} 'Half' 'void (half)'
+  Half(D); // expected-warning{{implicit conversion loses floating-point precision: 'double' to 'half'}}
+}