[HLSL] Convert vectors to bool for unary !

[llvm-beanz]

HLSL extends C++'s requirement that unary ! apply to boolean arguments and produces boolean results to apply to vectors. This change implements implicit conversion for non-boolean vector operands to boolean vectors.

I've noted this behavior in the issue tracking writing the HLSL specification section on unary operators
(microsoft/hlsl-specs#686).

Fixes #162913

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-hlsl

Author: Chris B (llvm-beanz)

Changes

HLSL extends C++'s requirement that unary ! apply to boolean arguments and produces boolean results to apply to vectors. This change implements implicit conversion for non-boolean vector operands to boolean vectors.

I've noted this behavior in the issue tracking writing the HLSL specification section on unary operators
(microsoft/hlsl-specs#686).

Fixes #162913

---

Full diff: https://github.com/llvm/llvm-project/pull/163857.diff

3 Files Affected:

• (modified) clang/lib/Sema/SemaExpr.cpp (+14)
• (added) clang/test/CodeGenHLSL/Operators/logical-not.hlsl (+33)
• (added) clang/test/SemaHLSL/Operators/logical-not.hlsl (+53)

diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
index 3e0e9bb5e39e5..e42d30cb5acc8 100644
--- a/clang/lib/Sema/SemaExpr.cpp
+++ b/clang/lib/Sema/SemaExpr.cpp
@@ -15944,6 +15944,20 @@ ExprResult Sema::CreateBuiltinUnaryOp(SourceLocation OpLoc,
             return ExprError(Diag(OpLoc, diag::err_typecheck_unary_expr)
                              << resultType << Input.get()->getSourceRange());
         }
+      } else if (Context.getLangOpts().HLSL && resultType->isVectorType() &&
+                 !resultType->hasBooleanRepresentation()) {
+        // HLSL unary logical not behaves like C++, which states that the
+        // operand is onverted to bool and the result is bool, however HLSL
+        // extends this property to vectors.
+        const VectorType *VTy = resultType->castAs<VectorType>();
+        resultType =
+            Context.getExtVectorType(Context.BoolTy, VTy->getNumElements());
+
+        Input = ImpCastExprToType(
+                    Input.get(), resultType,
+                    ScalarTypeToBooleanCastKind(VTy->getElementType()))
+                    .get();
+        break;
       } else if (resultType->isExtVectorType()) {
         if (Context.getLangOpts().OpenCL &&
             Context.getLangOpts().getOpenCLCompatibleVersion() < 120) {
diff --git a/clang/test/CodeGenHLSL/Operators/logical-not.hlsl b/clang/test/CodeGenHLSL/Operators/logical-not.hlsl
new file mode 100644
index 0000000000000..0f9d0677d8610
--- /dev/null
+++ b/clang/test/CodeGenHLSL/Operators/logical-not.hlsl
@@ -0,0 +1,33 @@
+// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.6-library -disable-llvm-passes -emit-llvm -finclude-default-header -fnative-half-type -o - %s | FileCheck %s
+
+// CHECK-LABEL: case1
+// CHECK: [[ToBool:%.*]] = icmp ne <2 x i32> {{.*}}, zeroinitializer
+// CHECK-NEXT: [[BoolCmp:%.*]] = icmp eq <2 x i1> [[ToBool]], zeroinitializer
+// CHECK-NEXT: {{.*}} = zext <2 x i1> [[BoolCmp]] to <2 x i32>
+export uint32_t2 case1(uint32_t2 b) {
+    return !b;
+}
+
+// CHECK-LABEL: case2
+// CHECK: [[ToBool:%.*]] = icmp ne <3 x i32> {{.*}}, zeroinitializer
+// CHECK-NEXT: [[BoolCmp:%.*]] = icmp eq <3 x i1> [[ToBool]], zeroinitializer
+// CHECK-NEXT: {{.*}} = zext <3 x i1> [[BoolCmp]] to <3 x i32>
+export int32_t3 case2(int32_t3 b) {
+    return !b;
+}
+
+// CHECK-LABEL: case3
+// CHECK: [[ToBool:%.*]] = fcmp reassoc nnan ninf nsz arcp afn une half {{.*}}, 0xH0000
+// CHECK-NEXT: [[BoolCmp:%.*]] = xor i1 [[ToBool]], true
+// CHECK-NEXT: {{.*}} = uitofp i1 [[BoolCmp]] to half
+export float16_t case3(float16_t b) {
+    return !b;
+}
+
+// CHECK-LABEL: case4
+// CHECK: [[ToBool:%.*]] = fcmp reassoc nnan ninf nsz arcp afn une <4 x float> {{.*}}, zeroinitializer
+// CHECK-NEXT: [[BoolCmp:%.*]] = icmp eq <4 x i1> [[ToBool]], zeroinitializer
+// CHECK-NEXT: {{.*}} = uitofp <4 x i1> [[BoolCmp]] to <4 x float>
+export float4 case4(float4 b) {
+    return !b;
+}
diff --git a/clang/test/SemaHLSL/Operators/logical-not.hlsl b/clang/test/SemaHLSL/Operators/logical-not.hlsl
new file mode 100644
index 0000000000000..d06ca3982be05
--- /dev/null
+++ b/clang/test/SemaHLSL/Operators/logical-not.hlsl
@@ -0,0 +1,53 @@
+// RUN: %clang_cc1 -finclude-default-header -triple  dxil-pc-shadermodel6.6-library %s -fnative-half-type -ast-dump -ast-dump-filter=case | FileCheck %s
+
+// CHECK-LABEL: FunctionDecl {{.*}} used case1 'uint32_t2 (uint32_t2)'
+// CHECK-NEXT: ParmVarDecl {{.*}} used b 'uint32_t2':'vector<uint32_t, 2>'
+// CHECK-NEXT: CompoundStmt
+// CHECK-NEXT: ReturnStmt
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<uint32_t, 2>' <IntegralCast>
+// CHECK-NEXT: UnaryOperator {{.*}} 'vector<bool, 2>' prefix '!' cannot overflow
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<bool, 2>' <IntegralToBoolean>
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'uint32_t2':'vector<uint32_t, 2>' <LValueToRValue>
+// CHECK-NEXT: DeclRefExpr {{.*}} 'uint32_t2':'vector<uint32_t, 2>' lvalue ParmVar {{.*}} 'b' 'uint32_t2':'vector<uint32_t, 2>'
+export uint32_t2 case1(uint32_t2 b) {
+    return !b;
+}
+
+// CHECK-LABEL: FunctionDecl {{.*}} used case2 'int32_t3 (int32_t3)'
+// CHECK-NEXT: ParmVarDecl {{.*}} used b 'int32_t3':'vector<int32_t, 3>'
+// CHECK-NEXT: CompoundStmt
+// CHECK-NEXT: ReturnStmt
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<int32_t, 3>' <IntegralCast>
+// CHECK-NEXT: UnaryOperator {{.*}} 'vector<bool, 3>' prefix '!' cannot overflow
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<bool, 3>' <IntegralToBoolean>
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'int32_t3':'vector<int32_t, 3>' <LValueToRValue>
+// CHECK-NEXT: DeclRefExpr {{.*}} 'int32_t3':'vector<int32_t, 3>' lvalue ParmVar {{.*}} 'b' 'int32_t3':'vector<int32_t, 3>'
+export int32_t3 case2(int32_t3 b) {
+    return !b;
+}
+
+// CHECK-LABEL: FunctionDecl {{.*}} used case3 'float16_t (float16_t)'
+// CHECK-NEXT: ParmVarDecl {{.*}} used b 'float16_t':'half'
+// CHECK-NEXT: CompoundStmt
+// CHECK-NEXT: ReturnStmt
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'float16_t':'half' <IntegralToFloating>
+// CHECK-NEXT: UnaryOperator {{.*}} 'bool' prefix '!' cannot overflow
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'bool' <FloatingToBoolean>
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'float16_t':'half' <LValueToRValue>
+// CHECK-NEXT: DeclRefExpr {{.*}} 'float16_t':'half' lvalue ParmVar {{.*}} 'b' 'float16_t':'half'
+export float16_t case3(float16_t b) {
+    return !b;
+}
+
+// CHECK-LABEL: FunctionDecl {{.*}} used case4 'float4 (float4)'
+// CHECK-NEXT: ParmVarDecl {{.*}} used b 'float4':'vector<float, 4>'
+// CHECK-NEXT: CompoundStmt
+// CHECK-NEXT: ReturnStmt
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<float, 4>' <IntegralToFloating>
+// CHECK-NEXT: UnaryOperator {{.*}} 'vector<bool, 4>' prefix '!' cannot overflow
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'vector<bool, 4>' <FloatingToBoolean>
+// CHECK-NEXT: ImplicitCastExpr {{.*}} 'float4':'vector<float, 4>' <LValueToRValue>
+// CHECK-NEXT: DeclRefExpr {{.*}} 'float4':'vector<float, 4>' lvalue ParmVar {{.*}} 'b' 'float4':'vector<float, 4>'
+export float4 case4(float4 b) {
+    return !b;
+}

[Icohedron]

LGTM. I tested a few of the affected DML operator tests with this patch and they passed!

[bogner]

I'm surprised that this is HLSL specific, though it obviously falls into the minutiae of a few different language extensions here.

Is the behaviour in C++ mode here intentional, or did it just follow the opencl behaviour from ext_vector? The change that introduced logical not for vector isn't very clear about this, but may be matching GCC, I'm not sure: https://reviews.llvm.org/D80979.

In any case, I think we should probably investigate whether we're doing the "right" thing in C++ mode separately from the change itself.

[llvm-beanz]

@erichkeane might be the right person to weigh in on the C/C++ side, but my guess is that C++ does what it currently does intentionally for compatibility with SSE where boolean true in vectors is expected to be all bits on (so sign-extending i1 gets the intended result).

[erichkeane]

We definitely do similar stuff for our ExtVector type, so I don't think of this as horribly inconsistent with the language. However, the base-vector type is 'owned' by GCC (Clang's version is the ExtVectorType), so I don't think we should be modifying the GCC version unless they are going to do the same. At that point, I don't think it should be HLSL specific.

That said, extending this 'splat-esque' behavior for ExtVectorType to HLSL should be non-controversial.