[SPIR-V] Replace pixel shader SV_CullPrimitive inputs with 0

docs/ReleaseNotes.md

@@ -26,6 +26,8 @@ The included licenses apply to the following files:
   descriptor heaps with a literal `ArrayStride`. The value must be a power of
   two in `[8, 256]`. Both flags require `-spirv` and
   `-fspv-use-descriptor-heap` and are rejected otherwise.
+- `SV_CullPrimitive` as a pixel shader input is now replaced with 0, matching
+  DXIL, instead of producing invalid SPIR-V.
 
 Place release notes for the upcoming release below this line and remove this
 line upon naming the release. Refer to previous for appropriate section names.

tools/clang/lib/SPIRV/DeclResultIdMapper.cpp

@@ -3881,6 +3881,16 @@ bool DeclResultIdMapper::createStageVars(StageVarDataBundle &stageVarData,
       return false;
     }
 
+    // DXIL replaces a pixel shader SV_CullPrimitive input with 0, and
+    // CullPrimitiveEXT can't be used in fragment shaders. An explicit
+    // vk::builtin still takes precedence.
+    if (semanticKind == hlsl::Semantic::Kind::CullPrimitive &&
+        sigPointKind == hlsl::SigPoint::Kind::PSIn &&
+        !stageVarData.decl->hasAttr<VKBuiltInAttr>()) {
+      *value = spvBuilder.getConstantNull(stageVarData.type);
+      return true;
+    }
+
     // Special handling of certain mappings between HLSL semantics and
     // SPIR-V builtins:
     // * SV_CullDistance/SV_ClipDistance are outsourced to GlPerVertex.
@@ -4786,10 +4796,6 @@ SpirvVariable *DeclResultIdMapper::createSpirvStageVar(
   // be used as MSOut.
   case hlsl::Semantic::Kind::CullPrimitive: {
     switch (sigPointKind) {
-    case hlsl::SigPoint::Kind::PSIn:
-      stageVar->setIsSpirvBuiltin();
-      return spvBuilder.addStageBuiltinVar(type, sc, BuiltIn::CullPrimitiveEXT,
-                                           isPrecise, srcLoc);
     case hlsl::SigPoint::Kind::MSPOut:
       stageVar->setIsSpirvBuiltin();
       return spvBuilder.addStageBuiltinVar(type, sc, BuiltIn::CullPrimitiveEXT,

tools/clang/test/CodeGenSPIRV/meshshading.ext.cullprimitive.ps.hlsl

@@ -0,0 +1,42 @@
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -DCULL_TYPE=bool | FileCheck %s
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -DCULL_TYPE=uint | FileCheck %s
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -DCULL_TYPE=bool2 | FileCheck %s
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -DCULL_ARRAY | FileCheck %s
+
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -fcgl -DCULL_TYPE=bool | FileCheck %s --check-prefix=BOOL
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -fcgl -DCULL_TYPE=uint | FileCheck %s --check-prefix=UINT
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -fcgl -DCULL_TYPE=bool2 | FileCheck %s --check-prefix=BOOL2
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -fcgl -DCULL_ARRAY | FileCheck %s --check-prefix=ARRAY
+// RUN: %dxc -T ps_6_5 -E main %s -spirv -fcgl -DCULL_BUILTIN | FileCheck %s --check-prefix=BUILTIN
+
+// SV_CullPrimitive as a pixel shader input is replaced with 0, like DXIL.
+// DXIL doesn't check the type here, so aggregates are replaced too.
+
+// CHECK-NOT: CullPrimitiveEXT
+// CHECK: OpEntryPoint Fragment %main "main" %out_var_SV_Target
+// CHECK-NOT: CullPrimitiveEXT
+
+// BOOL: [[null:%[0-9]+]] = OpConstantNull %bool
+// BOOL: OpStore %param_var_cull [[null]]
+
+// UINT: [[null:%[0-9]+]] = OpConstantNull %uint
+// UINT: OpStore %param_var_cull [[null]]
+
+// BOOL2: [[null:%[0-9]+]] = OpConstantNull %v2bool
+// BOOL2: OpStore %param_var_cull [[null]]
+
+// ARRAY: [[null:%[0-9]+]] = OpConstantNull %_arr_bool_uint_2
+// ARRAY: OpStore %param_var_cull [[null]]
+
+// An explicit vk::builtin is not replaced.
+// BUILTIN-NOT: OpConstantNull
+// BUILTIN: OpLoad %bool %HelperInvocation
+
+#ifdef CULL_ARRAY
+float4 main(bool cull[2] : SV_CullPrimitive) : SV_Target { return cull[1]; }
+#elif defined(CULL_BUILTIN)
+float4 main([[vk::builtin("HelperInvocation")]] bool cull : SV_CullPrimitive)
+    : SV_Target { return cull; }
+#else
+float4 main(CULL_TYPE cull : SV_CullPrimitive) : SV_Target { return cull.x; }
+#endif

tools/clang/test/CodeGenSPIRV/meshshading.ext.cullprimitive.ps.shared-struct.hlsl

@@ -0,0 +1,64 @@
+// RUN: %dxc -T ms_6_5 -E ms_main %s -spirv -fspv-target-env=vulkan1.3 -DCULL_FIRST | FileCheck %s --check-prefix=MS
+// RUN: %dxc -T ms_6_5 -E ms_main %s -spirv -fspv-target-env=vulkan1.3 | FileCheck %s --check-prefix=MS
+// RUN: %dxc -T ps_6_5 -E ps_main %s -spirv -DCULL_FIRST | FileCheck %s --check-prefix=PS
+// RUN: %dxc -T ps_6_5 -E ps_main %s -spirv | FileCheck %s --check-prefix=PS
+// RUN: %dxc -T ps_6_5 -E ps_main %s -spirv -fcgl -DCULL_FIRST | FileCheck %s --check-prefix=FIRST
+// RUN: %dxc -T ps_6_5 -E ps_main %s -spirv -fcgl | FileCheck %s --check-prefix=LATER
+
+// The other members keep the mesh shader's locations, and the 0 lands in the
+// right member, whether SV_CullPrimitive is first or not.
+
+// MS: OpDecorate {{%[0-9]+}} BuiltIn CullPrimitiveEXT
+// MS-DAG: OpDecorate %out_var_PRIMID Location 0
+// MS-DAG: OpDecorate %out_var_COLOR0 Location 1
+
+// PS-NOT: CullPrimitiveEXT
+// PS-DAG: OpDecorate %in_var_PRIMID Location 0
+// PS-DAG: OpDecorate %in_var_COLOR0 Location 1
+// PS-NOT: CullPrimitiveEXT
+
+// FIRST: [[null:%[0-9]+]] = OpConstantNull %bool
+// FIRST: [[id:%[0-9]+]] = OpLoad %uint %in_var_PRIMID
+// FIRST: [[col:%[0-9]+]] = OpLoad %v4float %in_var_COLOR0
+// FIRST: OpCompositeConstruct %Prim [[null]] [[id]] [[col]]
+
+// LATER: [[null:%[0-9]+]] = OpConstantNull %bool
+// LATER: [[id:%[0-9]+]] = OpLoad %uint %in_var_PRIMID
+// LATER: [[col:%[0-9]+]] = OpLoad %v4float %in_var_COLOR0
+// LATER: OpCompositeConstruct %Prim [[id]] [[null]] [[col]]
+
+#ifdef CULL_FIRST
+struct Prim {
+  bool cull : SV_CullPrimitive;
+  uint id : PRIMID;
+  float4 col : COLOR0;
+};
+#else
+struct Prim {
+  uint id : PRIMID;
+  bool cull : SV_CullPrimitive;
+  float4 col : COLOR0;
+};
+#endif
+
+struct Vert {
+  float4 pos : SV_Position;
+};
+
+[outputtopology("triangle")]
+[numthreads(1, 1, 1)]
+void ms_main(out vertices Vert verts[3], out indices uint3 tris[1],
+             out primitives Prim prims[1]) {
+  SetMeshOutputCounts(3, 1);
+  verts[0].pos = 0;
+  verts[1].pos = 0;
+  verts[2].pos = 0;
+  tris[0] = uint3(0, 1, 2);
+  prims[0].cull = false;
+  prims[0].id = 7;
+  prims[0].col = 1;
+}
+
+float4 ps_main(Vert v, Prim p) : SV_Target {
+  return p.cull ? 0 : p.col * p.id;
+}