Scope native barriers to shaders with shared execution

src/graphics/shader/recompiler/backend/spirv/spirvEmitterFlow.cpp

@@ -559,6 +559,10 @@ void EmitVoid(ValueEmitContext&) {}
 
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
+	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
+		// Independent graphics invocations have no native workgroup left to synchronize.
+		return;
+	}
 	const auto memory_scope = tessellation ? spv::ScopeInvocation : spv::ScopeWorkgroup;
 	const auto semantics    = tessellation ? spv::MemorySemanticsMaskNone
 	                                       : spv::MemorySemanticsAcquireReleaseMask |

src/graphics/shader/recompiler/backend/spirv/spirvEmitterMesh.cpp

@@ -1,4 +1,4 @@
-#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h"
+#include "graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h"
 
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
@@ -127,10 +127,7 @@ void EmitMeshEntryPoint(EmitterState& state) {
 	state.builder.AddFunction(spv::OpFunctionCall, TypeVoid(state), state.builder.AllocateId(),
 	                          state.mesh_guest_func);
 	// All guest waves finish before the uniform Vulkan allocation and output stores.
-	state.builder.AddFunction(spv::OpControlBarrier, ConstantU32(state, spv::ScopeWorkgroup),
+	EmitBarrier(state);
-	                          ConstantU32(state, spv::ScopeWorkgroup),
-	                          ConstantU32(state, spv::MemorySemanticsAcquireReleaseMask |
-	                                                 spv::MemorySemanticsWorkgroupMemoryMask));
 	const auto vertices   = MeshLoad(state, state.mesh_allocation, spv::StorageClassWorkgroup,
 	                                 TypeU32(state), ConstantU32(state, 0));
 	const auto primitives = MeshLoad(state, state.mesh_allocation, spv::StorageClassWorkgroup,

tests/shaderCfgTests.cpp

@@ -1679,45 +1679,14 @@ void TestNewShaderRecompilerSMovB32() {
   CheckSpirvBinaryValidates(result.spirv);
 }
 
-void TestNewShaderRecompilerSoppMarkers() {
+void TestShaderStageBarriers() {
   const uint32_t shader[] = {
-      EncodeSopp(0x00, 3),    // s_nop 3
-      EncodeSopp(0x0c, 0),    // s_waitcnt 0
-      EncodeSopp(0x10, 0x0f), // s_sendmsg 15
-      EncodeSopp(0x16, 0x2a), // s_ttracedata 42
-      EncodeSopp(0x20, 1),    // s_inst_prefetch 1
       EncodeSopp(0x0a, 0),    // s_barrier
       EncodeSopp(0x01, 0),    // s_endpgm
   };
 
-  auto options = MakeCompileOptions(ShaderType::Compute);
+  const auto result =
-  options.dump_ir = true;
+      RecompileForTest(shader, MakeCompileOptions(ShaderType::Compute));
-
-  auto result = RecompileForTest(shader, options);
-  Check((result.decoded_dump.find("s_nop 0x00000003") != std::string::npos),
-        "new decoder did not decode SOPP s_nop");
-  Check((result.decoded_dump.find("s_waitcnt 0x00000000") != std::string::npos),
-        "new decoder did not decode SOPP s_waitcnt");
-  Check((result.decoded_dump.find("s_sendmsg 0x0000000f") != std::string::npos),
-        "new decoder did not decode SOPP s_sendmsg");
-  Check((result.decoded_dump.find("s_ttracedata 0x0000002a") != std::string::npos),
-        "new decoder did not decode SOPP s_ttracedata");
-  Check((result.decoded_dump.find("s_inst_prefetch 0x00000001") != std::string::npos),
-        "new decoder did not decode SOPP s_inst_prefetch");
-  Check((result.decoded_dump.find("s_barrier") != std::string::npos),
-        "new decoder did not decode SOPP s_barrier");
-  Check((result.ir_dump.find("ControlNop null, 0x00000003") != std::string::npos),
-        "SOPP s_nop did not lower to an IR marker");
-  Check((result.ir_dump.find("Waitcnt null, 0x00000000") != std::string::npos),
-        "SOPP s_waitcnt did not lower to an IR marker");
-  Check((result.ir_dump.find("Sendmsg null, 0x0000000f") != std::string::npos),
-        "SOPP s_sendmsg did not lower to an IR marker");
-  Check((result.ir_dump.find("TtraceData null, 0x0000002a") != std::string::npos),
-        "SOPP s_ttracedata did not lower to an IR marker");
-  Check((result.ir_dump.find("InstPrefetch null, 0x00000001") != std::string::npos),
-        "SOPP s_inst_prefetch did not lower to an IR marker");
-  Check((result.ir_dump.find("Barrier null") != std::string::npos),
-        "SOPP s_barrier did not lower to an IR marker");
   Check(SpirvContainsOpcode(result.spirv, 224),
         "SPIR-V binary does not contain OpControlBarrier");
   Check(
@@ -1725,6 +1694,12 @@ void TestNewShaderRecompilerSoppMarkers() {
           result.spirv.end(),
       "SPIR-V barrier does not use workgroup acquire-release memory semantics");
   CheckSpirvBinaryValidates(result.spirv);
+
+  const auto vertex_result =
+      RecompileForTest(shader, MakeCompileOptions(ShaderType::Vertex));
+  CheckSpirvBinaryValidates(vertex_result.spirv);
+  Check(!SpirvContainsOpcode(vertex_result.spirv, 224),
+        "independent vertex invocations retained a workgroup barrier");
 }
 
 void TestNewShaderRecompilerSopkWaitcntMarkers() {
@@ -13443,6 +13418,7 @@ int main() {
   TestNewShaderRecompilerSpirvSizeBaselines();
   TestDemandDrivenSpirvDeclarations();
   TestNewShaderRecompilerSMovB32();
+  TestShaderStageBarriers();
   TestNewShaderRecompilerClipDisabledPosition();
   TestNewShaderRecompilerAuxPositionExports();
   TestNewShaderRecompilerNativeWideScalarMemoryIr();