gl_arb_decompiler: Implement robust buffer operations

This emulates the behavior we get on GLSL with regular SSBOs with a
pointer + length pair. It aims to be consistent with the crashes we
might get.

Out of bounds stores are ignored. Atomics are ignored and return zero.
Reads return zero.

src/video_core/renderer_opengl/gl_arb_decompiler.cpp

@@ -376,9 +376,11 @@ private:
         std::string temporary = AllocTemporary();
         std::string address;
         std::string_view opname;
+        bool robust = false;
         if (const auto gmem = std::get_if<GmemNode>(&*operation[0])) {
             address = GlobalMemoryPointer(*gmem);
             opname = "ATOM";
+            robust = true;
         } else if (const auto smem = std::get_if<SmemNode>(&*operation[0])) {
             address = fmt::format("shared_mem[{}]", Visit(smem->GetAddress()));
             opname = "ATOMS";
@@ -386,7 +388,15 @@ private:
             UNREACHABLE();
             return "{0, 0, 0, 0}";
         }
+        if (robust) {
+            AddLine("IF NE.x;");
+        }
         AddLine("{}.{}.{} {}, {}, {};", opname, op, type, temporary, Visit(operation[1]), address);
+        if (robust) {
+            AddLine("ELSE;");
+            AddLine("MOV.S {}, 0;", temporary);
+            AddLine("ENDIF;");
+        }
         return temporary;
     }
 
@@ -980,10 +990,9 @@ void ARBDecompiler::DeclareLocalMemory() {
 }
 
 void ARBDecompiler::DeclareGlobalMemory() {
-    const std::size_t num_entries = ir.GetGlobalMemory().size();
+    const size_t num_entries = ir.GetGlobalMemory().size();
     if (num_entries > 0) {
-        const std::size_t num_vectors = Common::AlignUp(num_entries, 2) / 2;
+        AddLine("PARAM c[{}] = {{ program.local[0..{}] }};", num_entries, num_entries - 1);
-        AddLine("PARAM c[{}] = {{ program.local[0..{}] }};", num_vectors, num_vectors - 1);
     }
 }
 
@@ -1363,7 +1372,8 @@ std::string ARBDecompiler::Visit(const Node& node) {
 
     if (const auto gmem = std::get_if<GmemNode>(&*node)) {
         std::string temporary = AllocTemporary();
-        AddLine("LOAD.U32 {}, {};", temporary, GlobalMemoryPointer(*gmem));
+        AddLine("MOV {}, 0;", temporary);
+        AddLine("LOAD.U32 {} (NE.x), {};", temporary, GlobalMemoryPointer(*gmem));
         return temporary;
     }
 
@@ -1441,18 +1451,21 @@ std::string ARBDecompiler::BuildAoffi(Operation operation) {
 }
 
 std::string ARBDecompiler::GlobalMemoryPointer(const GmemNode& gmem) {
+    // Read a bindless SSBO, return its address and set CC accordingly
+    // address = c[binding].xy
+    // length  = c[binding].z
     const u32 binding = global_memory_names.at(gmem.GetDescriptor());
-    const char result_swizzle = binding % 2 == 0 ? 'x' : 'y';
 
     const std::string pointer = AllocLongVectorTemporary();
     std::string temporary = AllocTemporary();
 
-    const u32 local_index = binding / 2;
+    AddLine("PK64.U {}, c[{}];", pointer, binding);
-    AddLine("PK64.U {}, c[{}];", pointer, local_index);
     AddLine("SUB.U {}, {}, {};", temporary, Visit(gmem.GetRealAddress()),
             Visit(gmem.GetBaseAddress()));
     AddLine("CVT.U64.U32 {}.z, {};", pointer, temporary);
-    AddLine("ADD.U64 {}.x, {}.{}, {}.z;", pointer, pointer, result_swizzle, pointer);
+    AddLine("ADD.U64 {}.x, {}.x, {}.z;", pointer, pointer, pointer);
+    // Compare offset to length and set CC
+    AddLine("SLT.U.CC RC.x, {}, c[{}].z;", temporary, binding);
     return fmt::format("{}.x", pointer);
 }
 
@@ -1552,7 +1565,9 @@ std::string ARBDecompiler::Assign(Operation operation) {
         ResetTemporaries();
         return {};
     } else if (const auto gmem = std::get_if<GmemNode>(&*dest)) {
+        AddLine("IF NE.x;");
         AddLine("STORE.U32 {}, {};", Visit(src), GlobalMemoryPointer(*gmem));
+        AddLine("ENDIF;");
         ResetTemporaries();
         return {};
     } else {

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -139,16 +139,12 @@ void oglEnable(GLenum cap, bool state) {
     (state ? glEnable : glDisable)(cap);
 }
 
-void UpdateBindlessPointers(GLenum target, GLuint64EXT* pointers, std::size_t num_entries) {
+void UpdateBindlessSSBOs(GLenum target, const BindlessSSBO* ssbos, size_t num_ssbos) {
-    if (num_entries == 0) {
+    if (num_ssbos == 0) {
         return;
     }
-    if (num_entries % 2 == 1) {
+    glProgramLocalParametersI4uivNV(target, 0, static_cast<GLsizei>(num_ssbos),
-        pointers[num_entries] = 0;
+                                    reinterpret_cast<const GLuint*>(ssbos));
-    }
-    const GLsizei num_vectors = static_cast<GLsizei>((num_entries + 1) / 2);
-    glProgramLocalParametersI4uivNV(target, 0, num_vectors,
-                                    reinterpret_cast<const GLuint*>(pointers));
 }
 
 } // Anonymous namespace
@@ -900,11 +896,11 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
 }
 
 void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, Shader* shader) {
-    static constexpr std::array PARAMETER_LUT = {
+    static constexpr std::array PARAMETER_LUT{
         GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV,          GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
         GL_TESS_EVALUATION_PROGRAM_PARAMETER_BUFFER_NV, GL_GEOMETRY_PROGRAM_PARAMETER_BUFFER_NV,
-        GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV};
+        GL_FRAGMENT_PROGRAM_PARAMETER_BUFFER_NV,
-
+    };
     MICROPROFILE_SCOPE(OpenGL_UBO);
     const auto& stages = maxwell3d.state.shader_stages;
     const auto& shader_stage = stages[stage_index];
@@ -1007,8 +1003,8 @@ void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, Shader* sh
     const auto& cbufs{maxwell3d.state.shader_stages[stage_index]};
     const auto& entries{shader->GetEntries().global_memory_entries};
 
-    std::array<GLuint64EXT, 32> pointers;
+    std::array<BindlessSSBO, 32> ssbos;
-    ASSERT(entries.size() < pointers.size());
+    ASSERT(entries.size() < ssbos.size());
 
     const bool assembly_shaders = device.UseAssemblyShaders();
     u32 binding = assembly_shaders ? 0 : device.GetBaseBindings(stage_index).shader_storage_buffer;
@@ -1016,11 +1012,11 @@ void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, Shader* sh
         const GPUVAddr addr{cbufs.const_buffers[entry.cbuf_index].address + entry.cbuf_offset};
         const GPUVAddr gpu_addr{gpu_memory.Read<u64>(addr)};
         const u32 size{gpu_memory.Read<u32>(addr + 8)};
-        SetupGlobalMemory(binding, entry, gpu_addr, size, &pointers[binding]);
+        SetupGlobalMemory(binding, entry, gpu_addr, size, &ssbos[binding]);
         ++binding;
     }
     if (assembly_shaders) {
-        UpdateBindlessPointers(TARGET_LUT[stage_index], pointers.data(), entries.size());
+        UpdateBindlessSSBOs(TARGET_LUT[stage_index], ssbos.data(), entries.size());
     }
 }
 
@@ -1028,29 +1024,32 @@ void RasterizerOpenGL::SetupComputeGlobalMemory(Shader* kernel) {
     const auto& cbufs{kepler_compute.launch_description.const_buffer_config};
     const auto& entries{kernel->GetEntries().global_memory_entries};
 
-    std::array<GLuint64EXT, 32> pointers;
+    std::array<BindlessSSBO, 32> ssbos;
-    ASSERT(entries.size() < pointers.size());
+    ASSERT(entries.size() < ssbos.size());
 
     u32 binding = 0;
     for (const auto& entry : entries) {
         const GPUVAddr addr{cbufs[entry.cbuf_index].Address() + entry.cbuf_offset};
         const GPUVAddr gpu_addr{gpu_memory.Read<u64>(addr)};
         const u32 size{gpu_memory.Read<u32>(addr + 8)};
-        SetupGlobalMemory(binding, entry, gpu_addr, size, &pointers[binding]);
+        SetupGlobalMemory(binding, entry, gpu_addr, size, &ssbos[binding]);
         ++binding;
     }
     if (device.UseAssemblyShaders()) {
-        UpdateBindlessPointers(GL_COMPUTE_PROGRAM_NV, pointers.data(), entries.size());
+        UpdateBindlessSSBOs(GL_COMPUTE_PROGRAM_NV, ssbos.data(), ssbos.size());
     }
 }
 
 void RasterizerOpenGL::SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& entry,
-                                         GPUVAddr gpu_addr, std::size_t size,
+                                         GPUVAddr gpu_addr, size_t size, BindlessSSBO* ssbo) {
-                                         GLuint64EXT* pointer) {
+    const size_t alignment{device.GetShaderStorageBufferAlignment()};
-    const std::size_t alignment{device.GetShaderStorageBufferAlignment()};
     const auto info = buffer_cache.UploadMemory(gpu_addr, size, alignment, entry.is_written);
     if (device.UseAssemblyShaders()) {
-        *pointer = info.address + info.offset;
+        *ssbo = BindlessSSBO{
+            .address = static_cast<GLuint64EXT>(info.address + info.offset),
+            .length = static_cast<GLsizei>(size),
+            .padding = 0,
+        };
     } else {
         glBindBufferRange(GL_SHADER_STORAGE_BUFFER, binding, info.handle, info.offset,
                           static_cast<GLsizeiptr>(size));

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -53,6 +53,13 @@ namespace OpenGL {
 struct ScreenInfo;
 struct DrawParameters;
 
+struct BindlessSSBO {
+    GLuint64EXT address;
+    GLsizei length;
+    GLsizei padding;
+};
+static_assert(sizeof(BindlessSSBO) * CHAR_BIT == 128);
+
 class RasterizerOpenGL : public VideoCore::RasterizerAccelerated {
 public:
     explicit RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window, Tegra::GPU& gpu,
@@ -126,7 +133,7 @@ private:
 
     /// Configures a global memory buffer.
     void SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& entry, GPUVAddr gpu_addr,
-                           std::size_t size, GLuint64EXT* pointer);
+                           size_t size, BindlessSSBO* ssbo);
 
     /// Configures the current textures to use for the draw command.
     void SetupDrawTextures(std::size_t stage_index, Shader* shader);