gl: Fix context race condition when preparing optimized interpreter variants

- The main thread may bind the program before the compiler context has published the binding instructions
- Once bound, late bindings are not updated so the program remains with uninitialized bindings and rendering breaks

rpcs3/Emu/RSX/GL/GLShaderInterpreter.cpp

@@ -7,7 +7,6 @@
 
 #include "Emu/RSX/rsx_methods.h"
 #include "Emu/RSX/Overlays/Shaders/shader_loading_dialog.h"
-#include "Emu/RSX/Program/ShaderInterpreter.h"
 #include "Emu/RSX/Program/GLSLCommon.h"
 
 namespace gl
@@ -108,7 +107,8 @@ namespace gl
 			}
 
 			auto data = new interpreter::cached_program();
-			data->flags |= CACHED_PIPE_UNOPTIMIZED;
+			data->flags = base_pipeline->second->flags | CACHED_PIPE_UNOPTIMIZED;
+			data->build_compiler_options = base_pipeline->second->build_compiler_options;
 			data->allocator = base_pipeline->second->allocator;
 			data->vertex_shader = base_pipeline->second->vertex_shader;
 			data->fragment_shader = base_pipeline->second->fragment_shader;
@@ -203,6 +203,13 @@ namespace gl
 					m_current_interpreter->flags |= CACHED_PIPE_RECOMPILING;
 					build_program_async(opt, {});
 				}
+
+				if (m_current_interpreter->flags & CACHED_PIPE_UNINITIALIZED)
+				{
+					m_current_interpreter->prog->sync();
+					init_program(m_current_interpreter, m_current_interpreter->build_compiler_options);
+				}
+
 				return m_current_interpreter->prog.get();
 			}
 		}
@@ -491,7 +498,8 @@ namespace gl
 			attach(*data->fragment_shader).
 			link();
 
-		post_init_hook(data, compiler_options);
+		init_program(data, compiler_options);
+		store_program(data, compiler_options);
 		return data;
 	}
 
@@ -510,8 +518,11 @@ namespace gl
 
 		auto storage_hook = [=, this](std::unique_ptr<glsl::program>& prog)
 		{
+			// NOTE: We need to do the program bindings in the consumer's context, so we skip the post-init hook and just set a flag
+			// The consumer will handle initializing the bindings
+			// The synchronization problem doesn't matter much on Windows driver, but Mesa drivers actually care about it.
 			data->prog = std::move(prog);
-			post_init_hook(data, compiler_options);
+			store_program(data, compiler_options);
 
 			if (callback)
 			{
@@ -528,7 +539,7 @@ namespace gl
 		);
 	}
 
-	void shader_interpreter::post_init_hook(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options)
+	void shader_interpreter::init_program(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options)
 	{
 		data->prog->uniforms[0] = GL_STREAM_BUFFER_START + 0;
 		data->prog->uniforms[1] = GL_STREAM_BUFFER_START + 1;
@@ -551,6 +562,13 @@ namespace gl
 			}
 		}
 
+		data->flags &= ~CACHED_PIPE_UNINITIALIZED;
+	}
+
+	void shader_interpreter::store_program(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options)
+	{
+		data->build_compiler_options = compiler_options;
+
 		std::lock_guard lock(m_program_cache_lock);
 		m_program_cache[compiler_options] = data;
 	}

rpcs3/Emu/RSX/GL/GLShaderInterpreter.h

@@ -1,6 +1,7 @@
 #pragma once
 #include "glutils/program.h"
 #include "../Program/ProgramStateCache.h"
+#include "../Program/ShaderInterpreter.h"
 #include "../Common/TextureUtils.h"
 
 #include <unordered_map>
@@ -64,7 +65,11 @@ namespace gl
 
 		struct cached_program
 		{
-			u32 flags = 0;
+			u32 flags = program_common::interpreter::CACHED_PIPE_UNINITIALIZED;
+
+			// Compiler options mask - May not always match the storage compiler options in case of compatible pipelines
+			// However the storage mask must be a subset of this options mask
+			u32 build_compiler_options = 0;
 
 			std::shared_ptr<glsl::shader> vertex_shader;
 			std::shared_ptr<glsl::shader> fragment_shader;
@@ -92,7 +97,8 @@ namespace gl
 
 		std::shared_ptr<interpreter::cached_program> build_program(u64 compiler_options);
 		void build_program_async(u64 compiler_options, async_build_callback_t callback);
-		void post_init_hook(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options);
+		void init_program(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options);
+		void store_program(const std::shared_ptr<interpreter::cached_program>& data, u64 compiler_options);
 
 		std::shared_ptr<interpreter::cached_program> m_current_interpreter;
 

rpcs3/Emu/RSX/Program/ShaderInterpreter.h

@@ -47,8 +47,9 @@ namespace program_common
 
 		enum cached_pipeline_flags : u32
 		{
-			CACHED_PIPE_UNOPTIMIZED = (1 << 0),
+			CACHED_PIPE_UNOPTIMIZED   = (1 << 0),
-			CACHED_PIPE_RECOMPILING = (1 << 1),
+			CACHED_PIPE_RECOMPILING   = (1 << 1),
+			CACHED_PIPE_UNINITIALIZED = (1 << 2),
 		};
 
 		[[maybe_unused]] static std::string get_vertex_interpreter()