From d013582223d24619a302eca32cfa72f6284366ab Mon Sep 17 00:00:00 2001
From: wheremyfoodat <44909372+wheremyfoodat@users.noreply.github.com>
Date: Wed, 17 Jul 2024 21:53:19 +0300
Subject: [PATCH] Shadergen: Optimize caching

---
 include/PICA/pica_frag_config.hpp    |  5 +++--
 src/core/renderer_gl/renderer_gl.cpp | 20 +++++++++++++-------
 2 files changed, 16 insertions(+), 9 deletions(-)

diff --git a/include/PICA/pica_frag_config.hpp b/include/PICA/pica_frag_config.hpp
index 59f13757..9e13b3b5 100644
--- a/include/PICA/pica_frag_config.hpp
+++ b/include/PICA/pica_frag_config.hpp
@@ -24,8 +24,9 @@ namespace PICA {
 		u32 texUnitConfig;
 		u32 texEnvUpdateBuffer;
 
-		// There's 6 TEV stages, and each one is configured via 5 word-sized registers
-		std::array<u32, 5 * 6> tevConfigs;
+		// There's 6 TEV stages, and each one is configured via 4 word-sized registers
+		// (+ the constant color register, which we don't include here, otherwise we'd generate too many shaders)
+		std::array<u32, 4 * 6> tevConfigs;
 	};
 
 	// Config used for identifying unique fragment pipeline configurations
diff --git a/src/core/renderer_gl/renderer_gl.cpp b/src/core/renderer_gl/renderer_gl.cpp
index 0c33b898..249d8484 100644
--- a/src/core/renderer_gl/renderer_gl.cpp
+++ b/src/core/renderer_gl/renderer_gl.cpp
@@ -793,13 +793,19 @@ OpenGL::Program& RendererGL::getSpecializedShader() {
 	texConfig.texUnitConfig = regs[InternalRegs::TexUnitCfg];
 	texConfig.texEnvUpdateBuffer = regs[InternalRegs::TexEnvUpdateBuffer];
 
-	// Set up TEV stages
-	std::memcpy(&texConfig.tevConfigs[0 * 5], &regs[InternalRegs::TexEnv0Source], 5 * sizeof(u32));
-	std::memcpy(&texConfig.tevConfigs[1 * 5], &regs[InternalRegs::TexEnv1Source], 5 * sizeof(u32));
-	std::memcpy(&texConfig.tevConfigs[2 * 5], &regs[InternalRegs::TexEnv2Source], 5 * sizeof(u32));
-	std::memcpy(&texConfig.tevConfigs[3 * 5], &regs[InternalRegs::TexEnv3Source], 5 * sizeof(u32));
-	std::memcpy(&texConfig.tevConfigs[4 * 5], &regs[InternalRegs::TexEnv4Source], 5 * sizeof(u32));
-	std::memcpy(&texConfig.tevConfigs[5 * 5], &regs[InternalRegs::TexEnv5Source], 5 * sizeof(u32));
+	// Set up TEV stages. Annoyingly we can't just memcpy as the TEV registers are arranged like
+	// {Source, Operand, Combiner, Color, Scale} and we want to skip the color register since it's uploaded via UBO
+#define setupTevStage(stage)                                                                                    \
+	std::memcpy(&texConfig.tevConfigs[stage * 4], &regs[InternalRegs::TexEnv##stage##Source], 3 * sizeof(u32)); \
+	texConfig.tevConfigs[stage * 4 + 3] = regs[InternalRegs::TexEnv##stage##Source + 5];
+
+	setupTevStage(0);
+	setupTevStage(1);
+	setupTevStage(2);
+	setupTevStage(3);
+	setupTevStage(4);
+	setupTevStage(5);
+#undef setupTevStage
 
 	CachedProgram& programEntry = shaderCache[fsConfig];
 	OpenGL::Program& program = programEntry.program;