Get rid of the bool, not worth it.

Core/Util/PPGeDraw.cpp

@@ -1327,7 +1327,7 @@ bool PPGeImage::Load() {
 	unsigned char *textureData;
 	int success;
 	if (filename_.empty()) {
-		success = pngLoadPtr(Memory::GetPointerRange(png_, size_), size_, &width_, &height_, &textureData);
+		success = pngLoadPtr(Memory::GetPointerRange(png_, (u32)size_), size_, &width_, &height_, &textureData);
 	} else {
 		std::vector<u8> pngData;
 		if (pspFileSystem.ReadEntireFile(filename_, pngData) < 0) {

GPU/Common/FragmentShaderGenerator.cpp

@@ -260,7 +260,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				if (texFunc == GE_TEXFUNC_BLEND) {
 					WRITE(p, "float3 u_texenv : register(c%i);\n", CONST_PS_TEXENV);
 				}
-				WRITE(p, "bool u_texAlpha : register(b%i);\n", CONST_PS_TEXALPHA);  // NOTE! "b" register, not "c"!
+				WRITE(p, "float u_texNoAlpha : register(c%i);\n", CONST_PS_TEX_NO_ALPHA);
 			}
 			WRITE(p, "float3 u_fogcolor : register(c%i);\n", CONST_PS_FOGCOLOR);
 			if (texture3D) {
@@ -354,7 +354,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				WRITE(p, "uniform sampler2D tex;\n");
 			}
 			*uniformMask |= DIRTY_TEXALPHA;
-			WRITE(p, "uniform bool u_texAlpha;\n");
+			WRITE(p, "uniform float u_texNoAlpha;\n");
 		}
 
 		if (readFramebufferTex) {
@@ -824,7 +824,7 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 			WRITE(p, "  vec4 p = v_color0;\n");
 
 			if (texFunc != GE_TEXFUNC_REPLACE) {
-				WRITE(p, "  if (!u_texAlpha) { t.a = 1.0; }\n");
+				WRITE(p, "  t.a = max(t.a, u_texNoAlpha);\n");
 			}
 
 			switch (texFunc) {
@@ -838,7 +838,9 @@ bool GenerateFragmentShader(const FShaderID &id, char *buffer, const ShaderLangu
 				WRITE(p, "  vec4 v = vec4(mix(p.rgb, u_texenv.rgb, t.rgb), p.a * t.a) + s;\n");
 				break;
 			case GE_TEXFUNC_REPLACE:
-				WRITE(p, "  vec4 v = (u_texAlpha ? t : vec4(t.rgb, p.a)) + s;\n");
+				WRITE(p, "  vec4 r = t;\n");
+				WRITE(p, "  r.a = mix(r.a, p.a, u_texNoAlpha);\n");
+				WRITE(p, "  vec4 v = r + s;\n");
 				break;
 			case GE_TEXFUNC_ADD:
 			case GE_TEXFUNC_UNKNOWN1:

GPU/Common/FragmentShaderGenerator.h

@@ -36,12 +36,12 @@ struct FShaderID;
 #define CONST_PS_TEXCLAMP 8
 #define CONST_PS_TEXCLAMPOFF 9
 #define CONST_PS_MIPBIAS 10
+#define CONST_PS_TEX_NO_ALPHA 11
 
 // For stencil upload
-#define BCONST_PS_STENCILVALUE 11
+#define BCONST_PS_STENCILVALUE 12
 
 // D3D9 bool constants, they have their own register space.
-#define CONST_PS_TEXALPHA 0
 
 
 // Can technically be deduced from the fragment shader ID, but this is safer.

GPU/Common/ShaderUniforms.cpp

@@ -199,7 +199,7 @@ void BaseUpdateUniforms(UB_VS_FS_Base *ub, uint64_t dirtyUniforms, bool flipView
 	}
 
 	if (dirtyUniforms & DIRTY_TEXALPHA) {
-		ub->texAlpha = gstate.isTextureAlphaUsed() ? 1 : 0;
+		ub->texNoAlpha = gstate.isTextureAlphaUsed() ? 0.0f : 1.0f;
 	}
 
 	if (dirtyUniforms & DIRTY_STENCILREPLACEVALUE) {

GPU/Common/ShaderUniforms.h

@@ -17,7 +17,7 @@ enum : uint64_t {
 	DIRTY_MATDIFFUSE | DIRTY_MATSPECULAR | DIRTY_MATEMISSIVE | DIRTY_AMBIENT,
 };
 
-// Currently 448 bytes.
+// Currently 496 bytes.
 // Every line here is a 4-float.
 struct alignas(16) UB_VS_FS_Base {
 	float proj[16];
@@ -39,7 +39,7 @@ struct alignas(16) UB_VS_FS_Base {
 	float blendFixB[3]; float rotation;
 	float texClamp[4];
 	float texClampOffset[2]; float fogCoef[2];
-	uint32_t texAlpha; float pad[3];
+	float texNoAlpha; float pad[3];
 	// VR stuff is to go here, later. For normal drawing, we can then get away
 	// with just uploading the first 448 bytes of the struct (up to and including fogCoef).
 };
@@ -66,7 +66,7 @@ R"(  mat4 u_proj;
   vec4 u_texclamp;
   vec2 u_texclampoff;
   vec2 u_fogcoef;
-  bool u_texAlpha; float pad0; float pad1; float pad2;
+  float u_texNoAlpha; float pad0; float pad1; float pad2;
 )";
 
 // 512 bytes. Would like to shrink more. Some colors only have 8-bit precision and we expand

GPU/Directx9/ShaderManagerDX9.cpp

@@ -285,7 +285,8 @@ void ShaderManagerDX9::PSUpdateUniforms(u64 dirtyUniforms) {
 		PSSetFloat(CONST_PS_STENCILREPLACE, (float)gstate.getStencilTestRef() * (1.0f / 255.0f));
 	}
 	if (dirtyUniforms & DIRTY_TEXALPHA) {
-		PSSetBool(CONST_PS_TEXALPHA, gstate.isTextureAlphaUsed());
+		// NOTE: Reversed value, more efficient in shader.
+		PSSetFloat(CONST_PS_TEX_NO_ALPHA, gstate.isTextureAlphaUsed() ? 0.0f : 1.0f);
 	}
 	if (dirtyUniforms & DIRTY_SHADERBLEND) {
 		PSSetColorUniform3(CONST_PS_BLENDFIXA, gstate.getFixA());

GPU/GLES/ShaderManagerGLES.cpp

@@ -152,7 +152,7 @@ LinkedShader::LinkedShader(GLRenderManager *render, VShaderID VSID, Shader *vs,
 	queries.push_back({ &u_uvscaleoffset, "u_uvscaleoffset" });
 	queries.push_back({ &u_texclamp, "u_texclamp" });
 	queries.push_back({ &u_texclampoff, "u_texclampoff" });
-	queries.push_back({ &u_texAlpha, "u_texAlpha" });
+	queries.push_back({ &u_texNoAlpha, "u_texNoAlpha" });
 	queries.push_back({ &u_lightControl, "u_lightControl" });
 
 	for (int i = 0; i < 4; i++) {
@@ -446,7 +446,7 @@ void LinkedShader::UpdateUniforms(const ShaderID &vsid, bool useBufferedRenderin
 		SetColorUniform3(render_, &u_texenv, gstate.texenvcolor);
 	}
 	if (dirty & DIRTY_TEXALPHA) {
-		SetBoolUniform(render_, &u_texAlpha, gstate.isTextureAlphaUsed());
+		render_->SetUniformF1(&u_texNoAlpha, gstate.isTextureAlphaUsed() ? 0.0f : 1.0f);
 	}
 	if (dirty & DIRTY_ALPHACOLORREF) {
 		if (shaderLanguage.bitwiseOps) {

GPU/GLES/ShaderManagerGLES.h

@@ -101,7 +101,7 @@ public:
 	int u_uvscaleoffset;
 	int u_texclamp;
 	int u_texclampoff;
-	int u_texAlpha;
+	int u_texNoAlpha;
 
 	// Lighting
 	int u_lightControl;