Add stencil blend mode as apparently the shaded mode uses indices whiles stencil uses mask/alpha..

src/r_compiler/fixedfunction/drawtrianglecodegen.cpp

@@ -581,10 +581,13 @@ SSAVec4i DrawTriangleCodegen::ProcessPixel32(SSAVec4i bg, SSAInt *varying)
 		fg = Sample32(uvoffset);
 		output = blend_revsub(shade_bgra_simple(fg, currentlight), bg, srcalpha, calc_blend_bgalpha(fg, destalpha));
 		break;
-	case TriBlendMode::Shaded:
+	case TriBlendMode::Stencil:
 		fg = Sample32(uvoffset);
 		output = blend_stencil(shade_bgra_simple(SSAVec4i::unpack(color), currentlight), fg[3], bg, srcalpha, destalpha);
 		break;
+	case TriBlendMode::Shaded:
+		output = blend_stencil(shade_bgra_simple(SSAVec4i::unpack(color), currentlight), Sample8(uvoffset), bg, srcalpha, destalpha);
+		break;
 	case TriBlendMode::TranslateCopy:
 		fg = TranslateSample32(uvoffset);
 		output = blend_copy(shade_bgra_simple(fg, currentlight));
@@ -662,6 +665,9 @@ SSAInt DrawTriangleCodegen::ProcessPixel8(SSAInt bg, SSAInt *varying)
 		fg = ToBgra(Shade8(Sample8(uvoffset)));
 		output = ToPal8(blend_revsub(fg, ToBgra(bg), srcalpha, calc_blend_bgalpha(fg, destalpha)));
 		break;
+	case TriBlendMode::Stencil:
+		output = ToPal8(blend_stencil(ToBgra(Shade8(color)), (Sample8(uvoffset) == SSAInt(0)).select(SSAInt(0), SSAInt(256)), ToBgra(bg), srcalpha, destalpha));
+		break;
 	case TriBlendMode::Shaded:
 		output = ToPal8(blend_stencil(ToBgra(Shade8(color)), Sample8(uvoffset), ToBgra(bg), srcalpha, destalpha));
 		break;

src/r_compiler/llvmdrawers.cpp

@@ -127,7 +127,7 @@ LLVMDrawers *LLVMDrawers::Instance()
 
 LLVMDrawersImpl::LLVMDrawersImpl()
 {
-	int version = 4; // Increment this number if the drawer codegen is modified (forces recreation of the module).
+	int version = 5; // Increment this number if the drawer codegen is modified (forces recreation of the module).
 	std::string targetCPU = mProgram.GetTargetCPU();
 	bool loaded = mProgram.LoadCachedModule(version, targetCPU);
 	if (!loaded)

src/r_compiler/llvmdrawers.h

@@ -281,7 +281,8 @@ enum class TriBlendMode
 	Add,            // blend_add(shade(fg), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
 	Sub,            // blend_sub(shade(fg), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
 	RevSub,         // blend_revsub(shade(fg), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
-	Shaded,         // blend_stencil(shade(color), fg, bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
+	Stencil,        // blend_stencil(shade(color), fg.a, bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
+	Shaded,         // blend_stencil(shade(color), fg.index, bg, srcalpha, calc_blend_bgalpha(fg, destalpha))
 	TranslateCopy,  // blend_copy(shade(translate(fg)))
 	TranslateAlphaBlend, // blend_alpha_blend(shade(translate(fg)), bg)
 	TranslateAdd,   // blend_add(shade(translate(fg)), bg, srcalpha, calc_blend_bgalpha(fg, destalpha))

src/r_poly_decal.cpp

@@ -130,6 +130,8 @@ void RenderPolyDecal::Render(const TriMatrix &worldToClip, DBaseDecal *decal, co
 
 	PolyDrawArgs args;
 	args.uniforms.flags = 0;
+	args.SetColormap(front->ColorMap);
+	args.SetTexture(tex, decal->Translation, true);
 	if (fullbrightSprite || fixedlightlev >= 0 || fixedcolormap)
 	{
 		args.uniforms.light = 256;
@@ -150,8 +152,8 @@ void RenderPolyDecal::Render(const TriMatrix &worldToClip, DBaseDecal *decal, co
 		args.uniforms.color = ((uint32_t)decal->AlphaColor) >> 24;
 	}
 	
-	args.uniforms.srcalpha = 256;
+	args.uniforms.srcalpha = (uint32_t)(decal->Alpha * 256.0 + 0.5);
-	args.uniforms.destalpha = 0;
+	args.uniforms.destalpha = 256 - args.uniforms.srcalpha;
 
 	args.objectToClip = &worldToClip;
 	args.vinput = vertices;
@@ -160,8 +162,6 @@ void RenderPolyDecal::Render(const TriMatrix &worldToClip, DBaseDecal *decal, co
 	args.ccw = true;
 	args.stenciltestvalue = 0;
 	args.stencilwritevalue = 1;
-	args.SetTexture(tex);
-	args.SetColormap(front->ColorMap);
 	//mode = R_SetPatchStyle (decal->RenderStyle, (float)decal->Alpha, decal->Translation, decal->AlphaColor);
 	PolyTriangleDrawer::draw(args, TriDrawVariant::DrawSubsector, TriBlendMode::Shaded);
 }

src/r_poly_sprite.cpp

@@ -185,14 +185,14 @@ void RenderPolySprite::Render(const TriMatrix &worldToClip, AActor *thing, subse
 		args.uniforms.destalpha = (uint32_t)(256 - thing->Alpha * 256);
 		args.uniforms.srcalpha = (uint32_t)(thing->Alpha * 256);
 		args.uniforms.color = 0xff000000 | thing->fillcolor;
-		blendmode = TriBlendMode::Shaded;
+		blendmode = TriBlendMode::Stencil;
 	}
 	else if (thing->RenderStyle == LegacyRenderStyles[STYLE_AddStencil])
 	{
 		args.uniforms.destalpha = 256;
 		args.uniforms.srcalpha = (uint32_t)(thing->Alpha * 256) * 2; // Don't know this needs to be multiplied by two..
 		args.uniforms.color = 0xff000000 | thing->fillcolor;
-		blendmode = TriBlendMode::Shaded;
+		blendmode = TriBlendMode::Stencil;
 	}
 	else if (thing->RenderStyle == LegacyRenderStyles[STYLE_Shaded])
 	{
@@ -215,6 +215,11 @@ void RenderPolySprite::Render(const TriMatrix &worldToClip, AActor *thing, subse
 		blendmode = args.translation ? TriBlendMode::TranslateAdd : TriBlendMode::Add;
 	}
 	
+	if (blendmode == TriBlendMode::Shaded)
+	{
+		args.SetTexture(tex, thing->Translation, true);
+	}
+	
 	if (!r_swtruecolor)
 	{
 		uint32_t r = (args.uniforms.color >> 16) & 0xff;

src/r_poly_triangle.h

@@ -67,7 +67,7 @@ public:
 		translation = nullptr;
 	}
 
-	void SetTexture(FTexture *texture, uint32_t translationID)
+	void SetTexture(FTexture *texture, uint32_t translationID, bool forcePal = false)
 	{
 		if (translationID != 0xffffffff && translationID != 0)
 		{
@@ -85,8 +85,17 @@ public:
 				return;
 			}
 		}
-
+		
-		SetTexture(texture);
+		if (forcePal)
+		{
+			textureWidth = texture->GetWidth();
+			textureHeight = texture->GetHeight();
+			texturePixels = texture->GetPixels();
+		}
+		else
+		{
+			SetTexture(texture);
+		}
 	}
 
 	void SetColormap(FSWColormap *base_colormap)