softgpu: Fix self-render detect in Ridge Racer.

When we flush we mark all pending writes zero, but we rely on this being
set to detect self-render.

TRANSFORM_ALL was wrong as well, sometimes clearing BINNER_RANGE.

GPU/Software/BinManager.cpp

@@ -197,16 +197,16 @@ void BinManager::UpdateState(bool throughMode) {
 			Flush("tex");
 
 		// Okay, now update what's pending.
-		constexpr uint32_t mirrorMask = 0x0FFFFFFF & ~0x00600000;
-		const uint32_t bpp = state.pixelID.FBFormat() == GE_FORMAT_8888 ? 4 : 2;
-		pendingWrites_[0].Expand(gstate.getFrameBufAddress() & mirrorMask, bpp, gstate.FrameBufStride(), scissorTL, scissorBR);
-		if (state.pixelID.depthWrite)
-			pendingWrites_[1].Expand(gstate.getDepthBufAddress() & mirrorMask, 2, gstate.DepthBufStride(), scissorTL, scissorBR);
+		MarkPendingWrites(state);
 
 		ClearDirty(SoftDirty::BINNER_RANGE);
 	} else if (pendingOverlap_) {
-		if (HasTextureWrite(state))
+		if (HasTextureWrite(state)) {
 			Flush("tex");
+
+			// We need the pending writes set, which flushing cleared.  Set them again.
+			MarkPendingWrites(state);
+		}
 	}
 
 	if (HasDirty(SoftDirty::BINNER_OVERLAP)) {
@@ -282,6 +282,17 @@ void BinManager::MarkPendingReads(const Rasterizer::RasterizerState &state) {
 	}
 }
 
+void BinManager::MarkPendingWrites(const Rasterizer::RasterizerState &state) {
+	DrawingCoords scissorTL(gstate.getScissorX1(), gstate.getScissorY1());
+	DrawingCoords scissorBR(std::min(gstate.getScissorX2(), gstate.getRegionX2()), std::min(gstate.getScissorY2(), gstate.getRegionY2()));
+
+	constexpr uint32_t mirrorMask = 0x0FFFFFFF & ~0x00600000;
+	const uint32_t bpp = state.pixelID.FBFormat() == GE_FORMAT_8888 ? 4 : 2;
+	pendingWrites_[0].Expand(gstate.getFrameBufAddress() & mirrorMask, bpp, gstate.FrameBufStride(), scissorTL, scissorBR);
+	if (state.pixelID.depthWrite)
+		pendingWrites_[1].Expand(gstate.getDepthBufAddress() & mirrorMask, 2, gstate.DepthBufStride(), scissorTL, scissorBR);
+}
+
 inline void BinDirtyRange::Expand(uint32_t newBase, uint32_t bpp, uint32_t stride, DrawingCoords &tl, DrawingCoords &br) {
 	const uint32_t w = br.x - tl.x + 1;
 	const uint32_t h = br.y - tl.y + 1;

GPU/Software/BinManager.h

@@ -267,6 +267,7 @@ private:
 	int mostThreads_ = 0;
 
 	void MarkPendingReads(const Rasterizer::RasterizerState &state);
+	void MarkPendingWrites(const Rasterizer::RasterizerState &state);
 	bool HasTextureWrite(const Rasterizer::RasterizerState &state);
 	BinCoords Scissor(BinCoords range);
 	BinCoords Range(const VertexData &v0, const VertexData &v1, const VertexData &v2);

GPU/Software/SoftGpu.h

@@ -64,17 +64,17 @@ enum class SoftDirty : uint64_t {
 	PIXEL_DITHER = 1ULL << 3,
 	PIXEL_WRITEMASK = 1ULL << 4,
 	PIXEL_CACHED = 1ULL << 5,
-	PIXEL_ALL = 63ULL << 0,
+	PIXEL_ALL = 0b111111ULL << 0,
 
 	SAMPLER_BASIC = 1ULL << 6,
 	SAMPLER_TEXLIST = 1ULL << 7,
 	SAMPLER_CLUT = 1ULL << 8,
-	SAMPLER_ALL = 7ULL << 6,
+	SAMPLER_ALL = 0b111ULL << 6,
 
 	RAST_BASIC = 1ULL << 9,
 	RAST_TEX = 1ULL << 10,
 	RAST_OFFSET = 1ULL << 11,
-	RAST_ALL = 7ULL << 9,
+	RAST_ALL = 0b111ULL << 9,
 
 	LIGHT_BASIC = 1ULL << 12,
 	LIGHT_MATERIAL = 1ULL << 13,
@@ -82,13 +82,13 @@ enum class SoftDirty : uint64_t {
 	LIGHT_1 = 1ULL << 15,
 	LIGHT_2 = 1ULL << 16,
 	LIGHT_3 = 1ULL << 17,
-	LIGHT_ALL = 63ULL << 12,
+	LIGHT_ALL = 0b111111ULL << 12,
 
 	TRANSFORM_BASIC = 1ULL << 18,
 	TRANSFORM_MATRIX = 1ULL << 19,
 	TRANSFORM_VIEWPORT = 1ULL << 20,
 	TRANSFORM_FOG = 1ULL << 21,
-	TRANSFORM_ALL = 31ULL << 18,
+	TRANSFORM_ALL = 0b1111ULL << 18,
 
 	BINNER_RANGE = 1ULL << 22,
 	BINNER_OVERLAP = 1ULL << 23,