GPU: Normalize framebuf addresses.

In VRAM, always store without mirror.  In RAM, always store without
cache/kernel bits.

Core/Debugger/MemBlockInfo.cpp

@@ -399,12 +399,18 @@ void FlushPendingMemInfo() {
 	pendingNotifyMaxAddr2 = 0;
 }
 
+static inline uint32_t NormalizeAddress(uint32_t addr) {
+	if ((addr & 0x3F000000) == 0x04000000)
+		return addr & 0x041FFFFF;
+	return addr & 0x3FFFFFFF;
+}
+
 void NotifyMemInfoPC(MemBlockFlags flags, uint32_t start, uint32_t size, uint32_t pc, const char *tagStr, size_t strLength) {
 	if (size == 0) {
 		return;
 	}
 	// Clear the uncached and kernel bits.
-	start &= ~0xC0000000;
+	start = NormalizeAddress(start);
 
 	bool needFlush = false;
 	// When the setting is off, we skip smaller info to keep things fast.
@@ -450,7 +456,7 @@ void NotifyMemInfo(MemBlockFlags flags, uint32_t start, uint32_t size, const cha
 }
 
 std::vector<MemBlockInfo> FindMemInfo(uint32_t start, uint32_t size) {
-	start &= ~0xC0000000;
+	start = NormalizeAddress(start);
 
 	if (pendingNotifyMinAddr1 < start + size && pendingNotifyMaxAddr1 >= start)
 		FlushPendingMemInfo();
@@ -466,7 +472,7 @@ std::vector<MemBlockInfo> FindMemInfo(uint32_t start, uint32_t size) {
 }
 
 std::vector<MemBlockInfo> FindMemInfoByFlag(MemBlockFlags flags, uint32_t start, uint32_t size) {
-	start &= ~0xC0000000;
+	start = NormalizeAddress(start);
 
 	if (pendingNotifyMinAddr1 < start + size && pendingNotifyMaxAddr1 >= start)
 		FlushPendingMemInfo();
@@ -486,7 +492,7 @@ std::vector<MemBlockInfo> FindMemInfoByFlag(MemBlockFlags flags, uint32_t start,
 }
 
 static const char *FindWriteTagByFlag(MemBlockFlags flags, uint32_t start, uint32_t size) {
-	start &= ~0xC0000000;
+	start = NormalizeAddress(start);
 
 	if (pendingNotifyMinAddr1 < start + size && pendingNotifyMaxAddr1 >= start)
 		FlushPendingMemInfo();

GPU/Common/FramebufferManagerCommon.cpp

@@ -112,7 +112,9 @@ void FramebufferManagerCommon::BeginFrame() {
 }
 
 void FramebufferManagerCommon::SetDisplayFramebuffer(u32 framebuf, u32 stride, GEBufferFormat format) {
-	displayFramebufPtr_ = framebuf;
+	displayFramebufPtr_ = framebuf & 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(displayFramebufPtr_))
+		displayFramebufPtr_ = framebuf & 0x041FFFFF;
 	displayStride_ = stride;
 	displayFormat_ = format;
 	GPUDebug::NotifyDisplay(framebuf, stride, format);
@@ -121,6 +123,8 @@ void FramebufferManagerCommon::SetDisplayFramebuffer(u32 framebuf, u32 stride, G
 
 VirtualFramebuffer *FramebufferManagerCommon::GetVFBAt(u32 addr) const {
 	addr &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(addr))
+		addr &= 0x041FFFFF;
 	VirtualFramebuffer *match = nullptr;
 	for (auto vfb : vfbs_) {
 		if (vfb->fb_address == addr) {
@@ -134,6 +138,9 @@ VirtualFramebuffer *FramebufferManagerCommon::GetVFBAt(u32 addr) const {
 }
 
 VirtualFramebuffer *FramebufferManagerCommon::GetExactVFB(u32 addr, int stride, GEBufferFormat format) const {
+	addr &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(addr))
+		addr &= 0x041FFFFF;
 	VirtualFramebuffer *newest = nullptr;
 	for (auto vfb : vfbs_) {
 		if (vfb->fb_address == addr && vfb->fb_stride == stride && vfb->fb_format == format) {
@@ -150,6 +157,9 @@ VirtualFramebuffer *FramebufferManagerCommon::GetExactVFB(u32 addr, int stride,
 }
 
 VirtualFramebuffer *FramebufferManagerCommon::ResolveVFB(u32 addr, int stride, GEBufferFormat format) {
+	addr &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(addr))
+		addr &= 0x041FFFFF;
 	// Find the newest one matching addr and stride.
 	VirtualFramebuffer *newest = nullptr;
 	for (auto vfb : vfbs_) {
@@ -235,7 +245,7 @@ void FramebufferManagerCommon::EstimateDrawingSize(u32 fb_address, int fb_stride
 		// The majority of the time, these are equal.  If not, let's check what we know.
 		u32 nearest_address = 0xFFFFFFFF;
 		for (auto vfb : vfbs_) {
-			const u32 other_address = vfb->fb_address & 0x3FFFFFFF;
+			const u32 other_address = vfb->fb_address;
 			if (other_address > fb_address && other_address < nearest_address) {
 				nearest_address = other_address;
 			}
@@ -296,10 +306,11 @@ void FramebufferManagerCommon::EstimateDrawingSize(u32 fb_address, int fb_stride
 }
 
 void GetFramebufferHeuristicInputs(FramebufferHeuristicParams *params, const GPUgstate &gstate) {
-	params->fb_address = (gstate.getFrameBufRawAddress() & 0x3FFFFFFF) | 0x04000000;  // GetFramebufferHeuristicInputs is only called from rendering, and thus, it's VRAM.
+	// GetFramebufferHeuristicInputs is only called from rendering, and thus, it's VRAM.
+	params->fb_address = gstate.getFrameBufRawAddress() | 0x04000000;
 	params->fb_stride = gstate.FrameBufStride();
 
-	params->z_address = (gstate.getDepthBufRawAddress() & 0x3FFFFFFF) | 0x04000000;
+	params->z_address = gstate.getDepthBufRawAddress() | 0x04000000;
 	params->z_stride = gstate.DepthBufStride();
 
 	if (params->z_address == params->fb_address) {
@@ -555,7 +566,7 @@ void FramebufferManagerCommon::SetDepthFrameBuffer(bool isClearingDepth) {
 	bool newlyUsingDepth = (currentRenderVfb_->usageFlags & FB_USAGE_RENDER_DEPTH) == 0;
 	currentRenderVfb_->usageFlags |= FB_USAGE_RENDER_DEPTH;
 
-	uint32_t boundDepthBuffer = gstate.getDepthBufAddress() & 0x3FFFFFFF;
+	uint32_t boundDepthBuffer = gstate.getDepthBufRawAddress();
 	if (currentRenderVfb_->z_address != boundDepthBuffer) {
 		WARN_LOG_N_TIMES(z_reassign, 5, G3D, "Framebuffer at %08x/%d has switched associated depth buffer from %08x to %08x, updating.",
 			currentRenderVfb_->fb_address, currentRenderVfb_->fb_stride, currentRenderVfb_->z_address, boundDepthBuffer);
@@ -1034,6 +1045,8 @@ void FramebufferManagerCommon::NotifyVideoUpload(u32 addr, int size, int stride,
 void FramebufferManagerCommon::UpdateFromMemory(u32 addr, int size) {
 	// Take off the uncached flag from the address. Not to be confused with the start of VRAM.
 	addr &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(addr))
+		addr &= 0x041FFFFF;
 	// TODO: Could go through all FBOs, but probably not important?
 	// TODO: Could also check for inner changes, but video is most important.
 	// TODO: This shouldn't care if it's a display framebuf or not, should work exactly the same.
@@ -1354,9 +1367,9 @@ void FramebufferManagerCommon::CopyDisplayToOutput(bool reallyDirty) {
 		// Let's search for a framebuf within this range. Note that we also look for
 		// "framebuffers" sitting in RAM (created from block transfer or similar) so we only take off the kernel
 		// and uncached bits of the address when comparing.
-		const u32 addr = fbaddr & 0x3FFFFFFF;
+		const u32 addr = fbaddr;
 		for (auto v : vfbs_) {
-			const u32 v_addr = v->fb_address & 0x3FFFFFFF;
+			const u32 v_addr = v->fb_address;
 			const u32 v_size = ColorBufferByteSize(v);
 
 			if (v->fb_format != displayFormat_ || v->fb_stride != displayStride_) {
@@ -1628,6 +1641,10 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 
 	dst &= 0x3FFFFFFF;
 	src &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(dst))
+		dst &= 0x041FFFFF;
+	if (Memory::IsVRAMAddress(src))
+		src &= 0x041FFFFF;
 
 	// TODO: Merge the below into FindTransferFramebuffer.
 	// Or at least this should be like the other ones, gathering possible candidates
@@ -1645,7 +1662,7 @@ bool FramebufferManagerCommon::NotifyFramebufferCopy(u32 src, u32 dst, int size,
 		}
 
 		// We only remove the kernel and uncached bits when comparing.
-		const u32 vfb_address = vfb->fb_address & 0x3FFFFFFF;
+		const u32 vfb_address = vfb->fb_address;
 		const u32 vfb_size = ColorBufferByteSize(vfb);
 		const u32 vfb_bpp = BufferFormatBytesPerPixel(vfb->fb_format);
 		const u32 vfb_byteStride = vfb->fb_stride * vfb_bpp;
@@ -1758,6 +1775,8 @@ std::string BlockTransferRect::ToString() const {
 // for depth data yet.
 bool FramebufferManagerCommon::FindTransferFramebuffer(u32 basePtr, int stride_pixels, int x_pixels, int y, int w_pixels, int h, int bpp, bool destination, BlockTransferRect *rect) {
 	basePtr &= 0x3FFFFFFF;
+	if (Memory::IsVRAMAddress(basePtr))
+		basePtr &= 0x041FFFFF;
 	rect->vfb = nullptr;
 
 	if (!stride_pixels) {
@@ -1781,11 +1800,11 @@ bool FramebufferManagerCommon::FindTransferFramebuffer(u32 basePtr, int stride_p
 		// Check for easily detected depth copies for logging purposes.
 		// Depth copies are not that useful though because you manually need to account for swizzle, so
 		// not sure if games will use them.
-		if ((vfb->z_address & 0x3FFFFFFF) == basePtr) {
+		if (vfb->z_address == basePtr) {
 			WARN_LOG_N_TIMES(z_xfer, 5, G3D, "FindTransferFramebuffer: found matching depth buffer, %08x (dest=%d, bpp=%d)", basePtr, (int)destination, bpp);
 		}
 
-		const u32 vfb_address = vfb->fb_address & 0x3FFFFFFF;
+		const u32 vfb_address = vfb->fb_address;
 		const u32 vfb_size = ColorBufferByteSize(vfb);
 
 		if (basePtr < vfb_address || basePtr >= vfb_address + vfb_size) {
@@ -1898,7 +1917,8 @@ VirtualFramebuffer *FramebufferManagerCommon::CreateRAMFramebuffer(uint32_t fbAd
 	// create a new one each frame.
 	VirtualFramebuffer *vfb = new VirtualFramebuffer{};
 	vfb->fbo = nullptr;
-	vfb->fb_address = fbAddress;  // NOTE - not necessarily in VRAM!
+	uint32_t mask = Memory::IsVRAMAddress(fbAddress) ? 0x041FFFFF : 0x3FFFFFFF;
+	vfb->fb_address = fbAddress & mask;  // NOTE - not necessarily in VRAM!
 	vfb->fb_stride = stride;
 	vfb->z_address = 0;  // marks that if anyone tries to render to this framebuffer, it should be dropped and recreated.
 	vfb->z_stride = 0;
@@ -2543,7 +2563,7 @@ void FramebufferManagerCommon::PackFramebufferSync(VirtualFramebuffer *vfb, int
 		return;
 	}
 
-	const u32 fb_address = vfb->fb_address & 0x3FFFFFFF;
+	const u32 fb_address = vfb->fb_address;
 
 	Draw::DataFormat destFormat = channel == RASTER_COLOR ? GEFormatToThin3D(vfb->fb_format) : GEFormatToThin3D(GE_FORMAT_DEPTH16);
 	const int dstBpp = (int)DataFormatSizeInBytes(destFormat);

GPU/Common/FramebufferManagerCommon.h

@@ -355,7 +355,9 @@ class FramebufferManagerCommon {
 
 	bool MayIntersectFramebuffer(u32 start) const {
 		// Clear the cache/kernel bits.
-		start = start & 0x3FFFFFFF;
+		start &= 0x3FFFFFFF;
+		if (Memory::IsVRAMAddress(start))
+			start &= 0x041FFFFF;
 		// Most games only have two framebuffers at the start.
 		if (start >= framebufRangeEnd_ || start < PSP_GetVidMemBase()) {
 			return false;

GPU/Common/TextureCacheCommon.cpp

@@ -688,7 +688,7 @@ bool TextureCacheCommon::GetBestFramebufferCandidate(const TextureDefinition &en
 		// Should avoid problems when pingponging two nearby buffers, like in Wipeout Pure in #15927.
 		if (candidate.channel == RASTER_COLOR &&
 			(candidate.match.yOffset != 0 || candidate.match.xOffset != 0) &&
-			(candidate.fb->fb_address & 0x1FFFFF) == (gstate.getFrameBufAddress() & 0x1FFFFF)) {
+			candidate.fb->fb_address == (gstate.getFrameBufRawAddress() | 0x04000000)) {
 			relevancy -= 2;
 		}
 
@@ -845,10 +845,8 @@ void TextureCacheCommon::HandleTextureChange(TexCacheEntry *const entry, const c
 }
 
 void TextureCacheCommon::NotifyFramebuffer(VirtualFramebuffer *framebuffer, FramebufferNotification msg) {
-	const u32 mirrorMask = 0x00600000;
 	const u32 fb_addr = framebuffer->fb_address;
-
-	const u32 z_addr = framebuffer->z_address & ~mirrorMask;  // Probably unnecessary.
+	const u32 z_addr = framebuffer->z_address;
 
 	const u32 fb_bpp = BufferFormatBytesPerPixel(framebuffer->fb_format);
 	const u32 z_bpp = 2;  // No other format exists.
@@ -931,7 +929,7 @@ bool TextureCacheCommon::MatchFramebuffer(
 	uint32_t fb_stride_in_bytes = fb_stride * BufferFormatBytesPerPixel(fb_format);
 	uint32_t tex_stride_in_bytes = entry.bufw * textureBitsPerPixel[entry.format] / 8;  // Note, we're looking up bits here so need to divide by 8.
 
-	u32 addr = fb_address & 0x3FFFFFFF;
+	u32 addr = fb_address;
 	u32 texaddr = entry.addr + texaddrOffset;
 
 	bool texInVRAM = Memory::IsVRAMAddress(texaddr);
@@ -943,10 +941,10 @@ bool TextureCacheCommon::MatchFramebuffer(
 	}
 
 	if (texInVRAM) {
-		const u32 mirrorMask = 0x00600000;
+		const u32 mirrorMask = 0x041FFFFF;
 
-		addr &= ~mirrorMask;
-		texaddr &= ~mirrorMask;
+		addr &= mirrorMask;
+		texaddr &= mirrorMask;
 	}
 
 	const bool noOffset = texaddr == addr;
@@ -1205,8 +1203,8 @@ void TextureCacheCommon::LoadClut(u32 clutAddr, u32 loadBytes) {
 
 	if (Memory::IsValidAddress(clutAddr)) {
 		if (Memory::IsVRAMAddress(clutAddr)) {
-			// Clear the uncached bit, etc. to match framebuffers.
-			const u32 clutLoadAddr = clutAddr & 0x3FFFFFFF;
+			// Clear the uncached and mirror bits, etc. to match framebuffers.
+			const u32 clutLoadAddr = clutAddr & 0x041FFFFF;
 			const u32 clutLoadEnd = clutLoadAddr + loadBytes;
 			static const u32 MAX_CLUT_OFFSET = 4096;
 
@@ -1217,7 +1215,7 @@ void TextureCacheCommon::LoadClut(u32 clutAddr, u32 loadBytes) {
 
 			VirtualFramebuffer *chosenFramebuffer = nullptr;
 			for (VirtualFramebuffer *framebuffer : framebuffers) {
-				const u32 fb_address = framebuffer->fb_address & 0x3FFFFFFF;
+				const u32 fb_address = framebuffer->fb_address;
 				const u32 fb_bpp = BufferFormatBytesPerPixel(framebuffer->fb_format);
 				int offset = clutLoadAddr - fb_address;
 

GPU/Directx9/FramebufferManagerDX9.cpp

@@ -140,7 +140,7 @@
 			return;
 		}
 
-		const u32 fb_address = vfb->fb_address & 0x3FFFFFFF;
+		const u32 fb_address = vfb->fb_address;
 		const int dstBpp = vfb->fb_format == GE_FORMAT_8888 ? 4 : 2;
 
 		// We always need to convert from the framebuffer native format.

GPU/GPUState.h

@@ -213,7 +213,7 @@ struct GPUgstate {
 	GEBufferFormat FrameBufFormat() const { return static_cast<GEBufferFormat>(framebufpixformat & 3); }
 	int FrameBufStride() const { return fbwidth&0x7FC; }
 	u32 getDepthBufRawAddress() const { return zbptr & 0x1FFFF0; }
-	u32 getDepthBufAddress() const { return 0x44000000 | getDepthBufRawAddress(); }
+	u32 getDepthBufAddress() const { return 0x44600000 | getDepthBufRawAddress(); }
 	int DepthBufStride() const { return zbwidth&0x7FC; }
 
 	// Pixel Pipeline

GPU/Software/BinManager.cpp

@@ -286,7 +286,7 @@ 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;
+	constexpr uint32_t mirrorMask = 0x041FFFFF;
 	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)
@@ -538,7 +538,7 @@ bool BinManager::HasPendingWrite(uint32_t start, uint32_t stride, uint32_t w, ui
 	if (!Memory::IsVRAMAddress(start))
 		return false;
 	// Ignore mirrors for overlap detection.
-	start &= 0x0FFFFFFF & ~0x00600000;
+	start &= 0x041FFFFF;
 
 	uint32_t size = stride * (h - 1) + w;
 	for (const auto &range : pendingWrites_) {
@@ -569,7 +569,7 @@ bool BinManager::HasPendingWrite(uint32_t start, uint32_t stride, uint32_t w, ui
 bool BinManager::HasPendingRead(uint32_t start, uint32_t stride, uint32_t w, uint32_t h) {
 	if (Memory::IsVRAMAddress(start)) {
 		// Ignore VRAM mirrors.
-		start &= 0x0FFFFFFF & ~0x00600000;
+		start &= 0x041FFFFF;
 	} else {
 		// Ignore only regular RAM mirrors.
 		start &= 0x3FFFFFFF;