optimizations (#200)

* Use direct file read and direct buffer allocation for small IoBuffer materalization
* Reduce range of materialized data in CompositeBuffer reading
CompressedBuffer header reading often only need a small part and not the whole file
* reduce lock contention in IoBuffer::Materialize
* Reduce parsing of compressed headers
Validate header type at decompression
* faster CreateDirectories - start from leaf going up and recurse back
* optimized BufferHeader::IsValid
* Add ValidateCompressedHeader to use when we don't need the actual compressed data
Validate that we always get compressed data in CidStore::AddChunk
* changelog

1 files changed, 33 insertions, 4 deletions

diff --git a/zencore/iobuffer.cpp b/zencore/iobuffer.cpp
index 16dd22a58..2522daf35 100644
--- a/zencore/iobuffer.cpp
+++ b/zencore/iobuffer.cpp
@@ -34,7 +34,7 @@ namespace zen {
 //////////////////////////////////////////////////////////////////////////
 
 void
-IoBufferCore::AllocateBuffer(size_t InSize, size_t Alignment)
+IoBufferCore::AllocateBuffer(size_t InSize, size_t Alignment) const
 {
 #if ZEN_PLATFORM_WINDOWS
 	if (((InSize & 0xffFF) == 0) && (Alignment == 0x10000))
@@ -238,7 +238,7 @@ IoBufferExtendedCore::~IoBufferExtendedCore()
 	m_DataPtr = nullptr;
 }
 
-static constexpr size_t MappingLockCount = 64;
+static constexpr size_t MappingLockCount = 128;
 static_assert(IsPow2(MappingLockCount), "MappingLockCount must be power of two");
 
 static RwLock g_MappingLocks[MappingLockCount];
@@ -247,7 +247,7 @@ static RwLock&
 MappingLockForInstance(const IoBufferExtendedCore* instance)
 {
 	intptr_t base		= (intptr_t)instance;
-	size_t	 lock_index = ((base >> 8) ^ (base >> 16)) & (MappingLockCount - 1u);
+	size_t	 lock_index = ((base >> 5) ^ (base >> 13)) & (MappingLockCount - 1u);
 	return g_MappingLocks[lock_index];
 }
 
@@ -271,9 +271,38 @@ IoBufferExtendedCore::Materialize() const
 
 	if (m_DataBytes == 0)
 	{
-		m_Flags.fetch_or(NewFlags, std::memory_order_release);
 		// Fake a "valid" pointer, nobody should read this as size is zero
 		m_DataPtr = reinterpret_cast<uint8_t*>(&m_MmapHandle);
+		m_Flags.fetch_or(NewFlags, std::memory_order_release);
+		return;
+	}
+
+	const size_t DisableMMapSizeLimit = 0x1000ull;
+
+	if (m_DataBytes < DisableMMapSizeLimit)
+	{
+		AllocateBuffer(m_DataBytes, sizeof(void*));
+		NewFlags |= kIsOwnedByThis;
+
+#if ZEN_PLATFORM_WINDOWS
+		OVERLAPPED Ovl{};
+
+		Ovl.Offset	   = DWORD(m_FileOffset & 0xffff'ffffu);
+		Ovl.OffsetHigh = DWORD(m_FileOffset >> 32);
+
+		DWORD dwNumberOfBytesRead = 0;
+		BOOL  Success			  = ::ReadFile(m_FileHandle, (void*)m_DataPtr, DWORD(m_DataBytes), &dwNumberOfBytesRead, &Ovl);
+
+		ZEN_ASSERT(Success);
+		ZEN_ASSERT(dwNumberOfBytesRead == m_DataBytes);
+#else
+		static_assert(sizeof(off_t) >= sizeof(uint64_t), "sizeof(off_t) does not support large files");
+		int	 Fd		   = int(uintptr_t(m_FileHandle));
+		int	 BytesRead = pread(Fd, (void*)m_DataPtr, m_DataBytes, m_FileOffset);
+		bool Success   = (BytesRead > 0);
+#endif	// ZEN_PLATFORM_WINDOWS
+
+		m_Flags.fetch_or(NewFlags, std::memory_order_release);
 		return;
 	}