Changed TlsData::get to read TLS slots directly from TEB on Windows x86/x64. (#461)

include/bx/inline/thread.inl

@@ -0,0 +1,73 @@
+/*
+ * Copyright 2010-2026 Branimir Karadzic. All rights reserved.
+ * License: https://github.com/bkaradzic/bx/blob/master/LICENSE
+ */
+
+#ifndef BX_THREAD_H_HEADER_GUARD
+#	error "Must be included from bx/thread.h!"
+#endif // BX_THREAD_H_HEADER_GUARD
+
+#define BX_TLS_DATA_TEB (BX_PLATFORM_WINDOWS && BX_CPU_X86 && !BX_CRT_NONE)
+
+#if BX_TLS_DATA_TEB
+#	if BX_COMPILER_MSVC
+#		if BX_ARCH_64BIT
+extern "C" unsigned __int64 __readgsqword(unsigned long _offset);
+#			pragma intrinsic(__readgsqword)
+#		else
+extern "C" unsigned long __readfsdword(unsigned long _offset);
+#			pragma intrinsic(__readfsdword)
+#		endif // BX_ARCH_64BIT
+#	endif // BX_COMPILER_MSVC
+
+namespace bx
+{
+	constexpr uint32_t kTebTlsSlotsOffset          = BX_ARCH_64BIT ? 0x1480 : 0xe10;
+	constexpr uint32_t kTebTlsExpansionSlotsOffset = BX_ARCH_64BIT ? 0x1780 : 0xf94;
+	constexpr uint32_t kTebTlsSlotsNum             = 64;
+
+#	if BX_COMPILER_MSVC && BX_ARCH_64BIT
+	inline void* readTeb(uint32_t _offset)
+	{
+		return (void*)__readgsqword(_offset);
+	}
+#	elif BX_COMPILER_MSVC
+	inline void* readTeb(uint32_t _offset)
+	{
+		return (void*)__readfsdword(_offset);
+	}
+#	elif BX_ARCH_64BIT
+	inline void* readTeb(uint32_t _offset)
+	{
+		void* ptr;
+		__asm__ volatile ("movq %%gs:(%1), %0" : "=r" (ptr) : "r" (uintptr_t(_offset) ) );
+		return ptr;
+	}
+#	else
+	inline void* readTeb(uint32_t _offset)
+	{
+		void* ptr;
+		__asm__ volatile ("movl %%fs:(%1), %0" : "=r" (ptr) : "r" (uintptr_t(_offset) ) );
+		return ptr;
+	}
+#	endif // BX_COMPILER_*
+
+	inline void* TlsData::get() const
+	{
+		const uint32_t slot = *(const uint32_t*)m_internal;
+
+		if (slot < kTebTlsSlotsNum) [[likely]]
+		{
+			return readTeb(kTebTlsSlotsOffset + slot*sizeof(void*) );
+		}
+
+		void** expansionSlots = (void**)readTeb(kTebTlsExpansionSlotsOffset);
+
+		return NULL == expansionSlots
+			? NULL
+			: expansionSlots[slot - kTebTlsSlotsNum]
+			;
+	}
+
+} // namespace bx
+#endif // BX_TLS_DATA_TEB

include/bx/thread.h

@@ -97,6 +97,8 @@ namespace bx
 
 } // namespace bx
 
+#include "inline/thread.inl"
+
 #endif
 
 #endif // BX_THREAD_H_HEADER_GUARD

src/thread.cpp

@@ -358,11 +358,13 @@ namespace bx
 		BX_ASSERT(0 != result, "Failed to free TLS index (err: 0x%08x).", GetLastError() ); BX_UNUSED(result);
 	}
 
+#if !BX_TLS_DATA_TEB
 	void* TlsData::get() const
 	{
 		TlsDataInternal* ti = (TlsDataInternal*)m_internal;
 		return TlsGetValue(ti->m_id);
 	}
+#endif // !BX_TLS_DATA_TEB
 
 	void TlsData::set(void* _ptr)
 	{

tests/thread_test.cpp

@@ -72,4 +72,97 @@ TEST_CASE("Thread", "")
 	}
 }
 
+static constexpr uint32_t kNumTlsData = 80;
+
+struct TlsDataCtx
+{
+	bx::TlsData* tls;
+	uint32_t     seed;
+	uint32_t     numErrors;
+};
+
+static int32_t tlsDataThread(bx::Thread* _thread, void* _userData)
+{
+	BX_UNUSED(_thread);
+
+	TlsDataCtx& ctx = *(TlsDataCtx*)_userData;
+
+	for (uint32_t ii = 0; ii < kNumTlsData; ++ii)
+	{
+		ctx.numErrors += NULL != ctx.tls[ii].get();
+		ctx.tls[ii].set( (void*)uintptr_t(ctx.seed + ii) );
+	}
+
+	for (uint32_t jj = 0; jj < 1000; ++jj)
+	{
+		for (uint32_t ii = 0; ii < kNumTlsData; ++ii)
+		{
+			ctx.numErrors += uintptr_t(ctx.seed + ii) != uintptr_t(ctx.tls[ii].get() );
+		}
+	}
+
+	return bx::kExitSuccess;
+}
+
+TEST_CASE("TlsData", "")
+{
+	bx::TlsData* tls = new bx::TlsData[kNumTlsData];
+
+	for (uint32_t ii = 0; ii < kNumTlsData; ++ii)
+	{
+		tls[ii].set( (void*)uintptr_t(0x1000 + ii) );
+	}
+
+	TlsDataCtx ctx[4];
+	bx::Thread thread[4];
+
+	for (uint32_t ii = 0; ii < BX_COUNTOF(thread); ++ii)
+	{
+		ctx[ii] = { tls, (ii + 1) << 16, 0 };
+		thread[ii].init(tlsDataThread, &ctx[ii]);
+	}
+
+	for (uint32_t ii = 0; ii < BX_COUNTOF(thread); ++ii)
+	{
+		thread[ii].shutdown();
+		REQUIRE(0 == ctx[ii].numErrors);
+	}
+
+	for (uint32_t ii = 0; ii < kNumTlsData; ++ii)
+	{
+		REQUIRE(uintptr_t(0x1000 + ii) == uintptr_t(tls[ii].get() ) );
+	}
+
+	delete [] tls;
+}
+
+static thread_local void* s_threadLocal;
+
+BX_NO_INLINE void* getTlsData(const bx::TlsData& _tls)
+{
+	return _tls.get();
+}
+
+BX_NO_INLINE void* getThreadLocal()
+{
+	return s_threadLocal;
+}
+
+TEST_CASE("TlsData-bench", "[.][!benchmark]")
+{
+	bx::TlsData tls;
+	tls.set(&tls);
+	s_threadLocal = &tls;
+
+	BENCHMARK("TlsData::get")
+	{
+		return getTlsData(tls);
+	};
+
+	BENCHMARK("thread_local")
+	{
+		return getThreadLocal();
+	};
+}
+
 #endif // BX_CONFIG_SUPPORTS_THREADING