diff --git a/azure-pipelines.yml b/azure-pipelines.yml
index 377bedbf..72783400 100644
--- a/azure-pipelines.yml
+++ b/azure-pipelines.yml
@@ -46,6 +46,14 @@ jobs:
         BuildType: release
         cmakeExtraArgs: -DCMAKE_BUILD_TYPE=Release -A Win32
         MSBuildConfiguration: Release
+      Debug Fixed TLS:
+        BuildType: debug
+        cmakeExtraArgs: -DCMAKE_BUILD_TYPE=Debug -DMI_DEBUG_FULL=ON -DMI_WIN_USE_FIXED_TLS=ON
+        MSBuildConfiguration: Debug
+      Release Fixed TLS:
+        BuildType: release
+        cmakeExtraArgs: -DCMAKE_BUILD_TYPE=Release -DMI_WIN_USE_FIXED_TLS=ON
+        MSBuildConfiguration: Release
   steps:
   - task: CMake@1
     inputs:
diff --git a/ide/vs2022/mimalloc-test-stress.vcxproj b/ide/vs2022/mimalloc-test-stress.vcxproj
index 128a4ff6..d6af71ce 100644
--- a/ide/vs2022/mimalloc-test-stress.vcxproj
+++ b/ide/vs2022/mimalloc-test-stress.vcxproj
@@ -282,8 +282,8 @@
     </ClCompile>
   </ItemGroup>
   <ItemGroup>
-    <ProjectReference Include="mimalloc-override-dll.vcxproj">
-      <Project>{abb5eae7-b3e6-432e-b636-333449892ea7}</Project>
+    <ProjectReference Include="mimalloc-lib.vcxproj">
+      <Project>{abb5eae7-b3e6-432e-b636-333449892ea6}</Project>
     </ProjectReference>
   </ItemGroup>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
diff --git a/include/mimalloc/prim.h b/include/mimalloc/prim.h
index 56b73544..ab085c8b 100644
--- a/include/mimalloc/prim.h
+++ b/include/mimalloc/prim.h
@@ -210,19 +210,18 @@ static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexce
 
 // On windows we can store the thread-local heap at a fixed TLS slot to avoid
 // thread-local initialization checks in the fast path.
-// We always use the second user TLS slot (the first one is always allocated already),
-// and at initialization (`windows/prim.c`) we call TlsAlloc and verify
-// we indeed get the second slot (and fail otherwise).
-// Todo: we could make the Tls slot completely dynamic but that would require
-// an extra read of the static Tls slot instead of using a constant offset.
+// We allocate a user TLS slot at process initialization (see `windows/prim.c`)
+// and store the offset `_mi_win_tls_offset`.
 #define MI_HAS_TLS_SLOT      2              // 2 = we can reliably initialize the slot (saving a test on each malloc)
 
+extern mi_decl_hidden size_t _mi_win_tls_offset;
+
 #if MI_WIN_USE_FIXED_TLS > 1
 #define MI_TLS_SLOT     (MI_WIN_USE_FIXED_TLS)
 #elif MI_SIZE_SIZE == 4
-#define MI_TLS_SLOT     (0x0E18)            // Second User TLS slot <https://en.wikipedia.org/wiki/Win32_Thread_Information_Block>
+#define MI_TLS_SLOT     (0x0E10 + _mi_win_tls_offset)  // User TLS slots <https://en.wikipedia.org/wiki/Win32_Thread_Information_Block>
 #else
-#define MI_TLS_SLOT     (0x1488)            // Second User TLS slot <https://en.wikipedia.org/wiki/Win32_Thread_Information_Block>
+#define MI_TLS_SLOT     (0x1480 + _mi_win_tls_offset)  // User TLS slots <https://en.wikipedia.org/wiki/Win32_Thread_Information_Block>
 #endif
 
 static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
@@ -271,9 +270,6 @@ static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexce
 
 
 // defined in `init.c`; do not use these directly
-#ifdef _MSC_VER
-__declspec(selectany)  // make it part of the comdat section to have faster TLS access (issue #1078)
-#endif
 extern mi_decl_hidden mi_decl_thread mi_heap_t* _mi_heap_default;  // default heap to allocate from
 extern mi_decl_hidden bool _mi_process_is_initialized;             // has mi_process_init been called?
 
diff --git a/src/prim/windows/prim.c b/src/prim/windows/prim.c
index 96259040..780cae18 100644
--- a/src/prim/windows/prim.c
+++ b/src/prim/windows/prim.c
@@ -646,22 +646,29 @@ bool _mi_prim_random_buf(void* buf, size_t buf_len) {
 // Process & Thread Init/Done
 //----------------------------------------------------------------
 
+#if MI_WIN_USE_FIXED_TLS==1
+mi_decl_cache_align size_t _mi_win_tls_offset = sizeof(void*);  // use 2nd slot by default
+#endif
+
 static void NTAPI mi_win_main(PVOID module, DWORD reason, LPVOID reserved) {
   MI_UNUSED(reserved);
   MI_UNUSED(module);
-  #if MI_TLS_SLOT >= 2
+  #if MI_HAS_TLS_SLOT >= 2  // we must initialize the TLS slot before any allocation
+  #if MI_WIN_USE_FIXED_TLS==1
   if (reason==DLL_PROCESS_ATTACH) {
     const DWORD tls_slot = TlsAlloc();
-    if (tls_slot != 1) { 
-      _mi_error_message(EFAULT, "unable to allocate the second TLS slot (rebuild without MI_WIN_USE_FIXED_TLS?)\n"); 
+    if (tls_slot == TLS_OUT_OF_INDEXES) { 
+      _mi_error_message(EFAULT, "unable to allocate the a TLS slot (rebuild without MI_WIN_USE_FIXED_TLS?)\n"); 
     }
+    _mi_win_tls_offset = (size_t)tls_slot * sizeof(void*);
   }
+  #endif
   if (reason==DLL_PROCESS_ATTACH || reason==DLL_THREAD_ATTACH) {
     if (mi_prim_get_default_heap() == NULL) {
       _mi_heap_set_default_direct((mi_heap_t*)&_mi_heap_empty);
     }
-    #if MI_DEBUG
-    void* const p = TlsGetValue(1);
+    #if MI_DEBUG && MI_WIN_USE_FIXED_TLS==1
+    void* const p = TlsGetValue((DWORD)(_mi_win_tls_offset / sizeof(void*)));
     mi_assert_internal(p == (void*)&_mi_heap_empty);
     #endif  
   }
@@ -827,7 +834,7 @@ static void NTAPI mi_win_main(PVOID module, DWORD reason, LPVOID reserved) {
   #endif
   mi_decl_export void _mi_redirect_entry(DWORD reason) {
     // called on redirection; careful as this may be called before DllMain
-    #if MI_TLS_SLOT >= 2
+    #if MI_HAS_TLS_SLOT >= 2 // we must initialize the TLS slot before any allocation
     if ((reason==DLL_PROCESS_ATTACH || reason==DLL_THREAD_ATTACH) && mi_prim_get_default_heap() == NULL) {
       _mi_heap_set_default_direct((mi_heap_t*)&_mi_heap_empty);
     }