diff --git a/lib/std/os/windows/tls.zig b/lib/std/os/windows/tls.zig
new file mode 100644
index 0000000000000000000000000000000000000000..f98e22b620c10a746b77e0467c37a0c2b71de19c
--- /dev/null
+++ b/lib/std/os/windows/tls.zig
@@ -0,0 +1,50 @@
+const std = @import("std");
+const builtin = @import("builtin");
+const windows = std.os.windows;
+
+export var _tls_index: u32 = std.os.windows.TLS_OUT_OF_INDEXES;
+export var _tls_start: ?*anyopaque linksection(".tls") = null;
+export var _tls_end: ?*anyopaque linksection(".tls$ZZZ") = null;
+export var __xl_a: windows.PIMAGE_TLS_CALLBACK linksection(".CRT$XLA") = null;
+export var __xl_z: windows.PIMAGE_TLS_CALLBACK linksection(".CRT$XLZ") = null;
+
+comptime {
+    if (builtin.cpu.arch == .x86 and builtin.abi == .msvc and builtin.zig_backend != .stage2_c) {
+        // The __tls_array is the offset of the ThreadLocalStoragePointer field
+        // in the TEB block whose base address held in the %fs segment.
+        asm (
+            \\ .global __tls_array
+            \\ __tls_array = 0x2C
+        );
+    }
+}
+
+// TODO this is how I would like it to be expressed
+//export const _tls_used linksection(".rdata$T") = std.os.windows.IMAGE_TLS_DIRECTORY {
+//    .StartAddressOfRawData = @intFromPtr(&_tls_start),
+//    .EndAddressOfRawData = @intFromPtr(&_tls_end),
+//    .AddressOfIndex = @intFromPtr(&_tls_index),
+//    .AddressOfCallBacks = @intFromPtr(__xl_a),
+//    .SizeOfZeroFill = 0,
+//    .Characteristics = 0,
+//};
+// This is the workaround because we can't do @intFromPtr at comptime like that.
+pub const IMAGE_TLS_DIRECTORY = extern struct {
+    StartAddressOfRawData: *?*anyopaque,
+    EndAddressOfRawData: *?*anyopaque,
+    AddressOfIndex: *u32,
+    AddressOfCallBacks: [*:null]windows.PIMAGE_TLS_CALLBACK,
+    SizeOfZeroFill: u32,
+    Characteristics: u32,
+};
+export const _tls_used linksection(".rdata$T") = IMAGE_TLS_DIRECTORY{
+    .StartAddressOfRawData = &_tls_start,
+    .EndAddressOfRawData = &_tls_end,
+    .AddressOfIndex = &_tls_index,
+    // __xl_a is just a global variable containing a null pointer; the actual callbacks sit in
+    // between __xl_a and __xl_z. So we need to skip over __xl_a here. If there are no callbacks,
+    // this just means we point to __xl_z (the null terminator).
+    .AddressOfCallBacks = @as([*:null]windows.PIMAGE_TLS_CALLBACK, @ptrCast(&__xl_a)) + 1,
+    .SizeOfZeroFill = 0,
+    .Characteristics = 0,
+};
diff --git a/lib/std/start.zig b/lib/std/start.zig
index 2a072832bb9ccd50af11eabb85ca7740158aea58..ca6a5416eb57689a349f63fd9a159fefdc337e29 100644
--- a/lib/std/start.zig
+++ b/lib/std/start.zig
@@ -176,7 +176,7 @@ fn _DllMainCRTStartup(
     lpReserved: std.os.windows.LPVOID,
 ) callconv(std.os.windows.WINAPI) std.os.windows.BOOL {
     if (!builtin.single_threaded and !builtin.link_libc) {
-        _ = @import("start_windows_tls.zig");
+        _ = @import("os/windows/tls.zig");
     }
 
     if (@hasDecl(root, "DllMain")) {
@@ -434,7 +434,7 @@ fn _start() callconv(.Naked) noreturn {
 fn WinStartup() callconv(std.os.windows.WINAPI) noreturn {
     @setAlignStack(16);
     if (!builtin.single_threaded and !builtin.link_libc) {
-        _ = @import("start_windows_tls.zig");
+        _ = @import("os/windows/tls.zig");
     }
 
     std.debug.maybeEnableSegfaultHandler();
@@ -445,7 +445,7 @@ fn WinStartup() callconv(std.os.windows.WINAPI) noreturn {
 fn wWinMainCRTStartup() callconv(std.os.windows.WINAPI) noreturn {
     @setAlignStack(16);
     if (!builtin.single_threaded and !builtin.link_libc) {
-        _ = @import("start_windows_tls.zig");
+        _ = @import("os/windows/tls.zig");
     }
 
     std.debug.maybeEnableSegfaultHandler();
diff --git a/lib/std/start_windows_tls.zig b/lib/std/start_windows_tls.zig
deleted file mode 100644
index 48880b4811d3536a78b6e7dccbf48bf561fbc282..0000000000000000000000000000000000000000
--- a/lib/std/start_windows_tls.zig
+++ /dev/null
@@ -1,48 +0,0 @@
-const std = @import("std");
-const builtin = @import("builtin");
-
-export var _tls_index: u32 = std.os.windows.TLS_OUT_OF_INDEXES;
-export var _tls_start: u8 linksection(".tls") = 0;
-export var _tls_end: u8 linksection(".tls$ZZZ") = 0;
-export var __xl_a: std.os.windows.PIMAGE_TLS_CALLBACK linksection(".CRT$XLA") = null;
-export var __xl_z: std.os.windows.PIMAGE_TLS_CALLBACK linksection(".CRT$XLZ") = null;
-
-comptime {
-    if (builtin.target.cpu.arch == .x86 and builtin.zig_backend != .stage2_c) {
-        // The __tls_array is the offset of the ThreadLocalStoragePointer field
-        // in the TEB block whose base address held in the %fs segment.
-        asm (
-            \\ .global __tls_array
-            \\ __tls_array = 0x2C
-        );
-    }
-}
-
-// TODO this is how I would like it to be expressed
-// TODO also note, ReactOS has a +1 on StartAddressOfRawData and AddressOfCallBacks. Investigate
-// why they do that.
-//export const _tls_used linksection(".rdata$T") = std.os.windows.IMAGE_TLS_DIRECTORY {
-//    .StartAddressOfRawData = @intFromPtr(&_tls_start),
-//    .EndAddressOfRawData = @intFromPtr(&_tls_end),
-//    .AddressOfIndex = @intFromPtr(&_tls_index),
-//    .AddressOfCallBacks = @intFromPtr(__xl_a),
-//    .SizeOfZeroFill = 0,
-//    .Characteristics = 0,
-//};
-// This is the workaround because we can't do @intFromPtr at comptime like that.
-pub const IMAGE_TLS_DIRECTORY = extern struct {
-    StartAddressOfRawData: *anyopaque,
-    EndAddressOfRawData: *anyopaque,
-    AddressOfIndex: *anyopaque,
-    AddressOfCallBacks: *anyopaque,
-    SizeOfZeroFill: u32,
-    Characteristics: u32,
-};
-export const _tls_used linksection(".rdata$T") = IMAGE_TLS_DIRECTORY{
-    .StartAddressOfRawData = &_tls_start,
-    .EndAddressOfRawData = &_tls_end,
-    .AddressOfIndex = &_tls_index,
-    .AddressOfCallBacks = @as(*anyopaque, @ptrCast(&__xl_a)),
-    .SizeOfZeroFill = 0,
-    .Characteristics = 0,
-};