diff --git a/src/audio/SDL_audio.c b/src/audio/SDL_audio.c
index 2ffd216c0..05674eb86 100644
--- a/src/audio/SDL_audio.c
+++ b/src/audio/SDL_audio.c
@@ -27,6 +27,7 @@
 #include "SDL_audio_c.h"
 #include "SDL_audiomem.h"
 #include "SDL_sysaudio.h"
+#include "../thread/SDL_systhread.h"
 
 #define _THIS SDL_AudioDevice *_this
 
@@ -1191,19 +1192,15 @@ open_audio_device(const char *devname, int iscapture,
     /* Start the audio thread if necessary */
     if (!current_audio.impl.ProvidesOwnCallbackThread) {
         /* Start the audio thread */
+
+        /* !!! FIXME: we don't force the audio thread stack size here because it calls into user code, but maybe we should? */
+        /* buffer queueing callback only needs a few bytes, so make the stack tiny. */
         char name[64];
+        const size_t stacksize = (device->spec.callback == SDL_BufferQueueDrainCallback) ? 64 * 1024 : 0;
+
         SDL_snprintf(name, sizeof (name), "SDLAudioDev%d", (int) device->id);
-/* !!! FIXME: this is nasty. */
-#if defined(__WIN32__) && !defined(HAVE_LIBC)
-#undef SDL_CreateThread
-#if SDL_DYNAMIC_API
-        device->thread = SDL_CreateThread_REAL(SDL_RunAudio, name, device, NULL, NULL);
-#else
-        device->thread = SDL_CreateThread(SDL_RunAudio, name, device, NULL, NULL);
-#endif
-#else
-        device->thread = SDL_CreateThread(SDL_RunAudio, name, device);
-#endif
+        device->thread = SDL_CreateThreadInternal(SDL_RunAudio, name, stacksize, device);
+
         if (device->thread == NULL) {
             SDL_CloseAudioDevice(device->id);
             SDL_SetError("Couldn't create audio thread");
diff --git a/src/audio/pulseaudio/SDL_pulseaudio.c b/src/audio/pulseaudio/SDL_pulseaudio.c
index 6b11e066f..00aad1b99 100644
--- a/src/audio/pulseaudio/SDL_pulseaudio.c
+++ b/src/audio/pulseaudio/SDL_pulseaudio.c
@@ -46,6 +46,7 @@
 #include "../SDL_audio_c.h"
 #include "SDL_pulseaudio.h"
 #include "SDL_loadso.h"
+#include "../thread/SDL_systhread.h"
 
 #if (PA_API_VERSION < 12)
 /** Return non-zero if the passed state is one of the connected states */
@@ -646,7 +647,7 @@ PULSEAUDIO_DetectDevices()
     WaitForPulseOperation(hotplug_mainloop, PULSEAUDIO_pa_context_get_source_info_list(hotplug_context, SourceInfoCallback, NULL));
 
     /* ok, we have a sane list, let's set up hotplug notifications now... */
-    hotplug_thread = SDL_CreateThread(HotplugThread, "PulseHotplug", NULL);
+    hotplug_thread = SDL_CreateThreadInternal(HotplugThread, "PulseHotplug", 256 * 1024, NULL);
 }
 
 static void
diff --git a/src/haptic/windows/SDL_xinputhaptic.c b/src/haptic/windows/SDL_xinputhaptic.c
index aefbabbda..3a2bbfb01 100644
--- a/src/haptic/windows/SDL_xinputhaptic.c
+++ b/src/haptic/windows/SDL_xinputhaptic.c
@@ -33,6 +33,7 @@
 #include "SDL_xinputhaptic_c.h"
 #include "../../core/windows/SDL_xinput.h"
 #include "../../joystick/windows/SDL_windowsjoystick_c.h"
+#include "../thread/SDL_systhread.h"
 
 /*
  * Internal stuff.
@@ -205,17 +206,8 @@ SDL_XINPUT_HapticOpenFromUserIndex(SDL_Haptic *haptic, const Uint8 userid)
     }
 
     SDL_snprintf(threadName, sizeof(threadName), "SDLXInputDev%d", (int)userid);
+    haptic->hwdata->thread = SDL_CreateThreadInternal(SDL_RunXInputHaptic, threadName, 64 * 1024, haptic->hwdata);
 
-#if defined(__WIN32__) && !defined(HAVE_LIBC)  /* !!! FIXME: this is nasty. */
-#undef SDL_CreateThread
-#if SDL_DYNAMIC_API
-    haptic->hwdata->thread = SDL_CreateThread_REAL(SDL_RunXInputHaptic, threadName, haptic->hwdata, NULL, NULL);
-#else
-    haptic->hwdata->thread = SDL_CreateThread(SDL_RunXInputHaptic, threadName, haptic->hwdata, NULL, NULL);
-#endif
-#else
-    haptic->hwdata->thread = SDL_CreateThread(SDL_RunXInputHaptic, threadName, haptic->hwdata);
-#endif
     if (haptic->hwdata->thread == NULL) {
         SDL_DestroyMutex(haptic->hwdata->mutex);
         SDL_free(haptic->effects);
diff --git a/src/joystick/psp/SDL_sysjoystick.c b/src/joystick/psp/SDL_sysjoystick.c
index e5247254f..1c7180e77 100644
--- a/src/joystick/psp/SDL_sysjoystick.c
+++ b/src/joystick/psp/SDL_sysjoystick.c
@@ -34,9 +34,9 @@
 
 #include "SDL_events.h"
 #include "SDL_error.h"
-#include "SDL_thread.h"
 #include "SDL_mutex.h"
 #include "SDL_timer.h"
+#include "../thread/SDL_systhread.h"
 
 /* Current pad state */
 static SceCtrlData pad = { .Lx = 0, .Ly = 0, .Buttons = 0 };
@@ -116,7 +116,7 @@ int SDL_SYS_JoystickInit(void)
         return SDL_SetError("Can't create input semaphore");
     }
     running = 1;
-    if((thread = SDL_CreateThread(JoystickUpdate, "JoySitckThread",NULL)) == NULL) {
+    if((thread = SDL_CreateThreadInternal(JoystickUpdate, "JoystickThread", 4096, NULL)) == NULL) {
         return SDL_SetError("Can't create input thread");
     }
 
diff --git a/src/joystick/windows/SDL_windowsjoystick.c b/src/joystick/windows/SDL_windowsjoystick.c
index 7123c6151..74a25f965 100644
--- a/src/joystick/windows/SDL_windowsjoystick.c
+++ b/src/joystick/windows/SDL_windowsjoystick.c
@@ -35,13 +35,13 @@
 #include "SDL_error.h"
 #include "SDL_assert.h"
 #include "SDL_events.h"
-#include "SDL_thread.h"
 #include "SDL_timer.h"
 #include "SDL_mutex.h"
 #include "SDL_events.h"
 #include "SDL_hints.h"
 #include "SDL_joystick.h"
 #include "../SDL_sysjoystick.h"
+#include "../thread/SDL_systhread.h"
 #if !SDL_EVENTS_DISABLED
 #include "../../events/SDL_events_c.h"
 #endif
@@ -301,18 +301,9 @@ SDL_SYS_JoystickInit(void)
     SDL_SYS_JoystickDetect();
 
     if (!s_threadJoystick) {
-        s_bJoystickThreadQuit = SDL_FALSE;
         /* spin up the thread to detect hotplug of devices */
-#if defined(__WIN32__) && !defined(HAVE_LIBC)
-#undef SDL_CreateThread
-#if SDL_DYNAMIC_API
-        s_threadJoystick= SDL_CreateThread_REAL(SDL_JoystickThread, "SDL_joystick", NULL, NULL, NULL);
-#else
-        s_threadJoystick= SDL_CreateThread(SDL_JoystickThread, "SDL_joystick", NULL, NULL, NULL);
-#endif
-#else
-        s_threadJoystick = SDL_CreateThread(SDL_JoystickThread, "SDL_joystick", NULL);
-#endif
+        s_bJoystickThreadQuit = SDL_FALSE;
+        s_threadJoystick = SDL_CreateThreadInternal(SDL_JoystickThread, "SDL_joystick", 64 * 1024, NULL);
     }
     return SDL_SYS_NumJoysticks();
 }
diff --git a/src/main/haiku/SDL_BeApp.cc b/src/main/haiku/SDL_BeApp.cc
index 36c2a1ab8..32fd6e892 100644
--- a/src/main/haiku/SDL_BeApp.cc
+++ b/src/main/haiku/SDL_BeApp.cc
@@ -31,9 +31,9 @@
 
 #include "SDL_BApp.h"	/* SDL_BApp class definition */
 #include "SDL_BeApp.h"
-#include "SDL_thread.h"
 #include "SDL_timer.h"
 #include "SDL_error.h"
+#include "../thread/SDL_systhread.h"
 
 #include "../../video/haiku/SDL_BWin.h"
 
@@ -62,7 +62,7 @@ SDL_InitBeApp(void)
 {
     /* Create the BApplication that handles appserver interaction */
     if (SDL_BeAppActive <= 0) {
-        SDL_AppThread = SDL_CreateThread(StartBeApp, "SDLApplication", NULL);
+        SDL_AppThread = SDL_CreateThreadInternal(StartBeApp, "SDLApplication", 0, NULL);
         if (SDL_AppThread == NULL) {
             return SDL_SetError("Couldn't create BApplication thread");
         }
diff --git a/src/thread/SDL_systhread.h b/src/thread/SDL_systhread.h
index f13f3e203..05a012536 100644
--- a/src/thread/SDL_systhread.h
+++ b/src/thread/SDL_systhread.h
@@ -60,6 +60,11 @@ extern SDL_TLSData *SDL_SYS_GetTLSData();
 /* Set the thread local storage for this thread */
 extern int SDL_SYS_SetTLSData(SDL_TLSData *data);
 
+/* This is for internal SDL use, so we don't need #ifdefs everywhere. */
+extern SDL_Thread *
+SDL_CreateThreadInternal(int (SDLCALL * fn) (void *), const char *name,
+                         const size_t stacksize, void *data);
+
 #endif /* _SDL_systhread_h */
 
 /* vi: set ts=4 sw=4 expandtab: */
diff --git a/src/thread/SDL_thread.c b/src/thread/SDL_thread.c
index 1306ae00d..ae865790a 100644
--- a/src/thread/SDL_thread.c
+++ b/src/thread/SDL_thread.c
@@ -423,6 +423,16 @@ SDL_CreateThread(int (SDLCALL * fn) (void *),
 #endif
 }
 
+SDL_Thread *
+SDL_CreateThreadInternal(int (SDLCALL * fn) (void *), const char *name,
+                         const size_t stacksize, void *data) {
+#ifdef SDL_PASSED_BEGINTHREAD_ENDTHREAD
+    return SDL_CreateThreadWithStackSize(fn, name, stacksize, data, NULL, NULL);
+#else
+    return SDL_CreateThreadWithStackSize(fn, name, stacksize, data);
+#endif
+}
+
 SDL_threadID
 SDL_GetThreadID(SDL_Thread * thread)
 {
diff --git a/src/thread/SDL_thread_c.h b/src/thread/SDL_thread_c.h
index 7749183d0..554325d6d 100644
--- a/src/thread/SDL_thread_c.h
+++ b/src/thread/SDL_thread_c.h
@@ -90,19 +90,6 @@ extern SDL_TLSData *SDL_Generic_GetTLSData();
  */
 extern int SDL_Generic_SetTLSData(SDL_TLSData *data);
 
-/* !!! FIXME: for 2.1, remove this and make stack size part of SDL_CreateThread. */
-#ifdef SDL_PASSED_BEGINTHREAD_ENDTHREAD
-SDL_Thread *SDLCALL
-SDL_CreateThreadWithStackSize(int (SDLCALL * fn) (void *),
-                 const char *name, const size_t stacksize, void *data,
-                 pfnSDL_CurrentBeginThread pfnBeginThread,
-                 pfnSDL_CurrentEndThread pfnEndThread)
-#else
-SDL_Thread *SDLCALL
-SDL_CreateThreadWithStackSize(int (SDLCALL * fn) (void *),
-                const char *name, const size_t stacksize, void *data)
-#endif
-
 #endif /* _SDL_thread_c_h */
 
 /* vi: set ts=4 sw=4 expandtab: */
diff --git a/src/timer/SDL_timer.c b/src/timer/SDL_timer.c
index ea24af84d..abe968e86 100644
--- a/src/timer/SDL_timer.c
+++ b/src/timer/SDL_timer.c
@@ -24,7 +24,7 @@
 #include "SDL_timer_c.h"
 #include "SDL_atomic.h"
 #include "SDL_cpuinfo.h"
-#include "SDL_thread.h"
+#include "../thread/SDL_systhread.h"
 
 /* #define DEBUG_TIMERS */
 
@@ -221,17 +221,9 @@ SDL_TimerInit(void)
         }
 
         SDL_AtomicSet(&data->active, 1);
-        /* !!! FIXME: this is nasty. */
-#if defined(__WIN32__) && !defined(HAVE_LIBC)
-#undef SDL_CreateThread
-#if SDL_DYNAMIC_API
-        data->thread = SDL_CreateThread_REAL(SDL_TimerThread, name, data, NULL, NULL);
-#else
-        data->thread = SDL_CreateThread(SDL_TimerThread, name, data, NULL, NULL);
-#endif
-#else
-        data->thread = SDL_CreateThread(SDL_TimerThread, name, data);
-#endif
+
+        /* Timer threads use a callback into the app, so we can't set a limited stack size here. */
+        data->thread = SDL_CreateThreadInternal(SDL_TimerThread, name, 0, data);
         if (!data->thread) {
             SDL_TimerQuit();
             return -1;
diff --git a/src/video/cocoa/SDL_cocoamousetap.m b/src/video/cocoa/SDL_cocoamousetap.m
index 0f29daf75..0ab98ed06 100644
--- a/src/video/cocoa/SDL_cocoamousetap.m
+++ b/src/video/cocoa/SDL_cocoamousetap.m
@@ -32,8 +32,8 @@
 #if SDL_MAC_NO_SANDBOX
 
 #include "SDL_keyboard.h"
-#include "SDL_thread.h"
 #include "SDL_cocoavideo.h"
+#include "../thread/SDL_systhread.h"
 
 #include "../../events/SDL_mouse_c.h"
 
@@ -202,7 +202,7 @@ Cocoa_InitMouseEventTap(SDL_MouseData* driverdata)
 
     tapdata->runloopStartedSemaphore = SDL_CreateSemaphore(0);
     if (tapdata->runloopStartedSemaphore) {
-        tapdata->thread = SDL_CreateThread(&Cocoa_MouseTapThread, "Event Tap Loop", tapdata);
+        tapdata->thread = SDL_CreateThreadInternal(&Cocoa_MouseTapThread, "Event Tap Loop", 64 * 1024, tapdata);
         if (!tapdata->thread) {
             SDL_DestroySemaphore(tapdata->runloopStartedSemaphore);
         }
diff --git a/src/video/psp/SDL_pspevents.c b/src/video/psp/SDL_pspevents.c
index c1a095dbb..931aeaf76 100644
--- a/src/video/psp/SDL_pspevents.c
+++ b/src/video/psp/SDL_pspevents.c
@@ -31,8 +31,8 @@
 #include "../../events/SDL_keyboard_c.h"
 #include "SDL_pspvideo.h"
 #include "SDL_pspevents_c.h"
-#include "SDL_thread.h"
 #include "SDL_keyboard.h"
+#include "../thread/SDL_systhread.h"
 #include <psphprm.h>
 
 #ifdef PSPIRKEYB
@@ -264,7 +264,7 @@ void PSP_EventInit(_THIS)
         return;
     }
     running = 1;
-    if((thread = SDL_CreateThread(EventUpdate, "PSPInputThread",NULL)) == NULL) {
+    if((thread = SDL_CreateThreadInternal(EventUpdate, "PSPInputThread", 4096, NULL)) == NULL) {
         SDL_SetError("Can't create input thread\n");
         return;
     }
diff --git a/src/video/winrt/SDL_winrtevents.cpp b/src/video/winrt/SDL_winrtevents.cpp
index e9df726d4..093adfd38 100644
--- a/src/video/winrt/SDL_winrtevents.cpp
+++ b/src/video/winrt/SDL_winrtevents.cpp
@@ -38,6 +38,7 @@ using Windows::UI::Core::CoreCursor;
 #include "../../core/winrt/SDL_winrtapp_xaml.h"
 #include "SDL_assert.h"
 #include "SDL_system.h"
+#include "../thread/SDL_systhread.h"
 
 extern "C" {
 #include "../SDL_sysvideo.h"
@@ -113,7 +114,7 @@ WINRT_CycleXAMLThread()
 
             _mutex = SDL_CreateMutex();
             _threadState = ThreadState_Running;
-            _XAMLThread = SDL_CreateThread(WINRT_XAMLThreadMain, "SDL/XAML App Thread", nullptr);
+            _XAMLThread = SDL_CreateThreadInternal(WINRT_XAMLThreadMain, "SDL/XAML App Thread", 0, nullptr);
 
             SDL_LockMutex(_mutex);
             while (_threadState != ThreadState_Yielding) {