diff --git a/demo/test.c b/demo/test.c
index 5ab9686c..2fa6e08d 100644
--- a/demo/test.c
+++ b/demo/test.c
@@ -2502,20 +2502,24 @@ struct thread_info {
    int ret;
 };
 
-static void *run_pthread(void *arg)
+static void run(struct thread_info *tinfo)
 {
-   struct thread_info *tinfo = arg;
-
    tinfo->ret = tinfo->t->fn();
 
+   if (mp_warray_free() == -2)
+      tinfo->ret = EXIT_FAILURE;
+}
+
+static void *run_pthread(void *arg)
+{
+   run(arg);
+
    return arg;
 }
 
 static unsigned long run_msvc(void *arg)
 {
-   struct thread_info *tinfo = arg;
-
-   tinfo->ret = tinfo->t->fn();
+   run(arg);
 
    return 0;
 }
@@ -2609,7 +2613,6 @@ static int unit_tests(int argc, char **argv)
    };
    struct thread_info test_threads[sizeof(test)/sizeof(test[0])], *res;
    unsigned long i, ok, fail, nop;
-   size_t n_threads = MP_HAS(MULTI_THREADED) ? sizeof(test) / sizeof(test[0]) : 1;
    uint64_t t;
    int j;
    ok = fail = nop = 0;
@@ -2620,8 +2623,7 @@ static int unit_tests(int argc, char **argv)
    mp_rand_source(s_mp_rand_jenkins);
 
    if (MP_HAS(MP_SMALL_STACK_SIZE)) {
-      printf("Small-stack enabled with %zu warray buffers\n\n", n_threads);
-      DO(mp_warray_init(n_threads, 1));
+      printf("Small-stack enabled\n\n");
    }
 
    if (MP_HAS(MULTI_THREADED)) {
diff --git a/mp_warray_free.c b/mp_warray_free.c
index 088efefc..f7470f81 100644
--- a/mp_warray_free.c
+++ b/mp_warray_free.c
@@ -9,19 +9,13 @@ MP_STATIC_ASSERT(warray_free_sz_does_not_overflow, (sizeof(mp_word) * MP_WARRAY)
 static int s_warray_free(void)
 {
    int ret = 0;
-   size_t n;
-   for (n = 0; n < s_mp_warray.allocated; ++n) {
-      if (s_mp_warray.l_used[n].warray) {
-         ret = -2;
-         goto ERR_OUT;
-      }
+   if (s_mp_warray.w_used)
+      return -2;
+   if (s_mp_warray.w_free) {
+      s_mp_zero_buf(s_mp_warray.w_free, sizeof(mp_word) * MP_WARRAY);
+      MP_FREE(s_mp_warray.w_free, sizeof(mp_word) * MP_WARRAY);
+      s_mp_warray.w_free = NULL;
    }
-   for (n = 0; n < s_mp_warray.allocated; ++n) {
-      MP_FREE(s_mp_warray.l_free[n].warray, sizeof(mp_word) * MP_WARRAY);
-      s_mp_warray.l_free[n].warray = NULL;
-   }
-   s_mp_warray_free(s_mp_warray.usable);
-ERR_OUT:
    return ret;
 }
 
diff --git a/mp_warray_init.c b/mp_warray_init.c
deleted file mode 100644
index c2509886..00000000
--- a/mp_warray_init.c
+++ /dev/null
@@ -1,46 +0,0 @@
-#include "tommath_private.h"
-#ifdef MP_WARRAY_INIT_C
-/* LibTomMath, multiple-precision integer library -- Tom St Denis */
-/* SPDX-License-Identifier: Unlicense */
-
-static mp_err s_warray_init(size_t n_alloc, bool preallocate)
-{
-   size_t n;
-   if (s_mp_warray.l_free != NULL || s_mp_warray.l_used != NULL) {
-      return MP_VAL;
-   }
-
-   s_mp_warray.l_free = MP_CALLOC(n_alloc, sizeof(*(s_mp_warray.l_free)));
-   s_mp_warray.l_used = MP_CALLOC(n_alloc, sizeof(*(s_mp_warray.l_used)));
-   if (s_mp_warray.l_free == NULL || s_mp_warray.l_used == NULL) {
-      s_mp_warray_free(n_alloc);
-      return MP_MEM;
-   }
-
-   if (preallocate) {
-      for (n = 0; n < n_alloc; ++n) {
-         s_mp_warray.l_free[n].warray = MP_CALLOC(MP_WARRAY, sizeof(mp_word));
-         if (s_mp_warray.l_free[n].warray == NULL) {
-            while (n > 0) {
-               n--;
-               MP_FREE(s_mp_warray.l_free[n].warray, MP_WARRAY * sizeof(mp_word));
-               s_mp_warray.l_free[n].warray = NULL;
-            }
-            s_mp_warray_free(n_alloc);
-            return MP_MEM;
-         }
-      }
-      s_mp_warray.allocated = n_alloc;
-   }
-
-   s_mp_warray.usable = n_alloc;
-   return MP_OKAY;
-}
-
-mp_err mp_warray_init(size_t n_alloc, bool preallocate)
-{
-   if (MP_HAS(MP_SMALL_STACK_SIZE)) return s_warray_init(n_alloc, preallocate);
-   return MP_ERR;
-}
-
-#endif
diff --git a/s_mp_cmpexch_n.c b/s_mp_cmpexch_n.c
deleted file mode 100644
index e8ef969c..00000000
--- a/s_mp_cmpexch_n.c
+++ /dev/null
@@ -1,43 +0,0 @@
-#include "tommath_private.h"
-#ifdef S_MP_CMPEXCH_N_C
-/* LibTomMath, multiple-precision integer library -- Tom St Denis */
-/* SPDX-License-Identifier: Unlicense */
-
-#ifdef __GNUC__
-#define S_CMPEXCH_N_GCC_C
-static bool s_cmpexch_n_gcc(void **ptr, void **expected, void *desired)
-{
-   return __atomic_compare_exchange_n(ptr, expected, desired, true, __ATOMIC_ACQ_REL, __ATOMIC_ACQUIRE);
-}
-#endif
-
-#ifdef _MSC_VER
-#define S_CMPEXCH_N_MSVC_C
-
-#ifndef _WIN32_WINNT
-#define _WIN32_WINNT 0x0501
-#endif
-#ifndef WINVER
-#define WINVER 0x0501
-#endif
-
-#define WIN32_LEAN_AND_MEAN
-#include <windows.h>
-
-static bool s_cmpexch_n_msvc(void **ptr, void **expected, void *desired)
-{
-   return InterlockedCompareExchangePointer(ptr, desired, *(expected));
-}
-#endif
-
-bool s_cmpexch_n_gcc(void **ptr, void **expected, void *desired);
-bool s_cmpexch_n_msvc(void **ptr, void **expected, void *desired);
-
-bool s_mp_cmpexch_n(void **ptr, void **expected, void *desired)
-{
-   if (MP_HAS(S_CMPEXCH_N_GCC)) return s_cmpexch_n_gcc(ptr, expected, desired);
-   if (MP_HAS(S_CMPEXCH_N_MSVC)) return s_cmpexch_n_msvc(ptr, expected, desired);
-   return false;
-}
-
-#endif
diff --git a/s_mp_warray.c b/s_mp_warray.c
index d181057c..1b8b068b 100644
--- a/s_mp_warray.c
+++ b/s_mp_warray.c
@@ -3,6 +3,6 @@
 /* LibTomMath, multiple-precision integer library -- Tom St Denis */
 /* SPDX-License-Identifier: Unlicense */
 
-st_warray s_mp_warray;
+mp_thread st_warray s_mp_warray = { 0 };
 
 #endif
diff --git a/s_mp_warray_free.c b/s_mp_warray_free.c
deleted file mode 100644
index 9d8b75eb..00000000
--- a/s_mp_warray_free.c
+++ /dev/null
@@ -1,17 +0,0 @@
-#include "tommath_private.h"
-#ifdef S_MP_WARRAY_FREE_C
-/* LibTomMath, multiple-precision integer library -- Tom St Denis */
-/* SPDX-License-Identifier: Unlicense */
-
-void s_mp_warray_free(size_t n)
-{
-   (void)n;
-   MP_FREE(s_mp_warray.l_free, n * sizeof(*(s_mp_warray.l_free)));
-   MP_FREE(s_mp_warray.l_used, n * sizeof(*(s_mp_warray.l_used)));
-   s_mp_warray.l_free = NULL;
-   s_mp_warray.l_used = NULL;
-   s_mp_warray.allocated = 0;
-   s_mp_warray.usable = 0;
-}
-
-#endif
diff --git a/s_mp_warray_get.c b/s_mp_warray_get.c
index 39176eb2..26b0d7c1 100644
--- a/s_mp_warray_get.c
+++ b/s_mp_warray_get.c
@@ -5,34 +5,14 @@
 
 void *s_mp_warray_get(void)
 {
-   void *ret = NULL;
-   size_t n;
-   if (s_mp_warray.usable == 0) {
-      if (mp_warray_init(1, false) != MP_OKAY)
-         return NULL;
+   if (s_mp_warray.w_used)
+      return NULL;
+   if (s_mp_warray.w_free == NULL) {
+      s_mp_warray.w_free = MP_CALLOC(MP_WARRAY, sizeof(mp_word));
    }
-   for (n = 0; n < s_mp_warray.allocated;) {
-      if (s_mp_warray.l_free[n].warray == NULL) {
-         n++;
-         continue;
-      }
-      ret = s_mp_warray.l_free[n].warray;
-      if (s_mp_cmpexch_n(&s_mp_warray.l_free[n].warray, &ret, NULL)) {
-         s_mp_warray.l_used[n].warray = ret;
-         goto LBL_OUT;
-      }
-      /* restart from the beginning if we missed a potential slot */
-      n = 0;
-   }
-   ret = NULL;
-   if (s_mp_warray.allocated + 1 > s_mp_warray.usable)
-      goto LBL_OUT;
-   ret = MP_CALLOC(MP_WARRAY, sizeof(mp_word));
-   if (ret != NULL)
-      s_mp_warray.l_used[s_mp_warray.allocated++].warray = ret;
-
-LBL_OUT:
-   return ret;
+   s_mp_warray.w_used = s_mp_warray.w_free;
+   s_mp_warray.w_free = NULL;
+   return s_mp_warray.w_used;
 }
 
 #endif
diff --git a/s_mp_warray_put.c b/s_mp_warray_put.c
index 4cf413d6..79e014ac 100644
--- a/s_mp_warray_put.c
+++ b/s_mp_warray_put.c
@@ -5,14 +5,10 @@
 
 void s_mp_warray_put(void *w)
 {
-   size_t n, allocated = s_mp_warray.allocated;
-   for (n = 0; n < allocated; ++n) {
-      if (s_mp_warray.l_used[n].warray == w) {
-         s_mp_warray.l_used[n].warray = NULL;
-         s_mp_warray.l_free[n].warray = w;
-         break;
-      }
-   }
+   if (s_mp_warray.w_free || s_mp_warray.w_used != w)
+      return;
+   s_mp_warray.w_free = w;
+   s_mp_warray.w_used = NULL;
 }
 
 #endif
diff --git a/tommath.h b/tommath.h
index 7da36d0e..1820d243 100644
--- a/tommath.h
+++ b/tommath.h
@@ -588,7 +588,6 @@ mp_err mp_fread(mp_int *a, int radix, FILE *stream) MP_WUR;
 mp_err mp_fwrite(const mp_int *a, int radix, FILE *stream) MP_WUR;
 #endif
 
-mp_err mp_warray_init(size_t n_alloc, bool preallocate);
 int mp_warray_free(void);
 
 #define mp_to_binary(M, S, N)  mp_to_radix((M), (S), (N), NULL, 2)
diff --git a/tommath_private.h b/tommath_private.h
index 36291a06..9c25f330 100644
--- a/tommath_private.h
+++ b/tommath_private.h
@@ -234,8 +234,6 @@ MP_PRIVATE mp_err s_mp_radix_size_overestimate(const mp_int *a, const int radix,
 MP_PRIVATE mp_err s_mp_fp_log(const mp_int *a, mp_int *c) MP_WUR;
 MP_PRIVATE mp_err s_mp_fp_log_d(const mp_int *a, mp_word *c) MP_WUR;
 
-MP_PRIVATE bool s_mp_cmpexch_n(void **ptr, void **expected, void *desired);
-
 #ifdef MP_SMALL_STACK_SIZE
 #define MP_SMALL_STACK_SIZE_C
 #define MP_ALLOC_WARRAY(name) *name = s_mp_warray_get()
@@ -247,19 +245,20 @@ MP_PRIVATE bool s_mp_cmpexch_n(void **ptr, void **expected, void *desired);
 #define MP_CHECK_WARRAY(name)
 #endif
 
-struct warray {
-   void *warray;
-};
+#if defined(_MSC_VER)
+#define mp_thread __declspec(thread)
+#elif defined(__GNUC__)
+#define mp_thread __thread
+#endif
+
 typedef struct {
-   struct warray *l_free, *l_used;
-   size_t allocated, usable;
+   void *w_free, *w_used;
 } st_warray;
 
-extern MP_PRIVATE st_warray s_mp_warray;
+extern MP_PRIVATE mp_thread st_warray s_mp_warray;
 
 MP_PRIVATE void *s_mp_warray_get(void);
 MP_PRIVATE void s_mp_warray_put(void *w);
-MP_PRIVATE void s_mp_warray_free(size_t n);
 
 #define MP_RADIX_MAP_REVERSE_SIZE 80u
 extern MP_PRIVATE const char s_mp_radix_map[];