diff --git a/retroarch.c b/retroarch.c
index 9bb97653c5..6c89fba7f1 100644
--- a/retroarch.c
+++ b/retroarch.c
@@ -2214,7 +2214,10 @@ static void check_rewind(void)
       if (cnt == 0)
 #endif
       {
+         RARCH_PERFORMANCE_INIT(rewind_serialize);
+         RARCH_PERFORMANCE_START(rewind_serialize);
          pretro_serialize(g_extern.state_buf, g_extern.state_size);
+         RARCH_PERFORMANCE_STOP(rewind_serialize);
          state_manager_push(g_extern.state_manager, g_extern.state_buf);
       }
    }
diff --git a/rewind.c b/rewind.c
index 5a7229f422..f0a39cf7cb 100644
--- a/rewind.c
+++ b/rewind.c
@@ -14,6 +14,7 @@
  */
 
 #include "rewind.h"
+#include "performance.h"
 #include <stdlib.h>
 #include <stdint.h>
 #include "boolean.h"
@@ -147,11 +148,45 @@ static void reassign_bottom(state_manager_t *state)
       state->bottom_ptr = (state->bottom_ptr + 1) & state->buf_size_mask;
 }
 
+#if __SSE2__
+#include <emmintrin.h>
+// There's no equivalent in libc, you'd think so ... std::mismatch exists, but it's not optimized at all. :(
+static unsigned find_mismatch(const uint32_t *a, const uint32_t *b, unsigned samples)
+{
+   unsigned i;
+   unsigned sse_samples = samples & ~3;
+   for (i = 0; i < sse_samples; i += 4)
+   {
+      __m128i v0 = _mm_loadu_si128((const __m128i*)(a + i));
+      __m128i v1 = _mm_loadu_si128((const __m128i*)(b + i));
+      __m128i c = _mm_cmpeq_epi32(v0, v1);
+      uint32_t mask = _mm_movemask_epi8(c);
+      if (mask != 0xffff) // Something has changed, figure out where.
+         return i + (__builtin_ctz(~mask) >> 2);
+   }
+
+   for (; i < samples; i++)
+      if (a[i] != b[i])
+         return i;
+
+   return samples;
+}
+#else
+static unsigned find_mismatch(const uint32_t *a, const uint32_t *b, unsigned samples)
+{
+   unsigned i;
+   for (i = 0; i < samples; i++)
+      if (a[i] != b[i])
+         return i;
+   return samples;
+}
+#endif
+
 static void generate_delta(state_manager_t *state, const void *data)
 {
-   uint64_t i;
+   size_t i;
    bool crossed = false;
-   const uint32_t *old_state = state->tmp_state;
+   uint32_t *old_state = state->tmp_state;
    const uint32_t *new_state = (const uint32_t*)data;
 
    state->buffer[state->top_ptr++] = 0; // For each separate delta, we have a 0 value sentinel in between.
@@ -163,20 +198,25 @@ static void generate_delta(state_manager_t *state, const void *data)
 
    for (i = 0; i < state->state_size; i++)
    {
-      uint64_t xor_ = old_state[i] ^ new_state[i];
+      unsigned avail = state->state_size - i;
+      unsigned pos = find_mismatch(old_state + i, new_state + i, avail);
+      if (pos == avail)
+         break;
+
+      i += pos;
 
       // If the data differs (xor != 0), we push that xor on the stack with index and xor.
       // This can be reversed by reapplying the xor.
       // This, if states don't really differ much, we'll save lots of space :)
       // Hopefully this will work really well with save states.
-      if (xor_)
-      {
-         state->buffer[state->top_ptr] = (i << 32) | xor_;
-         state->top_ptr = (state->top_ptr + 1) & state->buf_size_mask;
+      uint32_t xor_ = old_state[i] ^ new_state[i];
+      old_state[i] = new_state[i];
 
-         if (state->top_ptr == state->bottom_ptr)
-            crossed = true;
-      }
+      state->buffer[state->top_ptr] = ((uint64_t)i << 32) | xor_;
+      state->top_ptr = (state->top_ptr + 1) & state->buf_size_mask;
+
+      if (state->top_ptr == state->bottom_ptr)
+         crossed = true;
    }
 
    if (crossed)
@@ -185,8 +225,11 @@ static void generate_delta(state_manager_t *state, const void *data)
 
 bool state_manager_push(state_manager_t *state, const void *data)
 {
+   RARCH_PERFORMANCE_INIT(gen_delta);
+   RARCH_PERFORMANCE_START(gen_delta);
    generate_delta(state, data);
-   memcpy(state->tmp_state, data, state->state_size * sizeof(uint32_t));
+   RARCH_PERFORMANCE_STOP(gen_delta);
+
    state->first_pop = true;
 
    return true;