On x86 when bliting a software surface onto itself SDL_BlitSurface() internally sets the cpu direction flag, which causes following memcpy()s (or similar) to fail.

The x86 inline assembler version of the SDL macro SDL_revcpy() sets the direction flag ("std"), but does not clear it ("cld"). This is invalid according to GCC (inline assembler, which sets the direction flag, must reset it[0]) and the SYS V ABI (functions must leave with the direction flag cleared[1]). The macro is (indirectly, exact call sequence below) used in SDL_BlitSurface(), so this call sometimes returns with the direction flag set. This happens for bliting a software surface onto itself with the destination coordinates set right/down of the source coordinates (typical use of this is scrolling left/up). Later on other parts of the code (like inlined memcpy()) cause memory corruption.

Call sequence:
SDL_BlitSurface() (#define SDL_Blit_Surface SDL_BlitUpper)
SDL_BlitUpper()
SDL_BlitLower()
src->map->sw_blit() (function pointer to SDL_SoftBlit())
SDL_SoftBlit()
src->map->sw_data->blit() (function pointer to SDL_BlitCopyOverlap())
SDL_BlitCopyOverlap()
SDL_revcpy()


Here's a sample program, which demonstrates the problem. After the call to SDL_BlitSurface() the direction flag will be set, which is tested and an error is printed:
#include <SDL.h>
#include <stdio.h>
#include <stdlib.h>

static inline int TestDirectionFlag(void)
{
  unsigned eflags;
  __asm__ __volatile__("pushf\n\tpop %0" : "=r" (eflags));
  return (eflags & 0x400) != 0;
}

int main(void)
{
  SDL_Init(SDL_INIT_VIDEO);
  SDL_Surface* const s = SDL_SetVideoMode(640, 480, 16, SDL_SWSURFACE);

  SDL_Rect src_rect = { 0, 0, 630, 470 };
  SDL_Rect dst_rect = { 10, 10, 630, 470 };
  if (TestDirectionFlag())
  {
    fputs("direction flag set BEFORE SDL_BlitSurface()\n", stderr);
    abort();
  }
  SDL_BlitSurface(s, &src_rect, s, &dst_rect);
  if (TestDirectionFlag())
  {
    fputs("direction flag set AFTER SDL_BlitSurface()\n", stderr);
    abort();
  }

  SDL_Quit();
  return 0;
}


A Quick fix is to add a "cld" after the "rep movs" sequence in the SDL_revcpy() macro in SDL_stdinc.h:
--- include/SDL_stdinc.h.orig    2006-06-27 06:48:32.000000000 +0200
+++ include/SDL_stdinc.h 2008-01-27 16:39:16.000000000 +0100
@@ -319,6 +319,7 @@
         __asm__ __volatile__ (                          \
                 "std\n\t"                               \
                 "rep ; movsl\n\t"                       \
+                "cld\n\t"                               \
                 : "=&c" (u0), "=&D" (u1), "=&S" (u2)        \
                 : "0" (n >> 2),                           \
                   "1" (dstp+(n-4)), "2" (srcp+(n-4))    \


A better fix is to replace the macro by a simple call to memmove() (the non-x86 version of this is very slow and even the x86 assembler magic could be done better), i.e.:
#define SDL_revcpy(dst, src, len) memmove((dst), (src), (len))


[0] http://gcc.gnu.org/gcc-4.3/changes.html - "IA-32/x86-64", fourth bullet
[1] http://www.sco.com/developers/devspecs/abi386-4.pdf - page 38 EFLAGS