Re: Misalignment of local SSE variables in thread function

[Tim Prince <tprince at myrealbox dot com>]

Jonathan Kinsey wrote:
> This is a problem that I've seen in the archives and is marked as not
> being a problem with gcc.  I'm not sure what the solution is though, the
> simple test below runs fine using the Microsoft compiler (on windows)
> and fails using gcc (3.4.5) compiler.  There is no pthreads or glib code
> here.
>
> The output shows that the second call to f(), in the thread, has a
> misaligned s variable (which will crash in a SSE call).
>
> Jon
>
> #include <stdio.h>
> #include <process.h>
> #include <windows.h>
> #include <xmmintrin.h>
>
> void f(void *p)
> {
>   __m128 s;
>   printf("%d\n", ((int)&s) % 16);
> }
>
> int main(int argc, char ** argv)
> {
>   f(0);
>   _beginthread(f, 0, 0);
>
>   Sleep(2000);
>   return 0;
> }
>
>
>   
I don't remember what was said in the past, but any gcc function which 
uses SSE parallel instructions must be called from a gcc function which 
passes a 16-byte aligned stack.  So you may be able to overcome it by 
interposing a wrapper function at the thread entry point (which does not 
use parallel SSE), with all gcc functions compiled with normal options 
(not -Os, which sets a smaller value in -mpreferred-stack-alignment).  
To speculate further, possibly your pthreads library was compiled with 
options which don't pass 16-byte aligned stack.