Very strange ia64 code generation problem

[H. J. Lu]

On Sat, Jul 19, 2003 at 08:52:32AM -0700, H. J. Lu wrote:
> While working on ia64 intrinsics, I found a very strange ia64 code
> generation problem. The following input:
> 
> unsigned int
> bswap_int (unsigned int x, int s)
> {
>   return s ? (__extension__ ({ register unsigned int __v, __x = (x); { union { unsigned long long __u; long __m; } __c; __c.__u = ((unsigned long long) __x) << 32; __c.__m = ((long) __builtin_ia64_mux1 ((long) (__c.__m), (int)(0xb))); __v = (unsigned int) __c.__u; } __v; })) : x;
> }
> 
> "gcc -O" gives me
> 
> 	.file	"foo.i"
> 	.pred.safe_across_calls p1-p5,p16-p63
> 	.text
> 	.align 16
> 	.global bswap_int#
> 	.proc bswap_int#
> bswap_int:
> 	.prologue
> 	.body
> 	mov r8 = r32
> 	cmp4.eq p6, p7 = 0, r33
> 	;;
> 	(p7) shl r14 = r8, 32
> 	;;
> 	mux1 r8 = r14, 0xb
> 	^^^^^^^^^^^^^^^^^^
> 	br.ret.sptk.many rp
> 	;;
> 	.endp bswap_int#
> 	.ident	"GCC: (GNU) 3.3 20030715 (Red Hat Linux 3.3-14)"
> 

It turns out that it is a gcc bug. I opened a critical bug report:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=11599


H.J.