[Freemint-discuss] Strange bug with asm clobbers

[Thorsten O. <ad...@th...>]

Hi,

while trying to write a test program, i encountered a strange bug with gcc asm 
clobber list. The code in question is

	register unsigned short *start __asm__("a0");
	unsigned short *end;
	oldsp = Super(0);
	start = buf + sizeof(buf) / sizeof(buf[0]);
	__asm__ __volatile__(
		  ".dc.w 0xf320,0x4e71,0x4e71,0x4e71\n" /* fsave -(a0) 
*/
		  "move.l %1,%0\n"
		: "=a"(end)
		: "a"(start)
		: );
	SuperToUser(oldsp);

	printf("start: %08lx end: %08lx size: %ld\n", (long)end, 
(long)start, (long)start - (long)end);

And the generated code is:

[0000014a] 7000                      moveq.l    #0,d0
[0000014c] 2f00                      move.l     d0,-(a7)
[0000014e] 3f3c 0020                 move.w     #$0020,-(a7) ; Super
[00000152] 4e41                      trap       #1
[00000154] 5c8f                      addq.l     #6,a7
[00000156] 41f9 0001 94c8            lea.l      $000193E4,a0
[0000015c] f320                      fsave      -(a0)
[0000015e] 4e71                      nop
[00000160] 4e71                      nop
[00000162] 4e71                      nop
[00000164] 2648                      movea.l    a0,a3
[00000166] 2a4f                      movea.l    a7,a5         <-- apparently 
the backup_sp from SuperToUser macro
[00000168] 2f00                      move.l     d0,-(a7)
[0000016a] 3f3c 0020                 move.w     #$0020,-(a7) ; Super
[0000016e] 4e41                      trap       #1
[00000170] 2e4d                      movea.l    a5,a7
[00000172] 2008                      move.l     a0,d0         <-- using a0, 
although it was declared as being clobbered???
[00000174] 908b                      sub.l      a3,d0
[00000176] 2f00                      move.l     d0,-(a7)
[00000178] 2f08                      move.l     a0,-(a7)
[0000017a] 2f0b                      move.l     a3,-(a7)
[00000186] 487a ff3a                 pea.l      .LC0(pc)
[0000018a] 4eb9 0000 2998            jsr        _printf


As seen above, gcc seems to assume that "a0" still contains the value of 
start, although is was declared as being clobbered in the SuperToUser macro.

Did i miss something, or what is going wrong there?