Hello!
I have a problem with some inline assembly, which to me looks like a
GCC bug but I first wanted to check that I haven't done any obvious
mistake in my implementation. I see the problem both on MIPS and IA-32.
The code I have is
/* Assign 1 to v0 (the MIPS register used for passing function
results) */
static inline int inline_asm(void) {
register unsigned long __v0 asm("$2");
__asm__ volatile (
"li %0, 1\n"
: "=r" (__v0)
:
: "memory"
);
return (int) __v0;
}
void problem (void) {
int a = inline_asm();
int b = overwriter();
tst(a, b);
}
When disassembling the generated code, this generates the following
MIPS assembly code, with my comments
00000000 <problem>:
0: 27bdffe8 addiu sp,sp,-24
4: afbf0010 sw ra,16(sp)
8: 24020001 li v0,1 # Result of the inline assembly piece
c: 0c000000 jal overwriter # Overwrites v0
10: 00000000 nop # jal delay slot - v0 should have been saved here!
14: 00402021 move a0,v0 # GCC thinks v0 still contains the result of the inline assembly
18: 0c000000 jal tst
1c: 00402821 move a1,v0 # ... *and* the result of overwriter()!
20: 8fbf0010 lw ra,16(sp)
24: 00000000 nop
28: 03e00008 jr ra
2c: 27bd0018 addiu sp,sp,24
This bug seems to be portable - I rewrote the inline assembly part for
IA-32 - just a move to %eax - and %eax was indeed overwritten as well.
The versions where I have this problem is
mips-linux-gnu-gcc (GCC) 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)
gcc (GCC) 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)
but gcc-3.3 (IA-32) handles this just fine.
So the usual question: Did I miss something in the inline assembly
piece which can cause this problem or is this a proper GCC bug?
(The problem was first manifested in some inline assembly for the Cibyl
project, http://spel.bth.se/index.php/Cibyl, and the code is based on
the Linux system calls for MIPS)
--
// Simon
Attachment:
a.i
Description: Binary data
