Open Radar

 
28-Mar-2008 10:50 AM Alexander Strange:
Summary:
Given:
int a(int v)
{
    int temp[4];
    
    asm volatile("mov %0, 4+%1"
                 :
                 : "r"(v), "o"(*temp));
    
    return temp[1];
}

gcc/llvm will output 'mov %eax, 4+(%esp)'. Under other systems, this compiles to '4(%esp)', but Darwin compiles to 0(%esp), which is wrong.

Steps to Reproduce:
1. Under some other i386 system: gcc -O -fomit-frame-pointer -c asm-offset.c (attached)
2. Check with objdump -d:
Disassembly of section .text:

00000000 <a>:
   0:   83 ec 10                sub    $0x10,%esp
   3:   8b 44 24 14             mov    0x14(%esp),%eax
   7:   89 44 24 04             mov    %eax,0x4(%esp)
   b:   8b 44 24 04             mov    0x4(%esp),%eax
   f:   83 c4 10                add    $0x10,%esp
  12:   c3                      ret    


3. Under Darwin: gcc -arch i386 -O -fomit-frame-pointer -c asm-offset.c (attached)
4. Check the output with otool -tv.

Expected results:
The same assembly.

Actual results:
_a:
00000000	subl	$0x1c,%esp
00000003	movl	0x20(%esp),%eax
00000007	movl	%eax,0x00(%esp)
0000000b	movl	0x04(%esp),%eax
0000000f	addl	$0x1c,%esp
00000012	ret

Regression:
Checked with:
Apple Computer, Inc. version cctools-667.3~112, GNU assembler version 1.38
GNU assembler version 2.18 (x86_64-linux-gnu) using BFD version (GNU Binutils for Ubuntu) 2.18

Notes:
There seems to be a problem with the definition of "o" constraints in gcc - shouldn't it always output 0(%esp) in this case?