The only way to get the following sample code run properly is to compile it
with -O3 or -Os (this code has been design to illustrate a point, it results
from thourough examination of more complex code):
//---------------------------------------------------------
#include <stdio.h>
unsigned long hehe(int vv) {
unsigned long result;
asm volatile (
"movq $0x0102030405060708, %%rax\n\t"
"mov %1, %%eax\n\t"
: "=a" (result)
: "r" (vv)
);
return result;
}
int main(int argc, char** argv) {
printf("%lx\n", hehe(42));
}
//---------------------------------------------------------
the expected output is 42.
compiled with -O3 -Os the programs gives 42
compiled with no option the program gives 0x05060708
The reason seem to be that the compiler does not take into account the
dependency between %rax and %eax (where %eax contains the lower 32 bits of
%rax). Here is the wrong code generated:
unsigned long hehe(int vv) {
4004e4: 55 push %rbp
4004e5: 48 89 e5 mov %rsp,%rbp
4004e8: 89 7d ec mov %edi,-0x14(%rbp)
unsigned long result;
asm volatile (
4004eb: 8b 45 ec mov -0x14(%rbp),%eax
4004ee: 48 b8 08 07 06 05 04 mov $0x102030405060708,%rax
4004f5: 03 02 01
// the following line is crazy!
4004f8: 89 c0 mov %eax,%eax
4004fa: 48 89 45 f8 mov %rax,-0x8(%rbp)
"movq $0x0102030405060708, %%rax\n\t"
"mov %1, %%eax\n\t"
: "=a" (result)
: "r" (vv)
);
return result;
4004fe: 48 8b 45 f8 mov -0x8(%rbp),%rax
}
400502: c9 leaveq
400503: c3 retq
The following code faces the same type of problem but with a more radical
register allocation error:
inline int locateEndToken(char* data, unsigned long pos, unsigned long max) {
int result;
unsigned long tmp;
asm volatile (
"mov $0x0A0D20, %%rax\n\t"
"movq %%rax, %%xmm15\n\t"
"lea -16(%2,%3,1),%1\n\t"
"add %2, %4\n\t"
".align 16\n\t"
"1:\n\t"
"add $16, %1\n\t"
"cmp %4, %1\n\t"
"jge 2f\n\t"
"pcmpistri $0,(%1),%%xmm15\n\t"
"jnc 1b\n\t"
"lea 0x1(%1,%%rcx,1),%%rax\n\t"
"cmpb $0x020, -1(%%rax)\n\t"
"je 4f\n\t"
"2:\n\t"
"movq $-1, %%rax\n\t"
"jmp 3f\n\t"
"4:\n\t"
"sub %2, %%rax\n\t"
"3:\n\t"
: "=a" (result), "=r" (tmp) //%0, %1
: "r" (data), //%2
"r" (pos), //%3
"r" (max) //%4
: "%rcx" // pcmpistri modifies
rcx
);
return result;
}
Works perfectly in -O3 but allocates the same register for %1 and %4 in other
cicurmstances. Thus the test "cmp %4, %1" at the sixth line becomes "cmp %1,
%1" which do not represent the right functional test.
Bottom line, using ASM can get things 15 times faster, but the generated code
MUST be manually validated!!! This is a nightmare.
(tests have been done with 4.4.1, 4.4.2, 4.4.3)
--
Summary: Lack of dependency between %rax and %eax in inline
assembly
Product: gcc
Version: 4.4.3
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: ff at vedicis dot com
GCC host triplet: Debian 5.02 x86_64
GCC target triplet: X86_64
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43110
