configured with ../gcc-4.3-20080208/configure --enable-languages=c,c++,java
The same issue was also observed on i686-pc-linux x86-64 platforms.
>Description:
when an assignment to a variable is followed by 'cmpxchg' on the
variable, there exists a dependency between the two statements.
thus, the store has to complete before cmpxchg on the variable would
be issued.
The issue was not seen with -O0. It was seen when compiled with -O2.
I have not tried -O1, though.
For examples, when we have
inline bool LockByte( volatile unsigned char &flag ) {
volatile void *f = &flag;
if ( CompareAndSwap1(f,1,0) != 0) {
do {
__asm__ __volatile__("pause;");
} while ( CompareAndSwap1(f,1,0) != 0) ;
}
return 0;
}
main(...) {
...
char x = 0;
LockByte(x);
...
}
even with a single thread, LockByte() hangs i.e., spins forever.
It appears that the compiler misses the dependency and does not
generate the store to the memory location.
>How-To-Repeat:
Compile the following program with -O2 option. i.e.,
g++ -O2 <pgm-name>. It hangs. g++ <pgm-name> generates an executable
that completes and prints 'done'.
#include <assert.h>
#include <stdio.h>
#define DECL_CMPSWP(S,T,X) \
static inline T machine_cmpswp##S (volatile void *ptr, T value, T comparand ) \
{ \
T result; \
\
__asm__ __volatile__("lock\ncmpxchg" X " %2,%1" \
: "=a"(result), "=m"(*(T *)ptr) \
: "q"(value), "0"(comparand) \
: "memory"); \
return result; \
}
DECL_CMPSWP(1,unsigned char,"b");
#define CompareAndSwap1(a,b,c) machine_cmpswp1(a,b,c)
inline bool LockByte( volatile unsigned char &flag ) {
volatile void *f = &flag;
if ( CompareAndSwap1(f,1,0) != 0) {
do {
__asm__ __volatile__("pause;");
} while ( CompareAndSwap1(f,1,0) != 0) ;
}
return 0;
}
static void TestTinyLock() {
unsigned char flags[16];
for( int i=0; i<16; ++i )
flags[i] = i;
flags[8] = 0;
LockByte(flags[8]);
for( int i=0; i<16; ++i )
assert( flags[i]==(i==8?1:i) );
}
int main( int argc, char* argv[] ) {
try {
TestTinyLock();
} catch(...) {
}
printf("done\n");
return 0;
}
-----------------------------------------
'g++ -O2 -S' output
pushl %ecx
leal -16(%ebp), %eax
subl $88, %esp
movl %eax, -52(%ebp)
leal -84(%ebp), %eax
movl $___gxx_personality_sj0, -60(%ebp)
movl $LLSDA6, -56(%ebp)
movl $L13, -48(%ebp)
movl %esp, -44(%ebp)
movl %eax, (%esp)
call __Unwind_SjLj_Register
call ___main
xorl %eax, %eax
leal -32(%ebp), %edx
.p2align 4,,7
L2:
movb %al, (%edx,%eax)
addl $1, %eax
cmpl $16, %eax
jne L2
leal -24(%ebp), %ebx
xorl %eax, %eax
movl $1, %edx
/APP
# 44 "test_test.cpp" 1
lock
cmpxchgb %dl,(%ebx)
# 0 "" 2
/NO_APP
testb %al, %al
je L3
movl $1, %ecx
xorl %edx, %edx
.p2align 4,,7
L14:
/APP
# 52 "test_test.cpp" 1
pause;
# 0 "" 2
/NO_APP
movl %edx, %eax
/APP
# 44 "test_test.cpp" 1
lock
cmpxchgb %cl,(%ebx)
# 0 "" 2
/NO_APP
testb %al, %al
jne L14
L3:
movl $0, -88(%ebp)
jmp L8
.p2align 4,,7
L6:
movzbl %al, %eax
cmpl %edx, %eax
je L7
movl $LC0, 8(%esp)
movl $66, 4(%esp)
movl $LC1, (%esp)
movl $1, -80(%ebp)
call ___assert
L7:
addl $1, -88(%ebp)
cmpl $16, -88(%ebp)
je L9
L8:
>Fix:
The work-around is to declare the variabe to be 'volatile'
If we do
* (volatile unsigned char *) &flags[8] = 0;
instead, the generated executable completes.
--
Summary: possible incorrect opitmization due to missed dependency
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Severity: critical
Priority: P3
Component: c++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: kimwooyoung at gmail dot com
GCC build triplet: i686-pc-cygwin
GCC host triplet: i686-pc-cygwin
GCC target triplet: i686-pc-cygwin
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35396
