单个变量是否存在编译器内存障碍？

Question

Compiler memory barriers has the effect among other things to force the compiler to make sure all stack variables that are cached in registers are written to memory before the barrier.

For example, GCC has the following statement:

asm inline ("" : : : "memory");

Is there any way to tell a compiler (specifically GCC, but I'm interested in others as well) to do the same effect for only a specific variable? something like the following imagined construct:

int x;
...
asm inline ("" : : : "memory(x)");

With the expected behavior that the value of x and x only will be written to the corresponding memory location, if it happens to be cached in a register.

The reason for this is that I have a specific variable which I need to make sure is not cached in a register so that a hardware engine can read its value. However, a full compiler memory barrier will force the compiler to write to memory the value of all other variables that might be cached in register at that point in time as well which can amount to much more data then I need to write. I wondered if there something more specific.

Thanks in advance!

Answer 1

Try with { int y = x; *(volatile int*)&x = y; } { int y = x; *(volatile int*)&x = y; } { int y = x; *(volatile int*)&x = y; } and inspect the resulting assembly.

Answer 2

Since you are willing to work with gcc extensions you could use the extensions for atomic instructions for that feature:

__sync_bool_compare_and_swap(&myvar, 0, 0)

would set the value of the variable to 0 if it is already 0 :) and in addition imply a full sequential consistency on that memory location.

Answer 3

重新收集lkml上的一个线程，单变量仅编译器屏障的方法之一是：

#define forget(x) __asm__ __volatile__("":"=m"(x):"m"(x))

Answer 4

I guess you can achieve it by specifying your variable in the list of output values of asm :

__asm__ __volatile__ ("" : "=r" (x) : : )

See Extended Asm for some information.

UPD.

It may be better to use "g" constraint instead of "r" as more permissive.

__asm__ __volatile__ ("" : "=g" (x) : : )

Also, I've found another great howto for inline assembly.