| 1 | #include <stdio.h>
|
|---|
| 2 | #include <stdint.h> // uintptr_t
|
|---|
| 3 |
|
|---|
| 4 | #include "bench.h"
|
|---|
| 5 |
|
|---|
| 6 | // Does a "lock xchg" on entry but a simple "mov" on exit => cheaper as 0 contention. While it has much more code, the
|
|---|
| 7 | // bulk is never run.
|
|---|
| 8 |
|
|---|
| 9 | #define CALIGN __attribute__(( aligned (CACHE_ALIGN) ))
|
|---|
| 10 | #define CACHE_ALIGN 128
|
|---|
| 11 | #if defined( __i386 ) || defined( __x86_64 )
|
|---|
| 12 | #define Pause() __asm__ __volatile__ ( "pause" : : : )
|
|---|
| 13 | #elif defined( __ARM_ARCH )
|
|---|
| 14 | #define Pause() __asm__ __volatile__ ( "YIELD" : : : )
|
|---|
| 15 | #else
|
|---|
| 16 | #error unsupported architecture
|
|---|
| 17 | #endif
|
|---|
| 18 |
|
|---|
| 19 | typedef uintptr_t TYPE; // addressable word-size
|
|---|
| 20 | static volatile TYPE lock __attribute__(( aligned (128) )); // Intel recommendation
|
|---|
| 21 | static TYPE PAD CALIGN __attribute__(( unused )); // protect further false sharing
|
|---|
| 22 |
|
|---|
| 23 | static inline void spin_lock( volatile TYPE *lock ) {
|
|---|
| 24 | enum { SPIN_START = 4, SPIN_END = 64 * 1024, };
|
|---|
| 25 | unsigned int spin = SPIN_START;
|
|---|
| 26 |
|
|---|
| 27 | for ( unsigned int i = 1;; i += 1 ) {
|
|---|
| 28 | if ( *lock == 0 && __atomic_test_and_set( lock, __ATOMIC_ACQUIRE ) == 0 ) break;
|
|---|
| 29 | for ( volatile unsigned int s = 0; s < spin; s += 1 ) Pause(); // exponential spin
|
|---|
| 30 | //spin += spin; // powers of 2
|
|---|
| 31 | if ( i % 64 == 0 ) spin += spin; // slowly increase by powers of 2
|
|---|
| 32 | if ( spin > SPIN_END ) spin = SPIN_START; // prevent overflow
|
|---|
| 33 | } // for
|
|---|
| 34 | } // spin_lock
|
|---|
| 35 |
|
|---|
| 36 | static inline void spin_unlock( volatile TYPE *lock ) {
|
|---|
| 37 | __atomic_clear( lock, __ATOMIC_RELEASE );
|
|---|
| 38 | } // spin_unlock
|
|---|
| 39 |
|
|---|
| 40 | void __attribute__((noinline)) do_call() {
|
|---|
| 41 | spin_lock( &lock );
|
|---|
| 42 | // asm volatile ("");
|
|---|
| 43 | spin_unlock( &lock );
|
|---|
| 44 | }
|
|---|
| 45 |
|
|---|
| 46 | int main( int argc, char * argv[] ) {
|
|---|
| 47 | BENCH_START()
|
|---|
| 48 | BENCH(
|
|---|
| 49 | for (size_t i = 0; i < times; i++) {
|
|---|
| 50 | do_call();
|
|---|
| 51 | },
|
|---|
| 52 | result
|
|---|
| 53 | )
|
|---|
| 54 | printf( "%g\n", result );
|
|---|
| 55 | }
|
|---|
| 56 |
|
|---|
| 57 | // Local Variables: //
|
|---|
| 58 | // tab-width: 4 //
|
|---|
| 59 | // End: //
|
|---|
