diff options
| -rw-r--r-- | pu2/pu2.out.c | 36 |
1 files changed, 23 insertions, 13 deletions
diff --git a/pu2/pu2.out.c b/pu2/pu2.out.c index 8bc9f52..baae8f2 100644 --- a/pu2/pu2.out.c +++ b/pu2/pu2.out.c @@ -2,6 +2,7 @@ // COMPILE TIME FLAGS EXPLANATION // - NO_MMAP Prevents usage of mmap+hugepages, which is only supported on linux // - NO_ASM Prevents usage of inline assembly, useful for non-x86 platforms + // - NO_INTRINSICS Disables compiler intrinsics // - BAD_CLOCK Use if your system clock does not support nanoseconds. Uses microseconds instead. // - OLD_CLANG Use if your compiler version is not very up-to-date. // - SPEED_IMPROVEMENT Almost doubles clock rate for the VM with a few optimizations. @@ -11,6 +12,7 @@ /* -- BEGIN COMPILE TIME FLAGS -- */ // #define PU2_NO_MMAP // #define PU2_NO_ASM +// #define PU2_NO_INTRINSICS // #define PU2_BAD_CLOCK // #define PU2_OLD_CLANG #define PU2_SPEED_IMPROVEMENT @@ -87,8 +89,13 @@ // not implemented yet, was lazy // bake(int start, int end) {} - #define unlikely(JJ) __builtin_expect((JJ), 0) - #define likely(JJ) __builtin_expect((JJ), 1) + #ifndef PU2_NO_INTRINSICS + #define unlikely(JJ) __builtin_expect((JJ), 0) + #define likely(JJ) __builtin_expect((JJ), 1) + #else + #define unlikely(JJ) (JJ) + #define likely(JJ) (JJ) + #endif #ifndef PU2_NO_MMAP __attribute__((aligned(64))) int *m; @@ -103,7 +110,7 @@ typedef struct { const int hugepage_size = 2 * 1024 * 1024; -#ifndef PU2_NO_ASM + #ifndef PU2_NO_ASM // unused void set_affinity(int cpu) { cpu_set_t mask; @@ -112,7 +119,7 @@ void set_affinity(int cpu) { if (sched_setaffinity(0, sizeof(mask), &mask) != 0) { perror("sched_setaffinity failed"); } } -#endif + #endif void init_pu2() { // set_affinity(0); @@ -153,7 +160,7 @@ void __attribute__((hot, aligned(64))) pu2() { static int s1, s2, s3; static const __attribute__((aligned(64))) void *PRECOMPUTED_TABLE[] - = { &&PU2_END, &&C_ONE, &&C_TWO, &&C_THREE, &&C_FOUR, + = { &&PU2_END, &&C_ONE, &&C_TWO, &&C_THREE, &&C_FOUR, &&C_FIVE, &&C_SIX, &&C_SEVEN, &&C_EIGHT, &&C_NINE, &&C_TEN, &&C_ELEVEN, &&C_TWELVE, &&C_THIRTEEN, &&CB_FOURTEEN, &&CB_FIFTEEN, &&CB_SIXTEEN, &&CB_SEVENTEEN, &&CB_EIGHTEEN, [32] = 0 }; @@ -162,12 +169,14 @@ void __attribute__((hot, aligned(64))) pu2() { while (1) { PU2_START: inst = *M++; - // if (unlikely(inst.c > 18)) A; + // if (unlikely(inst.c > 18)) A; + #ifndef PU2_NO_INTRINSICS __builtin_prefetch(M + 1, 0, 3); + #endif goto *PRECOMPUTED_TABLE[ inst.c ]; -// C_ZERO: -// goto PU2_END; + // C_ZERO: + // goto PU2_END; C_ONE: #ifndef PU2_NO_ASM asm("mov %1, %0" : "=r"(a) : "r"(inst.v)); @@ -183,7 +192,8 @@ void __attribute__((hot, aligned(64))) pu2() { #endif goto PU2_START; C_THREE: - // if (unlikely(a == 7)) printf("conditional jmp %i (%i) [a=%i b=%i u=%i]\n", M - M2, u, a, b, u); + // if (unlikely(a == 7)) printf("conditional jmp %i (%i) [a=%i b=%i u=%i]\n", M - M2, u, a, + // b, u); if (likely(u)) { if (unlikely(a == 7)) M = M2 + (b >> 1); m[ a ] = b; @@ -233,8 +243,8 @@ void __attribute__((hot, aligned(64))) pu2() { M += inst.v; goto PU2_START; CB_FIFTEEN: - //if (unlikely(inst.v == 7)) M = M2 + (m[a] >> 1); - //if (unlikely(inst.v == a)) goto PU2_START; + // if (unlikely(inst.v == 7)) M = M2 + (m[a] >> 1); + // if (unlikely(inst.v == a)) goto PU2_START; m[ inst.v ] = m[ a ]; // asm("mov %1, %0" : "=r"(m[inst.v]) : "r"(m[a])); goto PU2_START; @@ -248,8 +258,8 @@ void __attribute__((hot, aligned(64))) pu2() { m[ a ] = inst.v; goto PU2_START; CB_EIGHTEEN: - //if (unlikely(a == 7)) M = M2 + (m[inst.v] >> 1); - //if (unlikely(inst.v == a)) goto PU2_START; + // if (unlikely(a == 7)) M = M2 + (m[inst.v] >> 1); + // if (unlikely(inst.v == a)) goto PU2_START; m[ a ] = m[ inst.v ]; // asm("mov %1, %0" : "=r"(m[a]) : "r"(m[inst.v])); goto PU2_START; |