From: Andrew Waterman Date: Mon, 29 Aug 2016 20:41:16 +0000 (-0700) Subject: Don't explicitly use atomics in rsort X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=3bef79ea985205ae9eb7dbf1c6a07169c2b1993d;p=riscv-tests.git Don't explicitly use atomics in rsort This is a pattern GCC should pick up for targets where AMOADD is faster than LW/ADD/SW. --- diff --git a/benchmarks/rsort/rsort.c b/benchmarks/rsort/rsort.c index fee68e5..dc53642 100644 --- a/benchmarks/rsort/rsort.c +++ b/benchmarks/rsort/rsort.c @@ -26,6 +26,12 @@ #define LOG_BASE 8 #define BASE (1 << LOG_BASE) +#if 0 +# define fetch_add(ptr, inc) __sync_fetch_and_add(ptr, inc) +#else +# define fetch_add(ptr, inc) ((*(ptr) += (inc)) - (inc)) +#endif + void sort(size_t n, type* arrIn, type* scratchIn) { size_t log_exp = 0; @@ -46,20 +52,20 @@ void sort(size_t n, type* arrIn, type* scratchIn) type a1 = p[1]; type a2 = p[2]; type a3 = p[3]; - __sync_fetch_and_add(&bucket[(a0 >> log_exp) % BASE], 1); - __sync_fetch_and_add(&bucket[(a1 >> log_exp) % BASE], 1); - __sync_fetch_and_add(&bucket[(a2 >> log_exp) % BASE], 1); - __sync_fetch_and_add(&bucket[(a3 >> log_exp) % BASE], 1); + fetch_add(&bucket[(a0 >> log_exp) % BASE], 1); + fetch_add(&bucket[(a1 >> log_exp) % BASE], 1); + fetch_add(&bucket[(a2 >> log_exp) % BASE], 1); + fetch_add(&bucket[(a3 >> log_exp) % BASE], 1); } for ( ; p < &arr[n]; p++) bucket[(*p >> log_exp) % BASE]++; size_t prev = bucket[0]; - prev += __sync_fetch_and_add(&bucket[1], prev); + prev += fetch_add(&bucket[1], prev); for (b = &bucket[2]; b < bucket + BASE; b += 2) { - prev += __sync_fetch_and_add(&b[0], prev); - prev += __sync_fetch_and_add(&b[1], prev); + prev += fetch_add(&b[0], prev); + prev += fetch_add(&b[1], prev); } static_assert(BASE % 2 == 0); @@ -73,10 +79,10 @@ void sort(size_t n, type* arrIn, type* scratchIn) size_t* pb1 = &bucket[(a1 >> log_exp) % BASE]; size_t* pb2 = &bucket[(a2 >> log_exp) % BASE]; size_t* pb3 = &bucket[(a3 >> log_exp) % BASE]; - type* s0 = scratch + __sync_fetch_and_add(pb0, -1); - type* s1 = scratch + __sync_fetch_and_add(pb1, -1); - type* s2 = scratch + __sync_fetch_and_add(pb2, -1); - type* s3 = scratch + __sync_fetch_and_add(pb3, -1); + type* s0 = scratch + fetch_add(pb0, -1); + type* s1 = scratch + fetch_add(pb1, -1); + type* s2 = scratch + fetch_add(pb2, -1); + type* s3 = scratch + fetch_add(pb3, -1); s0[-1] = a0; s1[-1] = a1; s2[-1] = a2;