https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84106
--- Comment #4 from Daniel Fruzynski <bugzi...@poradnik-webmastera.com> ---
Here are results of small benchmark executed on Xeon E5-2683 v3. Code was
compiled using gcc 4.8.5. This gcc version also splits loops. Manually
vectorized code is 3.5 times faster:
[out]
--------------------------------------------------
Benchmark Time CPU Iterations
--------------------------------------------------
BM_test1 25 ns 25 ns 26989634
BM_test2 7 ns 7 ns 94495591
[/out]
Benchmarko code:
[code]
#include <benchmark/benchmark.h>
#include "immintrin.h"
#define N 81
int a1[N] __attribute__((aligned(32)));
int a2[N] __attribute__((aligned(32)));
int a3[N] __attribute__((aligned(32)));
class Init
{
public:
Init()
{
for (int n = 0; n < N; n++)
{
a1[n] = n % 32;
}
}
} init;
static void BM_test1(benchmark::State& state)
{
for (auto _ : state)
{
for (int n = 0; n < N; n++)
{
a2[n] = a1[n];
a3[n] = 1 << a1[n];
}
benchmark::ClobberMemory();
}
}
BENCHMARK(BM_test1);
static void BM_test2(benchmark::State& state)
{
for (auto _ : state)
{
int n = 0;
for (; n < N - 7; n += 8)
{
__m256i v = _mm256_load_si256((__m256i*)(&a1[0] + n));
_mm256_store_si256((__m256i*)(&a2[0] + n), v);
v = _mm256_sllv_epi32(_mm256_set1_epi32(1), v);
_mm256_store_si256((__m256i*)(&a3[0] + n), v);
}
for (; n < N; n++)
{
a2[n] = a1[n];
a3[n] = 1 << a1[n];
}
benchmark::ClobberMemory();
}
}
BENCHMARK(BM_test2);
BENCHMARK_MAIN();
[/code]
