https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71414
Freddie Witherden <freddie at witherden dot org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |freddie at witherden dot org
--- Comment #11 from Freddie Witherden <freddie at witherden dot org> ---
I've been looking into this and the big difference appears to be that when
Clang unrolls the loop it does so using multiple accumulators (and indeed does
this without need to be told to unroll. Given:
double acc(double *x, int n)
{
double a = 0;
#pragma omp simd
for (int i = 0; i < n; i++)
a += x[i];
return a;
}
and compiling with clang -march=native -Ofast -fopenmp -S the core loop reads
as:
vaddpd (%rdi,%rsi,8), %ymm0, %ymm0
vaddpd 32(%rdi,%rsi,8), %ymm1, %ymm1
vaddpd 64(%rdi,%rsi,8), %ymm2, %ymm2
vaddpd 96(%rdi,%rsi,8), %ymm3, %ymm3
vaddpd 128(%rdi,%rsi,8), %ymm0, %ymm0
vaddpd 160(%rdi,%rsi,8), %ymm1, %ymm1
vaddpd 192(%rdi,%rsi,8), %ymm2, %ymm2
vaddpd 224(%rdi,%rsi,8), %ymm3, %ymm3
vaddpd 256(%rdi,%rsi,8), %ymm0, %ymm0
vaddpd 288(%rdi,%rsi,8), %ymm1, %ymm1
vaddpd 320(%rdi,%rsi,8), %ymm2, %ymm2
vaddpd 352(%rdi,%rsi,8), %ymm3, %ymm3
vaddpd 384(%rdi,%rsi,8), %ymm0, %ymm0
vaddpd 416(%rdi,%rsi,8), %ymm1, %ymm1
vaddpd 448(%rdi,%rsi,8), %ymm2, %ymm2
vaddpd 480(%rdi,%rsi,8), %ymm3, %ymm3
which is heavily unrolled and uses four separate accumulators to hide the
latency of the vector adds. Interestingly, one could argue that Clang is not
using enough registers given that Skylake can dual-issue adds and they have a
latency of 4 cycles (implying you want 8 separate accumulators).
GCC 10 with gcc -march=skylake -Ofast -fopenmp -S test.c -funroll-loops
vaddpd -224(%r8), %ymm1, %ymm2
vaddpd -192(%r8), %ymm2, %ymm3
vaddpd -160(%r8), %ymm3, %ymm4
vaddpd -128(%r8), %ymm4, %ymm5
vaddpd -96(%r8), %ymm5, %ymm6
vaddpd -64(%r8), %ymm6, %ymm7
vaddpd -32(%r8), %ymm7, %ymm0
which although it is unrolled, is not a useful unrolling due to the dependency
chain. Indeed, I would not be surprised if the performance is similar to the
unrolled code as the loop related cruft can be hidden.
