https://gcc.gnu.org/bugzilla/show_bug.cgi?id=125974
--- Comment #3 from mjr19 at cam dot ac.uk ---
Odd. I see, when the comment character is removed from the first line of the
loop,
gfortran-16 -O3 -mavx2 -fopenmp-simd -S v8.f90
.L4:
vmovupd (%rcx,%rax), %ymm0
vsubpd (%rdx,%rax), %ymm0, %ymm0
vpermilpd $5, %ymm0, %ymm0
vaddsubpd %ymm0, %ymm1, %ymm0
vmovupd %ymm0, (%rdx,%rax)
addq $32, %rax
cmpq %rsi, %rax
jne .L4
and
gfortran-16 -O3 -ffast-math -mavx2 -fopenmp-simd -S v8.f90
.L4:
vmovupd (%rdx,%rax), %ymm1
vaddpd %ymm1, %ymm2, %ymm0
vsubpd %ymm1, %ymm2, %ymm1
vshufpd $5, %ymm1, %ymm0, %ymm0
vpermilpd $5, (%rcx,%rax), %ymm1
vaddsubpd %ymm1, %ymm0, %ymm0
vmovupd %ymm0, (%rdx,%rax)
addq $32, %rax
cmpq %rsi, %rax
jne .L4
The -ffast-math version is not as bad as it looks. On machines I have tried, it
has either been equally fast, or about 5% slower, than the more concise code. I
have yet to find a machine on which it is quicker, but maybe I should stop
worrying if the performance difference is so slight.
