Hello all again,
I now have a question regarding the prefetch command generation for x86_64.
The option -fprefetch-loop-arrays generates prefetch commands for
non-vectorized code,
but not for vectorized one. Is that the intended functionality ? Is
there a way to get prefetching also
for the vectorized routines ?
Thanks, Thomas
Example: (compiled with g++-4.5.0)
g++ -O3 -fcx-fortran-rules -fprefetch-loop-arrays -mtune=core2
-march=core2 -mssse3 -S -c ../test_loop.cpp
The code for a complex multiplication loop done this way:
void f(std::complex<float> *a, std::complex<float> *b, std::complex<float> *r)
{
for(std::size_t s=0; s<N; s++)
r[s] = a[s]*b[s];
}
Is generated two-fold, one vectorized (.L3) and one not (L5):
.LFB1292:
leaq 32(%rdi), %rcx
leaq 32(%rdx), %rax
cmpq %rcx, %rdx
seta %r8b
cmpq %rdi, %rax
setb %cl
orb %r8b, %cl
jne .L9
.L2:
movq %rdx, %rax
xorl %edx, %edx
.p2align 4,,10
.p2align 3
.L5:
movss (%rdi), %xmm4
movss 4(%rdi), %xmm3
movss (%rsi), %xmm2
movss 4(%rsi), %xmm0
prefetcht0 80(%rdi)
movaps %xmm4, %xmm1
movaps %xmm3, %xmm5
mulss %xmm2, %xmm1
mulss %xmm0, %xmm5
mulss %xmm3, %xmm2
subss %xmm5, %xmm1
mulss %xmm4, %xmm0
movss %xmm1, (%rax)
addss %xmm2, %xmm0
incq %rdx
movss %xmm0, 4(%rax)
addq $8, %rdi
addq $8, %rsi
addq $8, %rax
cmpq $2097152, %rdx
jne .L5
rep
ret
.L9:
leaq 32(%rsi), %rcx
cmpq %rcx, %rdx
seta %cl
cmpq %rax, %rsi
seta %al
orb %al, %cl
je .L2
movq %rdx, %rax
xorps %xmm3, %xmm3
leaq 16777216(%rdi), %rdx
.p2align 4,,10
.p2align 3
.L3:
movaps %xmm3, %xmm2
movaps %xmm3, %xmm1
movlps (%rdi), %xmm2
movlps 16(%rdi), %xmm1
movhps 8(%rdi), %xmm2
movhps 24(%rdi), %xmm1
movaps %xmm2, %xmm0
movaps %xmm2, %xmm5
shufps $136, %xmm1, %xmm0
shufps $221, %xmm1, %xmm5
movaps %xmm3, %xmm2
movaps %xmm3, %xmm1
movlps (%rsi), %xmm2
movlps 16(%rsi), %xmm1
movhps 8(%rsi), %xmm2
movhps 24(%rsi), %xmm1
movaps %xmm2, %xmm4
movaps %xmm5, %xmm6
shufps $221, %xmm1, %xmm2
shufps $136, %xmm1, %xmm4
mulps %xmm2, %xmm6
movaps %xmm0, %xmm1
addq $32, %rdi
mulps %xmm4, %xmm1
mulps %xmm2, %xmm0
subps %xmm6, %xmm1
mulps %xmm5, %xmm4
movaps %xmm1, %xmm2
addps %xmm4, %xmm0
addq $32, %rsi
unpcklps %xmm0, %xmm2
unpckhps %xmm0, %xmm1
movlps %xmm2, (%rax)
movhps %xmm2, 8(%rax)
movlps %xmm1, 16(%rax)
movhps %xmm1, 24(%rax)
addq $32, %rax
cmpq %rdx, %rdi
jne .L3
rep
ret
