https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105617
--- Comment #20 from Mason <slash.tmp at free dot fr> ---
Doh! You're right.
I come from a background where overlapping/aliasing inputs are heresy,
thus got blindsided :(
This would be the optimal code, right?
add4i:
# rdi = dst, rsi = a, rdx = b
movq 0(%rdx), %r8
movq 8(%rdx), %rax
movq 16(%rdx), %rcx
movq 24(%rdx), %rdx
addq 0(%rsi), %r8
adcq 8(%rsi), %rax
adcq 16(%rsi), %rcx
adcq 24(%rsi), %rdx
movq %r8, 0(%rdi)
movq %rax, 8(%rdi)
movq %rcx, 16(%rdi)
movq %rdx, 24(%rdi)
ret
FWIW, trunk generates even nastier code for znver2:
add4i:
movq (%rdx), %rax
addq (%rsi), %rax
movq 8(%rsi), %rcx
adcq 8(%rdx), %rcx
movq 16(%rsi), %r8
adcq 16(%rdx), %r8
movq 24(%rdx), %rdx
adcq 24(%rsi), %rdx
vmovq %rax, %xmm2
vpinsrq $1, %rcx, %xmm2, %xmm0
vmovq %r8, %xmm1
vpinsrq $1, %rdx, %xmm1, %xmm1
vinserti128 $0x1, %xmm1, %ymm0, %ymm0
vmovdqu %ymm0, (%rdi)
vzeroupper
ret
