https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90579
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |missed-optimization
Status|UNCONFIRMED |NEW
Last reconfirmed| |2019-05-23
CC| |rguenth at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
.cfi_startproc
movslq %edi, %rax
addl $4, %edi
vbroadcastsd %xmm0, %ymm1
movslq %edi, %rdi
vmovddup %xmm0, %xmm0
vmulpd a(,%rax,8), %ymm1, %ymm1
vmulpd a(,%rdi,8), %xmm0, %xmm0
vmovupd %ymm1, r(%rip)
vmovups %xmm0, r+32(%rip)
vmovupd r+16(%rip), %ymm1
^^^ this one
vextractf128 $0x1, %ymm1, %xmm2
vunpckhpd %xmm2, %xmm2, %xmm0
vaddsd .LC0(%rip), %xmm0, %xmm0
vaddsd %xmm2, %xmm0, %xmm0
vunpckhpd %xmm1, %xmm1, %xmm2
vaddsd %xmm2, %xmm0, %xmm0
vaddsd %xmm1, %xmm0, %xmm0
vaddsd r+8(%rip), %xmm0, %xmm0
vaddsd r(%rip), %xmm0, %xmm0
vzeroupper
ret
unaligned accesses are prone to STLF issues but there's no easy
way out here, at least I don't see a good way of say, restricting
the 2nd loop vectorization to SSE.
Note when misaligning by a single element we'd have to disable
vectorization completely.
In some way this is a target issue since it allows unaligned
loads. If it would split them (we have a tunable for this)
we'd be fine here (by luck, until misalinging not by SSE vector size).
Similar cases can be made with placing unvectorized by element initializations
before a vectorized loop (possibly in another function).
Those STLF issues are just a bad "feature" of modern CPUs and the fix
is ultimatively in them...
