https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82189
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed|2017-09-12 00:00:00 |2021-8-11
--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
I'm not sure if it's "better" now. While we merge the store we now do not
vectorize the load at all. Basically it's a trade-off between giving up
when doing discovery on the whole store group or building sub-chains from
scalars. We do have heuristics in place that try to anticipate whether
splitting the group would succeed - if you change 'float' to 'double' we'll
go the splitting way ending up with
movupd (%rsi), %xmm1
unpcklpd %xmm0, %xmm0
divpd %xmm0, %xmm1
movups %xmm1, (%rdi)
movupd (%rdx), %xmm1
divpd %xmm0, %xmm1
movups %xmm1, 16(%rdi)
which is quite optimal. But for float we have
movss 4(%rdx), %xmm1
movss (%rdx), %xmm2
shufps $0, %xmm0, %xmm0
movss 4(%rsi), %xmm3
unpcklps %xmm1, %xmm2
movss (%rsi), %xmm1
unpcklps %xmm3, %xmm1
movlhps %xmm2, %xmm1
divps %xmm0, %xmm1
movups %xmm1, (%rdi)
that might be not too bad - it avoids STLF issues compared to doing two
8 byte loads and combining them.
The get-away is that SLP could be better in handling running into different
interleaving chains (or also into non-grouped loads).