vect: Tweak vect_better_loop_vinfo_p handling of variable VFs

vect: Tweak vect_better_loop_vinfo_p handling of variable VFs

This patch fixes a large lmbench performance regression with
128-bit SVE, compiled in length-agnostic mode.

vect_better_loop_vinfo_p (new in GCC 10) tries to estimate whether
a new loop_vinfo is cheaper than a previous one, with an in-built
preference for the old one.  For variable VF it prefers the old
loop_vinfo if it is cheaper for at least one VF.  However, we have
no idea how likely that VF is in practice.

Another extreme would be to do what most of the rest of the
vectoriser does, and rely solely on the constant estimated VF.
But as noted in the comment, this means that a one-unit cost
difference would be enough to pick the new loop_vinfo,
despite the target generally preferring the old loop_vinfo
where possible.  The cost model just isn't accurate enough
for that to produce good results as things stand: there might
not be any practical benefit to the new loop_vinfo at the
estimated VF, and it would be significantly worse for higher VFs.

The patch instead goes for a hacky compromise: make sure that the new
loop_vinfo is also no worse than the old loop_vinfo at double the
estimated VF.  For all but trivial loops, this ensures that the
new loop_vinfo is only chosen if it is better than the old one
by a non-trivial amount at the estimated VF.  It also avoids
putting too much faith in the VF estimate.

I realise this isn't great, but it's supposed to be a conservative fix
suitable for stage 4.  The only affected testcases are the ones for
pr89007-*.c, where Advanced SIMD is indeed preferred for 128-bit SVE
and is no worse for 256-bit SVE.

Part of the problem here is that if the new loop_vinfo is better,
we discard the old one and never consider using it even as an
epilogue loop.  This means that if we choose Advanced SIMD over SVE,
we're much more likely to have left-over scalar elements.

Another is that the estimate provided by estimated_poly_value might have
different probabilities attached.  E.g. when tuning for a particular core,
the estimate is probably accurate, but when tuning for generic code,
the estimate is more of a guess.  Relying solely on the estimate is
probably correct for the former but not for the latter.

Hopefully those are things that we could tackle in GCC 11.

2020-04-20  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* tree-vect-loop.c (vect_better_loop_vinfo_p): If old_loop_vinfo
	has a variable VF, prefer new_loop_vinfo if it is cheaper for the
	estimated VF and is no worse at double the estimated VF.

gcc/testsuite/
	* gcc.target/aarch64/sve/cost_model_8.c: New test.
	* gcc.target/aarch64/sve/cost_model_9.c: Likewise.
	* gcc.target/aarch64/sve/pr89007-1.c: Add -msve-vector-bits=512.
	* gcc.target/aarch64/sve/pr89007-2.c: Likewise.