lanes.
By special casing these we can often either reduce the total number of
shuffles significantly or reduce the number of (high latency on Haswell)
AVX2 shuffles that potentially cross 128-bit lanes. Even when these
don't actually cross lanes, they have much higher latency to support
that. Doing two of them and a blend is worse than doing a single insert
across the 128-bit lanes to blend and then doing a single interleaved
shuffle.
While this seems like a narrow case, it kept cropping up on me and the
difference is *huge* as you can see in many of the test cases. I first
hit this trying to perfectly fix the interleaving shuffle patterns used
by Halide for AVX2.
llvm-svn: 222533
between splitting a vector into 128-bit lanes and recombining them vs.
decomposing things into single-input shuffles and a final blend.
This handles a large number of cases in AVX1 where the cross-lane
shuffles would be much more expensive to represent even though we end up
with a fast blend at the root. Instead, we can do a better job of
shuffling in a single lane and then inserting it into the other lanes.
This fixes the remaining bits of Halide's regression captured in PR21281
for AVX1. However, the bug persists in AVX2 because I've made this
change reasonably conservative. The cases where it makes sense in AVX2
to split into 128-bit lanes are much more rare because we can often do
full permutations across all elements of the 256-bit vector. However,
the particular test case in PR21281 is an example of one of the rare
cases where it is *always* better to work in a single 128-bit lane. I'm
going to try to teach the logic to detect and form the good code even in
AVX2 next, but it will need to use a separate heuristic.
Finally, there is one pesky regression here where we previously would
craftily use vpermilps in AVX1 to shuffle both high and low halves at
the same time. We no longer pull that off, and not for any really good
reason. Ultimately, I think this is just another missing nuance to the
selection heuristic that I'll try to add in afterward, but this change
already seems strictly worth doing considering the magnitude of the
improvements in common matrix math shuffle patterns.
As always, please let me know if this causes a surprising regression for
you.
llvm-svn: 221861
that are unused.
This allows the combiner to delete math feeding shuffles where the math
isn't actually necessary. This improves some of the vperm2x128 tests
that regressed when the vector shuffle lowering started actually
generating vperm instructions rather than forcibly decomposing them.
Sadly, this isn't enough to get this *really* right because we still
form a completely unnecessary permutation. To fix that, we also need to
fold shuffles which just rearrange concatenated or inserted subvectors.
llvm-svn: 219086
shuffles using AVX and AVX2 instructions. This fixes PR21138, one of the
few remaining regressions impacting benchmarks from the new vector
shuffle lowering.
You may note that it "regresses" many of the vperm2x128 test cases --
these were actually "improved" by the naive lowering that the new
shuffle lowering previously did. This regression gave me fits. I had
this patch ready-to-go about an hour after flipping the switch but
wasn't sure how to have the best of both worlds here and thought the
correct solution might be a completely different approach to lowering
these vector shuffles.
I'm now convinced this is the correct lowering and the missed
optimizations shown in vperm2x128 are actually due to missing
target-independent DAG combines. I've even written most of the needed
DAG combine and will submit it shortly, but this part is ready and
should help some real-world benchmarks out.
llvm-svn: 219079
Update the entire regression test suite for the new shuffles. Remove
most of the old testing which was devoted to the old shuffle lowering
path and is no longer relevant really. Also remove a few other random
tests that only really exercised shuffles and only incidently or without
any interesting aspects to them.
Benchmarking that I have done shows a few small regressions with this on
LNT, zero measurable regressions on real, large applications, and for
several benchmarks where the loop vectorizer fires in the hot path it
shows 5% to 40% improvements for SSE2 and SSE3 code running on Sandy
Bridge machines. Running on AMD machines shows even more dramatic
improvements.
When using newer ISA vector extensions the gains are much more modest,
but the code is still better on the whole. There are a few regressions
being tracked (PR21137, PR21138, PR21139) but by and large this is
expected to be a win for x86 generated code performance.
It is also more correct than the code it replaces. I have fuzz tested
this extensively with ISA extensions up through AVX2 and found no
crashes or miscompiles (yet...). The old lowering had a few miscompiles
and crashers after a somewhat smaller amount of fuzz testing.
There is one significant area where the new code path lags behind and
that is in AVX-512 support. However, there was *extremely little*
support for that already and so this isn't a significant step backwards
and the new framework will probably make it easier to implement lowering
that uses the full power of AVX-512's table-based shuffle+blend (IMO).
Many thanks to Quentin, Andrea, Robert, and others for benchmarking
assistance. Thanks to Adam and others for help with AVX-512. Thanks to
Hal, Eric, and *many* others for answering my incessant questions about
how the backend actually works. =]
I will leave the old code path in the tree until the 3 PRs above are at
least resolved to folks' satisfaction. Then I will rip it (and 1000s of
lines of code) out. =] I don't expect this flag to stay around for very
long. It may not survive next week.
llvm-svn: 219046
floating point and integer domains.
Merge the AVX2 test into it and add an extra RUN line. Generate clean
FileCheck statements with my script. Remove the now merged AVX2 tests.
llvm-svn: 218903
Chandler Carruth