The IBM BG/Q supercomputer's A2 cores have a hardware prefetching unit, the
L1P, but it does not prefetch directly into the A2's L1 cache. Instead, it
prefetches into its own L1P buffer, and the latency to access that buffer is
significantly higher than that to the L1 cache (although smaller than the
latency to the L2 cache). As a result, especially when multiple hardware
threads are not actively busy, explicitly prefetching data into the L1 cache is
advantageous.
I've been using this pass out-of-tree for data prefetching on the BG/Q for well
over a year, and it has worked quite well. It is enabled by default only for
the BG/Q, but can be enabled for other cores as well via a command-line option.
Eventually, we might want to add some TTI interfaces and move this into
Transforms/Scalar (there is nothing particularly target dependent about it,
although only machines like the BG/Q will benefit from its simplistic
strategy).
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229966 91177308-0d34-0410-b5e6-96231b3b80d8
The new shuffle lowering has been the default for some time. I've
enabled the new legality testing by default with no really blocking
regressions. I've fuzz tested this very heavily (many millions of fuzz
test cases have passed at this point). And this cleans up a ton of code.
=]
Thanks again to the many folks that helped with this transition. There
was a lot of work by others that went into the new shuffle lowering to
make it really excellent.
In case you aren't using a diff algorithm that can handle this:
X86ISelLowering.cpp: 22 insertions(+), 2940 deletions(-)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229964 91177308-0d34-0410-b5e6-96231b3b80d8
is going well, remove the flag and the code for the old legality tests.
This is the first step toward removing the entire old vector shuffle
lowering. *Much* more code to delete coming up next.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229963 91177308-0d34-0410-b5e6-96231b3b80d8
reflects the fact that the x86 backend can in fact lower any shuffle you
want it to with reasonably high code quality.
My recent work on the new vector shuffle has made this regress *very*
little. The diff in the test cases makes me very, very happy.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229958 91177308-0d34-0410-b5e6-96231b3b80d8
The instructions were being generated on architectures that don't support avx512.
This reverts commit r229837.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229942 91177308-0d34-0410-b5e6-96231b3b80d8
This re-applies r223862, r224198, r224203, and r224754, which were
reverted in r228129 because they exposed Clang misalignment problems
when self-hosting.
The combine caused the crashes because we turned ISD::LOAD/STORE nodes
to ARMISD::VLD1/VST1_UPD nodes. When selecting addressing modes, we
were very lax for the former, and only emitted the alignment operand
(as in "[r1:128]") when it was larger than the standard alignment of
the memory type.
However, for ARMISD nodes, we just used the MMO alignment, no matter
what. In our case, we turned ISD nodes to ARMISD nodes, and this
caused the alignment operands to start being emitted.
And that's how we exposed alignment problems that were ignored before
(but I believe would have been caught with SCTRL.A==1?).
To fix this, we can just mirror the hack done for ISD nodes: only
take into account the MMO alignment when the access is overaligned.
Original commit message:
We used to only combine intrinsics, and turn them into VLD1_UPD/VST1_UPD
when the base pointer is incremented after the load/store.
We can do the same thing for generic load/stores.
Note that we can only combine the first load/store+adds pair in
a sequence (as might be generated for a v16f32 load for instance),
because other combines turn the base pointer addition chain (each
computing the address of the next load, from the address of the last
load) into independent additions (common base pointer + this load's
offset).
rdar://19717869, rdar://14062261.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229932 91177308-0d34-0410-b5e6-96231b3b80d8
In preparation for a future patch:
- rename isLoad to isLoadOp: the former is confusing, and can be taken
to refer to the fact that the node is an ISD::LOAD. (it isn't, yet.)
- change formatting here and there.
- add some comments.
- const-ify bools.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229929 91177308-0d34-0410-b5e6-96231b3b80d8
systematic lowering of v8i16.
This required a slight strategy shift to prefer unpack lowerings in more
places. While this isn't a cut-and-dry win in every case, it is in the
overwhelming majority. There are only a few places where the old
lowering would probably be a touch faster, and then only by a small
margin.
In some cases, this is yet another significant improvement.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229859 91177308-0d34-0410-b5e6-96231b3b80d8
addition to lowering to trees rooted in an unpack.
This saves shuffles and or registers in many various ways, lets us
handle another class of v4i32 shuffles pre SSE4.1 without domain
crosses, etc.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229856 91177308-0d34-0410-b5e6-96231b3b80d8
terribly complex partial blend logic.
This code path was one of the more complex and bug prone when it first
went in and it hasn't faired much better. Ultimately, with the simpler
basis for unpack lowering and support bit-math blending, this is
completely obsolete. In the worst case without this we generate
different but equivalent instructions. However, in many cases we
generate much better code. This is especially true when blends or pshufb
is available.
This does expose one (minor) weakness of the unpack lowering that I'll
try to address.
In case you were wondering, this is actually a big part of what I've
been trying to pull off in the recent string of commits.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229853 91177308-0d34-0410-b5e6-96231b3b80d8
needed, and significantly improve the SSSE3 path.
This makes the new strategy much more clear. If we can blend, we just go
with that. If we can't blend, we try to permute into an unpack so
that we handle cases where the unpack doing the blend also simplifies
the shuffle. If that fails and we've got SSSE3, we now call into
factored-out pshufb lowering code so that we leverage the fact that
pshufb can set up a blend for us while shuffling. This generates great
code, especially because we *know* we don't have a fast blend at this
point. Finally, we fall back on decomposing into permutes and blends
because we do at least have a bit-math-based blend if we need to use
that.
This pretty significantly improves some of the v8i16 code paths. We
never need to form pshufb for the single-input shuffles because we have
effective target-specific combines to form it there, but we were missing
its effectiveness in the blends.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229851 91177308-0d34-0410-b5e6-96231b3b80d8
them into permutes and a blend with the generic decomposition logic.
This works really well in almost every case and lets the code only
manage the expansion of a single input into two v8i16 vectors to perform
the actual shuffle. The blend-based merging is often much nicer than the
pack based merging that this replaces. The only place where it isn't we
end up blending between two packs when we could do a single pack. To
handle that case, just teach the v2i64 lowering to handle these blends
by digging out the operands.
With this we're down to only really random permutations that cause an
explosion of instructions.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229849 91177308-0d34-0410-b5e6-96231b3b80d8
v16i8 shuffles, and replace it with new facilities.
This uses precise patterns to match exact unpacks, and the new
generalized unpack lowering only when we detect a case where we will
have to shuffle both inputs anyways and they terminate in exactly
a blend.
This fixes all of the blend horrors that I uncovered by always lowering
blends through the vector shuffle lowering. It also removes *sooooo*
much of the crazy instruction sequences required for v16i8 lowering
previously. Much cleaner now.
The only "meh" aspect is that we sometimes use pshufb+pshufb+unpck when
it would be marginally nicer to use pshufb+pshufb+por. However, the
difference there is *tiny*. In many cases its a win because we re-use
the pshufb mask. In others, we get to avoid the pshufb entirely. I've
left a FIXME, but I'm dubious we can really do better than this. I'm
actually pretty happy with this lowering now.
For SSE2 this exposes some horrors that were really already there. Those
will have to fixed by changing a different path through the v16i8
lowering.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229846 91177308-0d34-0410-b5e6-96231b3b80d8
on things not being marked as either custom or legal, but we now do
custom lowering of more VSELECT nodes. To cope with this, manually
replicate the legality tests here. These have to stay in sync with the
set of tests used in the custom lowering of VSELECT.
Ideally, we wouldn't do any of this combine-based-legalization when we
have an actual custom legalization step for VSELECT, but I'm not going
to be able to rewrite all of that today.
I don't have a test case for this currently, but it was found when
compiling a number of the test-suite benchmarks. I'll try to reduce
a test case and add it.
This should at least fix the test-suite fallout on build bots.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229844 91177308-0d34-0410-b5e6-96231b3b80d8
lowering paths. I'm going to be leveraging this to simplify a lot of the
overly complex lowering of v8 and v16 shuffles in pre-SSSE3 modes.
Sadly, this isn't profitable on v4i32 and v2i64. There, the float and
double blending instructions for pre-SSE4.1 are actually pretty good,
and we can't beat them with bit math. And once SSE4.1 comes around we
have direct blending support and this ceases to be relevant.
Also, some of the test cases look odd because the domain fixer
canonicalizes these to floating point domain. That's OK, it'll use the
integer domain when it matters and some day I may be able to update
enough of LLVM to canonicalize the other way.
This restores almost all of the regressions from teaching x86's vselect
lowering to always use vector shuffle lowering for blends. The remaining
problems are because the v16 lowering path is still doing crazy things.
I'll be re-arranging that strategy in more detail in subsequent commits
to finish recovering the performance here.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229836 91177308-0d34-0410-b5e6-96231b3b80d8
First, don't combine bit masking into vector shuffles (even ones the
target can handle) once operation legalization has taken place. Custom
legalization of vector shuffles may exist for these patterns (making the
predicate return true) but that custom legalization may in some cases
produce the exact bit math this matches. We only really want to handle
this prior to operation legalization.
However, the x86 backend, in a fit of awesome, relied on this. What it
would do is mark VSELECTs as expand, which would turn them into
arithmetic, which this would then match back into vector shuffles, which
we would then lower properly. Amazing.
Instead, the second change is to teach the x86 backend to directly form
vector shuffles from VSELECT nodes with constant conditions, and to mark
all of the vector types we support lowering blends as shuffles as custom
VSELECT lowering. We still mark the forms which actually support
variable blends as *legal* so that the custom lowering is bypassed, and
the legal lowering can even be used by the vector shuffle legalization
(yes, i know, this is confusing. but that's how the patterns are
written).
This makes the VSELECT lowering much more sensible, and in fact should
fix a bunch of bugs with it. However, as you'll see in the test cases,
right now what it does is point out the *hilarious* deficiency of the
new vector shuffle lowering when it comes to blends. Fortunately, my
very next patch fixes that. I can't submit it yet, because that patch,
somewhat obviously, forms the exact and/or pattern that the DAG combine
is matching here! Without this patch, teaching the vector shuffle
lowering to produce the right code infloops in the DAG combiner. With
this patch alone, we produce terrible code but at least lower through
the right paths. With both patches, all the regressions here should be
fixed, and a bunch of the improvements (like using 2 shufps with no
memory loads instead of 2 andps with memory loads and an orps) will
stay. Win!
There is one other change worth noting here. We had hilariously wrong
vectorization cost estimates for vselect because we fell through to the
code path that assumed all "expand" vector operations are scalarized.
However, the "expand" lowering of VSELECT is vector bit math, most
definitely not scalarized. So now we go back to the correct if horribly
naive cost of "1" for "not scalarized". If anyone wants to add actual
modeling of shuffle costs, that would be cool, but this seems an
improvement on its own. Note the removal of 16 and 32 "costs" for doing
a blend. Even in SSE2 we can blend in fewer than 16 instructions. ;] Of
course, we don't right now because of OMG bad code, but I'm going to fix
that. Next patch. I promise.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229835 91177308-0d34-0410-b5e6-96231b3b80d8
Previously, subtarget features were a bitfield with the underlying type being uint64_t.
Since several targets (X86 and ARM, in particular) have hit or were very close to hitting this bound, switching the features to use a bitset.
No functional change.
Differential Revision: http://reviews.llvm.org/D7065
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229831 91177308-0d34-0410-b5e6-96231b3b80d8
accordingly. This changes the constructors of a number of classes
that don't need to know the subtarget's 64-bitness.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229787 91177308-0d34-0410-b5e6-96231b3b80d8
This involved moving two non-subtarget dependent features (64-bitness
and the driver interface) to the NVPTX target machine and updating
the uses (or migrating around the subtarget use for ease of review).
Otherwise use the cached subtarget or create a default subtarget
based on the TargetMachine cpu and feature string for the module
level assembler emission.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229785 91177308-0d34-0410-b5e6-96231b3b80d8
VOP2 declares vsrc1, but VOP3 declares src1.
We can't use the same "ins" if the operands have different names in VOP2
and VOP3 encodings.
This fixes a hang in geometry shaders which spill M0 on VI.
(BTW it doesn't look like M0 needs spilling and the spilling seems
duplicated 3 times)
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@229752 91177308-0d34-0410-b5e6-96231b3b80d8
