Resubmit after fixing test/Transforms/LoopVectorize/ARM/mve-gather-scatter-tailpred.ll
Previous commit message...
This is a resubmit of 3e5ce4 (which was reverted by 7fe41ac). The original commit caused a PPC build bot failure we never really got to the bottom of. I can't reproduce the issue, and the bot owner was non-responsive. In the meantime, we stumbled across an issue which seems possibly related, and worked around a latent bug in 80e8025. My best guess is that the original patch exposed that latent issue at higher frequency, but it really is just a guess.
Original commit message follows...
If we know that the scalar epilogue is required to run, modify the CFG to end the middle block with an unconditional branch to scalar preheader. This is instead of a conditional branch to either the preheader or the exit block.
The motivation to do this is to support multiple exit blocks. Specifically, the current structure forces us to identify immediate dominators and *which* exit block to branch from in the middle terminator. For the multiple exit case - where we know require scalar will hold - these questions are ill formed.
This is the last change needed to support multiple exit loops, but since the diffs are already large enough, I'm going to land this, and then enable separately. You can think of this as being NFCIish prep work, but the changes are a bit too involved for me to feel comfortable tagging the review that way.
Differential Revision: https://reviews.llvm.org/D94892
The loop vectorizer will currently assume a large trip count when
calculating which of several vectorization factors are more profitable.
That is often not a terrible assumption to make as small trip count
loops will usually have been fully unrolled. There are cases however
where we will try to vectorize them, and especially when folding the
tail by masking can incorrectly choose to vectorize loops that are not
beneficial, due to the folded tail rounding the iteration count up for
the vectorized loop.
The motivating example here has a trip count of 5, so either performs 5
scalar iterations or 2 vector iterations (with VF=4). At a high enough
trip count the vectorization becomes profitable, but the rounding up to
2 vector iterations vs only 5 scalar makes it unprofitable.
This adds an alternative cost calculation when we know the max trip
count and are folding tail by masking, rounding the iteration count up
to the correct number for the vector width. We still do not account for
anything like setup cost or the mixture of vector and scalar loops, but
this is at least an improvement in a few cases that we have had
reported.
Differential Revision: https://reviews.llvm.org/D101726
D99674 stopped the folding of certain select operations into and/or, due
to incorrect folding in the presence of poison. D97360 added some costs
to attempt to account for the change, but only worked at the getUserCost
level, not the getCmpSelInstrCost that the vectorizer will use directly.
This adds similar logic into the vectorizer to handle these logical
and/or selects, treating them like and/or directly.
This fixes 60% performance regressions from code like the attached test
case.
Differential Revision: https://reviews.llvm.org/D99884
The scalarization overhead was set deliberately high for MVE, whilst the
codegen was new. It helps protect us against the negative ramifications
of mixing scalar and vector instructions. This decreases that,
especially for floating point where the cost of extracting/inserting
lane elements can be low. For integer the cost is still fairly high due
to the cross-register-bank copy, but is no longer n^2 in the length of
the vector.
In general, this will decrease the cost of scalarizing floats and long
integer vectors. i64 increase in cost, having a high cost before and
after this patch. For floats this allows up to start doing things like
vectorizing fdiv instructions, even if they are scalarized.
Differential Revision: https://reviews.llvm.org/D98245
This test shows a case where we can potentially scalarize the store in a
predicated loop, creating a lot of instructions that would be much
slower than scalar.
A v8i32 compare will produce a v8i1 predicate, but during codegen the
v8i32 will be split into two v4i32, potentially requiring two v4i1
predicates to be merged into a single v8i1. Because this merging of two
v4i1's into a v8i1 is very expensive, we need to make the cost of the
compare equally high.
This patch adds the cost of that to ARMTTIImpl::getCmpSelInstrCost.
Because we don't know whether the user of the predicate can be split,
and the cost model is mostly pre-instruction, we may be pessimistic but
that should only be for larger and legal types. This also adds min/max
detection to the costmodel where it can be detected, to keep those in
line with the cost of simple min/max instructions. Otherwise for the
most part, costs that were already expensive have become more expensive.
Differential Revision: https://reviews.llvm.org/D96692
The vector reduction intrinsics started life as experimental ops, so backend support
was lacking. As part of promoting them to 1st-class intrinsics, however, codegen
support was added/improved:
D58015
D90247
So I think it is safe to now remove this complication from IR.
Note that we still have an IR-level codegen expansion pass for these as discussed
in D95690. Removing that is another step in simplifying the logic. Also note that
x86 was already unconditionally forming reductions in IR, so there should be no
difference for x86.
I spot checked a couple of the tests here by running them through opt+llc and did
not see any asm diffs.
If we do find functional differences for other targets, it should be possible
to (at least temporarily) restore the shuffle IR with the ExpandReductions IR
pass.
Differential Revision: https://reviews.llvm.org/D96552
This adds cost modelling for the inloop vectorization added in
745bf6cf44. Up until now they have been modelled as the original
underlying instruction, usually an add. This happens to works OK for MVE
with instructions that are reducing into the same type as they are
working on. But MVE's instructions can perform the equivalent of an
extended MLA as a single instruction:
%sa = sext <16 x i8> A to <16 x i32>
%sb = sext <16 x i8> B to <16 x i32>
%m = mul <16 x i32> %sa, %sb
%r = vecreduce.add(%m)
->
R = VMLADAV A, B
There are other instructions for performing add reductions of
v4i32/v8i16/v16i8 into i32 (VADDV), for doing the same with v4i32->i64
(VADDLV) and for performing a v4i32/v8i16 MLA into an i64 (VMLALDAV).
The i64 are particularly interesting as there are no native i64 add/mul
instructions, leading to the i64 add and mul naturally getting very
high costs.
Also worth mentioning, under NEON there is the concept of a sdot/udot
instruction which performs a partial reduction from a v16i8 to a v4i32.
They extend and mul/sum the first four elements from the inputs into the
first element of the output, repeating for each of the four output
lanes. They could possibly be represented in the same way as above in
llvm, so long as a vecreduce.add could perform a partial reduction. The
vectorizer would then produce a combination of in and outer loop
reductions to efficiently use the sdot and udot instructions. Although
this patch does not do that yet, it does suggest that separating the
input reduction type from the produced result type is a useful concept
to model. It also shows that a MLA reduction as a single instruction is
fairly common.
This patch attempt to improve the costmodelling of in-loop reductions
by:
- Adding some pattern matching in the loop vectorizer cost model to
match extended reduction patterns that are optionally extended and/or
MLA patterns. This marks the cost of the reduction instruction correctly
and the sext/zext/mul leading up to it as free, which is otherwise
difficult to tell and may get a very high cost. (In the long run this
can hopefully be replaced by vplan producing a single node and costing
it correctly, but that is not yet something that vplan can do).
- getExtendedAddReductionCost is added to query the cost of these
extended reduction patterns.
- Expanded the ARM costs to account for these expanded sizes, which is a
fairly simple change in itself.
- Some minor alterations to allow inloop reduction larger than the highest
vector width and i64 MVE reductions.
- An extra InLoopReductionImmediateChains map was added to the vectorizer
for it to efficiently detect which instructions are reductions in the
cost model.
- The tests have some updates to show what I believe is optimal
vectorization and where we are now.
Put together this can greatly improve performance for reduction loop
under MVE.
Differential Revision: https://reviews.llvm.org/D93476
It turns out the vectorizer calls the getIntrinsicInstrCost functions
with a scalar return type and vector VF. This updates the costmodel to
handle that, still producing the correct vector costs.
A vectorizer test is added to show it vectorizing at the correct factor
again.
This patch makes SLP and LV emit operations with initial vectors set to poison constant instead of undef.
This is a part of efforts for using poison vector instead of undef to represent "doesn't care" vector.
The goal is to make nice shufflevector optimizations valid that is currently incorrect due to the tricky interaction between undef and poison (see https://bugs.llvm.org/show_bug.cgi?id=44185 ).
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D94061
Creating in-loop reductions relies on IR references to map
IR values to VPValues after interleave group creation.
Make sure we re-add the updated member to the plan, so the look-ups
still work as expected
This fixes a crash reported after D90562.
As mentioned in D93793, there are quite a few places where unary `IRBuilder::CreateShuffleVector(X, Mask)` can be used
instead of `IRBuilder::CreateShuffleVector(X, Undef, Mask)`.
Let's update them.
Actually, it would have been more natural if the patches were made in this order:
(1) let them use unary CreateShuffleVector first
(2) update IRBuilder::CreateShuffleVector to use poison as a placeholder value (D93793)
The order is swapped, but in terms of correctness it is still fine.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D93923
This patch updates IRBuilder to create insertelement/shufflevector using poison as a placeholder.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93793
Temporarily revert commit 8b1c4e310c.
After 8b1c4e310c the compile-time for `MultiSource/Benchmarks/MiBench/consumer-lame`
dramatically increases with -O3 & LTO, causing issues for builders with
that configuration.
I filed PR48553 with a smallish reproducer that shows a 10-100x compile
time increase.
... so just ensure that we pass DomTreeUpdater it into it.
Fixes DomTree preservation for a large number of tests,
all of which are marked as such so that they do not regress.
When it comes to the scalar cost of any predicated block, the loop
vectorizer by default regards this predication as a sign that it is
looking at an if-conversion and divides the scalar cost of the block by
2, assuming it would only be executed half the time. This however makes
no sense if the predication has been introduced to tail predicate the
loop.
Original patch by Anna Welker
Differential Revision: https://reviews.llvm.org/D86452
If we have two unknown sizes and one GEP operand and one non-GEP
operand, then we currently simply return MayAlias. The comment says
we can't do anything useful ... but we can! We can still check that
the underlying objects are different (and do so for the GEP-GEP case).
To reduce the compile-time impact, this a) checks this early, before
doing the relatively expensive GEP decomposition that will not be
used and b) doesn't do the check if the other operand is a phi or
select. In that case, the phi/select will already recurse, so this
would just do two slightly different recursive walks that arrive at
the same roots.
Compile-time is still a bit of a mixed bag: https://llvm-compile-time-tracker.com/compare.php?from=624af932a808b363a888139beca49f57313d9a3b&to=845356e14adbe651a553ed11318ddb5e79a24bcd&stat=instructions
On average this is a small improvement, but sqlite with ThinLTO has
a 0.5% regression (lencod has a 1% improvement).
The BasicAA test case checks this by using two memsets with unknown
size. However, the more interesting case where this is useful is
the LoopVectorize test case, as analysis of accesses in loops tends
to always us unknown sizes.
Differential Revision: https://reviews.llvm.org/D92401
Use -0.0 instead of 0.0 as the start value. The previous use of 0.0
was fine for all existing uses of this function though, as it is
always generated with fast flags right now, and thus nsz.
This expands upon the inloop reductions added in e9761688e41cb9e976,
allowing them to be inserted into tail folded loops. Reductions are
generates with the form:
x = select(mask, vecop, zero)
v = vecreduce.add(x)
c = add chain, v
Where zero here is chosen as the identity value for add reductions. The
backend is then expected to fold the select and the vecreduce into a
single predicated instruction.
Most of the code is fairly straight forward, except for the creation of
blockmasks which need to ensure they are created in dominance order. The
order they are added is altered to be after any phis, keeping the
requirements for the underlying IR.
Differential Revision: https://reviews.llvm.org/D84451
We currently collect the ICmp and Add from an induction variable,
marking them as dead so that vplan values are not created for them. This
extends that to include any single use trunk from the ICmp, which allows
the Add to more readily be removed too.
This can help with costing vplan nodes, as the ICmp and Add are more
reliably removed and are not double-counted.
Differential Revision: https://reviews.llvm.org/D88873
We have been running tests/benchmarks downstream with tail-predication enabled
for some time now and this behaves as expected: we are not aware of any
correctness issues, and this performs better across the board than with
tail-predication disabled. Time to flip the switch!
Differential Revision: https://reviews.llvm.org/D88093
This allows the backend to tell the vectorizer to produce inloop
reductions through a TTI hook.
For the moment on ARM under MVE this means allowing integer add
reductions of the correct size. In the future this can include integer
min/max too, under -Os.
Differential Revision: https://reviews.llvm.org/D75512
This implements 2 different vectorisation fallback strategies if tail-folding
fails: 1) don't vectorise at all, or 2) vectorise using a scalar epilogue. This
can be controlled with option -prefer-predicate-over-epilogue, that has been
changed to take a numeric value corresponding to the tail-folding preference
and preferred fallback.
Patch by: Pierre van Houtryve, Sjoerd Meijer.
Differential Revision: https://reviews.llvm.org/D79783
MVE Gather scatter codegeneration is looking a lot better than it used
to, but still has some issues. The instructions we currently model as 1
cycle per element, which is a bit low for some cases. Increasing the
cost by the MVECostFactor brings them in-line with our other instruction
costs. This will have the effect of only generating then when the extra
benefit is more likely to overcome some of the issues. Notably in
running out of registers and vectorizing loops that could otherwise be
SLP vectorized.
In the short-term whilst we look at other ways of dealing with those
more directly, we can increase the costs of gathers to make them more
likely to be beneficial when created.
Differential Revision: https://reviews.llvm.org/D86444
This adapts LV to the new semantics of get.active.lane.mask as discussed in
D86147, which means that the LV now emits intrinsic get.active.lane.mask with
the loop tripcount instead of the backedge-taken count as its second argument.
The motivation for this is described in D86147.
Differential Revision: https://reviews.llvm.org/D86304
If gather/scatters are enabled, ARMTargetTransformInfo now allows
tail predication for loops with a much wider range of strides, up
to anything that is loop invariant.
Differential Revision: https://reviews.llvm.org/D85410
As part of D84741, this adds a target hook for the
preferPredicatedReductionSelect option and makes use
of it under MVE, allowing us to tail predicate most
reduction loops.
Differential Revision: https://reviews.llvm.org/D85980
This patch uses the feature added in D79162 to fix the cost of a
sext/zext of a masked load, or a trunc for a masked store.
Previously, those were considered cheap or even free, but it's
not the case as we cannot split the load in the same way we would for
normal loads.
This updates the costs to better reflect reality, and adds a test for it
in test/Analysis/CostModel/ARM/cast.ll.
It also adds a vectorizer test that showcases the improvement: in some
cases, the vectorizer will now choose a smaller VF when
tail-predication is enabled, which results in better codegen. (Because
if it were to use a higher VF in those cases, the code we see above
would be generated, and the vmovs would block tail-predication later in
the process, resulting in very poor codegen overall)
Original Patch by Pierre van Houtryve
Differential Revision: https://reviews.llvm.org/D79163
It was getting difficult to see which test was in which file, so this
reorganises the test files so that now all filenames start with tail-folding-*
followed by a more descriptive name what that group of tests check.
This patch enables the LoopVectorizer to build a phi of pointer
type and provide the vector loads and stores with vector type
getelementptrs built from the pointer induction variable, which
produces much less instructions than the previous approach of
creating scalar getelementpointers and glue them together to a
vector.
Differential Revision: https://reviews.llvm.org/D81267
If a vector body has live-out values, it is probably a reduction, which needs a
final reduction step after the loop. MVE has a VADDV instruction to reduce
integer vectors, but doesn't have an equivalent one for float vectors. A
live-out value that is not recognised as reduction later in the optimisation
pipeline will result in the tail-predicated loop to be reverted to a
non-predicated loop and this is very expensive, i.e. it has a significant
performance impact, which is what we hope to avoid with fine tuning the ARM TTI
hook preferPredicateOverEpilogue implementation.
Differential Revision: https://reviews.llvm.org/D82953
This refactors option -disable-mve-tail-predication to take different arguments
so that we have 1 option to control tail-predication rather than several
different ones.
This is also a prep step for D82953, in which we want to reject reductions
unless that is requested with this option.
Differential Revision: https://reviews.llvm.org/D83133
This adjusts the MVE fp16 cost model, similar to how we already do for
integer casts. It uses the base cost of 1 per cvt for most fp extend /
truncates, but adjusts it for loads and stores where we know that a
extending load has been used to get the load into the correct lane, and
only an MVE VCVTB is then needed.
Differential Revision: https://reviews.llvm.org/D81813
This alters getMemoryOpCost to use the Base TargetTransformInfo version
that includes some additional checks for whether extending loads are
legal. This will generally have the effect of making <2 x ..> and some
<4 x ..> loads/stores more expensive, which in turn should help favour
larger vector factors.
Notably it alters the cost of a <4 x half>, which with the current
codegen will be expensive if it is not extended.
Differential Revision: https://reviews.llvm.org/D82456
This patch enables the LoopVectorizer to build a phi of pointer
type and provide the vector loads and stores with vector type
getelementptrs built from the pointer induction variable, which
produces much less instructions than the previous approach of
creating scalar getelementpointers and glue them together to a
vector.
Differential Revision: https://reviews.llvm.org/D81267
This emits new IR intrinsic @llvm.get.active.mask for tail-folded vectorised
loops if the intrinsic is supported by the backend, which is checked by
querying TargetTransform hook emitGetActiveLaneMask.
This intrinsic creates a mask representing active and inactive vector lanes,
which is used by the masked load/store instructions that are created for
tail-folded loops. The semantics of @llvm.get.active.mask are described here in
LangRef:
https://llvm.org/docs/LangRef.html#llvm-get-active-lane-mask-intrinsics
This intrinsic is also used to provide a hint to the backend. That is, the
second argument of the intrinsic represents the back-edge taken count of the
loop. For MVE, for example, we use that to set up tail-predication, which is a
new form of predication in MVE for vector loops that implicitely predicates the
last vector loop iteration by implicitely setting active/inactive lanes, i.e.
the tail loop is predicated. In order to set up a tail-predicated vector loop,
we need to know the number of data elements processed by the vector loop, which
corresponds the the tripcount of the scalar loop, which we can now reconstruct
using @llvm.get.active.mask.
Differential Revision: https://reviews.llvm.org/D79100
Similar to a recent change to the X86 backend, this changes things so
that we always produce a reduction intrinsics for all reduction types,
not just the legal ones. This gives a better chance in the backend to
custom lower them to something more suitable for MVE. Especially for
something like fadd the in-order reduction produced during DAG lowering
is already better than the shuffles produced in the midend, and we can
do even better with a bit of custom lowering.
Differential Revision: https://reviews.llvm.org/D81398
This was reverted because of a miscompilation. At closer inspection, the
problem was actually visible in a changed llvm regression test too. This
one-line follow up fix/recommit will splat the IV, which is what we are trying
to avoid if unnecessary in general, if tail-folding is requested even if all
users are scalar instructions after vectorisation. Because with tail-folding,
the splat IV will be used by the predicate of the masked loads/stores
instructions. The previous version omitted this, which caused the
miscompilation. The original commit message was:
If tail-folding of the scalar remainder loop is applied, the primary induction
variable is splat to a vector and used by the masked load/store vector
instructions, thus the IV does not remain scalar. Because we now mark
that the IV does not remain scalar for these cases, we don't emit the vector IV
if it is not used. Thus, the vectoriser produces less dead code.
Thanks to Ayal Zaks for the direction how to fix this.
- Specifically check for sext/zext users which have 'long' form NEON
instructions.
- Add more entries to the table for sext/zexts so that we can report
more accurately the number of vmovls required for NEON.
- Pass the instruction to the pass implementation.
Differential Revision: https://reviews.llvm.org/D79561
If tail-folding of the scalar remainder loop is applied, the primary induction
variable is splat to a vector and used by the masked load/store vector
instructions, thus the IV does not remain scalar. Because we now mark
that the IV does not remain scalar for these cases, we don't emit the vector IV
if it is not used. Thus, the vectoriser produces less dead code.
Thanks to Ayal Zaks for the direction how to fix this.
Differential Revision: https://reviews.llvm.org/D78911
D77635 added support to recognise primary induction variables for counting-down
loops. This allows us to fold the scalar tail loop into the main vector body,
which we need for MVE tail-predication. This adds some ARM tail-folding test
cases that we want to support.
This test was extracted from D76838, which implemented a different approach to
reverse and thus find a primary induction variable.
The codegen for splitting a llvm.vector.reduction intrinsic into parts
will be better than the codegen for the generic reductions. This will
only directly effect when vectorization factors are specified by the
user.
Also added tests to make sure the codegen for larger reductions is OK.
Differential Revision: https://reviews.llvm.org/D72257
Architecturally, it's allowed to have MVE-I without an FPU, thus
-mfpu=none should not disable MVE-I, or moves to/from FP-registers.
This patch removes `+/-fpregs` from features unconditionally added to
target feature list, depending on FPU and moves the logic to Clang
driver, where the negative form (`-fpregs`) is conditionally added to
the target features list for the cases of `-mfloat-abi=soft`, or
`-mfpu=none` without either `+mve` or `+mve.fp`. Only the negative
form is added by the driver, the positive one is derived from other
features in the backend.
Differential Revision: https://reviews.llvm.org/D71843
This addresses a vectorisation regression for tail-folded loops that are
counting down, e.g. loops as simple as this:
void foo(char *A, char *B, char *C, uint32_t N) {
while (N > 0) {
*C++ = *A++ + *B++;
N--;
}
}
These are loops that can be vectorised, but when tail-folding is requested, it
can't find a primary induction variable which we do need for predicating the
loop. As a result, the loop isn't vectorised at all, which it is able to do
when tail-folding is not attempted. So, this adds a check for the primary
induction variable where we decide how to lower the scalar epilogue. I.e., when
there isn't a primary induction variable, a scalar epilogue loop is allowed
(i.e. don't request tail-folding) so that vectorisation could still be
triggered.
Having this check for the primary induction variable make sense anyway, and in
addition, in a follow-up of this I will look into discovering earlier the
primary induction variable for counting down loops, so that this can also be
tail-folded.
Differential revision: https://reviews.llvm.org/D72324
With the extra optimisations we have done, these should now be fine to
enable by default. Which is what this patch does.
Differential Revision: https://reviews.llvm.org/D70968
This attempts to teach the cost model in Arm that code such as:
%s = shl i32 %a, 3
%a = and i32 %s, %b
Can under Arm or Thumb2 become:
and r0, r1, r2, lsl #3
So the cost of the shift can essentially be free. To do this without
trying to artificially adjust the cost of the "and" instruction, it
needs to get the users of the shl and check if they are a type of
instruction that the shift can be folded into. And so it needs to have
access to the actual instruction in getArithmeticInstrCost, which if
available is added as an extra parameter much like getCastInstrCost.
We otherwise limit it to shifts with a single user, which should
hopefully handle most of the cases. The list of instruction that the
shift can be folded into include ADC, ADD, AND, BIC, CMP, EOR, MVN, ORR,
ORN, RSB, SBC and SUB. This translates to Add, Sub, And, Or, Xor and
ICmp.
Differential Revision: https://reviews.llvm.org/D70966
This adds some extra cost model tests for shifts, and does some minor
adjustments to some Neon code to make it clear as to what it applies to.
Both NFC.
Alas, using half the available vector registers in a single instruction
is just too much for the register allocator to handle. The mve-vldst4.ll
test here fails when these instructions are enabled at present. This
patch disables the generation of VLD4 and VST4 by adding a
mve-max-interleave-factor option, which we currently default to 2.
Differential Revision: https://reviews.llvm.org/D71109
Follow-up of cb47b8783: don't query TTI->preferPredicateOverEpilogue when
option -prefer-predicate-over-epilog is set to false, i.e. when we prefer not
to predicate the loop.
Differential Revision: https://reviews.llvm.org/D70382
Now that we have the intrinsics, we can add VLD2/4 and VST2/4 lowering
for MVE. This works the same way as Neon, recognising the load/shuffles
combination and converting them into intrinsics in a pre-isel pass,
which just calls getMaxSupportedInterleaveFactor, lowerInterleavedLoad
and lowerInterleavedStore.
The main difference to Neon is that we do not have a VLD3 instruction.
Otherwise most of the code works very similarly, with just some minor
differences in the form of the intrinsics to work around. VLD3 is
disabled by making isLegalInterleavedAccessType return false for those
cases.
We may need some other future adjustments, such as VLD4 take up half the
available registers so should maybe cost more. This patch should get the
basics in though.
Differential Revision: https://reviews.llvm.org/D69392
This is a follow up of d90804d, to also flag fmcp instructions as instructions
that we do not support in tail-predicated vector loops.
Differential Revision: https://reviews.llvm.org/D70295
The vectoriser queries TTI->preferPredicateOverEpilogue to determine if
tail-folding is preferred for a loop, but it was not respecting loop hint
'predicate' that can disable this, which has now been added. This showed that
we were incorrectly initialising loop hint 'vectorize.predicate.enable' with 0
(i.e. FK_Disabled) but this should have been FK_Undefined, which has been
fixed.
Differential Revision: https://reviews.llvm.org/D70125
This implements TTI hook 'preferPredicateOverEpilogue' for MVE. This is a
first version and it operates on single block loops only. With this change, the
vectoriser will now determine if tail-folding scalar remainder loops is
possible/desired, which is the first step to generate MVE tail-predicated
vector loops.
This is disabled by default for now. I.e,, this is depends on option
-disable-mve-tail-predication, which is off by default.
I will follow up on this soon with a patch for the vectoriser to respect loop
hint 'vectorize.predicate.enable'. I.e., with this loop hint set to Disabled,
we don't want to tail-fold and we shouldn't query this TTI hook, which is
done in D70125.
Differential Revision: https://reviews.llvm.org/D69845
We have two ways to steer creating a predicated vector body over creating a
scalar epilogue. To force this, we have 1) a command line option and 2) a
pragma available. This adds a third: a target hook to TargetTransformInfo that
can be queried whether predication is preferred or not, which allows the
vectoriser to make the decision without forcing it.
While this change behaves as a non-functional change for now, it shows the
required TTI plumbing, usage of this new hook in the vectoriser, and the
beginning of an ARM MVE implementation. I will follow up on this with:
- a complete MVE implementation, see D69845.
- a patch to disable this, i.e. we should respect "vector_predicate(disable)"
and its corresponding loophint.
Differential Revision: https://reviews.llvm.org/D69040
Add generic DAG combine for extending masked loads.
Allow us to generate sext/zext masked loads which can access v4i8,
v8i8 and v4i16 memory to produce v4i32, v8i16 and v4i32 respectively.
Differential Revision: https://reviews.llvm.org/D68337
llvm-svn: 375085
Now that the vectorizer can do tail-folding (rL367592), and the ARM backend
understands MVE masked loads/stores (rL371932), it's time to add the MVE
tail-folding equivalent of the X86 tests that I added.
llvm-svn: 371996
Masked loads and store fit naturally with MVE, the instructions being easily
predicated. This adds lowering for the simple cases of masked loads and stores.
It does not yet deal with widening/narrowing or pre/post inc, and so is
currently behind an option.
The llvm masked load intrinsic will accept a "passthru" value, dictating the
values used for the zero masked lanes. In MVE the instructions write 0 to the
zero predicated lanes, so we need to match a passthru that isn't 0 (or undef)
with a select instruction to pull in the correct data after the load.
Differential Revision: https://reviews.llvm.org/D67186
llvm-svn: 371932
We don't yet know how to generate these instructions for MVE. And in the case
of VLD3, we don't even have the instruction. For the moment don't tell the
vectoriser that we have VLD4, just to end up serialising the results.
Differential Revision: https://reviews.llvm.org/D66009
llvm-svn: 369101
With enough codegen complete, we can now correctly report the number and size
of vector registers for MVE, allowing auto vectorisation. This also allows FP
auto-vectorization for MVE without -Ofast/-ffast-math, due to support for IEEE
FP arithmetic and parity between scalar and vector FP behaviour.
Patch by David Sherwood.
Differential Revision: https://reviews.llvm.org/D63728
llvm-svn: 368529
As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
Summary:
This is a fix for PR23997.
The loop vectorizer is not preserving the inbounds property of GEPs that it creates.
This is inhibiting some optimizations. This patch preserves the inbounds property in
the case where a load/store is being fed by an inbounds GEP.
Reviewers: mkuper, javed.absar, hsaito
Reviewed By: hsaito
Subscribers: dcaballe, hsaito, llvm-commits
Differential Revision: https://reviews.llvm.org/D46191
llvm-svn: 331269
This is a slight reduction of one of the benchmarks
that suffered with D43079. Cost model changes should
not cause this test to remain scalarized.
llvm-svn: 326221
Summary:
vectorizer-maximize-bandwidth is generally useful in terms of performance. I've tested the impact of changing this to default on speccpu benchmarks on sandybridge machines. The result shows non-negative impact:
spec/2006/fp/C++/444.namd 26.84 -0.31%
spec/2006/fp/C++/447.dealII 46.19 +0.89%
spec/2006/fp/C++/450.soplex 42.92 -0.44%
spec/2006/fp/C++/453.povray 38.57 -2.25%
spec/2006/fp/C/433.milc 24.54 -0.76%
spec/2006/fp/C/470.lbm 41.08 +0.26%
spec/2006/fp/C/482.sphinx3 47.58 -0.99%
spec/2006/int/C++/471.omnetpp 22.06 +1.87%
spec/2006/int/C++/473.astar 22.65 -0.12%
spec/2006/int/C++/483.xalancbmk 33.69 +4.97%
spec/2006/int/C/400.perlbench 33.43 +1.70%
spec/2006/int/C/401.bzip2 23.02 -0.19%
spec/2006/int/C/403.gcc 32.57 -0.43%
spec/2006/int/C/429.mcf 40.35 +0.27%
spec/2006/int/C/445.gobmk 26.96 +0.06%
spec/2006/int/C/456.hmmer 24.4 +0.19%
spec/2006/int/C/458.sjeng 27.91 -0.08%
spec/2006/int/C/462.libquantum 57.47 -0.20%
spec/2006/int/C/464.h264ref 46.52 +1.35%
geometric mean +0.29%
The regression on 453.povray seems real, but is due to secondary effects as all hot functions are bit-identical with and without the flag.
I started this patch to consult upstream opinions on this. It will be greatly appreciated if the community can help test the performance impact of this change on other architectures so that we can decided if this should be target-dependent.
Reviewers: hfinkel, mkuper, davidxl, chandlerc
Reviewed By: chandlerc
Subscribers: rengolin, sanjoy, javed.absar, bjope, dorit, magabari, RKSimon, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D33341
llvm-svn: 306933
Philip Reames