We were hitting an assertion because we were computing smaller type sizes for
instructions that cannot be demoted. The fix first determines the instructions
that will be demoted, and then applies the smaller type size to only those
instructions.
This should fix PR26239.
llvm-svn: 258705
This change attempts to produce vectorized integer expressions in bit widths
that are narrower than their scalar counterparts. The need for demotion arises
especially on architectures in which the small integer types (e.g., i8 and i16)
are not legal for scalar operations but can still be used in vectors. Like
similar work done within the loop vectorizer, we rely on InstCombine to perform
the actual type-shrinking. We use the DemandedBits analysis and
ComputeNumSignBits from ValueTracking to determine the minimum required bit
width of an expression.
Differential revision: http://reviews.llvm.org/D15815
llvm-svn: 258404
The fix uniques the bundle of getelementptr indices we are about to vectorize
since it's possible for the same index to be used by multiple instructions.
The original commit message is below.
[SLP] Vectorize the index computations of getelementptr instructions.
This patch seeds the SLP vectorizer with getelementptr indices. The primary
motivation in doing so is to vectorize gather-like idioms beginning with
consecutive loads (e.g., g[a[0] - b[0]] + g[a[1] - b[1]] + ...). While these
cases could be vectorized with a top-down phase, seeding the existing bottom-up
phase with the index computations avoids the complexity, compile-time, and
phase ordering issues associated with a full top-down pass. Only bundles of
single-index getelementptrs with non-constant differences are considered for
vectorization.
llvm-svn: 257918
This patch seeds the SLP vectorizer with getelementptr indices. The primary
motivation in doing so is to vectorize gather-like idioms beginning with
consecutive loads (e.g., g[a[0] - b[0]] + g[a[1] - b[1]] + ...). While these
cases could be vectorized with a top-down phase, seeding the existing bottom-up
phase with the index computations avoids the complexity, compile-time, and
phase ordering issues associated with a full top-down pass. Only bundles of
single-index getelementptrs with non-constant differences are considered for
vectorization.
Differential Revision: http://reviews.llvm.org/D14829
llvm-svn: 257800
This patch prevents us from unintentionally creating entries in the reductions
map for PHIs that are not actually reductions. This is currently not an issue
since we bail out if we encounter PHIs other than inductions or reductions.
However the behavior could become problematic as we add support for additional
recurrence types.
llvm-svn: 256930
When considering incoming values as part of a reduction phi, ensure the
incoming value is dominated by said phi.
Failing to ensure this property causes miscompiles.
Fixes PR25787.
Many thanks to Mattias Eriksson for reporting, reducing and analyzing the
problem for me.
Differential Revision: http://reviews.llvm.org/D15580
llvm-svn: 255792
(This is the third attempt to check in this patch, and the first two are r255454
and r255460. The once failed test file reg-usage.ll is now moved to
test/Transform/LoopVectorize/X86 directory with target datalayout and target
triple indicated.)
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255691
(This is the second attempt to check in this patch: REQUIRES: asserts is added
to reg-usage.ll now.)
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255460
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255454
GlobalsAA's assumptions that passes do not escape globals not previously
escaped is not violated by AlignmentFromAssumptions and SLPVectorizer. Marking
them as such allows GlobalsAA to be preserved until GVN in the LTO pipeline.
http://lists.llvm.org/pipermail/llvm-dev/2015-December/092972.html
Patch by Vaivaswatha Nagaraj!
llvm-svn: 255348
ScalarEvolution.h, in order to avoid cyclic dependencies between the Transform
and Analysis modules:
[LV][LAA] Add a layer over SCEV to apply run-time checked knowledge on SCEV expressions
Summary:
This change creates a layer over ScalarEvolution for LAA and LV, and centralizes the
usage of SCEV predicates. The SCEVPredicatedLayer takes the statically deduced knowledge
by ScalarEvolution and applies the knowledge from the SCEV predicates. The end goal is
that both LAA and LV should use this interface everywhere.
This also solves a problem involving the result of SCEV expression rewritting when
the predicate changes. Suppose we have the expression (sext {a,+,b}) and two predicates
P1: {a,+,b} has nsw
P2: b = 1.
Applying P1 and then P2 gives us {a,+,1}, while applying P2 and the P1 gives us
sext({a,+,1}) (the AddRec expression was changed by P2 so P1 no longer applies).
The SCEVPredicatedLayer maintains the order of transformations by feeding back
the results of previous transformations into new transformations, and therefore
avoiding this issue.
The SCEVPredicatedLayer maintains a cache to remember the results of previous
SCEV rewritting results. This also has the benefit of reducing the overall number
of expression rewrites.
Reviewers: mzolotukhin, anemet
Subscribers: jmolloy, sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D14296
llvm-svn: 255122
Summary:
This change creates a layer over ScalarEvolution for LAA and LV, and centralizes the
usage of SCEV predicates. The SCEVPredicatedLayer takes the statically deduced knowledge
by ScalarEvolution and applies the knowledge from the SCEV predicates. The end goal is
that both LAA and LV should use this interface everywhere.
This also solves a problem involving the result of SCEV expression rewritting when
the predicate changes. Suppose we have the expression (sext {a,+,b}) and two predicates
P1: {a,+,b} has nsw
P2: b = 1.
Applying P1 and then P2 gives us {a,+,1}, while applying P2 and the P1 gives us
sext({a,+,1}) (the AddRec expression was changed by P2 so P1 no longer applies).
The SCEVPredicatedLayer maintains the order of transformations by feeding back
the results of previous transformations into new transformations, and therefore
avoiding this issue.
The SCEVPredicatedLayer maintains a cache to remember the results of previous
SCEV rewritting results. This also has the benefit of reducing the overall number
of expression rewrites.
Reviewers: mzolotukhin, anemet
Subscribers: jmolloy, sanjoy, llvm-commits
Differential Revision: http://reviews.llvm.org/D14296
llvm-svn: 255115
The order in which instructions are truncated in truncateToMinimalBitwidths
effects code generation. Switch to a map with a determinisic order, since the
iteration order over a DenseMap is not defined.
This code is not hot, so the difference in container performance isn't
interesting.
Many thanks to David Blaikie for making me aware of MapVector!
Fixes PR25490.
Differential Revision: http://reviews.llvm.org/D14981
llvm-svn: 254179
Measurements primarily on AArch64 have shown this feature does not
significantly effect compile-time. The are no significant perf changes in LNT,
but for AArch64 at least, there are wins in third party benchmarks.
As discussed on llvm-dev, we're going to try turning this on by default and see
how other targets react to the change.
llvm-svn: 252733
Implemented as many of Michael's suggestions as were possible:
* clang-format the added code while it is still fresh.
* tried to change Value* to Instruction* in many places in computeMinimumValueSizes - unfortunately there are several places where Constants need to be handled so this wasn't possible.
* Reduce the pass list on loop-vectorization-factors.ll.
* Fix a bug where we were querying MinBWs for I->getOperand(0) but using MinBWs[I].
llvm-svn: 252469
The SLPVectorizer had a very crude way of trying to benefit
from associativity: it tried to optimize for splat/broadcast
or in order to have the same operator on the same side.
This is benefitial to the cost model and allows more vectorization
to occur.
This patch improve the logic and make the detection optimal (locally,
we don't look at the full tree but only at the immediate children).
Should fix https://llvm.org/bugs/show_bug.cgi?id=25247
Reviewers: mzolotukhin
Differential Revision: http://reviews.llvm.org/D13996
From: Mehdi Amini <mehdi.amini@apple.com>
llvm-svn: 252337
To be able to maximize the bandwidth during vectorization, this patch provides a new flag vectorizer-maximize-bandwidth. When it is turned on, the vectorizer will determine the vectorization factor (VF) using the smallest instead of widest type in the loop. To avoid increasing register pressure too much, estimates of the register usage for different VFs are calculated so that we only choose a VF when its register usage doesn't exceed the number of available registers.
This is the second attempt to submit this patch. The first attempt got a test failure on ARM. This patch is updated to try to fix the failure (more specifically, by handling the case when VF=1).
Differential revision: http://reviews.llvm.org/D8943
llvm-svn: 251850
Summary:
SCEV Predicates represent conditions that typically cannot be derived from
static analysis, but can be used to reduce SCEV expressions to forms which are
usable for different optimizers.
ScalarEvolution now has the rewriteUsingPredicate method which can simplify a
SCEV expression using a SCEVPredicateSet. The normal workflow of a pass using
SCEVPredicates would be to hold a SCEVPredicateSet and every time assumptions
need to be made a new SCEV Predicate would be created and added to the set.
Each time after calling getSCEV, the user will call the rewriteUsingPredicate
method.
We add two types of predicates
SCEVPredicateSet - implements a set of predicates
SCEVEqualPredicate - tests for equality between two SCEV expressions
We use the SCEVEqualPredicate to re-implement stride versioning. Every time we
version a stride, we will add a SCEVEqualPredicate to the context.
Instead of adding specific stride checks, LoopVectorize now adds a more
generic SCEV check.
We only need to add support for this in the LoopVectorizer since this is the
only pass that will do stride versioning.
Reviewers: mzolotukhin, anemet, hfinkel, sanjoy
Subscribers: sanjoy, hfinkel, rengolin, jmolloy, llvm-commits
Differential Revision: http://reviews.llvm.org/D13595
llvm-svn: 251800
To be able to maximize the bandwidth during vectorization, this patch provides a new flag vectorizer-maximize-bandwidth. When it is turned on, the vectorizer will determine the vectorization factor (VF) using the smallest instead of widest type in the loop. To avoid increasing register pressure too much, estimates of the register usage for different VFs are calculated so that we only choose a VF when its register usage doesn't exceed the number of available registers.
llvm-svn: 251592
Summary:
This change could be way off-piste, I'm looking for any feedback on whether it's an acceptable approach.
It never seems to be a problem to gobble up as many reduction values as can be found, and then to attempt to reduce the resulting tree. Some of the workloads I'm looking at have been aggressively unrolled by hand, and by selecting reduction widths that are not constrained by a vector register size, it becomes possible to profitably vectorize. My test case shows such an unrolling which SLP was not vectorizing (on neither ARM nor X86) before this patch, but with it does vectorize.
I measure no significant compile time impact of this change when combined with D13949 and D14063. There are also no significant performance regressions on ARM/AArch64 in SPEC or LNT.
The more principled approach I thought of was to generate several candidate tree's and use the cost model to pick the cheapest one. That seemed like quite a big design change (the algorithms seem very much one-shot), and would likely be a costly thing for compile time. This seemed to do the job at very little cost, but I'm worried I've misunderstood something!
Reviewers: nadav, jmolloy
Subscribers: mssimpso, llvm-commits, aemerson
Differential Revision: http://reviews.llvm.org/D14116
llvm-svn: 251428
Summary:
Currently, when the SLP vectorizer considers whether a phi is part of a reduction, it dismisses phi's whose incoming blocks are not the same as the block containing the phi. For the patterns I'm looking at, extending this rule to allow phis whose incoming block is a containing loop latch allows me to vectorize certain workloads.
There is no significant compile-time impact, and combined with D13949, no performance improvement measured in ARM/AArch64 in any of SPEC2000, SPEC2006 or LNT.
Reviewers: jmolloy, mcrosier, nadav
Subscribers: mssimpso, nadav, aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D14063
llvm-svn: 251425
Summary:
Certain workloads, in particular sum-of-absdiff loops, can be vectorized using SLP if it can treat select instructions as reduction values.
The test case is a bit awkward. The AArch64 cost model needs some tuning to not be so pessimistic about selects. I've had to tweak the SLP threshold here.
Reviewers: jmolloy, mzolotukhin, spatel, nadav
Subscribers: nadav, mssimpso, aemerson, llvm-commits
Differential Revision: http://reviews.llvm.org/D13949
llvm-svn: 251424
Vectorization of memory instruction (Load/Store) is possible when the pointer is coming from GEP. The GEP analysis allows to estimate the profit.
In some cases we have a "bitcast" between GEP and memory instruction.
I added code that skips the "bitcast".
http://reviews.llvm.org/D13886
llvm-svn: 251291
Besides the usual, I finally added an overload to
`BasicBlock::splitBasicBlock()` that accepts an `Instruction*` instead
of `BasicBlock::iterator`. Someone can go back and remove this overload
later (after updating the callers I'm going to skip going forward), but
the most common call seems to be
`BB->splitBasicBlock(BB->getTerminator(), ...)` and I'm not sure it's
better to add `->getIterator()` to every one than have the overload.
It's pretty hard to get the usage wrong.
llvm-svn: 250745
Originally I planned to use the same interface for masked gather/scatter and set isConsecutive to "false" in this case.
Now I'm implementing masked gather/scatter and see that the interface is inconvenient. I want to add interfaces isLegalMaskedGather() / isLegalMaskedScatter() instead of using the "Consecutive" parameter in the existing interfaces.
Differential Revision: http://reviews.llvm.org/D13850
llvm-svn: 250686
C semantics force sub-int-sized values (e.g. i8, i16) to be promoted to int
type (e.g. i32) whenever arithmetic is performed on them.
For targets with native i8 or i16 operations, usually InstCombine can shrink
the arithmetic type down again. However InstCombine refuses to create illegal
types, so for targets without i8 or i16 registers, the lengthening and
shrinking remains.
Most SIMD ISAs (e.g. NEON) however support vectors of i8 or i16 even when
their scalar equivalents do not, so during vectorization it is important to
remove these lengthens and truncates when deciding the profitability of
vectorization.
The algorithm this uses starts at truncs and icmps, trawling their use-def
chains until they terminate or instructions outside the loop are found (or
unsafe instructions like inttoptr casts are found). If the use-def chains
starting from different root instructions (truncs/icmps) meet, they are
unioned. The demanded bits of each node in the graph are ORed together to form
an overall mask of the demanded bits in the entire graph. The minimum bitwidth
that graph can be truncated to is the bitwidth minus the number of leading
zeroes in the overall mask.
The intention is that this algorithm should "first do no harm", so it will
never insert extra cast instructions. This is why the use-def graphs are
unioned, so that subgraphs with different minimum bitwidths do not need casts
inserted between them.
This algorithm works hard to reduce compile time impact. DemandedBits are only
queried if there are extends of illegal types and if a truncate to an illegal
type is seen. In the general case, this results in a simple linear scan of the
instructions in the loop.
No non-noise compile time impact was seen on a clang bootstrap build.
llvm-svn: 250032
The most important part required to make clang
devirtualization works ( ͡°͜ʖ ͡°).
The code is able to find non local dependencies, but unfortunatelly
because the caller can only handle local dependencies, I had to add
some restrictions to look for dependencies only in the same BB.
http://reviews.llvm.org/D12992
llvm-svn: 249196
