We use getExtractWithExtendCost to calculate the cost of extractelement and
s|zext together when computing the extract cost after vectorization, but we
calculate the cost of extractelement and s|zext separately when computing the
scalar cost which is larger than it should be.
Differential Revision: https://reviews.llvm.org/D45469
llvm-svn: 330143
We use two approaches for determining the minimum bitwidth.
* Demanded bits
* Value tracking
If demanded bits doesn't result in a narrower type, we then try value tracking.
We need this if we want to root SLP trees with the indices of getelementptr
instructions since all the bits of the indices are demanded.
But there is a missing piece though. We need to be able to distinguish "demanded
and shrinkable" from "demanded and not shrinkable". For example, the bits of %i
in
%i = sext i32 %e1 to i64
%gep = getelementptr inbounds i64, i64* %p, i64 %i
are demanded, but we can shrink %i's type to i32 because it won't change the
result of the getelementptr. On the other hand, in
%tmp15 = sext i32 %tmp14 to i64
%tmp16 = insertvalue { i64, i64 } undef, i64 %tmp15, 0
it doesn't make sense to shrink %tmp15 and we can skip the value tracking.
Ideas are from Matthew Simpson!
Differential Revision: https://reviews.llvm.org/D44868
llvm-svn: 329035
When building the SLP tree, we look for reuse among the vectorized tree
entries. However, each gather sequence is represented by a unique tree entry,
even though the sequence may be identical to another one. This means, for
example, that a gather sequence with two uses will be counted twice when
computing the cost of the tree. We should only count the cost of the definition
of a gather sequence rather than its uses. During code generation, the
redundant gather sequences are emitted, but we optimize them away with CSE. So
it looks like this problem just affects the cost model.
Differential Revision: https://reviews.llvm.org/D44742
llvm-svn: 328316
This patch provides an implementation of getArithmeticReductionCost for
AArch64. We can specialize the cost of add reductions since they are computed
using the 'addv' instruction.
Differential Revision: https://reviews.llvm.org/D44490
llvm-svn: 327702
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: 326217
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: 325717
There are too many perf regressions resulting from this, so we need to
investigate (and add tests for) targets like ARM and AArch64 before
trying to reinstate.
llvm-svn: 325658
This change was mentioned at least as far back as:
https://bugs.llvm.org/show_bug.cgi?id=26837#c26
...and I found a real program that is harmed by this:
Himeno running on AMD Jaguar gets 6% slower with SLP vectorization:
https://bugs.llvm.org/show_bug.cgi?id=36280
...but the change here appears to solve that bug only accidentally.
The div/rem costs for x86 look very wrong in some cases, but that's already true,
so we can fix those in follow-up patches. There's also evidence that more cost model
changes are needed to solve SLP problems as shown in D42981, but that's an independent
problem (though the solution may be adjusted after this change is made).
Differential Revision: https://reviews.llvm.org/D43079
llvm-svn: 325515
Summary:
For better vectorization result we should take into consideration the
cost of the user insertelement instructions when we try to
vectorize sequences that build the whole vector. I.e. if we have the
following scalar code:
```
<Scalar code>
insertelement <ScalarCode>, ...
```
we should consider the cost of the last `insertelement ` instructions as
the cost of the scalar code.
Reviewers: RKSimon, spatel, hfinkel, mkuper
Subscribers: javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D42657
llvm-svn: 324893
Summary:
Added more remarks to SLP pass, in particular "missed" optimization remarks.
Also proposed several tests for new functionality.
Patch by Vladimir Miloserdov!
For reference you may look at: https://reviews.llvm.org/rL302811
Reviewers: anemet, fhahn
Reviewed By: anemet
Subscribers: javed.absar, lattner, petecoup, yakush, llvm-commits
Differential Revision: https://reviews.llvm.org/D38367
llvm-svn: 318307
Summary:
The New Pass Manager infrastructure was forgetting to keep around the optimization remark yaml file that the compiler might have been producing. This meant setting the option to '-' for stdout worked, but setting it to a filename didn't give file output (presumably it was deleted because compilation didn't explicitly keep it). This change just ensures that the file is kept if compilation succeeds.
So far I have updated one of the optimization remark output tests to add a version with the new pass manager. It is my intention for this patch to also include changes to all tests that use `-opt-remark-output=` but I wanted to get the code patch ready for review while I was making all those changes.
Fixes https://bugs.llvm.org/show_bug.cgi?id=33951
Reviewers: anemet, chandlerc
Reviewed By: anemet, chandlerc
Subscribers: javed.absar, chandlerc, fhahn, llvm-commits
Differential Revision: https://reviews.llvm.org/D36906
llvm-svn: 311271
Patch tries to improve two-pass vectorization analysis, existing in SLP vectorizer. What it does:
1. Defines key nodes, that are the vectorization roots. Previously vectorization started if StoreInst or ReturnInst is found. For now, the vectorization started for all Instructions with no users and void types (Terminators, StoreInst) + CallInsts.
2. CmpInsts, InsertElementInsts and InsertValueInsts are stored in the
array. This array is processed only after the vectorization of the
first-after-these instructions key node is finished. Vectorization goes
in reverse order to try to vectorize as much code as possible.
Reviewers: mzolotukhin, Ayal, mkuper, gilr, hfinkel, RKSimon
Subscribers: ashahid, anemet, RKSimon, mssimpso, llvm-commits
Differential Revision: https://reviews.llvm.org/D29826
llvm-svn: 310260
Summary:
Patch tries to improve two-pass vectorization analysis, existing in SLP vectorizer. What it does:
1. Defines key nodes, that are the vectorization roots. Previously vectorization started if StoreInst or ReturnInst is found. For now, the vectorization started for all Instructions with no users and void types (Terminators, StoreInst) + CallInsts.
2. CmpInsts, InsertElementInsts and InsertValueInsts are stored in the array. This array is processed only after the vectorization of the first-after-these instructions key node is finished. Vectorization goes in reverse order to try to vectorize as much code as possible.
Reviewers: mzolotukhin, Ayal, mkuper, gilr, hfinkel, RKSimon
Subscribers: ashahid, anemet, RKSimon, mssimpso, llvm-commits
Differential Revision: https://reviews.llvm.org/D29826
llvm-svn: 310255
ARM Neon has native support for half-sized vector registers (64 bits). This
is beneficial for example for 2D and 3D graphics. This patch adds the option
to lower MinVecRegSize from 128 via a TTI in the SLP Vectorizer.
*** Performance Analysis
This change was motivated by some internal benchmarks but it is also
beneficial on SPEC and the LLVM testsuite.
The results are with -O3 and PGO. A negative percentage is an improvement.
The testsuite was run with a sample size of 4.
** SPEC
* CFP2006/482.sphinx3 -3.34%
A pretty hot loop is SLP vectorized resulting in nice instruction reduction.
This used to be a +22% regression before rL299482.
* CFP2000/177.mesa -3.34%
* CINT2000/256.bzip2 +6.97%
My current plan is to extend the fix in rL299482 to i16 which brings the
regression down to +2.5%. There are also other problems with the codegen in
this loop so there is further room for improvement.
** LLVM testsuite
* SingleSource/Benchmarks/Misc/ReedSolomon -10.75%
There are multiple small SLP vectorizations outside the hot code. It's a bit
surprising that it adds up to 10%. Some of this may be code-layout noise.
* MultiSource/Benchmarks/VersaBench/beamformer/beamformer -8.40%
The opt-viewer screenshot can be seen at F3218284. We start at a colder store
but the tree leads us into the hottest loop.
* MultiSource/Applications/lambda-0.1.3/lambda -2.68%
* MultiSource/Benchmarks/Bullet/bullet -2.18%
This is using 3D vectors.
* SingleSource/Benchmarks/Shootout-C++/Shootout-C++-lists +6.67%
Noise, binary is unchanged.
* MultiSource/Benchmarks/Ptrdist/anagram/anagram +4.90%
There is an additional SLP in the cold code. The test runs for ~1sec and
prints out over 2000 lines. This is most likely noise.
* MultiSource/Applications/aha/aha +1.63%
* MultiSource/Applications/JM/lencod/lencod +1.41%
* SingleSource/Benchmarks/Misc/richards_benchmark +1.15%
Differential Revision: https://reviews.llvm.org/D31965
llvm-svn: 303116
This caused PR33053.
Original commit message:
> The new experimental reduction intrinsics can now be used, so I'm enabling this
> for AArch64. We will need this for SVE anyway, so it makes sense to do this for
> NEON reductions as well.
>
> The existing code to match shufflevector patterns are replaced with a direct
> lowering of the reductions to AArch64-specific nodes. Tests updated with the
> new, simpler, representation.
>
> Differential Revision: https://reviews.llvm.org/D32247
llvm-svn: 303115
The approach I followed was to emit the remark after getTreeCost concludes
that SLP is profitable. I initially tried emitting them after the
vectorizeRootInstruction calls in vectorizeChainsInBlock but I vaguely
remember missing a few cases for example in HorizontalReduction::tryToReduce.
ORE is placed in BoUpSLP so that it's available from everywhere (notably
HorizontalReduction::tryToReduce).
We use the first instruction in the root bundle as the locator for the remark.
In order to get a sense how far the tree is spanning I've include the size of
the tree in the remark. This is not perfect of course but it gives you at
least a rough idea about the tree. Then you can follow up with -view-slp-tree
to really see the actual tree.
llvm-svn: 302811
The new experimental reduction intrinsics can now be used, so I'm enabling this
for AArch64. We will need this for SVE anyway, so it makes sense to do this for
NEON reductions as well.
The existing code to match shufflevector patterns are replaced with a direct
lowering of the reductions to AArch64-specific nodes. Tests updated with the
new, simpler, representation.
Differential Revision: https://reviews.llvm.org/D32247
llvm-svn: 302678
Summary:
If horizontal reduction tree starts from the binary operation that is
used in PHI node, but this PHI is not used in horizontal reduction, we
may end up with extra addition of this PHI node after vectorization.
Here is an example:
```
%phi = phi i32 [ %tmp, %end], ...
...
%tmp = add i32 %tmp1, %tmp2
end:
```
after vectorization we always have something like:
```
%phi = phi i32 [ %tmp, %end], ...
...
%red = extractelement <8 x 32> %vec.red, 0
%tmp = add i32 %red, %phi
end:
```
even if `%phi` is not used in reduction tree. Patch considers these PHI
nodes as extra arguments and considers them in the final result iff they
really used in reduction.
Reviewers: mkuper, hfinkel, mzolotukhin
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D30409
llvm-svn: 296606
The test case included with r279125 exposed an existing signed integer
overflow. Since getTreeCost can return INT_MAX, we can't sum this cost together
with other costs, such as getReductionCost.
This patch removes the possibility of assigning a cost of INT_MAX. Since we
were previously using INT_MAX as an indicator for "should not vectorize", we
now explicitly check this condition with "isTreeTinyAndNotFullyVectorizable"
before computing a cost.
This patch adds a run-line to the test case used for r279125 that ensures we
don't vectorize. Previously, this line would vectorize the test case by chance
due to undefined behavior in the cost calculation.
Differential Revision: https://reviews.llvm.org/D23723
llvm-svn: 279562
The test case included in r279125 exposed existing undefined behavior in the
SLP vectorizer that it did not introduce. This patch reapplies the original
patch, but modifies the test case to avoid hitting the undefined behavior. This
allows us to close PR28330 while keeping the UBSan bot happy. The undefined
behavior the original test uncovered will be addressed in a follow-on patch.
Reference: https://llvm.org/bugs/show_bug.cgi?id=28330
llvm-svn: 279370
We abort building vectorizable trees in some cases (e.g., if the maximum
recursion depth is reached, if the region size is too large, etc.). If this
happens for a reduction, we can be left with a root entry that needs to be
gathered. For these cases, we need make sure we actually set VectorizedValue to
the resulting vector.
This patch ensures we properly set VectorizedValue, and it also ensures the
insertelement sequence generated for the gathers is inserted at the correct
location.
Reference: https://llvm.org/bugs/show_bug.cgi?id=28330
Differential Revison: https://reviews.llvm.org/D23410
llvm-svn: 279125
This change adds a new hook for estimating the cost of vector extracts followed
by zero- and sign-extensions. The motivating example for this change is the
SMOV and UMOV instructions on AArch64. These instructions move data from vector
to general purpose registers while performing the corresponding extension
(sign-extend for SMOV and zero-extend for UMOV) at the same time. For these
operations, TargetTransformInfo can assume the extensions are free and only
report the cost of the vector extract. The SLP vectorizer has been updated to
make use of the new hook.
Differential Revision: http://reviews.llvm.org/D18523
llvm-svn: 267725
Commit r259357 was reverted because it caused PR26629. We were assuming all
roots of a vectorizable tree could be truncated to the same width, which is not
the case in general. This commit reapplies the patch along with a fix and a new
test case to ensure we don't regress because of this issue again. This should
fix PR26629.
llvm-svn: 261212
This patch is the second attempt to reapply commit r258404. There was bug in
the initial patch and subsequent fix (mentioned below).
The initial patch caused 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 and PR26307.
llvm-svn: 258929
This commit exposes a crash in computeKnownBits on the Chromium buildbots.
Reverting to investigate.
Reference: https://llvm.org/bugs/show_bug.cgi?id=26307
llvm-svn: 258812
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
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
Haicheng Wu