This allows relying on VPBasicBlock::insert to make sure insertion is
well formed, i.e. by updating the recipe's parent as well as other
potential invariants in the future.
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Original Pull Request: https://github.com/llvm/llvm-project/pull/84334
The patch has the same functionality (no test changes, no changes in
benchmarks) as the original patch, just has some compile time
improvements + fixes for xxhash unittest, discovered earlier in the
previous version of the patch.
Reviewers:
Pull Request: https://github.com/llvm/llvm-project/pull/84536
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Reviewers: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/84334
Instead of generating a <1 x i1> active lane mask intrinsic, generate
the equivalent scalar ICMP instead. This allows us to avoid
unnecessarily extracting the scalar part from the vector mask.
Fixes llvm#73894.
This improves overall analysis for minbitwidth in SLP. It allows to
analyze the trees with store/insertelement root nodes. Also, instead of
using single minbitwidth, detected from the very first analysis stage,
it tries to detect the best one for each trunc/ext subtree in the graph
and use it for the subtree.
Results in better code and less vector register pressure.
Metric: size..text
Program size..text
results results0 diff
test-suite :: SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant.test 92549.00 92609.00 0.1%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 663381.00 663493.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 663381.00 663493.00 0.0%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 307182.00 307214.00 0.0%
test-suite :: External/SPEC/CFP2017speed/638.imagick_s/638.imagick_s.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/538.imagick_r/538.imagick_r.test 1394420.00 1394484.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2040257.00 2040273.00 0.0%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 12396098.00 12395858.00 -0.0%
test-suite :: External/SPEC/CINT2006/445.gobmk/445.gobmk.test 909944.00 909768.00 -0.0%
SingleSource/Benchmarks/Adobe-C++/simple_types_loop_invariant - 4 scalar
instructions remain scalar (good).
Spec2017/x264 - the whole function idct4x4dc is vectorized using <16
x i16> instead of <16 x i32>, also zext/trunc are removed. In other
places last vector zext/sext removed and replaced by
extractelement + scalar zext/sext pair.
MultiSource/Benchmarks/Bullet/bullet - reduce or <4 x i32> replaced by
reduce or <4 x i8>
Spec2017/imagick - Removed extra zext from 2 packs of the operations.
Spec2017/parest - Removed extra zext, replaced by extractelement+scalar
zext
Spec2017/blender - the whole bunch of vector zext/sext replaced by
extractelement+scalar zext/sext, some extra code vectorized in smaller
types.
Spec2006/gobmk - fixed cost estimation, some small code remains scalar.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/78976
As part of the RemoveDIs project we need LLVM to insert instructions using
iterators wherever possible, so that the iterators can carry a bit of
debug-info. This commit implements some of that by updating the contents of
llvm/lib/Transforms/Utils to always use iterator-versions of instruction
constructors.
There are two general flavours of update:
* Almost all call-sites just call getIterator on an instruction
* Several make use of an existing iterator (scenarios where the code is
actually significant for debug-info)
The underlying logic is that any call to getFirstInsertionPt or similar
APIs that identify the start of a block need to have that iterator passed
directly to the insertion function, without being converted to a bare
Instruction pointer along the way.
Noteworthy changes:
* FindInsertedValue now takes an optional iterator rather than an
instruction pointer, as we need to always insert with iterators,
* I've added a few iterator-taking versions of some value-tracking and
DomTree methods -- they just unwrap the iterator. These are purely
convenience methods to avoid extra syntax in some passes.
* A few calls to getNextNode become std::next instead (to keep in the
theme of using iterators for positions),
* SeparateConstOffsetFromGEP has it's insertion-position field changed.
Noteworthy because it's not a purely localised spelling change.
All this should be NFC.
Given a scalable VF of the form <NumElts * VScale>, this patch adds the
ability to discharge a backedge test for a loop whose trip count is
between (NumElts, MinVScale*NumElts).
A couple of notes on this:
* Annoyingly, I could not figure out to write a test for this case. My
attempt is checked in as test32_i8 in f67ef1a, but LV uses a fixed
vector in that case, and ignored the force flags.
* This depends on 9eb5f94f to avoid appearing like a regression. Since
SCEV doesn't know any upper bound on vscale without the vscale_range
attribute (it doesn't query TTI), the ranges overflow on the multiply.
Arguably, this is fixing a bug in the current LV code since in theory
vscale can be large enough to overflow for real, but no actual target is
going to see that case.
Verifying CFG invariants of a block before verifying its contents allows
contents verification to rely on the CFG invariants (e.g. that there's a
vector loop region that can be retrieved).
This avoids extra checks in
https://github.com/llvm/llvm-project/pull/76172.
Exit early if MaxVF < MinVF. In that case, the loop body below will
never get entered. Note that this adjusts the condition from MaxVF <=
MinVF. If MaxVF == MinVF, vectorization may still be feasible (and the
loop below gets entered).
PR: https://github.com/llvm/llvm-project/pull/83283
Masked gather is very expensive operation and sometimes better to
represent it as a serie of consecutive/strided loads + insertsubvectors
sequences. Patch adds some basic estimation and if loads+insertsubvector
is cheaper, decides to represent it in this way rather than masked
gather.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/83481
If the graph incudes only single "gather" node with only
extractelements/undefs, which used only in insertelement-based
buildvector sequences, it still might be profitable to vectorize it.
Need to rely on the cost model, not throw this graph away immediately.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/83581
This was tripping off STL implementations that check for it (like libc++
with debug checking). The goal of this sort is to cluster operations on
the same values so preserve that property but sort everything else based
on the existing numbering.
When do the analysis for the (potential) masked gather node, we check
that not greater than half of the pointer operands are loop invariants
or potentially vectorizable.
Need to check actually, that we have a loop at first
and do better check for the potentially vectorizable
pointers.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/83472
If the buildvector node contains extract, which later should be combined
with some other nodes by shuffling, need to estimate the cost of this
shuffle before building the mask after shuffle.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/83442
Recent set of changes (PR #67725) in loop interleaving algorithm caused removal of the loop trip count threshold for allowing interleaving. Therefore configuration option interleave-small-loop-scalar-reduction is no longer needed.
Some optimizations are apply after UF and VF have been chosen. This
patch adds an extra print of the final VPlan just before
codegen/execution.
In the future, there will be additional transforms that are applied
later (interleaving for example).
PR: https://github.com/llvm/llvm-project/pull/82269
At the moment, some VPInstructions create only a single scalar value,
but use VPTransformatState's 'vector' storage for this value. Those
values are effectively uniform-per-VF (or in some cases
uniform-across-VF-and-UF). Using the vector/per-part storage doesn't
interact well with other recipes, that more accurately using (Part,
Lane) to look up scalar values and prevents VPInstructions creating
scalars from interacting with other recipes working with scalars.
This PR tries to unify handling of scalars by using (Part, 0) for scalar
values where only the first lane is demanded. This allows using
VPInstructions with other recipes like VPScalarCastRecipe and is also
needed when using VPInstructions in more cases otuside the vector loop
region to generate scalars.
Depends on https://github.com/llvm/llvm-project/pull/80269
Patch syncs the code in findReusedOrderedScalars with cost
estimation/codegen. It tries to use similar logic to better determine
best order.
Before, it just tried to find previously vectorized node without
checking if it is possible to use the vectorized value in the shuffle.
Now it relies on the more generalized version. If it determines, that
a single vector must be reordered (using same mechanism, as codegen and
cost estimation), it generates better order.
The comparison between new/ref ordering:
Metric: SLP.NumVectorInstructions
Program SLP.NumVectorInstructions
results results0 diff
test-suite :: MultiSource/Benchmarks/nbench/nbench.test 139.00 140.00 0.7%
test-suite :: MultiSource/Benchmarks/DOE-ProxyApps-C++/miniFE/miniFE.test 344.00 346.00 0.6%
test-suite :: MultiSource/Benchmarks/FreeBench/pifft/pifft.test 1293.00 1292.00 -0.1%
test-suite :: External/SPEC/CFP2017rate/511.povray_r/511.povray_r.test 5176.00 5170.00 -0.1%
test-suite :: External/SPEC/CFP2006/453.povray/453.povray.test 5173.00 5167.00 -0.1%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 11692.00 11660.00 -0.3%
test-suite :: External/SPEC/CINT2006/464.h264ref/464.h264ref.test 1621.00 1615.00 -0.4%
test-suite :: External/SPEC/CINT2006/403.gcc/403.gcc.test 795.00 792.00 -0.4%
test-suite :: External/SPEC/CFP2017rate/526.blender_r/526.blender_r.test 26499.00 26338.00 -0.6%
test-suite :: MultiSource/Benchmarks/Bullet/bullet.test 7343.00 7281.00 -0.8%
test-suite :: MultiSource/Applications/JM/ldecod/ldecod.test 1104.00 1094.00 -0.9%
test-suite :: MultiSource/Applications/JM/lencod/lencod.test 2216.00 2180.00 -1.6%
test-suite :: External/SPEC/CFP2006/433.milc/433.milc.test 787.00 637.00 -19.1%
Less 0% is better.
Most of the benchmarks see more vectorized code. The first ones just
have shuffles removed.
The ordering analysis still may require some improvements (e.g. for
alternate nodes), but this one should be produce better results.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/77529
A VPlan contains multiple live-ins without underlying IR, like VFxUF or
VectorTripCount. Trying to infer the scalar type of those causes a crash
at the moment.
Update VPTypeAnalysis to take a VPlan in its constructor and assign
types to those live-ins up front. All those live-ins share the type of
the canonical IV.
PR: https://github.com/llvm/llvm-project/pull/80723
Unify VPlan verifiers in verifyVPlanIsValid. This adds verification for
various properties on blocks to the verifier used for VPlans generated
by the inner loop vectorizer. It also adds def-use checks for the
verifier used in the VPlan native path.
This drops the separate flag to enable HCFG verification. Instead, all
VPlans are verified once they have been created, if assertions are
enabled.
This also removes VPWidenPHIRecipe from VPHeaderPHIRecipe; it is used to
model any phi node in the native path.
All values accessed via get are now part of VPTransformState, the ILV
reference in the comment has been removed a long time ago. Remove the
stale comment.
Vector truncations can be pretty expensive, especially on X86, whilst scalar truncations are often free.
If the cost of performing the add/mul/and/or/xor reduction is cheap enough on the pre-truncated type, then avoid the vector truncation entirely.
Fixes https://github.com/llvm/llvm-project/issues/81469
