Need to check that the signed operand has an extra sign bit to be sure
that we do not skip signedness, when trying to minimize bitwidth for
smin/smax intrinsics.
The shuffleToIdentity fold needs to be a bit more careful about the difference
between call instructions and intrinsics. The second can be handled, but the
first should result in bailing out. This patch also adds some extra intrinsic
tests from #91000.
Fixes#91078
Code checking stores to invariant addresses and reductions made an
incorrect assumption that the case of both a load & store to the same
invariant address does not need to be handled.
In some cases when vectorizing with runtime checks, there may be
dependences with a load and store to the same address, storing a
reduction value.
Update LAA to separately track if there was a store-store and a
load-store dependence with an invariant addresses.
Bail out early if there as a load-store dependence with invariant
address. If there was a store-store one, still apply the logic checking
if they all store a reduction.
This patch adds a basic version of a combine that attempts to remove
shuffles that when combined simplify away to an identity shuffle. For
example:
%ab = shufflevector <8 x half> %a, <8 x half> poison, <4 x i32> <i32 3,
i32 2, i32 1, i32 0>
%at = shufflevector <8 x half> %a, <8 x half> poison, <4 x i32> <i32 7,
i32 6, i32 5, i32 4>
%abt = fneg <4 x half> %at
%abb = fneg <4 x half> %ab
%r = shufflevector <4 x half> %abt, <4 x half> %abb, <8 x i32> <i32 7,
i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
By looking through the shuffles and fneg, it can be simplified to:
%r = fneg <8 x half> %a
The code tracks each lane starting from the original shuffle, keeping a
track of a vector of {src, idx}. As we propagate up through the
instructions we will either look through intermediate instructions
(binops and unops) or see a collections of lanes that all have the same
src and incrementing idx (an identity). We can also see a single value
with identical lanes, which we can treat like a splat.
Only the basic version is added here, handling identities, splats,
binops and unops. In follow-up patches other instructions can be added
such as constants, intrinsics, cmp/sel and zext/sext/trunc.
This reverts the revert commit c6e01627acf859.
This patch includes a fix for any-of reductions and epilogue
vectorization. Extra test coverage for the issue that caused the revert
has been added in bce3bfced5fe0b019 and an assertion has been added in
c7209cbb8be7a3c65813.
--------------------------------
Original commit message:
Update AnyOf reduction code generation to only keep track of the AnyOf
property in a boolean vector in the loop, only selecting either the new
or start value in the middle block.
The patch incorporates feedback from https://reviews.llvm.org/D153697.
This fixes the #62565, as now there aren't multiple uses of the
start/new values.
Fixes https://github.com/llvm/llvm-project/issues/62565
PR: https://github.com/llvm/llvm-project/pull/78304
Support for predicated vector reverse intrinsic was added some time ago.
Adds support for predicated reversed loads/stores in the loop
vectorizer.
Reviewers: fhahn
Reviewed By: fhahn
Pull Request: https://github.com/llvm/llvm-project/pull/88025
This patch adds an assert to createAndCollectMergePhiForReduction to
make sure there is a resume phi when vectorizing the epilogue loop. This
is needed to set the resume value from the main vector loop.
This assertion guards against the issue caused the revert of
https://github.com/llvm/llvm-project/pull/78304.
In some cases masked gather is less profitable than insert-subvector of
consecutive/strided stores. SLP has this kind of analysis, but need to
improve it by adding the cost of the GEP analysis.
Also, the GEP cost estimation for masked gather is fixed.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/90737
After minbitwidth analysis, and <v>, (power_of_2 - 1 const) can be
transformed into just an <v>, (all_ones const), which can be ignored at
the cost estimation and at the codegen. x264 benchmark has this pattern.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/90739
Replace relying on the underling CallInst for looking up the called
function and its types by instead adding the called function as operand,
in line with how called functions are handled in CallInst.
Operand bundles, metadata and fast-math flags are optionally used if
there's an underlying CallInst.
This enables creating VPWidenCallRecipes without requiring an underlying
IR instruction.
Adds transformation of consecutive vector store + reverse to strided
stores with stride -1, if it is profitable
Reviewers: RKSimon, preames
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/90464
Added a limit of 128 incoming values at max for PHIs nodes to be
vectorized plus improved performance by using logarithmic search instead
of linear if the number of incoming values is > 4.
The conditional branch from the loop latch will be replaced by a
single branch controlling the loop, so there is no extra overhead from
scalarization. This improves the cost esimates in some cases.
Metric: size..text
Program size..text
exp ref diff
test-suite :: MultiSource/Benchmarks/mediabench/gsm/toast/toast.test 42906.00 42986.00 0.2%
test-suite :: MultiSource/Benchmarks/MiBench/telecomm-gsm/telecomm-gsm.test 42909.00 42989.00 0.2%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 664581.00 664661.00 0.0%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 664581.00 664661.00 0.0%
Less is better.
Replaces `buildvector <p x in> + trunc <p x in> to <p x im>` sequences to
`buildvector <p x im> of { trunc in to im }` scalars, which is free in
most cases, results in better code.
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/88504
If the gather node matches the vectorized node, it must also match with
the scalars completely. Otherwise, need to revectorize the gather node
to generate correct code.
This patch is moving out following intrinsics:
* vector.interleave2/deinterleave2
* vector.reverse
* vector.splice
from the experimental namespace.
All these intrinsics exist in LLVM for more than a year now, and are
widely used, so should not be considered as experimental.
If we vectorize a loop with multiple exits, all exiting branches should
be considered uniform, as the resulting loop will be controlled by the
canonical IV only. Previously we were overestimating the cost of values
contributing to the other exits.
Before deleting extractelement instruction for vectorized GEP with
external users, need to check that all users vectorized before deleting
this extractelement.
We can try to vectorize long store sequences, if short ones were
unsuccessful because of the non-profitable vectorization. It should not
increase compile time significantly (stores are sorted already,
complexity is n x log n), but vectorize extra code.
Metric: size..text
Program size..text
results results0 diff
test-suite :: External/SPEC/CINT2006/400.perlbench/400.perlbench.test 1088012.00 1088236.00 0.0%
test-suite :: SingleSource/UnitTests/matrix-types-spec.test 480396.00 480476.00 0.0%
test-suite :: External/SPEC/CINT2017rate/525.x264_r/525.x264_r.test 664613.00 664661.00 0.0%
test-suite :: External/SPEC/CINT2017speed/625.x264_s/625.x264_s.test 664613.00 664661.00 0.0%
test-suite :: External/SPEC/CFP2017rate/510.parest_r/510.parest_r.test 2041105.00 2040961.00 -0.0%
test-suite :: MultiSource/Applications/JM/lencod/lencod.test 836563.00 836387.00 -0.0%
test-suite :: MultiSource/Benchmarks/7zip/7zip-benchmark.test 1035100.00 1032140.00 -0.3%
In all benchmarks extra code gets vectorized
Reviewers: RKSimon
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/88563
Addresses issue #88716.
Some function parameter names in the affected header files did not match
the parameter names in the definitions, or were listed in a different
order.
---------
Signed-off-by: Troy-Butler <squintik@outlook.com>
No need to try to vectorize single gather/buildvector with alternate
opcode graph, it is not profitable. In other cases, need to use last
instruction for inserting the vectorized code.
Ignore incoming values with constant false masks when trying to simplify
VPBlendRecipes.
As a follow-on optimization, we should also be able to drop all incoming
values with false masks by creating a new VPBlendRecipe with those
operands dropped.
PR: https://github.com/llvm/llvm-project/pull/89384
Ensure the getShuffleCost arguments/instruction args are populated - minor extension to #88743 to help improve shuffle costs for certain corner cases (e.g. shuffles of loads)
Introduced transformNodes() function to perform transformation of the
nodes (cost-based, instruction count based, etc.).
Implemented transformation of consecutive loads + reverse order to
strided loads with stride -1, if profitable.
Reviewers: RKSimon, preames, topperc
Reviewed By: RKSimon
Pull Request: https://github.com/llvm/llvm-project/pull/88530
The compiler should not take into account the type of the cmp
instruction, otherwise it may treat the size incorrectly and it may lead
to incorrect codegen.
Another step towards cleaning up shuffles that have been split, often across bitcasts between SSE intrinsic.
Strip shuffles entirely if we fold to an identity shuffle.
After e2a72fa583d9, def-use chains of EVL are modeled explicitly.
So there's no need for a custom check of its placement, as regular
def-use verification will catch mis-placements.
Need to check that at least single bit is cleared for unsigned nodes
before reducing their size. Otherwise they might be treated as signed in
signed nodes.
This reverts commit 74e07ab523122d6a8347b25770062ab331b6bb84.
It might be that Mask.getBitWidth() == Mask.countl_zero() (32 in my
case) and zero bitwidth2 causes the crash.
Introduce new subclasses of VPWidenMemoryRecipe for VP
(vector-predicated) loads and stores to address multiple TODOs from
https://github.com/llvm/llvm-project/pull/76172
Note that the introduction of the new recipes also improves code-gen for
VP gather/scatters by removing the redundant header mask. With the new
approach, it is not sufficient to look at users of the widened canonical
IV to find all uses of the header mask.
In some cases, a widened IV is used instead of separately widening the
canonical IV. To handle that, first collect all VPValues representing header
masks (by looking at users of both the canonical IV and widened inductions
that are canonical) and then checking all users (recursively) of those header
masks.
Depends on https://github.com/llvm/llvm-project/pull/87411.
PR: https://github.com/llvm/llvm-project/pull/87816
Need to check that at least single bit is cleared for unsigned nodes
before reducing their size. Otherwise they might be treated as signed in
signed nodes.
