In-loop vector reductions which use the llvm.fmuladd intrinsic involve
the creation of two recipes; a VPReductionRecipe for the fadd and a
VPInstruction for the fmul. If the call to llvm.fmuladd has fast-math flags
these should be propagated through to the fmul instruction, so an
interface setFastMathFlags has been added to the VPInstruction class to
enable this.
Differential Revision: https://reviews.llvm.org/D113125
This patch fixes PR52111. The problem is that LV propagates poison-generating flags (`nuw`/`nsw`, `exact`
and `inbounds`) in instructions that contribute to the address computation of widen loads/stores that are
guarded by a condition. It may happen that when the code is vectorized and the control flow within the loop
is linearized, these flags may lead to generating a poison value that is effectively used as the base address
of the widen load/store. The fix drops all the integer poison-generating flags from instructions that
contribute to the address computation of a widen load/store whose original instruction was in a basic block
that needed predication and is not predicated after vectorization.
Reviewed By: fhahn, spatel, nlopes
Differential Revision: https://reviews.llvm.org/D111846
A first step towards modeling preheader and exit blocks in VPlan as well.
Keeping the vector loop in a region allows for changing the VF as we
traverse region boundaries.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D113182
When asking how many parts are required for a scalable vector type
there are occasions when it cannot be computed. For example, <vscale x 1 x i3>
is one such vector for AArch64+SVE because at the moment no matter how we
promote the i3 type we never end up with a legal vector. This means
that getTypeConversion returns TypeScalarizeScalableVector as the
LegalizeKind, and then getTypeLegalizationCost returns an invalid cost.
This then causes BasicTTImpl::getNumberOfParts to dereference an invalid
cost, which triggers an assert. This patch changes getNumberOfParts to
return 0 for such cases, since the definition of getNumberOfParts in
TargetTransformInfo.h states that we can use a return value of 0 to represent
an unknown answer.
Currently, LoopVectorize.cpp is the only place where we need to check for
0 as a return value, because all other instances will not currently
ask for the number of parts for <vscale x 1 x iX> types.
In addition, I have changed the target-independent interface for
getNumberOfParts to return 1 and assume there is a single register
that can fit the type. The loop vectoriser has lots of tests that are
target-independent and they relied upon the 0 value to mean the
answer is known and that we are not scalarising the vector.
I have added tests here that show we correctly return an invalid cost
for VF=vscale x 1 when the loop contains unusual types such as i7:
Transforms/LoopVectorize/AArch64/sve-inductions-unusual-types.ll
Differential Revision: https://reviews.llvm.org/D113772
No need to count the final shuffle cost for the constants, gathering of
the constants is just a constant vector + extra inserts, if required.
Differential Revision: https://reviews.llvm.org/D113770
Need to adjust the types of GEPs indices when building the tree
entries/operands. Otherwise some of the nodes might differ and
vectorizer is unable to correctly find them and count their cost.
Differential Revision: https://reviews.llvm.org/D113792
A bunch of scalars can be treated as a splat not only if all elements
are the same but also if some of them are undefvalues.
Differential Revision: https://reviews.llvm.org/D113774
If the vector intrinsic has scalar argument, we currently still create
a tree entry for this argument. This entry is not used, just consumes
resources and increases the cost of the tree.
Differential Revision: https://reviews.llvm.org/D113806
The interface is a convenience function to ask if a block requires
predication when widening, but it's important that there are two
separate concepts to consider:
(A) The block was predicated in the original loop.
(B) The block was unpredicated in the original loop, but requires
predication because of tail folding.
In the case of (B) we know that at least one lane of the vector will
be executed, which means we can implementing a load from a uniform address
with a scalar load + splat (D112552). In the case of predication because
of (A), we cannot do this, because the scalar load itself requires
predication.
The name 'blockNeedsPredication' does not make the distinction between
(A) and (B), hence the reason to rename it.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D113392
Need to fix ther cost estimation for split loads, since we look at the
subregs already, no need to permute them, need just to estimate
subregister insert, if it is smaller than the real register. Also, using
split loads, it might be profitable already to vectorize smaller trees
with gathering of the loads.
Differential Revision: https://reviews.llvm.org/D107188
Unfortunately sinking recipes for first-order recurrences relies on
the original position of recipes. So if a recipes needs to be sunk after
an optimized induction, it needs to stay in the original position, until
sinking is done. This is causing PR52460.
To fix the crash, keep the recipes in the original position until
sink-after is done.
Post-commit follow-up to c45045bfd0 to address PR52460.
Changes VPReplicateRecipe to extract the last lane from an unconditional,
uniform store instruction. collectLoopUniforms will also add stores to
the list of uniform instructions where Legal->isUniformMemOp is true.
setCostBasedWideningDecision now sets the widening decision for
all uniform memory ops to Scalarize, where previously GatherScatter
may have been chosen for scalable stores.
This fixes an assert ("Cannot yet scalarize uniform stores") in
setCostBasedWideningDecision when we have a loop containing a
uniform i1 store and a scalable VF, which we cannot create a scatter for.
Reviewed By: sdesmalen, david-arm, fhahn
Differential Revision: https://reviews.llvm.org/D112725
This patch adds a function to verify general properties of VPlans. The
first check makes sure that all phi-like recipes are at the beginning of
a block, with no other recipes in between.
Note that this currently may not hold for VPBlendRecipes at the moment,
as other recipes may be inserted before the VPBlendRecipe during mask
creation.
Note that this patch depends on D111300 and D111301, which fix code that
breaks the checked invariant.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D111302
All phi-like recipes should be at the beginning of a VPBasicBlock with
no other recipes in between. Ensure that the recurrence-splicing recipe
is not added between phi-like recipes, but after them.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D111301
When targeting a specific CPU with scalable vectorization, the knowledge
of that particular CPU's vscale value can be used to tune the cost-model
and make the cost per lane less pessimistic.
If the target implements 'TTI.getVScaleForTuning()', the cost-per-lane
is calculated as:
Cost / (VScaleForTuning * VF.KnownMinLanes)
Otherwise, it assumes a value of 1 meaning that the behavior
is unchanged and calculated as:
Cost / VF.KnownMinLanes
Reviewed By: kmclaughlin, david-arm
Differential Revision: https://reviews.llvm.org/D113209
The common use case for calling createStepForVF is currently something
like:
Value *Step = createStepForVF(Builder, ConstantInt::get(Ty, UF), VF);
and it makes more sense to reduce overall lines of code and change the
function to let it create the constant instead. With my patch this
becomes:
Value *Step = createStepForVF(Builder, Ty, VF, UF);
and the ConstantInt is created instead createStepForVF. A side-effect of
this is that the code in createStepForVF is also becomes simpler.
As part of this patch I've also replaced some calls to getRuntimeVF
with calls to createStepForVF, i.e.
getRuntimeVF(Builder, Count->getType(), VFactor * UFactor) ->
createStepForVF(Builder, Count->getType(), VFactor, UFactor)
because this feels semantically better.
Differential Revision: https://reviews.llvm.org/D113122
At the moment in LoopVectorizationCostModel::selectEpilogueVectorizationFactor
we bail out if the main vector loop uses a scalable VF. This patch adds
support for generating epilogue vector loops using a fixed-width VF when the
main vector loop uses a scalable VF.
I've changed LoopVectorizationCostModel::selectEpilogueVectorizationFactor
so that we convert the scalable VF into a fixed-width VF and do profitability
checks on that instead. In addition, since the scalable and fixed-width VFs
live in different VPlans that means I had to change the calls to
LVP.hasPlanWithVFs so that we only pass in the fixed-width VF.
New tests added here:
Transforms/LoopVectorize/AArch64/sve-epilog-vect.ll
Differential Revision: https://reviews.llvm.org/D109432
The public API for this functionality is forgetValue(). There was
only one call from LoopVectorize, which was directly next to a
forgetValue() call and as such redundant.
The recipe produces exactly one VPValue and can inherit directly from
it. This is in line with other recipes and avoids having to use
getVPSingleValue.
This patch updates VPReductionRecipe::execute so that the fast-math
flags associated with the underlying instruction of the VPRecipe are
propagated through to the reductions which are created.
Differential Revision: https://reviews.llvm.org/D112548
We never expect the runtime VF to be negative so we should use
the uitofp instruction instead of sitofp.
Differential revision: https://reviews.llvm.org/D112610
This patch updates recipe creation to ensure all
VPWidenIntOrFpInductionRecipes are in the header block. At the moment,
new induction recipes can be created in different blocks when trying to
optimize casts and induction variables.
Having all induction recipes in the header makes it easier to
analyze/transform them in VPlan.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D111300
Gathered loads/extractelements/extractvalue instructions should be
checked if they can represent a vector reordering node too and their
order should ve taken into account for better graph reordering analysis/
Also, if the gather node has reused scalars, they must be reordered
instead of the scalars themselves.
Differential Revision: https://reviews.llvm.org/D112454
Gathered loads/extractelements/extractvalue instructions should be
checked if they can represent a vector reordering node too and their
order should ve taken into account for better graph reordering analysis/
Also, if the gather node has reused scalars, they must be reordered
instead of the scalars themselves.
Differential Revision: https://reviews.llvm.org/D112454
This patch changes the definition of getStepVector from:
Value *getStepVector(Value *Val, int StartIdx, Value *Step, ...
to
Value *getStepVector(Value *Val, Value *StartIdx, Value *Step, ...
because:
1. it seems inconsistent to pass some values as Value* and some as
integer, and
2. future work will require the StartIdx to be an expression made up
of runtime calculations of the VF.
In widenIntOrFpInduction I've changed the code to pass in the
value returned from getRuntimeVF, but the presence of the assert:
assert(!VF.isScalable() && "scalable vectors not yet supported.");
means that currently this code path is only exercised for fixed-width
VFs and so the patch is still NFC.
Differential revision: https://reviews.llvm.org/D111882
Gathered loads/extractelements/extractvalue instructions should be
checked if they can represent a vector reordering node too and their
order should ve taken into account for better graph reordering analysis/
Also, if the gather node has reused scalars, they must be reordered
instead of the scalars themselves.
Differential Revision: https://reviews.llvm.org/D112454
Need to emit select(cmp) instructions for poison-safe forms of select
ops. Currently alive reports that `Target is more poisonous than source`
for operations we generating for such instructions.
https://alive2.llvm.org/ce/z/FiNiAA
Differential Revision: https://reviews.llvm.org/D112562
I have removed LoopVectorizationPlanner::setBestPlan, since this
function is quite aggressive because it deletes all other plans
except the one containing the <VF,UF> pair required. The code is
currently written to assume that all <VF,UF> pairs will live in the
same vplan. This is overly restrictive, since scalable VFs live in
different plans to fixed-width VFS. When we add support for
vectorising epilogue loops when the main loop uses scalable vectors
then we will the vplan for the main loop will be different to the
epilogue.
Instead I have added a new function called
LoopVectorizationPlanner::getBestPlanFor
that returns the best vplan for the <VF,UF> pair requested and leaves
all the vplans untouched. We then pass this best vplan to
LoopVectorizationPlanner::executePlan
which now takes an additional VPlanPtr argument.
Differential revision: https://reviews.llvm.org/D111125
Use RdxDesc->getOpcode instead of getUnderlingInstr()->getOpcode.
Move the code which finds Kind and IsOrdered to be outside the for loop
since neither of these change with the vector part.
Differential Revision: https://reviews.llvm.org/D112547
The final reduction nodes should not be reordered, the order does not
matter for reductions. Also, it might be profitable to vectorize smaller
reduction trees, reduction cost may compensate small tree cost.
Part of D111574
Differential Revision: https://reviews.llvm.org/D112467
Need to change the order of the reduction/binops args pair vectorization
attempts. Need to try to find the reduction at first and postpone
vectorization of binops args. This may help to find more reduction
patterns and vectorize them.
Part of D111574.
Differential Revision: https://reviews.llvm.org/D112224
At the moment a dummy entry block is created at the beginning of VPlan
construction. This dummy block is later removed again.
This means it is not easy to identify the VPlan header block in a
general fashion, because during recipe creation it is the single
successor of the entry block, while later it is the entry block.
To make getting the header easier, just skip creating the dummy block.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D111299
shuf (bo X, Y), (bo X, W) --> bo (shuf X), (shuf Y, W)
This is motivated by an example in D111800
(although that patch avoids the problem for that particular example).
The pattern is shown in reduced form with:
https://llvm.org/PR52178https://alive2.llvm.org/ce/z/d8zB4D
There is no difference on the PhaseOrdering test from D111800
because the aarch64 cost model says that the shuffle cost is 3 while
the fadd cost is 2.
Differential Revision: https://reviews.llvm.org/D111901
Vectorization of PHIs and stores very similar, it might be beneficial to
try to revectorize stores (like PHIs) if the total number of stores with
the same/alternate opcode is less than the vector size but number of
stores with the same type is larger than the vector size.
Differential Revision: https://reviews.llvm.org/D109831
Need to follow the order of the reused scalars from the
ReuseShuffleIndices mask rather than rely on the natural order.
Differential Revision: https://reviews.llvm.org/D111898
This simplifies the return value of addRuntimeCheck from a pair of
instructions to a single `Value *`.
The existing users of addRuntimeChecks were ignoring the first element
of the pair, hence there is not reason to track FirstInst and return
it.
Additionally all users of addRuntimeChecks use the second returned
`Instruction *` just as `Value *`, so there is no need to return an
`Instruction *`. Therefore there is no need to create a redundant
dummy `and X, true` instruction any longer.
Effectively this change should not impact the generated code because the
redundant AND will be folded by later optimizations. But it is easy to
avoid creating it in the first place and it allows more accurately
estimating the cost of the runtime checks.
Record widening decisions for memory operations within the planned recipes and
use the recorded decisions in code-gen rather than querying the cost model.
Differential Revision: https://reviews.llvm.org/D110479
This patch adds a pass option to only run transforms that scalarize
vector operations and do not create new vector instructions.
When running VectorCombine early in the pipeline introducing new vector
operations can have negative effects, like blocking loop or SLP
vectorization. To avoid regressions, restrict the early VectorCombine
run (when using -enable-matrix) to only perform scalarization and not
introduce new vector operations.
This is done as option to the pass directly, which is then set when
adding the pass to the pipeline. This is done for the new pass manager
only.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D111800
Need to check that either Idx is UndefMaskElem and value is UndefValue
or Idx is valid and value is the same as the scalar value in the node.
Differential Revision: https://reviews.llvm.org/D111802
This patch fixes another crash revealed by PR51614:
when *deciding* to vectorize with masked interleave groups, check if the access
is reverse (which is currently not supported).
Differential Revision: https://reviews.llvm.org/D108900
This patch continues unblocking optimizations that are blocked by pseudo probe instrumentation.
Not exactly like DbgIntrinsics, PseudoProbe intrinsic has other attributes (such as mayread, maywrite, mayhaveSideEffect) that can block optimizations. The issues fixed are:
- Flipped default param of getFirstNonPHIOrDbg API to skip pseudo probes
- Unblocked CSE by avoiding pseudo probe from clobbering memory SSA
- Unblocked induction variable simpliciation
- Allow empty loop deletion by treating probe intrinsic isDroppable
- Some refactoring.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D110847
collectLoopScalars collects pointer induction updates in ScalarPtrs, assuming
that the instruction will be scalar after vectorization. This may crash later
in VPReplicateRecipe::execute() if there there is another user of the instruction
other than the Phi node which needs to be widened.
This changes collectLoopScalars so that if there are any other users of
Update other than a Phi node, it is not added to ScalarPtrs.
Reviewed By: david-arm, fhahn
Differential Revision: https://reviews.llvm.org/D111294
At the moment, a VPValue is created for the backedge-taken count, which
is used by some recipes. To make it easier to identify the operands of
recipes using the backedge-taken count, print it at the beginning of the
VPlan if it is used.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D111298
This patch adds further support for vectorisation of loops that involve
selecting an integer value based on a previous comparison. Consider the
following C++ loop:
int r = a;
for (int i = 0; i < n; i++) {
if (src[i] > 3) {
r = b;
}
src[i] += 2;
}
We should be able to vectorise this loop because all we are doing is
selecting between two states - 'a' and 'b' - both of which are loop
invariant. This just involves building a vector of values that contain
either 'a' or 'b', where the final reduced value will be 'b' if any lane
contains 'b'.
The IR generated by clang typically looks like this:
%phi = phi i32 [ %a, %entry ], [ %phi.update, %for.body ]
...
%pred = icmp ugt i32 %val, i32 3
%phi.update = select i1 %pred, i32 %b, i32 %phi
We already detect min/max patterns, which also involve a select + cmp.
However, with the min/max patterns we are selecting loaded values (and
hence loop variant) in the loop. In addition we only support certain
cmp predicates. This patch adds a new pattern matching function
(isSelectCmpPattern) and new RecurKind enums - SelectICmp & SelectFCmp.
We only support selecting values that are integer and loop invariant,
however we can support any kind of compare - integer or float.
Tests have been added here:
Transforms/LoopVectorize/AArch64/sve-select-cmp.ll
Transforms/LoopVectorize/select-cmp-predicated.ll
Transforms/LoopVectorize/select-cmp.ll
Differential Revision: https://reviews.llvm.org/D108136
We need to be better at exposing the comparison predicate to getCmpSelInstrCost calls as some targets (e.g. X86 SSE) have very different costs for different comparisons (PR48337), and we can't always rely on the optional Instruction argument.
This initial commit requires explicit condition type and predicate arguments. The next step will be to review a lot of the existing getCmpSelInstrCost calls which have used BAD_ICMP_PREDICATE even when the predicate is known.
Differential Revision: https://reviews.llvm.org/D111024
Some initially gathered nodes missed the check for the reused scalars,
which leads to high gather cost. Such nodes still can be represented as
m gathers + shuffle instead of n gathers, where m < n.
Differential Revision: https://reviews.llvm.org/D111153
This patch is changing the InsertElement's placeholder to poison without changing the LSV's behavior.
Regardless of whether `StoreTy` is FixedVectorType or not, the poison value will be overwritten with a different value.
Therefore, whether the InsertElement's placeholder is poison or undef will not affect the result of the program.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D111005
D104809 changed `buildTree_rec` to check for extract element instructions
with scalable types. However, if the extract is extended or truncated,
these changes do not apply and we assert later on in isShuffle(), which
attempts to cast the type of the extract to FixedVectorType.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D110640
This patch adds further support for vectorisation of loops that involve
selecting an integer value based on a previous comparison. Consider the
following C++ loop:
int r = a;
for (int i = 0; i < n; i++) {
if (src[i] > 3) {
r = b;
}
src[i] += 2;
}
We should be able to vectorise this loop because all we are doing is
selecting between two states - 'a' and 'b' - both of which are loop
invariant. This just involves building a vector of values that contain
either 'a' or 'b', where the final reduced value will be 'b' if any lane
contains 'b'.
The IR generated by clang typically looks like this:
%phi = phi i32 [ %a, %entry ], [ %phi.update, %for.body ]
...
%pred = icmp ugt i32 %val, i32 3
%phi.update = select i1 %pred, i32 %b, i32 %phi
We already detect min/max patterns, which also involve a select + cmp.
However, with the min/max patterns we are selecting loaded values (and
hence loop variant) in the loop. In addition we only support certain
cmp predicates. This patch adds a new pattern matching function
(isSelectCmpPattern) and new RecurKind enums - SelectICmp & SelectFCmp.
We only support selecting values that are integer and loop invariant,
however we can support any kind of compare - integer or float.
Tests have been added here:
Transforms/LoopVectorize/AArch64/sve-select-cmp.ll
Transforms/LoopVectorize/select-cmp-predicated.ll
Transforms/LoopVectorize/select-cmp.ll
Differential Revision: https://reviews.llvm.org/D108136
This is analogous to D86156 (which preserves "lossy" BFI in loop
passes). Lossy means that the analysis preserved may not be up to date
with regards to new blocks that are added in loop passes, but BPI will
not contain stale pointers to basic blocks that are deleted by the loop
passes.
This is achieved through BasicBlockCallbackVH in BPI, which calls
eraseBlock that updates the data structures in BPI whenever a basic
block is deleted.
This patch does not have any changes in the upstream pipeline, since
none of the loop passes in the pipeline use BPI currently.
However, since BPI wasn't previously preserved in loop passes, the loop
predication pass was invoking BPI *on the entire
function* every time it ran in an LPM. This caused massive compile time
in our downstream LPM invocation which contained loop predication.
See updated test with an invocation of a loop-pipeline containing loop
predication and -debug-pass turned ON.
Reviewed-By: asbirlea, modimo
Differential Revision: https://reviews.llvm.org/D110438
Try to improve vectorization of the PHI nodes by trying to vectorize
similar instructions at the size of the widest possible vectors, then
aggregating with compatible type PHIs and trying to vectoriza again and
only if this failed, try smaller sizes of the vector factors for
compatible PHI nodes. This restores performance of several benchmarks
after tuning of the fp/int conversion instructions costs.
Differential Revision: https://reviews.llvm.org/D108740
The instruction extractelement/extractvalue are not required to
be scheduled since they only depend on the source vector/aggregate (with
constant indices), smae applies to the parent basic block checks.
Improves compile time and saves scheduling budget.
Differential Revision: https://reviews.llvm.org/D108703
ScalarizationResult's destructor makes sure ToFreeze is not ignored if
set. Currently, scalarizeLoadExtract has an early exit if the index is
not safe directly. But when it is SafeWithFreeze, we need to discard the
state first, otherwise we hit the assert in the destructor.
Fixes PR51992.
We see that it might otherwise do:
%10 = getelementptr {}**, <2 x {}***> %9, <2 x i32> <i32 10, i32 4>
%11 = bitcast <2 x {}***> %10 to <2 x i64*>
...
%27 = extractelement <2 x i64*> %11, i32 0
%28 = bitcast i64* %27 to <2 x i64>*
store <2 x i64> %22, <2 x i64>* %28, align 4, !tbaa !2
Which is an out-of-bounds store (the extractelement got offset 10
instead of offset 4 as intended). With the fix, we correctly generate
extractelement for i32 1 and generate correct code.
Differential Revision: https://reviews.llvm.org/D106613
Avoid relying on the default cost kinds in TTI calls (we already do this in other places in SLP) - noticed while trying to see how much work it'd be to extend D110242 and remove all remaining uses of default CostKind arguments.
This patch updates VectorCombine to use a worklist to allow iterative
simplifications where a combine enables other combines.
Suggested in D100302.
The main use case at the moment is foldSingleElementStore and
scalarizeLoadExtract working together to improve scalarization.
Note that we now also do not run SimplifyInstructionsInBlock on the
whole function if there have been changes. This means we fail to
remove/simplify instructions not related to any of the vector combines.
IMO this is fine, as simplifying the whole function seems more like a
workaround for not tracking the changed instructions.
Compile-time impact looks neutral:
NewPM-O3: +0.02%
NewPM-ReleaseThinLTO: -0.00%
NewPM-ReleaseLTO-g: -0.02%
http://llvm-compile-time-tracker.com/compare.php?from=52832cd917af00e2b9c6a9d1476ba79754dcabff&to=e66520a4637290550a945d528e3e59573485dd40&stat=instructions
Reviewed By: spatel, lebedev.ri
Differential Revision: https://reviews.llvm.org/D110171
This patch fixes the crash found by PR51614:
whenever doing tail folding, interleave groups must be considered under mask.
Another fix D108900 follows for targets that support masked loads and stores:
when *deciding* to vectorize with masked interleave groups, check if the access
is reverse - which is currently not supported; rather than (only) asserting when
computing cost and generating code.
Differential Revision: https://reviews.llvm.org/D108891
isValidAssumeForContext can provide better results with access to the
dominator tree in some cases. This patch adjusts computeConstantRange to
allow passing through a dominator tree.
The use VectorCombine is updated to pass through the DT to enable
additional scalarization.
Note that similar APIs like computeKnownBits already accept optional dominator
tree arguments.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D110175
Reworked reordering algorithm. Originally, the compiler just tried to
detect the most common order in the reordarable nodes (loads, stores,
extractelements,extractvalues) and then fully rebuilding the graph in
the best order. This was not effecient, since it required an extra
memory and time for building/rebuilding tree, double the use of the
scheduling budget, which could lead to missing vectorization due to
exausted scheduling resources.
Patch provide 2-way approach for graph reodering problem. At first, all
reordering is done in-place, it doe not required tree
deleting/rebuilding, it just rotates the scalars/orders/reuses masks in
the graph node.
The first step (top-to bottom) rotates the whole graph, similarly to the previous
implementation. Compiler counts the number of the most used orders of
the graph nodes with the same vectorization factor and then rotates the
subgraph with the given vectorization factor to the most used order, if
it is not empty. Then repeats the same procedure for the subgraphs with
the smaller vectorization factor. We can do this because we still need
to reshuffle smaller subgraph when buildiong operands for the graph
nodes with lasrger vectorization factor, we can rotate just subgraph,
not the whole graph.
The second step (bottom-to-top) scans through the leaves and tries to
detect the users of the leaves which can be reordered. If the leaves can
be reorder in the best fashion, they are reordered and their user too.
It allows to remove double shuffles to the same ordering of the operands in
many cases and just reorder the user operations instead. Plus, it moves
the final shuffles closer to the top of the graph and in many cases
allows to remove extra shuffle because the same procedure is repeated
again and we can again merge some reordering masks and reorder user nodes
instead of the operands.
Also, patch improves cost model for gathering of loads, which improves
x264 benchmark in some cases.
Gives about +2% on AVX512 + LTO (more expected for AVX/AVX2) for {625,525}x264,
+3% for 508.namd, improves most of other benchmarks.
The compile and link time are almost the same, though in some cases it
should be better (we're not doing an extra instruction scheduling
anymore) + we may vectorize more code for the large basic blocks again
because of saving scheduling budget.
Differential Revision: https://reviews.llvm.org/D105020
This is a first step towards addressing the last remaining limitation of
the VPlan version of sinkScalarOperands: the legacy version can
partially sink operands. For example, if a GEP has uniform users outside
the sink target block, then the legacy version will sink all scalar
GEPs, other than the one for lane 0.
This patch works towards addressing this case in the VPlan version by
detecting such cases and duplicating the sink candidate. All users
outside of the sink target will be updated to use the uniform clone.
Note that this highlights an issue with VPValue naming. If we duplicate
a replicate recipe, they will share the same underlying IR value and
both VPValues will have the same name ir<%gep>.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D104254
Added '-print-pipeline-passes' printing of parameters for those passes
declared with *_WITH_PARAMS macro in PassRegistry.def.
Note that it only prints the parameters declared inside *_WITH_PARAMS as
in a few cases there appear to be additional parameters not parsable.
The following passes are now covered (i.e. all of those with *_WITH_PARAMS in
PassRegistry.def).
LoopExtractorPass - loop-extract
HWAddressSanitizerPass - hwsan
EarlyCSEPass - early-cse
EntryExitInstrumenterPass - ee-instrument
LowerMatrixIntrinsicsPass - lower-matrix-intrinsics
LoopUnrollPass - loop-unroll
AddressSanitizerPass - asan
MemorySanitizerPass - msan
SimplifyCFGPass - simplifycfg
LoopVectorizePass - loop-vectorize
MergedLoadStoreMotionPass - mldst-motion
GVN - gvn
StackLifetimePrinterPass - print<stack-lifetime>
SimpleLoopUnswitchPass - simple-loop-unswitch
Differential Revision: https://reviews.llvm.org/D109310
Instead of discovering the sink-to block for each operand in the main
loop, the sink-to block can instead be directly queued with the
operands.
This simplifies processing in the main loop and is a NFC change split
off from D104254 as suggested there.
38b098be66 limited scalarization to indices that are known non-poison.
For certain patterns that restrict the range of an index, we can insert
a freeze of the original value, to prevent propagation of poison.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D107580
This patch simply replaces any unsigned VFs with ElementCounts. It's
still NFC because at the moment epilogue vectorisation is disabled
when the main vector loop uses scalable vectors.
Differential Revision: https://reviews.llvm.org/D109364
Users of VPValues are managed in a vector, so we need to be more
careful when iterating over users while updating them. For now, just
copy them.
Fixes 51798.
Pass the access type to getPtrStride(), so it is not determined
from the pointer element type. Many cases still fetch the element
type at a higher level though, so this only partially addresses
the issue.
For SVE, when scalarising the PHI instruction the whole vector part is
generated as opposed to creating instructions for each lane for fixed-
width vectors. However, in some cases the lane values may be needed
later (e.g for a load instruction) so we still need to calculate
these values to avoid extractelement being called on the vector part.
Differential Revision: https://reviews.llvm.org/D109445
This renames the primary methods for creating a zero value to `getZero`
instead of `getNullValue` and renames predicates like `isAllOnesValue`
to simply `isAllOnes`. This achieves two things:
1) This starts standardizing predicates across the LLVM codebase,
following (in this case) ConstantInt. The word "Value" doesn't
convey anything of merit, and is missing in some of the other things.
2) Calling an integer "null" doesn't make any sense. The original sin
here is mine and I've regretted it for years. This moves us to calling
it "zero" instead, which is correct!
APInt is widely used and I don't think anyone is keen to take massive source
breakage on anything so core, at least not all in one go. As such, this
doesn't actually delete any entrypoints, it "soft deprecates" them with a
comment.
Included in this patch are changes to a bunch of the codebase, but there are
more. We should normalize SelectionDAG and other APIs as well, which would
make the API change more mechanical.
Differential Revision: https://reviews.llvm.org/D109483
Store the used element type in the InductionDescriptor. For typed
pointers, it remains the pointer element type. For opaque pointers,
we always use an i8 element type, such that the step is a simple
offset.
A previous version of this patch instead tried to guess the element
type from an induction GEP, but this is not reliable, as the GEP
may be hidden (see @both in iv_outside_user.ll).
Differential Revision: https://reviews.llvm.org/D104795
The load store vectorizer currently uses isNoAlias() to determine
whether memory-accessing instructions should prevent vectorization.
However, this only works for loads and stores. Additionally, a
couple of intrinsics like assume are special-cased to be ignored.
Instead use getModRefInfo() to generically determine whether the
instruction accesses/modifies the relevant location. This will
automatically handle all inaccessiblememonly intrinsics correctly
(as well as other calls that don't modref for other reasons).
This requires generalizing the code a bit, as it was previously
only considering loads and stored in particular.
Differential Revision: https://reviews.llvm.org/D109020
SLPVectorizer currently uses AA::isNoAlias() to determine whether
two locations alias. This does not work if one of the instructions
is a call. Instead, we should check getModRefInfo(), which
determines whether an arbitrary instruction modifies or references
a given location.
Among other things, this prevents @llvm.experimental.noalias.scope.decl()
and other inaccessiblmemonly intrinsics from interfering with SLP
vectorization.
Differential Revision: https://reviews.llvm.org/D109012
After applying VPlan-to-VPlan transformations, using IR references to
query VPlan values may be incorrect, as the IR is not in sync with the
VPlan any longer.
To better detect such mis-matches, this patch introduces a new flag to
VPlans to indicate whether it is safe to query VPValues using IR values.
getVPValue is updated to assert if it is called when the flag indicates
it is not safe any longer.
There is an escape hatch via an extra argument, because there are 3
places that need to be fixed first. Those are
1. truncateToMinimalBitwidths
2. clearReductionWrapFlags
3. fixLCSSAPHIs
As a first step, this flag will help preventing new code from violating
this property.
Any suggestions with respect to naming very welcome!
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D108573
Reworked reordering algorithm. Originally, the compiler just tried to
detect the most common order in the reordarable nodes (loads, stores,
extractelements,extractvalues) and then fully rebuilding the graph in
the best order. This was not effecient, since it required an extra
memory and time for building/rebuilding tree, double the use of the
scheduling budget, which could lead to missing vectorization due to
exausted scheduling resources.
Patch provide 2-way approach for graph reodering problem. At first, all
reordering is done in-place, it doe not required tree
deleting/rebuilding, it just rotates the scalars/orders/reuses masks in
the graph node.
The first step (top-to bottom) rotates the whole graph, similarly to the previous
implementation. Compiler counts the number of the most used orders of
the graph nodes with the same vectorization factor and then rotates the
subgraph with the given vectorization factor to the most used order, if
it is not empty. Then repeats the same procedure for the subgraphs with
the smaller vectorization factor. We can do this because we still need
to reshuffle smaller subgraph when buildiong operands for the graph
nodes with lasrger vectorization factor, we can rotate just subgraph,
not the whole graph.
The second step (bottom-to-top) scans through the leaves and tries to
detect the users of the leaves which can be reordered. If the leaves can
be reorder in the best fashion, they are reordered and their user too.
It allows to remove double shuffles to the same ordering of the operands in
many cases and just reorder the user operations instead. Plus, it moves
the final shuffles closer to the top of the graph and in many cases
allows to remove extra shuffle because the same procedure is repeated
again and we can again merge some reordering masks and reorder user nodes
instead of the operands.
Also, patch improves cost model for gathering of loads, which improves
x264 benchmark in some cases.
Gives about +2% on AVX512 + LTO (more expected for AVX/AVX2) for {625,525}x264,
+3% for 508.namd, improves most of other benchmarks.
The compile and link time are almost the same, though in some cases it
should be better (we're not doing an extra instruction scheduling
anymore) + we may vectorize more code for the large basic blocks again
because of saving scheduling budget.
Differential Revision: https://reviews.llvm.org/D105020
Reworked reordering algorithm. Originally, the compiler just tried to
detect the most common order in the reordarable nodes (loads, stores,
extractelements,extractvalues) and then fully rebuilding the graph in
the best order. This was not effecient, since it required an extra
memory and time for building/rebuilding tree, double the use of the
scheduling budget, which could lead to missing vectorization due to
exausted scheduling resources.
Patch provide 2-way approach for graph reodering problem. At first, all
reordering is done in-place, it doe not required tree
deleting/rebuilding, it just rotates the scalars/orders/reuses masks in
the graph node.
The first step (top-to bottom) rotates the whole graph, similarly to the previous
implementation. Compiler counts the number of the most used orders of
the graph nodes with the same vectorization factor and then rotates the
subgraph with the given vectorization factor to the most used order, if
it is not empty. Then repeats the same procedure for the subgraphs with
the smaller vectorization factor. We can do this because we still need
to reshuffle smaller subgraph when buildiong operands for the graph
nodes with lasrger vectorization factor, we can rotate just subgraph,
not the whole graph.
The second step (bottom-to-top) scans through the leaves and tries to
detect the users of the leaves which can be reordered. If the leaves can
be reorder in the best fashion, they are reordered and their user too.
It allows to remove double shuffles to the same ordering of the operands in
many cases and just reorder the user operations instead. Plus, it moves
the final shuffles closer to the top of the graph and in many cases
allows to remove extra shuffle because the same procedure is repeated
again and we can again merge some reordering masks and reorder user nodes
instead of the operands.
Also, patch improves cost model for gathering of loads, which improves
x264 benchmark in some cases.
Gives about +2% on AVX512 + LTO (more expected for AVX/AVX2) for {625,525}x264,
+3% for 508.namd, improves most of other benchmarks.
The compile and link time are almost the same, though in some cases it
should be better (we're not doing an extra instruction scheduling
anymore) + we may vectorize more code for the large basic blocks again
because of saving scheduling budget.
Differential Revision: https://reviews.llvm.org/D105020
The instruction extractelement/extractvalue are not required to
be scheduled since they only depend on the source vector/aggregate (with
constant indices), smae applies to the parent basic block checks.
Improves compile time and saves scheduling budget.
Differential Revision: https://reviews.llvm.org/D108703
Adjusting the reduction recipes still relies on references to the
original IR, which can become outdated by the first-order recurrence
handling. Until reduction recipe construction does not require IR
references, move it before first-order recurrence handling, to prevent a
crash as exposed by D106653.
This reverts commit f4122398e7 to
investigate a crash exposed by it.
The patch breaks building the code below with `clang -O2 --target=aarch64-linux`
int a;
double b, c;
void d() {
for (; a; a++) {
b += c;
c = a;
}
}
I have added a new TTI interface called enableOrderedReductions() that
controls whether or not ordered reductions should be enabled for a
given target. By default this returns false, whereas for AArch64 it
returns true and we rely upon the cost model to make sensible
vectorisation choices. It is still possible to override the new TTI
interface by setting the command line flag:
-force-ordered-reductions=true|false
I have added a new RUN line to show that we use ordered reductions by
default for SVE and Neon:
Transforms/LoopVectorize/AArch64/strict-fadd.ll
Transforms/LoopVectorize/AArch64/scalable-strict-fadd.ll
Differential Revision: https://reviews.llvm.org/D106653
Removed AArch64 usage of the getMaxVScale interface, replacing it with
the vscale_range(min, max) IR Attribute.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D106277
LoopLoadElimination, LoopVersioning and LoopVectorize currently
fetch MemorySSA when construction LoopAccessAnalysis. However,
LoopAccessAnalysis does not actually use MemorySSA and we can pass
nullptr instead.
This saves one MemorySSA calculation in the default pipeline, and
thus improves compile-time.
Differential Revision: https://reviews.llvm.org/D108074
Previously we emitted a "does not support scalable vectors"
remark for all targets whenever vectorisation is attempted. This
pollutes the output for architectures that don't support scalable
vectors and is likely confusing to the user.
Instead this patch introduces a debug message that reports when
scalable vectorisation is allowed by the target and only issues
the previous remark when scalable vectorisation is specifically
requested, for example:
#pragma clang loop vectorize_width(2, scalable)
Differential Revision: https://reviews.llvm.org/D108028
Teach LV to use masked-store to support interleave-store-group with
gaps (instead of scatters/scalarization).
The symmetric case of using masked-load to support
interleaved-load-group with gaps was introduced a while ago, by
https://reviews.llvm.org/D53668; This patch completes the store-scenario
leftover from D53668, and solves PR50566.
Reviewed by: Ayal Zaks
Differential Revision: https://reviews.llvm.org/D104750
After refactoring the phi recipes, we can now iterate over all header
phis in a VPlan to detect reductions when it comes to fixing them up
when tail folding.
This reduces the coupling with the cost model & legal by using the
information directly available in VPlan. It also removes a call to
getOrAddVPValue, which references the original IR value which may
become outdated after VPlan transformations.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100102
This patch adds more instructions to the Uniforms list, for example certain
intrinsics that are uniform by definition or whose operands are loop invariant.
This list includes:
1. The intrinsics 'experimental.noalias.scope.decl' and 'sideeffect', which
are always uniform by definition.
2. If intrinsics 'lifetime.start', 'lifetime.end' and 'assume' have
loop invariant input operands then these are also uniform too.
Also, in VPRecipeBuilder::handleReplication we check if an instruction is
uniform based purely on whether or not the instruction lives in the Uniforms
list. However, there are certain cases where calls to some intrinsics can
be effectively treated as uniform too. Therefore, we now also treat the
following cases as uniform for scalable vectors:
1. If the 'assume' intrinsic's operand is not loop invariant, then we
are free to treat this as uniform anyway since it's only a performance
hint. We will get the benefit for the first lane.
2. When the input pointers for 'lifetime.start' and 'lifetime.end' are loop
variant then for scalable vectors we assume these still ultimately come
from the broadcast of an alloca. We do not support scalable vectorisation
of loops containing alloca instructions, hence the alloca itself would
be invariant. If the pointer does not come from an alloca then the
intrinsic itself has no effect.
I have updated the assume test for fixed width, since we now treat it
as uniform:
Transforms/LoopVectorize/assume.ll
I've also added new scalable vectorisation tests for other intriniscs:
Transforms/LoopVectorize/scalable-assume.ll
Transforms/LoopVectorize/scalable-lifetime.ll
Transforms/LoopVectorize/scalable-noalias-scope-decl.ll
Differential Revision: https://reviews.llvm.org/D107284
All information to fix-up the reduction phi nodes in the vectorized loop
is available in VPlan now. This patch moves the code to do so, to make
this clearer. Fixing up the loop exit value still relies on other
information and remains outside of VPlan for now.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100113
If the vectorized insertelements instructions form indentity subvector
(the subvector at the beginning of the long vector), it is just enough
to extend the vector itself, no need to generate inserting subvector
shuffle.
Differential Revision: https://reviews.llvm.org/D107494
Since all operands to ExtractValue must be loop-invariant when we deem
the loop vectorizable, we can consider ExtractValue to be uniform.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D107286
We can only trust the range of the index if it is guaranteed
non-poison.
Fixes PR50949.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D107364
This patch adds more instructions to the Uniforms list, for example certain
intrinsics that are uniform by definition or whose operands are loop invariant.
This list includes:
1. The intrinsics 'experimental.noalias.scope.decl' and 'sideeffect', which
are always uniform by definition.
2. If intrinsics 'lifetime.start', 'lifetime.end' and 'assume' have
loop invariant input operands then these are also uniform too.
Also, in VPRecipeBuilder::handleReplication we check if an instruction is
uniform based purely on whether or not the instruction lives in the Uniforms
list. However, there are certain cases where calls to some intrinsics can
be effectively treated as uniform too. Therefore, we now also treat the
following cases as uniform for scalable vectors:
1. If the 'assume' intrinsic's operand is not loop invariant, then we
are free to treat this as uniform anyway since it's only a performance
hint. We will get the benefit for the first lane.
2. When the input pointers for 'lifetime.start' and 'lifetime.end' are loop
variant then for scalable vectors we assume these still ultimately come
from the broadcast of an alloca. We do not support scalable vectorisation
of loops containing alloca instructions, hence the alloca itself would
be invariant. If the pointer does not come from an alloca then the
intrinsic itself has no effect.
I have updated the assume test for fixed width, since we now treat it
as uniform:
Transforms/LoopVectorize/assume.ll
I've also added new scalable vectorisation tests for other intriniscs:
Transforms/LoopVectorize/scalable-assume.ll
Transforms/LoopVectorize/scalable-lifetime.ll
Transforms/LoopVectorize/scalable-noalias-scope-decl.ll
Differential Revision: https://reviews.llvm.org/D107284
This change wasn't strictly necessary for D106164 and could be removed.
This patch addresses the post-commit comments from @fhahn on D106164, and
also changes sve-widen-gep.ll to use the same IR test as shown in
pointer-induction.ll.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D106878
If the vectorized insertelements instructions form indentity subvector
(the subvector at the beginning of the long vector), it is just enough
to extend the vector itself, no need to generate inserting subvector
shuffle.
Differential Revision: https://reviews.llvm.org/D107344
I'm renaming the flag because a future patch will add a new
enableOrderedReductions() TTI interface and so the meaning of this
flag will change to be one of forcing the target to enable/disable
them. Also, since other places in LoopVectorize.cpp use the word
'Ordered' instead of 'strict' I changed the flag to match.
Differential Revision: https://reviews.llvm.org/D107264
This patch updates VPInterleaveRecipe::print to print the actual defined
VPValues for load groups and the store VPValue operands for store
groups.
The IR references may become outdated while transforming the VPlan and
the defined and stored VPValues always are up-to-date.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D107223
Replace insertelement instructions for splats with just single
insertelement + broadcast shuffle. Also, try to merge these instructions
if they come from the same/shuffled gather node.
Differential Revision: https://reviews.llvm.org/D107104
For the nodes with reused scalars the user may be not only of the size
of the final shuffle but also of the size of the scalars themselves,
need to check for this. It is safe to just modify the check here, since
the order of the scalars themselves is preserved, only indeces of the
reused scalars are changed. So, the users with the same size as the
number of scalars in the node, will not be affected, they still will get
the operands in the required order.
Reported by @mstorsjo in D105020.
Differential Revision: https://reviews.llvm.org/D107080
If the instruction was previously deleted, it should not be treated as
an external user. This fixes cost estimation and removes dead
extractelement instructions.
Differential Revision: https://reviews.llvm.org/D107106
Need to check that the minimum acceptable vector factor is at least 2,
not 0, to avoid compiler crash during gathered loads analysis.
Differential Revision: https://reviews.llvm.org/D107058
Reworked reordering algorithm. Originally, the compiler just tried to
detect the most common order in the reordarable nodes (loads, stores,
extractelements,extractvalues) and then fully rebuilding the graph in
the best order. This was not effecient, since it required an extra
memory and time for building/rebuilding tree, double the use of the
scheduling budget, which could lead to missing vectorization due to
exausted scheduling resources.
Patch provide 2-way approach for graph reodering problem. At first, all
reordering is done in-place, it doe not required tree
deleting/rebuilding, it just rotates the scalars/orders/reuses masks in
the graph node.
The first step (top-to bottom) rotates the whole graph, similarly to the previous
implementation. Compiler counts the number of the most used orders of
the graph nodes with the same vectorization factor and then rotates the
subgraph with the given vectorization factor to the most used order, if
it is not empty. Then repeats the same procedure for the subgraphs with
the smaller vectorization factor. We can do this because we still need
to reshuffle smaller subgraph when buildiong operands for the graph
nodes with lasrger vectorization factor, we can rotate just subgraph,
not the whole graph.
The second step (bottom-to-top) scans through the leaves and tries to
detect the users of the leaves which can be reordered. If the leaves can
be reorder in the best fashion, they are reordered and their user too.
It allows to remove double shuffles to the same ordering of the operands in
many cases and just reorder the user operations instead. Plus, it moves
the final shuffles closer to the top of the graph and in many cases
allows to remove extra shuffle because the same procedure is repeated
again and we can again merge some reordering masks and reorder user nodes
instead of the operands.
Also, patch improves cost model for gathering of loads, which improves
x264 benchmark in some cases.
Gives about +2% on AVX512 + LTO (more expected for AVX/AVX2) for {625,525}x264,
+3% for 508.namd, improves most of other benchmarks.
The compile and link time are almost the same, though in some cases it
should be better (we're not doing an extra instruction scheduling
anymore) + we may vectorize more code for the large basic blocks again
because of saving scheduling budget.
Differential Revision: https://reviews.llvm.org/D105020
As suggested in D105008, move the code that fixes up the backedge value
for first order recurrences to VPlan::execute.
Now all that remains in fixFirstOrderRecurrences is the code responsible
for creating the exit values in the middle block.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D106244
This makes a couple of changes to the costing of MLA reduction patterns,
to more accurately cost various patterns that can come up from
vectorization.
- The Arm implementation of getExtendedAddReductionCost is altered to
only provide costs for legal or smaller types. Larger than legal types
need to be split, which currently does not work very well, especially
for predicated reductions where the predicate may be legal but needs to
be split. Currently we limit it to legal or smaller input types.
- The getReductionPatternCost has learnt that reduce(ext(mul(ext, ext))
is a pattern that can come up, and can be treated the same as
reduce(mul(ext, ext)) providing the extension types match.
- And it has been adjusted to not count the ext in reduce(mul(ext, ext))
as part of a reduce(mul) pattern.
Together these changes help to more accurately cost the mla reductions
in cases such as where the extend types don't match or the extend
opcodes are different, picking better vector factors that don't result
in expanded reductions.
Differential Revision: https://reviews.llvm.org/D106166
Consider the following loop:
void foo(float *dst, float *src, int N) {
for (int i = 0; i < N; i++) {
dst[i] = 0.0;
for (int j = 0; j < N; j++) {
dst[i] += src[(i * N) + j];
}
}
}
When we are not building with -Ofast we may attempt to vectorise the
inner loop using ordered reductions instead. In addition we also try
to select an appropriate interleave count for the inner loop. However,
when choosing a VF=1 the inner loop will be scalar and there is existing
code in selectInterleaveCount that limits the interleave count to 2
for reductions due to concerns about increasing the critical path.
For ordered reductions this problem is even worse due to the additional
data dependency, and so I've added code to simply disable interleaving
for scalar ordered reductions for now.
Test added here:
Transforms/LoopVectorize/AArch64/strict-fadd-vf1.ll
Differential Revision: https://reviews.llvm.org/D106646
The loop vectorizer may decide to use tail folding when the trip-count
is low. When that happens, scalable VFs are no longer a candidate,
since tail folding/predication is not yet supported for scalable vectors.
This can be re-enabled in a future patch.
Reviewed By: kmclaughlin
Differential Revision: https://reviews.llvm.org/D106657
Invalid costs can be used to avoid vectorization with a given VF, which is
used for scalable vectors to avoid things that the code-generator cannot
handle. If we override the cost using the -force-target-instruction-cost
option of the LV, we would override this mechanism, rendering the flag useless.
This change ensures the cost is only overriden when the original cost that
was calculated is valid. That allows the flag to be used in combination
with the -scalable-vectorization option.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D106677
Scalarization for scalable vectors is not (yet) supported, so the
LV discards a VF when scalarization is chosen as the widening
decision. It should therefore not assert that the VF is not scalable
when it computes the decision to scalarize.
The code can get here when both the interleave-cost, gather/scatter cost
and scalarization-cost are all illegal. This may e.g. happen for SVE
when the VF=1, to avoid generating `<vscale x 1 x eltty>` types that
the code-generator cannot yet handle.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D106656
This fixes an issue that was found in D105199, where a GEP instruction
is used both as the address of a store, as well as the value of a store.
For the former, the value is scalar after vectorization, but the latter
(as value) requires widening.
Other code in that function seems to prevent similar cases from happening,
but it seems this case was missed.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D106164
This reverts the revert commit b1777b04dc.
The patch originally got reverted due to a crash:
https://bugs.chromium.org/p/chromium/issues/detail?id=1232798#c2
The underlying issue was that we were not using the stored values from
the modified memory recipes, but the out-of-date values directly from
the IR (accessed via the VPlan). This should be fixed in d995d6376. A
reduced version of the reproducer has been added in 93664503be.
Need to fix several cost-related problems. The final type may be defined
incorrectly because of to early definition (we may end up with the wider
type), the CommonCost should not be redefined in ExtractElements
cost related calculations and the shuffle of the final insertelements
vectors should be calculated as a cost of single vector permutations
+ costs of two vector permutations for other n-1 incoming vectors.
Differential Revision: https://reviews.llvm.org/D106578
Fixes more casts to `<FixedVectorType>` for the cases where the
instruction is a Insert/ExtractElementInst.
For fixed-width, this part of truncateToMinimalBitWidths is tested by
AArch64/type-shrinkage-insertelt.ll. I attempted to write a test case for this part
of truncateToMinimalBitWidths which uses scalable vectors, but was unable to add
one. The tests in type-shrinkage-insertelt.ll rely on scalarization to create extract
element instructions for instance, which is not possible for scalable vectors.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D106163
Need to fix several cost-related problems. The final type may be defined
incorrectly because of to early definition (we may end up with the wider
type), the CommonCost should not be redefined in ExtractElements
cost related calculations and the shuffle of the final insertelements
vectors should be calculated as a cost of single vector permutations
+ costs of two vector permutations for other n-1 incoming vectors.
Differential Revision: https://reviews.llvm.org/D106578
Instead of getting the VPValue for the stored IR values through the
current plan, use the stored value of the recipes directly.
This way, the correct VPValues are used if the store recipes have been
modified in the VPlan and the IR value is not correct any longer. This
can happen, e.g. due to D105008.
I have added a new FastMathFlags parameter to getArithmeticReductionCost
to indicate what type of reduction we are performing:
1. Tree-wise. This is the typical fast-math reduction that involves
continually splitting a vector up into halves and adding each
half together until we get a scalar result. This is the default
behaviour for integers, whereas for floating point we only do this
if reassociation is allowed.
2. Ordered. This now allows us to estimate the cost of performing
a strict vector reduction by treating it as a series of scalar
operations in lane order. This is the case when FP reassociation
is not permitted. For scalable vectors this is more difficult
because at compile time we do not know how many lanes there are,
and so we use the worst case maximum vscale value.
I have also fixed getTypeBasedIntrinsicInstrCost to pass in the
FastMathFlags, which meant fixing up some X86 tests where we always
assumed the vector.reduce.fadd/mul intrinsics were 'fast'.
New tests have been added here:
Analysis/CostModel/AArch64/reduce-fadd.ll
Analysis/CostModel/AArch64/sve-intrinsics.ll
Transforms/LoopVectorize/AArch64/strict-fadd-cost.ll
Transforms/LoopVectorize/AArch64/sve-strict-fadd-cost.ll
Differential Revision: https://reviews.llvm.org/D105432
This patch avoids computing discounts for predicated instructions when the
VF is scalable.
There is no support for vectorization of loops with division because the
vectorizer cannot guarantee that zero divisions will not happen.
This loop now does not use VF scalable
```
for (long long i = 0; i < n; i++)
if (cond[i])
a[i] /= b[i];
```
Differential Revision: https://reviews.llvm.org/D101916
Currently the Instruction cost of getReductionPatternCost returns an
Invalid cost to specify "did not find the pattern". This changes that to
return an Optional with None specifying not found, allowing Invalid to
mean an infinite cost as is used elsewhere.
Differential Revision: https://reviews.llvm.org/D106140
This patch removes the assertion when VF is scalable and replaces
getKnownMinValue() by getFixedValue(), so it still guards the code against
scalable vector types.
The assertions were used to guarantee that getknownMinValue were not used for
scalable vectors.
Differential Revision: https://reviews.llvm.org/D106359
This patch adds a VPFirstOrderRecurrencePHIRecipe, to further untangle
VPWidenPHIRecipe into distinct recipes for distinct use cases/lowering.
See D104989 for a new recipe for reduction phis.
This patch also introduces a new `FirstOrderRecurrenceSplice`
VPInstruction opcode, which is used to make the forming of the vector
recurrence value explicit in VPlan. This more accurately models def-uses
in VPlan and also simplifies code-generation. Now, the vector recurrence
values are created at the right place during VPlan-codegeneration,
rather than during post-VPlan fixups.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D105008
The incoming values for PHI nodes may come from unreachable BasicBlocks,
need to handle this case.
Differential Revision: https://reviews.llvm.org/D106264
Part of D105020. Also, fixed FIXMEs that need to use wider vector type
when trying to calculate the cost of reused scalars. This may cause
regressions unless D100486 is landed to improve the cost estimations
for long vectors shuffling.
Differential Revision: https://reviews.llvm.org/D106060
The cost of the InsertSubvector shuffle kind cost is not complete and
may end up with just extracts + inserts costs in many cases. Added
a workaround to represent it as a generic PermuteSingleSrc, which is
still pessimistic but better than InsertSubvector.
Differential Revision: https://reviews.llvm.org/D105827
This patch returns an Invalid cost from getInstructionCost() for alloca
instructions if the VF is scalable, as otherwise loops which contain
these instructions will crash when attempting to scalarize the alloca.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D105824
The original patch was:
https://reviews.llvm.org/D105806
There were some issues with undeterministic behaviour of the sorting
function, which led to scalable-call.ll passing and/or failing. This
patch fixes the issue by numbering all instructions in the array first,
and using that number as the order, which should provide a consistent
ordering.
This reverts commit a607f64118.
This bug was introduced with D105730 / 25ee55c0ba .
If we are not converting all of the operations of a reduction
into a vector op, we need to preserve the existing select form
of the remaining ops. Otherwise, we are potentially leaking
poison where it did not in the original code.
Alive2 agrees that the version that freezes some inputs
and then falls back to scalar is correct:
https://alive2.llvm.org/ce/z/erF4K2
This change enables vectorization of multiple exit loops when the exit count is statically computable. That requirement - shared with the rest of LV - in turn requires each exit to be analyzeable and to dominate the latch.
The majority of work to support this was done in a set of previous patches. In particular,, 72314466 avoids having multiple edges from the middle block to the exits, and 4b33b2387 which added support for non-latch single exit and multiple exits with a single exiting block. As a result, this change is basically just removing a bailout and adjusting some tests now that the prerequisite work is done and has stuck in tree for a bit.
Differential Revision: https://reviews.llvm.org/D105817
The sort function for emitting an OptRemark was not deterministic,
which caused scalable-call.ll to fail on some buildbots. This patch
fixes that.
This patch also fixes an issue where `Instruction::comesBefore()`
is called when two Instructions are in different basic blocks,
which would otherwise cause an assertion failure.
This patch emits remarks for instructions that have invalid costs for
a given set of vectorization factors. Some example output:
t.c:4:19: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1): load
dst[i] = sinf(src[i]);
^
t.c:4:14: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1, vscale x 2, vscale x 4): call to llvm.sin.f32
dst[i] = sinf(src[i]);
^
t.c:4:12: remark: Instruction with invalid costs prevented vectorization at VF=(vscale x 1): store
dst[i] = sinf(src[i]);
^
Reviewed By: fhahn, kmclaughlin
Differential Revision: https://reviews.llvm.org/D105806
The cost of the InsertSubvector shuffle kind cost is not complete and
may end up with just extracts + inserts costs in many cases. Added
a workaround to represent it as a generic PermuteSingleSrc, which is
still pessimistic but better than InsertSubvector.
Differential Revision: https://reviews.llvm.org/D105827
This has been a work-in-progress for a long time...we finally have all of
the pieces in place to handle vectorization of compare code as shown in:
https://llvm.org/PR41312
To do this (see PhaseOrdering tests), we converted SimplifyCFG and
InstCombine to the poison-safe (select) forms of the logic ops, so now we
need to have SLP recognize those patterns and insert a freeze op to make
a safe reduction:
https://alive2.llvm.org/ce/z/NH54Ah
We get the minimal patterns with this patch, but the PhaseOrdering tests
show that we still need adjustments to get the ideal IR in some or all of
the motivating cases.
Differential Revision: https://reviews.llvm.org/D105730
The const version of VPValue::getVPValue still had a default value for
the value index. Remove the default value and use getVPSingleValue
instead, which is the proper function.
Instead of performing the isMoreProfitable() operation on
InstructionCost::CostTy the operation is performed on InstructionCost
directly, so that it can handle the case where one of the costs is
Invalid.
This patch also changes the CostTy to be int64_t, so that the type is
wide enough to deal with multiplications with e.g. `unsigned MaxTripCount`.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D105113
This makes it clearer when we have encountered the extra arg.
Also, we may need to adjust the way the operand iteration
works when handling logical and/or.
This is NFC-intended currently (so no test diffs). The motivation
is to eventually allow matching for poison-safe logical-and and
logical-or (these are in the form of a select-of-bools).
( https://llvm.org/PR41312 )
Those patterns will not have all of the same constraints as min/max
in the form of cmp+sel. We may also end up removing the cmp+sel
min/max matching entirely (if we canonicalize to intrinsics), so
this will make that step easier.
This patch removes the IsPairwiseForm flag from the Reduction Cost TTI
hooks, along with some accompanying code for pattern matching reductions
from trees starting at extract elements. IsPairWise is now assumed to be
false, which was the predominant way that the value was used from both
the Loop and SLP vectorizers. Since the adjustments such as D93860, the
SLP vectorizer has not relied upon this distinction between paiwise and
non-pairwise reductions.
This also removes some code that was detecting reductions trees starting
from extract elements inside the costmodel. This case was
double-counting costs though, adding the individual costs on the
individual instruction _and_ the total cost of the reduction. Removing
it changes the costs in llvm/test/Analysis/CostModel/X86/reduction.ll to
not double count. The cost of reduction intrinsics is still tested
through the various tests in
llvm/test/Analysis/CostModel/X86/reduce-xyz.ll.
Differential Revision: https://reviews.llvm.org/D105484
Patch tries to improve the vectorization of stores. Originally, we just
check the type and the base pointer of the store.
Patch adds some extra checks to avoid non-profitable vectorization
cases. It includes analysis of the scalar values to be stored and
triggers the vectorization attempt only if the scalar values have
same/alt opcode and are from same basic block, i.e. we don't end up
immediately with the gather node, which is not profitable.
This also improves compile time by filtering out non-profitable cases.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D104122
Revived D101297 in its original form + added some changes in X86
legalization cehcking for masked gathers.
This solution is the most stable and the most correct one. We have to
check the legality before trying to build the masked gather in SLP.
Without this check we have incorrect cost (for SLP) in case if the masked gather
is not legal/slower than the gather. And we're missing some
vectorization opportunities.
This can be fixed in the cost model, but in this case we need to add
special checks for the cost of GEPs for ScatterVectorize node, add
special check for small trees, etc., i.e. there are a lot of corner
cases here and there, which insrease code base and make it harder to
maintain the code.
> Can't we rely on cost model to deal with this? This can be profitable for futher vectorization, when we can start from such gather loads as seed.
The question from D101297. Actually, no, it can't. Actually, simple
gather may give us better result, especially after we started
vectorization of insertelements. Plus, like I said before, the cost for
non-legal masked gathers leads to missed vectorization opportunities.
Differential Revision: https://reviews.llvm.org/D105042
The reduction matching was probably only dealing with binops
when it was written, but we have now generalized it to handle
select and intrinsics too, so assert on that too.
Resubmit after the following changes:
* Fix a latent bug related to unrolling with required epilogue (see e49d65f). I believe this is the cause of the prior PPC buildbot failure.
* Disable non-latch exits for epilogue vectorization to be safe (9ffa90d)
* Split out assert movement (600624a) to reduce churn if this gets reverted again.
Previous commit message (try 3)
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
When skimming through old review discussion, I noticed a post commit comment on an earlier patch which had gone unaddressed. Better late (4 months), than never right?
I'm not aware of an active problem with the combination of non-latch exits and epilogue vectorization, but the interaction was not considered and I'm not modivated to make epilogue vectorization work with early exits. If there were a bug in the interaction, it would be pretty hard to hit right now (as we canonicalize towards bottom tested loops), but an upcoming change to allow multiple exit loops will greatly increase the chance for error. Thus, let's play it safe for now.
Compare type IDs and DFS numbering for basic block instead of addresses
to fix non-determinism.
Differential Revision: https://reviews.llvm.org/D105031
This reverts commit 706bbfb35b.
The committed version moves the definition of VPReductionPHIRecipe out
of an ifdef only intended for ::print helpers. This should resolve the
build failures that caused the revert
This patch adds a TTI function, isElementTypeLegalForScalableVector, to query
whether it is possible to vectorize a given element type. This is called by
isLegalToVectorizeInstTypesForScalable to reject scalable vectorization if
any of the instruction types in the loop are unsupported, e.g:
int foo(__int128_t* ptr, int N)
#pragma clang loop vectorize_width(4, scalable)
for (int i=0; i<N; ++i)
ptr[i] = ptr[i] + 42;
This example currently crashes if we attempt to vectorize since i128 is not a
supported type for scalable vectorization.
Reviewed By: sdesmalen, david-arm
Differential Revision: https://reviews.llvm.org/D102253
This reverts commit 3fed6d443f,
bbcbf21ae6 and
6c3451cd76.
The changes causing build failures with certain configurations, e.g.
https://lab.llvm.org/buildbot/#/builders/67/builds/3365/steps/6/logs/stdio
lib/libLLVMVectorize.a(LoopVectorize.cpp.o): In function `llvm::VPRecipeBuilder::tryToCreateWidenRecipe(llvm::Instruction*, llvm::ArrayRef<llvm::VPValue*>, llvm::VFRange&, std::unique_ptr<llvm::VPlan, std::default_delete<llvm::VPlan> >&) [clone .localalias.8]':
LoopVectorize.cpp:(.text._ZN4llvm15VPRecipeBuilder22tryToCreateWidenRecipeEPNS_11InstructionENS_8ArrayRefIPNS_7VPValueEEERNS_7VFRangeERSt10unique_ptrINS_5VPlanESt14default_deleteISA_EE+0x63b): undefined reference to `vtable for llvm::VPReductionPHIRecipe'
collect2: error: ld returned 1 exit status
This patch is a first step towards splitting up VPWidenPHIRecipe into
separate recipes for the 3 distinct cases they model:
1. reduction phis,
2. first-order recurrence phis,
3. pointer induction phis.
This allows untangling the code generation and allows us to reduce the
reliance on LoopVectorizationCostModel during VPlan code generation.
Discussed/suggested in D100102, D100113, D104197.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D104989
Splits `getSmallestAndWidestTypes` into two functions, one of which now collects
a list of all element types found in the loop (`ElementTypesInLoop`). This ensures we do not
have to iterate over all instructions in the loop again in other places, such as in D102253
which disables scalable vectorization of a loop if any of the instructions use invalid types.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D105437
The function vectorizeChainsInBlock does not support scalable vector,
because function like canReuseExtract and isCommutative in the code
path assert with scalable vectors.
This patch avoids vectorizing blocks that have extract instructions with scalable
vector..
Differential Revision: https://reviews.llvm.org/D104809
This API is not compatible with opaque pointers, the method
accepting an explicit pointer element type should be used instead.
Thankfully there were few in-tree users. The BPF case still ends
up using the pointer element type for now and needs something like
D105407 to avoid doing so.
Same as other CreateLoad-style APIs, these need an explicit type
argument to support opaque pointers.
Differential Revision: https://reviews.llvm.org/D105395
The compiler should not ignore UndefValue when gathering the scalars,
otherwise the resulting code may be less defined than the original one.
Also, grouped scalars to insert them at first to reduce the analysis in
further passes.
Differential Revision: https://reviews.llvm.org/D105275
In lots of places we were calling setDebugLocFromInst and passing
in the same Builder member variable found in InnerLoopVectorizer.
I personally found this confusing so I've changed the interface
to take an Optional<IRBuilder<> *> and we can now pass in None
when we want to use the class member variable.
Differential Revision: https://reviews.llvm.org/D105100
If we unroll a loop in the vectorizer (without vectorizing), and the cost model requires a epilogue be generated for correctness, the code generation must actually do so.
The included test case on an unmodified opt will access memory one past the expected bound. As a result, this patch is fixing a latent miscompile.
Differential Revision: https://reviews.llvm.org/D103700
This patch fixes a crash when the target instruction for sinking is
dead. In that case, no recipe is created and trying to get the recipe
for it results in a crash. To ensure all sink targets are alive, find &
use the first previous alive instruction.
Note that the case where the sink source is dead is already handled.
Found by
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=35320
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D104603
Previously in setCostBasedWideningDecision if we encountered an
invariant store we just assumed that we could scalarize the store
and called getUniformMemOpCost to get the associated cost.
However, for scalable vectors this is not an option because it is
not currently possibly to scalarize the store. At the moment we
crash in VPReplicateRecipe::execute when trying to scalarize the
store.
Therefore, I have changed setCostBasedWideningDecision so that if
we are storing a scalable vector out to a uniform address and the
target supports scatter instructions, then we should use those
instead.
Tests have been added here:
Transforms/LoopVectorize/AArch64/sve-inv-store.ll
Differential Revision: https://reviews.llvm.org/D104624
Currently we will allow loops with a fixed width VF of 1 to vectorize
if the -enable-strict-reductions flag is set. However, the loop vectorizer
will not use ordered reductions if `VF.isScalar()` and the resulting
vectorized loop will be out of order.
This patch removes `VF.isVector()` when checking if ordered reductions
should be used. Also, instead of converting the FAdds to reductions if the
VF = 1, operands of the FAdds are changed such that the order is preserved.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D104533
Sinking scalar operands into predicated-triangle regions may allow
merging regions. This patch adds a VPlan-to-VPlan transform that tries
to merge predicate-triangle regions after sinking.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100260
This patch updates VPWidenPHI recipes for first-order recurrences to
also track the incoming value from the back-edge. Similar to D99294,
which did the same for reductions.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D104197
Make getPointersDiff() and sortPtrAccesses() compatible with opaque
pointers by explicitly passing in the element type instead of
determining it from the pointer element type.
The SLPVectorizer result is slightly non-optimal in that unnecessary
pointer bitcasts are added.
Differential Revision: https://reviews.llvm.org/D104784
Perform better analysis when trying to vectorize PHIs.
1. Do not try to vectorize vector PHIs.
2. Do deeper analysis for more profitable nodes for the vectorization.
Before we just tried to vectorize the PHIs of the same type. Patch
improves this and tries to vectorize PHIs with incoming values which
come from the same basic block, have the same and/or alternative
opcodes.
It allows to save the compile time and provides better vectorization
results in general.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103638
This really isn't talking about vectors in general,
but only about either fixed or scalable vectors,
and it's pretty confusing to see it state
that there aren't any vectors :)
At the moment, we create insertelement instructions directly after
LastInst when inserting scalar values in a vector in
VPTransformState::get.
This results in invalid IR when LastInst is a phi, followed by another
phi. In that case, the new instructions should be inserted just after
the last PHI node in the block.
At the moment, I don't think the problematic case can be triggered, but
it can happen once predicate regions are merged and multiple
VPredInstPHI recipes are in the same block (D100260).
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D104188
This can be seen as a follow up to commit 0ee439b705,
that changed the second argument of __powidf2, __powisf2 and
__powitf2 in compiler-rt from si_int to int. That was to align with
how those runtimes are defined in libgcc.
One thing that seem to have been missing in that patch was to make
sure that the rest of LLVM also handle that the argument now depends
on the size of int (not using the si_int machine mode for 32-bit).
When using __builtin_powi for a target with 16-bit int clang crashed.
And when emitting libcalls to those rtlib functions, typically when
lowering @llvm.powi), the backend would always prepare the exponent
argument as an i32 which caused miscompiles when the rtlib was
compiled with 16-bit int.
The solution used here is to use an overloaded type for the second
argument in @llvm.powi. This way clang can use the "correct" type
when lowering __builtin_powi, and then later when emitting the libcall
it is assumed that the type used in @llvm.powi matches the rtlib
function.
One thing that needed some extra attention was that when vectorizing
calls several passes did not support that several arguments could
be overloaded in the intrinsics. This patch allows overload of a
scalar operand by adding hasVectorInstrinsicOverloadedScalarOpd, with
an entry for powi.
Differential Revision: https://reviews.llvm.org/D99439
As Eli mentioned post-commit in D103378, the result of the freeze may
still be out-of-range according to Alive2. So for now, just limit the
transform to indices that are non-poison.
It was found by chance revealing discrepancy between comment (few lines above),
the condition and how re-ordering of instruction is done inside the if statement
it guards. The condition was always evaluated to true.
Differential Revision: https://reviews.llvm.org/D104064
We were passing the RecurrenceDescriptor by value to most of the reduction analysis methods, despite it being rather bulky with TrackingVH members (that can be costly to copy). In all these cases we're only using the RecurrenceDescriptor for rather basic purposes (access to types/kinds etc.).
Differential Revision: https://reviews.llvm.org/D104029
This fixes the concern in single element store scalarization that the
alignment of new store may be larger than it should be. It calculates
the largest alignment if index is constant, and a safe one if not.
Reviewed By: lebedev.ri, spatel
Differential Revision: https://reviews.llvm.org/D103419
First we refactor the code which does no wrapping add sequences
match: we need to allow different operand orders for
the key add instructions involved in the match.
Then we use the refactored code trying 4 variants of matching operands.
Originally the code relied on the fact that the matching operands
of the two last add instructions of memory index calculations
had the same LHS argument. But which operand is the same
in the two instructions is actually not essential, so now we allow
that to be any of LHS or RHS of each of the two instructions.
This increases the chances of vectorization to happen.
Reviewed By: volkan
Differential Revision: https://reviews.llvm.org/D103912
As noted in https://bugs.llvm.org/show_bug.cgi?id=46666, the current behavior of assuming if-conversion safety if a loop is annotated parallel (`!llvm.loop.parallel_accesses`), is not expectable, the documentation for this behavior was since removed from the LangRef again, and can lead to invalid reads.
This was observed in POCL (https://github.com/pocl/pocl/issues/757) and would require similar workarounds in current work at hipSYCL.
The question remains why this was initially added and what the implications of removing this optimization would be.
Do we need an alternative mechanism to propagate the information about legality of if-conversion?
Or is the idea that conditional loads in `#pragma clang loop vectorize(assume_safety)` can be executed unmasked without additional checks flawed in general?
I think this implication is not part of what a user of that pragma (and corresponding metadata) would expect and thus dangerous.
Only two additional tests failed, which are adapted in this patch. Depending on the further direction force-ifcvt.ll should be removed or further adapted.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D103907
There is no need to schedule insertelement instructions. The compiler
did not schedule them before it started support their vectorization and
it should not do it after. We pre-schedule them manually when finding
a build vector sequence.
Disabling scheduling of insertelement instructions improves compile
time and vectorization of the very large basic blocks by saving
scheduling budget for other instructions.
Differential Revision: https://reviews.llvm.org/D104026
```
llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:8024:19: warning: loop variable 'VF' of type 'const llvm::ElementCount' creates a copy from type 'const llvm::ElementCount' [-Wrange-loop-analysis]
for (const auto VF : VFCandidates) {
^
llvm-project/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp:8024:8: note: use reference type 'const llvm::ElementCount &' to prevent copying
for (const auto VF : VFCandidates) {
^~~~~~~~~~~~~~~
&
1 warning generated.
```
Differential Revision: https://reviews.llvm.org/D103970
1. Better sorting of scalars to be gathered. Trying to insert
constants/arguments/instructions-out-of-loop at first and only then
the instructions which are inside the loop. It improves hoisting of
invariant insertelements instructions.
2. Better detection of shuffle candidates in gathering function.
3. The cost of insertelement for constants is 0.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103458
If the `-enable-strict-reductions` flag is set to true, then currently we will
always choose to vectorize the loop with strict in-order reductions. This is
not necessary where we allow the reordering of FP operations, such as
when loop hints are passed via metadata.
This patch moves useOrderedReductions so that we can also check whether
loop hints allow reordering, in which case we should use the default
behaviour of vectorizing with unordered reductions.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D103814
The non-DOT printing does not include the successors of VPregionBlocks.
This patch use the same style for printing successors as for
VPBasicBlock.
I think the printing of successors could be a bit improved further, as
at the moment it is hard to ensure a check line matches all successors.
But that can be done as follow-up.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D103515
This patch marks the induction increment of the main induction variable
of the vector loop as NUW when not folding the tail.
If the tail is not folded, we know that End - Start >= Step (either
statically or through the minimum iteration checks). We also know that both
Start % Step == 0 and End % Step == 0. We exit the vector loop if %IV +
%Step == %End. Hence we must exit the loop before %IV + %Step unsigned
overflows and we can mark the induction increment as NUW.
This should make SCEV return more precise bounds for the created vector
loops, used by later optimizations, like late unrolling.
At the moment quite a few tests still need to be updated, but before
doing so I'd like to get initial feedback to make sure I am not missing
anything.
Note that this could probably be further improved by using information
from the original IV.
Attempt of modeling of the assumption in Alive2:
https://alive2.llvm.org/ce/z/H_DL_g
Part of a set of fixes required for PR50412.
Reviewed By: mkazantsev
Differential Revision: https://reviews.llvm.org/D103255
No need to recalculate the cost of extractelements, just no need to
compensate the cost of all extractelements, need to check before if this
is actually going to be removed at the vectorization. Also, no need to
generate new extractelement instruction, we may just regenerate the
original one. It may improve the final vectorization.
Differential Revision: https://reviews.llvm.org/D102933
tryToVectorizeList function allows to reorder only 2 scalars. Patch
allows to reorder >2 scalars. Also, to avoid possible regressions, it
allows extra vectorization of the remaining parts of the scalars
elements if possible.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D103247
As noticed by NAKAMURA Takumi back in 2017, we cannot use
properlyDominates for std::stable_sort as properlyDominates only
partially orders blocks. That is, for blocks A, B, C, D, where A
dominates B and C dominates D, we have A == C, B == C, but A < B. This
is not a valid comparison function for std::stable_sort and causes
different results between libstdc++ and libc++. This change uses DFS
numbering to give deterministic results for all reachable blocks.
Unreachable blocks are ignored already, so do not need special
consideration.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103441
This patch uses the calculated maximum scalable VFs to build VPlans,
cost them and select a suitable scalable VF.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D98722
llvm::getLoadStoreType was added recently and has the same implementation
as 'getMemInstValueType' in LoopVectorize.cpp. Since there is no
value in having two implementations, this patch removes the custom LV
implementation in favor of the generic one defined in Instructions.h.
As the existing test unreachable.ll shows, we should be doing more
work to avoid entering unreachable blocks: we should not stop
vectorization just because a PHI incoming value from an unreachable
block cannot be vectorized. We know that particular value will never
be used so we can just replace it with poison.
Implemented better scheme for perfect/shuffled matches of the gather
nodes which allows to fix the performance regressions introduced by
earlier patches. Starting detecting matches for broadcast nodes and
extractelement gathering.
Differential Revision: https://reviews.llvm.org/D102920
If the index itself is already poison, the poison propagates through
instructions clamping the index to a valid range. This still causes
introducing a load of poison, as flagged by Alive2 and pointed out
at 575e2aff55.
This patch updates the code to freeze the index, unless it is proven to
not be poison.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D103378
Update isFirstOrderRecurrence to explore all uses of a recurrence phi
and check if we can sink them. If there are multiple users to sink, they
are all mapped to the previous instruction.
Fixes PR44286 (and another PR or two).
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D84951
For uniform ReplicateRecipes, only the first lane should be used, so
sinking them would mean we have to compute the value of the first lane
multiple times. Also, at the moment, sinking them causes a crash because
the value of the first lane is re-used by all users.
Reported post-commit for D100258.
SLP vectorizer should not consider in sertelements with multiple uses as
a part of high level build vector, it must be considered as
a terminating insertelement in the vector build, otherwise it may
produce incorrect code.
Differential Revision: https://reviews.llvm.org/D103164
When loop hints are passed via metadata, the allowReordering function
in LoopVectorizationLegality will allow the order of floating point
operations to be changed:
bool allowReordering() const {
// When enabling loop hints are provided we allow the vectorizer to change
// the order of operations that is given by the scalar loop. This is not
// enabled by default because can be unsafe or inefficient.
The -enable-strict-reductions flag introduced in D98435 will currently only
vectorize reductions in-loop if hints are used, since canVectorizeFPMath()
will return false if reordering is not allowed.
This patch changes canVectorizeFPMath() to query whether it is safe to
vectorize the loop with ordered reductions if no hints are used. For
testing purposes, an additional flag (-hints-allow-reordering) has been
added to disable the reordering behaviour described above.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D101836
We can only scalarize memory accesses if we know the index is valid.
This patch adjusts canScalarizeAcceess to fall back to
computeConstantRange to check if the index is known to be valid.
Reviewed By: nlopes
Differential Revision: https://reviews.llvm.org/D102476
This patch adds a first VPlan-based implementation of sinking of scalar
operands.
The current version traverse a VPlan once and processes all operands of
a predicated REPLICATE recipe. If one of those operands can be sunk,
it is moved to the block containing the predicated REPLICATE recipe.
Continue with processing the operands of the sunk recipe.
The initial version does not re-process candidates after other recipes
have been sunk. It also cannot partially sink induction increments at
the moment. The VPlan only contains WIDEN-INDUCTION recipes and if the
induction is used for example in a GEP, only the first lane is used and
in the lowered IR the adds for the other lanes can be sunk into the
predicated blocks.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100258
This reverts commit 94d54155e2.
This fixes a sanitizer failure by moving scalarizeLoadExtract(I)
before foldSingleElementStore(I), which may remove instructions.
This patch adds a new combine that tries to scalarize chains of
`extractelement (load %ptr), %idx` to `load (gep %ptr, %idx)`. This is
profitable when extracting only a few elements out of a large vector.
At the moment, `store (extractelement (load %ptr), %idx), %ptr`
operations on large vectors result in huge code in the backend.
This can easily be triggered by using the matrix extension, e.g.
https://clang.godbolt.org/z/qsccPdPf4
This should complement D98240.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D100273
External insertelement users can be represented as a result of shuffle
of the vectorized element and noconsecutive insertlements too. Added
support for handling non-consecutive insertelements.
Differential Revision: https://reviews.llvm.org/D101555
If we gather extract elements and they actually are just shuffles, it
might be profitable to vectorize them even if the tree is tiny.
Differential Revision: https://reviews.llvm.org/D101460
In InnerLoopVectorizer::setDebugLocFromInst we were previously
asserting that the VF is not scalable. This is because we want to
use the number of elements to create a duplication factor for the
debug profiling data. However, for scalable vectors we only know the
minimum number of elements. I've simply removed the assert for now
and added a FIXME saying that we assume vscale is always 1. When
vscale is not 1 it just means that the profiling data isn't as
accurate, but shouldn't cause any functional problems.
This patch adds a new option to the LoopVectorizer to control how
scalable vectors can be used.
Initially, this suggests three levels to control scalable
vectorization, although other more aggressive options can be added in
the future.
The possible options are:
- Disabled: Disables vectorization with scalable vectors.
- Enabled: Vectorize loops using scalable vectors or fixed-width
vectors, but favors fixed-width vectors when the cost
is a tie.
- Preferred: Like 'Enabled', but favoring scalable vectors when the
cost-model is inconclusive.
Reviewed By: paulwalker-arm, vkmr
Differential Revision: https://reviews.llvm.org/D101945
This patch implements first part of Flow Sensitive SampleFDO (FSAFDO).
It has the following changes:
(1) disable current discriminator encoding scheme,
(2) new hierarchical discriminator for FSAFDO.
For this patch, option "-enable-fs-discriminator=true" turns on the new
functionality. Option "-enable-fs-discriminator=false" (the default)
keeps the current SampleFDO behavior. When the fs-discriminator is
enabled, we insert a flag variable, namely, llvm_fs_discriminator, to
the object. This symbol will checked by create_llvm_prof tool, and used
to generate a profile with FS-AFDO discriminators enabled. If this
happens, for an extbinary format profile, create_llvm_prof tool
will add a flag to profile summary section.
Differential Revision: https://reviews.llvm.org/D102246
Currently all AA analyses marked as preserved are stateless, not taking
into account their dependent analyses. So there's no need to mark them
as preserved, they won't be invalidated unless their analyses are.
SCEVAAResults was the one exception to this, it was treated like a
typical analysis result. Make it like the others and don't invalidate
unless SCEV is invalidated.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D102032
This allows cast/dyn_cast'ing from VPUser to recipes. This is needed
because there are VPUsers that are not recipes.
Reviewed By: gilr, a.elovikov
Differential Revision: https://reviews.llvm.org/D100257
This patch introduces a new class, MaxVFCandidates, that holds the
maximum vectorization factors that have been computed for both scalable
and fixed-width vectors.
This patch is intended to be NFC for fixed-width vectors, although
considering a scalable max VF (which is disabled by default) pessimises
tail-loop elimination, since it can no longer determine if any chosen VF
(less than fixed/scalable MaxVFs) is guaranteed to handle all vector
iterations if the trip-count is known. This issue will be addressed in
a future patch.
Reviewed By: fhahn, david-arm
Differential Revision: https://reviews.llvm.org/D98721
This reverts commit 6d3e3ae8a9.
Still seeing PPC build bot failures, and one arm self host bot failing. I'm officially stumped, and need help from a bot owner to reduce.
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
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
As discussed in D102437, the VF argument to isScalarWithPredication
seems redundant, so this is intended to be a non-functional change. It
seems wrong to query the widening decision at this point. Removing the
operand and code to get the widening decision causes no unit/regression
tests to fail. I've also found no issues running the LLVM test-suite.
This subsequently removes the VF argument from isPredicatedInst as well,
since it is no longer required.
Add new type of tree node for `InsertElementInst` chain forming vector.
These instructions could be either removed, or replaced by shuffles during
vectorization and we can add this node to cost model, so naturally estimating
their cost, getting rid of `CompensateCost` tricks and reducing further work
for InstCombine. This fixes PR40522 and PR35732 in a natural way. Also this
patch is the first step towards revectorization of partially vectorization
(to fix PR42022 completely). After adding inserts to tree the next step is
to add vector instructions there (for instance, to merge `store <2 x float>`
and `store <2 x float>` to `store <4 x float>`).
Fixes PR40522 and PR35732.
Differential Revision: https://reviews.llvm.org/D98714
This change enables cases for which the index value for the first
load/store instruction in a pair could be a function argument. This
allows using llvm.assume to provide known bits information in such
cases.
Patch by Viacheslav Nikolaev. Thanks!
Differential Revision: https://reviews.llvm.org/D101680
In InnerLoopVectorizer::widenPHIInstruction there are cases where we have
to scalarise a pointer induction variable after vectorisation. For scalable
vectors we already deal with the case where the pointer induction variable
is uniform, but we currently crash if not uniform. For fixed width vectors
we calculate every lane of the scalarised pointer induction variable for a
given VF, however this cannot work for scalable vectors. In this case I
have added support for caching the whole vector value for each unrolled
part so that we can always extract an arbitrary element. Additionally, we
still continue to cache the known minimum number of lanes too in order
to improve code quality by avoiding an extractelement operation.
I have adapted an existing test `pointer_iv_mixed` from the file:
Transforms/LoopVectorize/consecutive-ptr-uniforms.ll
and added it here for scalable vectors instead:
Transforms/LoopVectorize/AArch64/sve-widen-phi.ll
Differential Revision: https://reviews.llvm.org/D101294
Vector single element update optimization is landed in 2db4979. But the
scope needs restriction. This patch restricts the index to inbounds and
vector must be fixed sized. In future, we may use value tracking to
relax constant restrictions.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D102146
If the simplified VPValue is a recipe, we need to register it for Instr,
in case it needs to be recorded. The way this is handled in general may
change soon, following some post-commit comments.
This fixes PR50298.
The test example from https://llvm.org/PR50256 (and reduced here)
shows that we can match a load combine candidate even when there
are no "or" instructions. We can avoid that by confirming that we
do see an "or". This doesn't apply when matching an or-reduction
because that match begins from the operands of the reduction.
Differential Revision: https://reviews.llvm.org/D102074
Need to remove the old code for avoiding double counting of the gather
nodes with perfect diamond matches within the tree after we started
detecting perfect/shuffled matching in the previous patch D100495. We
may skip the cost for such nodes completely.
Differential Revision: https://reviews.llvm.org/D102023
The comment incorrectly states that the PHI is recorded. That's not
accurate, only the recipe for the incoming value is recorded.
Suggested post-commit for 4ba8720f88.
Currently sinking a replicate region into another replicate region is
not supported. Add an assert, to make the problem more obvious, should
it occur.
Discussed post-commit for ccebf7a109.
The function fixReduction used to assert/crash for scalable vector when
a vector reduce could be done with a smaller vector.
This patch removes this assertion as it is safe to use scalable vector for
vector reduce and truncate.
Differential Revision: https://reviews.llvm.org/D101260
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
Adds support for scalable vectorization of loops containing first-order recurrences, e.g:
```
for(int i = 0; i < n; i++)
b[i] = a[i] + a[i - 1]
```
This patch changes fixFirstOrderRecurrence for scalable vectors to take vscale into
account when inserting into and extracting from the last lane of a vector.
CreateVectorSplice has been added to construct a vector for the recurrence, which
returns a splice intrinsic for scalable types. For fixed-width the behaviour
remains unchanged as CreateVectorSplice will return a shufflevector instead.
The tests included here are the same as test/Transform/LoopVectorize/first-order-recurrence.ll
Reviewed By: david-arm, fhahn
Differential Revision: https://reviews.llvm.org/D101076
LoopVectorize has a fairly deeply baked in design problem where it will try to query analysis (primarily SCEV, but also ValueTracking) in the midst of mutating IR. In particular, the intermediate IR state does not represent the semantics of the original (or final) program.
Fixing this for real is hard, but all of the cases seen so far share a common symptom. In cases seen to date, the analysis being queried is the computation of the original loop's trip count. We can fix this particular instance of the issue by simply computing the trip count early, and caching it.
I want to be really clear that this is nothing but a workaround. It does nothing to fix the root issue, and at best, delays the time until we have to fix this for real. Florian and I have discussed an eventual solution in the review comments for https://reviews.llvm.org/D100663, but it's a lot of work.
Test taken from https://reviews.llvm.org/D100663.
Differential Revision: https://reviews.llvm.org/D101487
This patch updates the code that sinks recipes required for first-order
recurrences to properly handle replicate-regions. At the moment, the
code would just move the replicate recipe out of its replicate-region,
producing an invalid VPlan.
When sinking a recipe in a replicate-region, we have to sink the whole
region. To do that, we first need to split the block at the target
recipe and move the region in between.
This patch also adds a splitAt helper to VPBasicBlock to split a
VPBasicBlock at a given iterator.
Fixes PR50009.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100751
This patch updates the code handling reduction recipes to also keep
track of the incoming value from the latch in the recipe. This is needed
to model the def-use chains completely in VPlan, so that it is possible
to replace the incoming value with an arbitrary VPValue.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D99294
Need to check if target allows/supports masked gathers before trying to
estimate its cost, otherwise we may fail to vectorize some of the
patterns because of too pessimistic cost model.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D101297
Need to check if target allows/supports masked gathers before trying to
estimate its cost, otherwise we may fail to vectorize some of the
patterns because of too pessimistic cost model.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D101297
As we gradually move more elements of LV to VPlan, we are trying to
reduce the number of places that still has to check IR of the original
loop.
This patch adjusts the code to fix cross iteration phis to get the PHIs
to fix directly from the VPlan that is executed. We still need the
original PHI to check for first-order recurrences, but we can get rid of
that once we model that explicitly in VPlan as well.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D99293
This patch introduces a helper to obtain an iterator range for the
PHI-like recipes in a block.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D100101
If the extracts from the non-power-2 vectors are recognized as shuffles,
need some extra checks to not crash cost calculations if trying to gext
the ecost for subvector extracts. In this case need to check carefully
that we do not exit out of bounds of the original vector, otherwise the
TTI's cost model will crash on assert.
Differential Revision: https://reviews.llvm.org/D101477
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
As suggested in D99294, this adds a getVPSingleValue helper to use for
recipes that are guaranteed to define a single value. This replaces uses
of getVPValue() which used to default to I = 0.
This patch causes the loop vectorizer to not interleave loops that have
nounroll loop hints (llvm.loop.unroll.disable and llvm.loop.unroll_count(1)).
Note that if a particular interleave count is being requested
(through llvm.loop.interleave_count), it will still be honoured, regardless
of the presence of nounroll hints.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D101374
This patch fixes a crash encountered when vectorising the following loop:
void foo(float *dst, float *src, long long n) {
for (long long i = 0; i < n; i++)
dst[i] = -src[i];
}
using scalable vectors. I've added a test to
Transforms/LoopVectorize/AArch64/sve-basic-vec.ll
as well as cleaned up the other tests in the same file.
Differential Revision: https://reviews.llvm.org/D98054
If the first tree element is vectorize and the second is gather, it
still might be profitable to vectorize it if the gather node contains
less scalars to vectorize than the original tree node. It might be
profitable to use shuffles.
Differential Revision: https://reviews.llvm.org/D101397
This patch simplifies the calculation of certain costs in
getInstructionCost when isScalarAfterVectorization() returns a true value.
There are a few places where we multiply a cost by a number N, i.e.
unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1;
return N * TTI.getArithmeticInstrCost(...
After some investigation it seems that there are only these cases that occur
in practice:
1. VF is a scalar, in which case N = 1.
2. VF is a vector. We can only get here if: a) the instruction is a
GEP/bitcast/PHI with scalar uses, or b) this is an update to an induction
variable that remains scalar.
I have changed the code so that N is assumed to always be 1. For GEPs
the cost is always 0, since this is calculated later on as part of the
load/store cost. PHI nodes are costed separately and were never previously
multiplied by VF. For all other cases I have added an assert that none of
the users needs scalarising, which didn't fire in any unit tests.
Only one test required fixing and I believe the original cost for the scalar
add instruction to have been wrong, since only one copy remains after
vectorisation.
I have also added a new test for the case when a pointer PHI feeds directly
into a store that will be scalarised as we were previously never testing it.
Differential Revision: https://reviews.llvm.org/D99718
This patch also refactors the way the feasible max VF is calculated,
although this is NFC for fixed-width vectors.
After this change scalable VF hints are no longer truncated/clamped
to a shorter scalable VF, nor does it drop the 'scalable flag' from
the suggested VF to vectorize with a similar VF that is fixed.
Instead, the hint is ignored which means the vectorizer is free
to find a more suitable VF, using the CostModel to determine the
best possible VF.
Reviewed By: c-rhodes, fhahn
Differential Revision: https://reviews.llvm.org/D98509
When using the -enable-strict-reductions flag where UF>1 we generate multiple
Phi nodes, though only one of these is used as an input to the vector.reduce.fadd
intrinsics. The unused Phi nodes are removed later by instcombine.
This patch changes widenPHIInstruction/fixReduction to only generate
one Phi, and adds an additional test for unrolling to strict-fadd.ll
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D100570
This patch simplifies the calculation of certain costs in
getInstructionCost when isScalarAfterVectorization() returns a true value.
There are a few places where we multiply a cost by a number N, i.e.
unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1;
return N * TTI.getArithmeticInstrCost(...
After some investigation it seems that there are only these cases that occur
in practice:
1. VF is a scalar, in which case N = 1.
2. VF is a vector. We can only get here if: a) the instruction is a
GEP/bitcast/PHI with scalar uses, or b) this is an update to an induction
variable that remains scalar.
I have changed the code so that N is assumed to always be 1. For GEPs
the cost is always 0, since this is calculated later on as part of the
load/store cost. PHI nodes are costed separately and were never previously
multiplied by VF. For all other cases I have added an assert that none of
the users needs scalarising, which didn't fire in any unit tests.
Only one test required fixing and I believe the original cost for the scalar
add instruction to have been wrong, since only one copy remains after
vectorisation.
I have also added a new test for the case when a pointer PHI feeds directly
into a store that will be scalarised as we were previously never testing it.
Differential Revision: https://reviews.llvm.org/D99718
This patch simplifies VPSlotTracker by using the recursive traversal
iterator to traverse all blocks in a VPlan in reverse post-order when
numbering VPValues in a plan.
This depends on a fix to RPOT (D100169). It also extends the traversal
unit tests to check RPOT.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D100176
When iterating over const blocks, the base type in the lambdas needs
to use const VPBlockBase *, otherwise it cannot be used with input
iterators over const VPBlockBase.
Also adjust the type of the input iterator range to const &, as it
does not take ownership of the input range.
This patch adds a blocksOnly helpers which take an iterator range
over VPBlockBase * or const VPBlockBase * and returns an interator
range that only include BlockTy blocks. The accesses are casted to
BlockTy.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D101093
This patch adds a new iterator to traverse through VPRegionBlocks and a
GraphTraits specialization using the iterator to traverse through
VPRegionBlocks.
Because there is already a GraphTraits specialization for VPBlockBase *
and co, a new VPBlockRecursiveTraversalWrapper helper is introduced.
This allows us to provide a new GraphTraits specialization for that
type. Users can use the new recursive traversal by using this wrapper.
The graph trait visits both the entry block of a region, as well as all
its successors. Exit blocks of a region implicitly have their parent
region's successors. This ensures all blocks in a region are visited
before any blocks in a successor region when doing a reverse post-order
traversal of the graph.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D100175
We can skip check for undefs trying to find perfect/shuffled tree
entries matching, they can be ignored completely improving the final
cost/vectorization results.
Differential Revision: https://reviews.llvm.org/D101061
This commit fixes a bug where the loop vectoriser fails to predicate
loads/stores when interleaving for targets that support masked
loads and stores.
Code such as:
1 void foo(int *restrict data1, int *restrict data2)
2 {
3 int counter = 1024;
4 while (counter--)
5 if (data1[counter] > data2[counter])
6 data1[counter] = data2[counter];
7 }
... could previously be transformed in such a way that the predicated
store implied by:
if (data1[counter] > data2[counter])
data1[counter] = data2[counter];
... was lost, resulting in miscompiles.
This bug was causing some tests in llvm-test-suite to fail when built
for SVE.
Differential Revision: https://reviews.llvm.org/D99569
1. No need to call `areAllUsersVectorized` as later the cost is
calculated only if the instruction has one use and gets vectorized.
2. Need to calculate the cost of the dead extractelement more precisely,
taking the vector type of the vector operand, not the resulting
vector type.
Part of D57059.
Differential Revision: https://reviews.llvm.org/D99980
In quite a few cases in LoopVectorize.cpp we call createStepForVF
with a step value of 0, which leads to unnecessary generation of
llvm.vscale intrinsic calls. I've optimised IRBuilder::CreateVScale
and createStepForVF to return 0 when attempting to multiply
vscale by 0.
Differential Revision: https://reviews.llvm.org/D100763
SLP supports perfect diamond matching for the vectorized tree entries
but do not support it for gathered entries and does not support
non-perfect (shuffled) matching with 1 or 2 tree entries. Patch adds
support for this matching to improve cost of the vectorized tree.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D100495
SLP supports perfect diamond matching for the vectorized tree entries
but do not support it for gathered entries and does not support
non-perfect (shuffled) matching with 1 or 2 tree entries. Patch adds
support for this matching to improve cost of the vectorized tree.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D100495
SLP supports perfect diamond matching for the vectorized tree entries
but do not support it for gathered entries and does not support
non-perfect (shuffled) matching with 1 or 2 tree entries. Patch adds
support for this matching to improve cost of the vectorized tree.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D100495
Rather than maintaining two separate values, a `float` for the per-lane
cost and a Width for the VF, maintain a single VectorizationFactor which
comprises the two and also removes the need for converting an integer value
to float.
This simplifies the query when asking if one VF is more profitable than
another when we want to extend this for scalable vectors (which may
require additional options to determine if e.g. a scalable VF of the
some cost, is more profitable than a fixed VF of the same cost).
The patch isn't entirely NFC because it also fixes an issue in
selectEpilogueVectorizationFactor, where the cost passed to ProfitableVFs
no longer truncates the floating-point cost from `float` to `unsigned` to
then perform the calculation on the truncated cost. It now does
a cost comparison with the correct precision.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D100121
SLP supports perfect diamond matching for the vectorized tree entries
but do not support it for gathered entries and does not support
non-perfect (shuffled) matching with 1 or 2 tree entries. Patch adds
support for this matching to improve cost of the vectorized tree.
Differential Revision: https://reviews.llvm.org/D100495
Add an initial version of a helper to determine whether a recipe may
have side-effects.
Reviewed By: a.elovikov
Differential Revision: https://reviews.llvm.org/D100259
There were a few places in widenPHIInstruction where calculations of
offsets were failing to take the runtime calculation of VF into
account for scalable vectors. I've fixed those cases in this patch
as well as adding an assert that we should not be scalarising for
scalable vectors.
Tests are added here:
Transforms/LoopVectorize/AArch64/sve-widen-phi.ll
Differential Revision: https://reviews.llvm.org/D99254
There are a few places in LoopVectorize.cpp where we have been too
cautious in adding VF.isScalable() asserts and it can be confusing.
It also makes it more difficult to see the genuine places where
work needs doing to improve scalable vectorization support.
This patch changes getMemInstScalarizationCost to return an
invalid cost instead of firing an assert for scalable vectors. Also,
vectorizeInterleaveGroup had multiple asserts all for the same
thing. I have removed all but one assert near the start of the
function, and added a new assert that we aren't dealing with masks
for scalable vectors.
Differential Revision: https://reviews.llvm.org/D99727
Only attempt to propagateIRFlags if we have both SelectInst - afaict we shouldn't have matched a min/max reduction without both SelectInst, but static analyzer doesn't know that.
Main reason is preparation to transform AliasResult to class that contains
offset for PartialAlias case.
Reviewed By: asbirlea
Differential Revision: https://reviews.llvm.org/D98027
Instead of passing the start value and the defined value to
widenPHIInstruction, pass the VPWidenPHIRecipe directly, which can be
used to get both (and more in future patches).
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
No need to lookup through and/or try to vectorize operands of the
CmpInst instructions during attempts to find/vectorize min/max
reductions. Compiler implements postanalysis of the CmpInsts so we can
skip extra attempts in tryToVectorizeHorReductionOrInstOperands and save
compile time.
Differential Revision: https://reviews.llvm.org/D99950
Add the subclass, update a few places which check for the intrinsic to use idiomatic dyn_cast, and update the public interface of AssumptionCache to use the new class. A follow up change will do the same for the newer assumption query/bundle mechanisms.
Previously we could only vectorize FP reductions if fast math was enabled, as this allows us to
reorder FP operations. However, it may still be beneficial to vectorize the loop by moving
the reduction inside the vectorized loop and making sure that the scalar reduction value
be an input to the horizontal reduction, e.g:
%phi = phi float [ 0.0, %entry ], [ %reduction, %vector_body ]
%load = load <8 x float>
%reduction = call float @llvm.vector.reduce.fadd.v8f32(float %phi, <8 x float> %load)
This patch adds a new flag (IsOrdered) to RecurrenceDescriptor and makes use of the changes added
by D75069 as much as possible, which already teaches the vectorizer about in-loop reductions.
For now in-order reduction support is off by default and controlled with the `-enable-strict-reductions` flag.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D98435
Changes getRecurrenceIdentity to always return a neutral value of -0.0 for FAdd.
Reviewed By: dmgreen, spatel
Differential Revision: https://reviews.llvm.org/D98963
For VPWidenPHIRecipes that model all incoming values as VPValue
operands, print those operands instead of printing the original PHI.
D99294 updates recipes of reduction PHIs to use the VPValue for the
incoming value from the loop backedge, making use of this new printing.
During vectorization better to postpone the vectorization of the CmpInst
instructions till the end of the basic block. Otherwise we may vectorize
it too early and may miss some vectorization patterns, like reductions.
Reworked part of D57059
Differential Revision: https://reviews.llvm.org/D99796
This patch moves mapping of IR operands to VPValues out of
tryToCreateWidenRecipe. This allows using existing VPValue operands when
widening recipes directly, which will be introduced in future patches.
The ultimate reduction node may have multiple uses, but if the ultimate
reduction is min/max reduction and based on SelectInstruction, the
condition of this select instruction must have only single use.
Differential Revision: https://reviews.llvm.org/D99753
The motivation for this patch is to better estimate the cost of
extracelement instructions in cases were they are going to be free,
because the source vector can be used directly.
A simple example is
%v1.lane.0 = extractelement <2 x double> %v.1, i32 0
%v1.lane.1 = extractelement <2 x double> %v.1, i32 1
%a.lane.0 = fmul double %v1.lane.0, %x
%a.lane.1 = fmul double %v1.lane.1, %y
Currently we only consider the extracts free, if there are no other
users.
In this particular case, on AArch64 which can fit <2 x double> in a
vector register, the extracts should be free, independently of other
users, because the source vector of the extracts will be in a vector
register directly, so it should be free to use the vector directly.
The SLP vectorized version of noop_extracts_9_lanes is 30%-50% faster on
certain AArch64 CPUs.
It looks like this does not impact any code in
SPEC2000/SPEC2006/MultiSource both on X86 and AArch64 with -O3 -flto.
This originally regressed after D80773, so if there's a better
alternative to explore, I'd be more than happy to do that.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D99719
1. Need to cleanup InstrElementSize map for each new tree, otherwise might
use sizes from the previous run of the vectorization attempt.
2. No need to include into analysis the instructions from the different basic
blocks to save compile time.
Differential Revision: https://reviews.llvm.org/D99677
Use SetVector instead of SmallPtrSet to track values with uniform use. Doing this
can help avoid non-determinism caused by iterating over unordered containers.
This bug was found with reverse iteration turning on,
--extra-llvm-cmake-variables="-DLLVM_REVERSE_ITERATION=ON".
Failing LLVM test consecutive-ptr-uniforms.ll .
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D99549
This marks FSIN and other operations to EXPAND for scalable
vectors, so that they are not assumed to be legal by the cost-model.
Depends on D97470
Reviewed By: dmgreen, paulwalker-arm
Differential Revision: https://reviews.llvm.org/D97471
Use SetVector instead of SmallPtrSet for external definitions created for VPlan.
Doing this can help avoid non-determinism caused by iterating over unordered containers.
This bug was found with reverse iteration turning on,
--extra-llvm-cmake-variables="-DLLVM_REVERSE_ITERATION=ON".
Failing LLVM-Unit test VPRecipeTest.dump.
Reviewed By: MaskRay
Differential Revision: https://reviews.llvm.org/D99544
This patch adds support for the vectorization of induction variables when
using scalable vectors, which required the following changes:
1. Removed assert from InnerLoopVectorizer::getStepVector.
2. Modified InnerLoopVectorizer::createVectorIntOrFpInductionPHI to use
a runtime determined value for VF and removed an assert.
3. Modified InnerLoopVectorizer::buildScalarSteps to work for scalable
vectors. I did this by calculating the full vector value for each Part
of the unroll factor (UF) and caching this in the VP state. This means
that we are always able to extract an arbitrary element from the vector
if necessary. In addition to this, I also permitted the caching of the
individual lane values themselves for the known minimum number of elements
in the same way we do for fixed width vectors. This is a further
optimisation that improves the code quality since it avoids unnecessary
extractelement operations when extracting the first lane.
4. Added an assert to InnerLoopVectorizer::widenPHIInstruction, since while
testing some code paths I noticed this is currently broken for scalable
vectors.
Various tests to support different cases have been added here:
Transforms/LoopVectorize/AArch64/sve-inductions.ll
Differential Revision: https://reviews.llvm.org/D98715
Re-apply 25fbe803d4, with a small update to emit the right remark
class.
Original message:
[LV] Move runtime pointer size check to LVP::plan().
This removes the need for the remaining doesNotMeet check and instead
directly checks if there are too many runtime checks for vectorization
in the planner.
A subsequent patch will adjust the logic used to decide whether to
vectorize with runtime to consider their cost more accurately.
Reviewed By: lebedev.ri
This is a 2nd try of:
3c8473ba53
which was reverted at:
a26312f9d4
because of crashing.
This version includes extra code and tests to avoid the known
crashing examples as discussed in PR49730.
Original commit message:
As noted in D98152, we need to patch SLP to avoid regressions when
we start canonicalizing to integer min/max intrinsics.
Most of the real work to make this possible was in:
7202f47508
Differential Revision: https://reviews.llvm.org/D98981
This removes the need for the remaining doesNotMeet check and instead
directly checks if there are too many runtime checks for vectorization
in the planner.
A subsequent patch will adjust the logic used to decide whether to
vectorize with runtime to consider their cost more accurately.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D98634
This reverts commit 3c8473ba53 and includes test diffs to
maintain testing status.
There's at least 1 place that was not updated with 7202f47508 ,
so we can crash mismatching select and intrinsics as shown in
PR49730.
This patch simplifies the calculation of certain costs in
getInstructionCost when isScalarAfterVectorization() returns a true value.
There are a few places where we multiply a cost by a number N, i.e.
unsigned N = isScalarAfterVectorization(I, VF) ? VF.getKnownMinValue() : 1;
return N * TTI.getArithmeticInstrCost(...
After some investigation it seems that there are only these cases that occur
in practice:
1. VF is a scalar, in which case N = 1.
2. VF is a vector. We can only get here if: a) the instruction is a
GEP/bitcast with scalar uses, or b) this is an update to an induction variable
that remains scalar.
I have changed the code so that N is assumed to always be 1. For GEPs
the cost is always 0, since this is calculated later on as part of the
load/store cost. For all other cases I have added an assert that none of the
users needs scalarising, which didn't fire in any unit tests.
Only one test required fixing and I believe the original cost for the scalar
add instruction to have been wrong, since only one copy remains after
vectorisation.
Differential Revision: https://reviews.llvm.org/D98512
The SCEV commit b46c085d2b [NFCI] SCEVExpander:
emit intrinsics for integral {u,s}{min,max} SCEV expressions
seems to reveal a new crash in SLPVectorizer.
SLP crashes expecting a SelectInst as an externally used value
but umin() call is found.
The patch relaxes the assumption to make the IR flag propagation safe.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D99328
We do not need to scan further if the upper end or lower end of the
basic block is reached already and the instruction is not found. It
means that the instruction is definitely in the lower part of basic
block or in the upper block relatively.
This should improve compile time for the very big basic blocks.
Differential Revision: https://reviews.llvm.org/D99266
This patch changes the interface to take a RegisterKind, to indicate
whether the register bitwidth of a scalar register, fixed-width vector
register, or scalable vector register must be returned.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D98874
We know if the loop contains FP instructions preventing vectorization
after we are done with legality checks. This patch updates the code the
check for un-vectorizable FP operations earlier, to avoid unnecessarily
running the cost model and picking a vectorization factor. It also makes
the code more direct and moves the check to a position where similar
checks are done.
I might be missing something, but I don't see any reason to handle this
check differently to other, similar checks.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D98633
Added getPointersDiff function to LoopAccessAnalysis and used it instead
direct calculatoin of the distance between pointers and/or
isConsecutiveAccess function in SLP vectorizer to improve compile time
and detection of stores consecutive chains.
Part of D57059
Differential Revision: https://reviews.llvm.org/D98967
Added getPointersDiff function to LoopAccessAnalysis and used it instead
direct calculatoin of the distance between pointers and/or
isConsecutiveAccess function in SLP vectorizer to improve compile time
and detection of stores consecutive chains.
Part of D57059
Differential Revision: https://reviews.llvm.org/D98967
As noted in D98152, we need to patch SLP to avoid regressions when
we start canonicalizing to integer min/max intrinsics.
Most of the real work to make this possible was in:
7202f47508
Differential Revision: https://reviews.llvm.org/D98981
In places where we create a ConstantVector whose elements are a
linear sequence of the form <start, start + 1, start + 2, ...>
I've changed the code to make use of CreateStepVector, which creates
a vector with the sequence <0, 1, 2, ...>, and a vector addition
operation. This patch is a non-functional change, since the output
from the vectoriser remains unchanged for fixed length vectors and
there are existing asserts that still fire when attempting to use
scalable vectors for vectorising induction variables.
In a later patch we will enable support for scalable vectors
in InnerLoopVectorizer::getStepVector(), which relies upon the new
stepvector intrinsic in IRBuilder::CreateStepVector.
Differential Revision: https://reviews.llvm.org/D97861
The name is included when printing in DOT mode. Also print it in non-DOT
mode after 93a9d2de8f.
This will become more important to distinguish different plans once
VPlans are gradually refined.
Make sure we use PowerOf2Floor instead of PowerOf2Ceil when
calculating max number of elements that fits inside a vector
register (otherwise we could end up creating vectors larger
than the maximum vector register size).
Also make sure we honor the min/max VF (as given by TTI or
cmd line parameters) when doing vectorizeStores.
Reviewed By: anton-afanasyev
Differential Revision: https://reviews.llvm.org/D97691
I foresee two uses for this:
1) It's easier to use those in debugger.
2) Once we start implementing more VPlan-to-VPlan transformations (especially
inner loop massaging stuff), using the vectorized LLVM IR as CHECK targets in
LIT test would become too obscure. I can imagine that we'd want to CHECK
against VPlan dumps after multiple transformations instead. That would be
easier with plain text dumps than with DOT format.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D96628
This reverts commit 6b053c9867.
The build is broken:
ld.lld: error: undefined symbol: llvm::VPlan::printDOT(llvm::raw_ostream&) const
>>> referenced by LoopVectorize.cpp
>>> LoopVectorize.cpp.o:(llvm::LoopVectorizationPlanner::printPlans(llvm::raw_ostream&)) in archive lib/libLLVMVectorize.a
I foresee two uses for this:
1) It's easier to use those in debugger.
2) Once we start implementing more VPlan-to-VPlan transformations (especially
inner loop massaging stuff), using the vectorized LLVM IR as CHECK targets in
LIT test would become too obscure. I can imagine that we'd want to CHECK
against VPlan dumps after multiple transformations instead. That would be
easier with plain text dumps than with DOT format.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D96628
If SLP vectorizer tries to extend the scheduling region and runs out of
the budget too early, but still extends the region to the new ending
instructions (i.e., it was able to extend the region for the first
instruction in the bundle, but not for the second), the compiler need to
recalculate dependecies in full, just like if the extending was
successfull. Without it, the schedule data chunks may end up with the
wrong number of (unscheduled) dependecies and it may end up with the
incorrect function, where the vectorized instruction does not dominate
on the extractelement instruction.
Differential Revision: https://reviews.llvm.org/D98531
This adds an Mask ArrayRef to getShuffleCost, so that if an exact mask
can be provided a more accurate cost can be provided by the backend.
For example VREV costs could be returned by the ARM backend. This should
be an NFC until then, laying the groundwork for that to be added.
Differential Revision: https://reviews.llvm.org/D98206
The `hasIrregularType` predicate checks whether an array of N values of type Ty is "bitcast-compatible" with a <N x Ty> vector.
The previous check returned invalid results in some cases where there's some padding between the array elements: eg. a 4-element array of u7 values is considered as compatible with <4 x u7>, even though the vector is only loading/storing 28 bits instead of 32.
The problem causes LLVM to generate incorrect code for some targets: for AArch64 the vector loads/stores are lowered in terms of ubfx/bfi, effectively losing the top (N * padding bits).
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D97465
This adds the cost of an i1 extract and a branch to the cost in
getMemInstScalarizationCost when the instruction is predicated. These
predicated loads/store would generate blocks of something like:
%c1 = extractelement <4 x i1> %C, i32 1
br i1 %c1, label %if, label %else
if:
%sa = extractelement <4 x i32> %a, i32 1
%sb = getelementptr inbounds float, float* %pg, i32 %sa
%sv = extractelement <4 x float> %x, i32 1
store float %sa, float* %sb, align 4
else:
So this increases the cost by the extract and branch. This is probably
still too low in many cases due to the cost of all that branching, but
there is already an existing hack increasing the cost using
useEmulatedMaskMemRefHack. It will increase the cost of a memop if it is
a load or there are more than one store. This patch improves the cost
for when there is only a single store, and hopefully at some point in
the future the hack can be removed.
Differential Revision: https://reviews.llvm.org/D98243
Current SLP pass has this piece of code that inserts a trunc instruction
after the vectorized instruction. In the case that the vectorized instruction
is a phi node and not the last phi node in the BB, the trunc instruction
will be inserted between two phi nodes, which will trigger verify problem
in debug version or unpredictable error in another pass.
This patch changes the algorithm to 'if the last vectorized instruction
is a phi, insert it after the last phi node in current BB' to fix this problem.
The motivation is to handle integer min/max reductions independently
of whether they are in the current cmp+sel form or the planned intrinsic
form.
We assumed that min/max included a select instruction, but we can
decouple that implementation detail by checking the instructions
themselves rather than relying on the recurrence (reduction) type.
Instead of maintaining a separate map from predicated instructions to
recipes, we can instead directly look at the VP operands. If the operand
comes from a predicated instruction, the operand will be a
VPPredInstPHIRecipe with a VPReplicateRecipe as its operand.
This patch adds support for reverse loop vectorization.
It is possible to vectorize the following loop:
```
for (int i = n-1; i >= 0; --i)
a[i] = b[i] + 1.0;
```
with fixed or scalable vector.
The loop-vectorizer will use 'reverse' on the loads/stores to make
sure the lanes themselves are also handled in the right order.
This patch adds support for scalable vector on IRBuilder interface to
create a reverse vector. The IR function
CreateVectorReverse lowers to experimental.vector.reverse for scalable vector
and keedp the original behavior for fixed vector using shuffle reverse.
Differential Revision: https://reviews.llvm.org/D95363
This patch fixes a crash when trying to get a scalar value using
VPTransformState::get() for uniform induction values or truncated
induction values. IVs and truncated IVs can be uniform and the updated
code accounts for that, fixing the crash.
This should fix
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=31981
Associative reduction matcher in SLP begins with select instruction but when
it reached call to llvm.umax (or alike) via def-use chain the latter also matched
as UMax kind. The routine's later code assumes matched instruction to be a select
and thus it merely died on the first encountered cast that did not fit.
Differential Revision: https://reviews.llvm.org/D98432
Add support to widen select instructions in VPlan native path by using a correct recipe when such instructions are encountered. This is already used by inner loop vectorizer.
Previously select instructions get handled by the wrong recipe and resulted in unreachable instruction errors like this one: https://bugs.llvm.org/show_bug.cgi?id=48139.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D97136
Add support to widen call instructions in VPlan native path by using a correct recipe when such instructions are encountered. This is already used by inner loop vectorizer.
Previously call instructions got handled by wrong recipes and resulted in unreachable instruction errors like this one: https://bugs.llvm.org/show_bug.cgi?id=48139.
Patch by Mauri Mustonen <mauri.mustonen@tuni.fi>
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D97278
There are certain loops like this below:
for (int i = 0; i < n; i++) {
a[i] = b[i] + 1;
*inv = a[i];
}
that can only be vectorised if we are able to extract the last lane of the
vectorised form of 'a[i]'. For fixed width vectors this already works since
we know at compile time what the final lane is, however for scalable vectors
this is a different story. This patch adds support for extracting the last
lane from a scalable vector using a runtime determined lane value. I have
added support to VPIteration for runtime-determined lanes that still permit
the caching of values. I did this by introducing a new class called VPLane,
which describes the lane we're dealing with and provides interfaces to get
both the compile-time known lane and the runtime determined value. Whilst
doing this work I couldn't find any explicit tests for extracting the last
lane values of fixed width vectors so I added tests for both scalable and
fixed width vectors.
Differential Revision: https://reviews.llvm.org/D95139
This code assumed that FP math was only permissable if it was
fully "fast", so it hard-coded "fast" when creating new instructions.
The underlying code already allows matching recurrences/reductions
that are only "reassoc", so this change should prevent the potential
miscompile seen in the test diffs (we created "fast" ops even though
none existed in the original code).
I don't know if we need to create the temporary IRBuilder objects
used here, so that could be follow-up clean-up.
There's an open question about whether we should require "nsz" in
addition to "reassoc" here. InstCombine uses that combo for its
reassociative folds, but I think codegen is not as strict.
Similar to b3a33553ae, but this shows a TODO and a potential
miscompile is already present.
We are tracking an FP instruction that does *not* have FMF (reassoc)
properties, so calling that "Unsafe" seems opposite of the common
reading.
I also removed one getter method by rolling the null check into
the access. Further simplification may be possible.
The motivation is to clean up the interactions between FMF and
function-level attributes in these classes and their callers.
The new test shows that there is an existing bug somewhere in
the callers. We assumed that the original code was fully 'fast'
and so we produced IR with 'fast' even though it was just 'reassoc'.
We are tracking an FP instruction that does *not* have FMF (reassoc)
properties, so calling that "Unsafe" seems opposite of the common
reading.
I also removed one getter method by rolling the null check into
the access. Further simplification seems possible.
The motivation is to clean up the interactions between FMF and
function-level attributes in these classes and their callers.
It is possible to merge reuse and reorder shuffles and reduce the total
cost of the vectorization tree/number of final instructions.
Differential Revision: https://reviews.llvm.org/D94992
Update the deletion order when destroying VPBasicBlocks. This ensures
recipes that depend on earlier ones in the block are removed first.
Otherwise this may cause issues when recipes have remaining users later
in the block.
This patch updates LV to generate the runtime checks just after cost
modeling, to allow a more precise estimate of the actual cost of the
checks. This information will be used in future patches to generate
larger runtime checks in cases where the checks only make up a small
fraction of the expected scalar loop execution time.
The runtime checks are created up-front in a temporary block to allow better
estimating the cost and un-linked from the existing IR. After deciding to
vectorize, the checks are moved backed. If deciding not to vectorize, the
temporary block is completely removed.
This patch is similar in spirit to D71053, but explores a different
direction: instead of delaying the decision on whether to vectorize in
the presence of runtime checks it instead optimistically creates the
runtime checks early and discards them later if decided to not
vectorize. This has the advantage that the cost-modeling decisions
can be kept together and can be done up-front and thus preserving the
general code structure. I think delaying (part) of the decision to
vectorize would also make the VPlan migration a bit harder.
One potential drawback of this patch is that we speculatively
generate IR which we might have to clean up later. However it seems like
the code required to do so is quite manageable.
Reviewed By: lebedev.ri, ebrevnov
Differential Revision: https://reviews.llvm.org/D75980
This reverts the revert commit 437f0bbcd5.
It adds a new toVPRecipeResult, which forces VPRecipeOrVPValueTy to be
constructed with a VPRecipeBase *. This should address ambiguous
constructor issues for recipe sub-types that also inherit from VPValue.
Generalize the return value of tryToCreateWidenRecipe to return either a
newly create recipe or an existing VPValue. Use this to avoid creating
unnecessary VPBlendRecipes.
Fixes PR44800.
As a followup to D95291, getOperandsScalarizationOverhead was still
using a VF as a vector factor if the arguments were scalar, and would
assert on certain matrix intrinsics with differently sized vector
arguments. This patch removes the VF arg, instead passing the Types
through directly. This should allow it to more accurately compute the
cost without having to guess at which operands will be vectorized,
something difficult with more complex intrinsics.
This adjusts one SVE test as it is now calling the wrong intrinsic vs
veccall. Without invalid InstructCosts the cost of the scalarized
intrinsic is too low. This should get fixed when the cost of
scalarization is accounted for with scalable types.
Differential Revision: https://reviews.llvm.org/D96287
getIntrinsicInstrCost takes a IntrinsicCostAttributes holding various
parameters of the intrinsic being costed. It can either be called with a
scalar intrinsic (RetTy==Scalar, VF==1), with a vector instruction
(RetTy==Vector, VF==1) or from the vectorizer with a scalar type and
vector width (RetTy==Scalar, VF>1). A RetTy==Vector, VF>1 is considered
an error. Both of the vector modes are expected to be treated the same,
but because this is confusing many backends end up getting it wrong.
Instead of trying work with those two values separately this removes the
VF parameter, widening the RetTy/ArgTys by VF used called from the
vectorizer. This keeps things simpler, but does require some other
modifications to keep things consistent.
Most backends look like this will be an improvement (or were not using
getIntrinsicInstrCost). AMDGPU needed the most changes to keep the code
from c230965ccf working. ARM removed the fix in
dfac521da1, webassembly happens to get a fixup for an SLP cost
issue and both X86 and AArch64 seem to now be using better costs from
the vectorizer.
Differential Revision: https://reviews.llvm.org/D95291
Pointer operand of scatter loads does not remain scalar in the tree (it
gest vectorized) and thus must not be marked as the scalar that remains
scalar in vectorized form.
Differential Revision: https://reviews.llvm.org/D96818
This patch extends VPWidenPHIRecipe to manage pairs of incoming
(VPValue, VPBasicBlock) in the VPlan native path. This is made possible
because we now directly manage defined VPValues for recipes.
By keeping both the incoming value and block in the recipe directly,
code-generation in the VPlan native path becomes independent of the
predecessor ordering when fixing up non-induction phis, which currently
can cause crashes in the VPlan native path.
This fixes PR45958.
Reviewed By: sguggill
Differential Revision: https://reviews.llvm.org/D96773
Now that all state for generated instructions is managed directly in
VPTransformState, VPCallBack is no longer needed. This patch updates the
last use of `getOrCreateScalarValue` to instead manage the value
directly in VPTransformState and removes VPCallback.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D95383
Floating point conversions inside vectorized loops have performance
implications but are very subtle. The user could specify a floating
point constant, or call a function without realizing that it will
force a change in the vector width. An example of this behaviour is
seen in https://godbolt.org/z/M3nT6c . The vectorizer should indicate
when this happens becuase it is most likely unintended behaviour.
This patch adds a simple check for this behaviour by following floating
point stores in the original loop and checking if a floating point
conversion operation occurs.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D95539
This patch enables scalable vectorization of loops with integer/fast reductions, e.g:
```
unsigned sum = 0;
for (int i = 0; i < n; ++i) {
sum += a[i];
}
```
A new TTI interface, isLegalToVectorizeReduction, has been added to prevent
reductions which are not supported for scalable types from vectorizing.
If the reduction is not supported for a given scalable VF,
computeFeasibleMaxVF will fall back to using fixed-width vectorization.
Reviewed By: david-arm, fhahn, dmgreen
Differential Revision: https://reviews.llvm.org/D95245
This patch updates codegen to use VPValues to manage the generated
scalarized instructions.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D92285
This patch fixes pr48832 by correctly generating the mask when a poison value is involved.
Consider this CFG (which is a part of the input):
```
for.body: ; preds = %for.cond
br i1 true, label %cond.false, label %land.rhs
land.rhs: ; preds = %for.body
br i1 poison, label %cond.end, label %cond.false
cond.false: ; preds = %for.body, %land.rhs
br label %cond.end
cond.end: ; preds = %land.rhs, %cond.false
%cond = phi i32 [ 0, %cond.false ], [ 1, %land.rhs ]
```
The path for.body -> land.rhs -> cond.end should be taken when 'select i1 false, i1 poison, i1 false' holds (which means it's never taken); but VPRecipeBuilder::createEdgeMask was emitting 'and i1 false, poison' instead.
The former one successfully blocks poison propagation whereas the latter one doesn't, making the condition poison and thus causing the miscompilation.
SimplifyCFG has a similar bug (which didn't expose a real-world bug yet), and a patch for this is also ongoing (see https://reviews.llvm.org/D95026).
Reviewed By: bjope
Differential Revision: https://reviews.llvm.org/D95217
Changes `getScalarizationOverhead` to return an invalid cost for scalable VFs
and adds some simple tests for loops containing a function for which
there is a vectorized variant available.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D96356
The individual recipes have been updated to manage their operands using
VPUser a while back. Now that the transition is done, we can instead
make VPRecipeBase a VPUser and get rid of the toVPUser helper.
This patch changes the VecDesc struct to use ElementCount
instead of an unsigned VF value, in preparation for
future work that adds support for vectorized versions of
math functions using scalable vectors. Since all I'm doing
in this patch is switching the type I believe it's a
non-functional change. I changed getWidestVF to now return
both the widest fixed-width and scalable VF values, but
currently the widest scalable value will be zero.
Differential Revision: https://reviews.llvm.org/D96011
This will be needed in the loop-vectorizer where the minimum VF
requested may be a scalable VF. getMinimumVF now takes an additional
operand 'IsScalableVF' that indicates whether a scalable VF is required.
Reviewed By: kparzysz, rampitec
Differential Revision: https://reviews.llvm.org/D96020
This patch is NFC and changes occurrences of `unsigned Width`
and `unsigned i` to work on type ElementCount instead.
This patch is a preparatory patch with the ultimate goal of making
`computeMaxVF()` return both a max fixed VF and a max scalable VF,
so that `selectVectorizationFactor()` can pick the most cost-effective
vectorization factor.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D96019
This patch is NFC and changes occurrences of `unsigned MaxVectorSize`
to work on type ElementCount.
This patch is a preparatory patch with the ultimate goal of making
`computeMaxVF()` return both a max fixed VF and a max scalable VF,
so that `selectVectorizationFactor()` can pick the most cost-effective
vectorization factor.
Reviewed By: kmclaughlin
Differential Revision: https://reviews.llvm.org/D96018
VP blocks keep track of a condition, which is a VPValue. This patch
updates VPBlockBase to manage the value using VPUser, so
replaceAllUsesWith properly updates the condition bit as well.
This is required to enable VP2VP transformations and it helps with
simplifying some of the code required to manage condition bits.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D95382
This reverts commit 502a67dd7f.
This expose a failure in test-suite build on PowerPC,
revert to unblock buildbot first,
Dave will re-commit in https://reviews.llvm.org/D96287.
Thanks Dave.
This patch updates some places where VectorLoopValueMap is accessed
directly to instead go through VPTransformState.
As we move towards managing created values exclusively in VPTransformState,
this ensures the use always can fetch the correct value.
This is in preparation for D92285, which switches to managing scalarized
values through VPValues.
In the future, the various fix* functions should be moved directly into
the VPlan codegen stage.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D95757
getIntrinsicInstrCost takes a IntrinsicCostAttributes holding various
parameters of the intrinsic being costed. It can either be called with a
scalar intrinsic (RetTy==Scalar, VF==1), with a vector instruction
(RetTy==Vector, VF==1) or from the vectorizer with a scalar type and
vector width (RetTy==Scalar, VF>1). A RetTy==Vector, VF>1 is considered
an error. Both of the vector modes are expected to be treated the same,
but because this is confusing many backends end up getting it wrong.
Instead of trying work with those two values separately this removes the
VF parameter, widening the RetTy/ArgTys by VF used called from the
vectorizer. This keeps things simpler, but does require some other
modifications to keep things consistent.
Most backends look like this will be an improvement (or were not using
getIntrinsicInstrCost). AMDGPU needed the most changes to keep the code
from c230965ccf working. ARM removed the fix in
dfac521da1, webassembly happens to get a fixup for an SLP cost
issue and both X86 and AArch64 seem to now be using better costs from
the vectorizer.
Differential Revision: https://reviews.llvm.org/D95291
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 NFCI-ish prep work, but the changes are a bit too involved for me to feel comfortable tagging the change that way.
Differential Revision: https://reviews.llvm.org/D94892
This patch updates the induction value creation to use VPValues of
recipes to map the created values. This should bring is one step closer
to being able to optimize induction recipes directly in VPlan.
Currently widenIntOrFpInduction also generates vector values for a cast
of the induction, if it exists. Make this explicit by adding the cast
instruction to the values defined by the recipe.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D92284
This patch adds constructors to VPIteration as a cleaner way of
initialising the struct and replaces existing constructions of
the form:
{Part, Lane}
with
VPIteration(Part, Lane)
I have also added a default constructor, which is used by VPlan.cpp
when deciding whether to replicate a block or not.
This refactoring will be required in a later patch that adds more
members and functions to VPIteration.
Differential Revision: https://reviews.llvm.org/D95676
This patch updates IRBuilder::CreateMaskedGather/Scatter to work
with ScalableVectorType and adds isLegalMaskedGather/Scatter functions
to AArch64TargetTransformInfo. In addition I've fixed up
isLegalMaskedLoad/Store to return true for supported scalar types,
since this is what the vectorizer asks for.
In LoopVectorize.cpp I've changed
LoopVectorizationCostModel::getInterleaveGroupCost to return an invalid
cost for scalable vectors, since currently this relies upon using shuffle
vector for reversing vectors. In addition, in
LoopVectorizationCostModel::setCostBasedWideningDecision I have assumed
that the cost of scalarising memory ops is infinitely expensive.
I have added some simple masked load/store and gather/scatter tests,
including cases where we use gathers and scatters for conditional invariant
loads and stores.
Differential Revision: https://reviews.llvm.org/D95350
Extend applyLoopGuards() to take into account conditions/assumes proving some
value %v to be divisible by D by rewriting %v to (%v / D) * D. This lets the
loop unroller and the loop vectorizer identify more loops as not requiring
remainder loops.
Differential Revision: https://reviews.llvm.org/D95521
This is another step (see D95452) towards correcting fast-math-flags
bugs in vector reductions.
There are multiple bugs visible in the test diffs, and this is still
not working as it should. We still use function attributes (rather
than FMF) to drive part of the logic, but we are not checking for
the correct FP function attributes.
Note that FMF may not be propagated optimally on selects (example
in https://llvm.org/PR35607 ). That's why I'm proposing to union the
FMF of a fcmp+select pair and avoid regressions on existing vectorizer
tests.
Differential Revision: https://reviews.llvm.org/D95690
D90687 introduced a crash:
llvm::LoopVectorizationCostModel::computeMaxVF(llvm::ElementCount, unsigned int):
Assertion `WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() &&
"No decisions should have been taken at this point"' failed.
when compiling the following C code:
typedef struct {
char a;
} b;
b *c;
int d, e;
int f() {
int g = 0;
for (; d; d++) {
e = 0;
for (; e < c[d].a; e++)
g++;
}
return g;
}
with:
clang -Os -target hexagon -mhvx -fvectorize -mv67 testcase.c -S -o -
This occurred since prior to D90687 computeFeasibleMaxVF would only be
called in computeMaxVF when a scalar epilogue was allowed, but now it's
always called. This causes the assert above since computeFeasibleMaxVF
collects all viable VFs larger than the default MaxVF, and for each VF
calculates the register usage which results in analysis being done the
assert above guards against. This can occur in computeFeasibleMaxVF if
TTI.shouldMaximizeVectorBandwidth and this target hook is implemented in
the hexagon backend to always return true.
Reported by @iajbar.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D94869
I am trying to untangle the fast-math-flags propagation logic
in the vectorizers (see a6f022127 for SLP).
The loop vectorizer has a mix of checking FP function attributes,
IR-level FMF, and just wrong assumptions.
I am trying to avoid regressions while fixing this, and I think
the IR-level logic is good enough for that, but it's hard to say
for sure. This would be the 1st step in the clean-up.
The existing test that I changed to include 'fast' actually shows
a miscompile: the function only had the equivalent of nnan, but we
created new instructions that had fast (all FMF set). This is
similar to the example in https://llvm.org/PR35538
Differential Revision: https://reviews.llvm.org/D95452
This gives the user control over which expander to use, which in turn
allows the user to decide what to do with the expanded instructions.
Used in D75980.
Reviewed By: lebedev.ri
Differential Revision: https://reviews.llvm.org/D94295
Now that VPRecipeBase inherits from VPDef, we can always use the new
VPValue for replacement, if the recipe defines one. Given the recipes
that are supported at the moment, all new recipes must have either 0 or
1 defined values.
a6f0221276 enabled intersection of FMF on reduction instructions,
so it is safe to ease the check here.
There is still some room to improve here - it looks like we
have nearly duplicate flags propagation logic inside of the
LoopUtils helper but it is limited targets that do not form
reduction intrinsics (they form the shuffle expansion).
Add an intrinsic type class to represent the
llvm.experimental.noalias.scope.decl intrinsic, to make code
working with it a bit nicer by hiding the metadata extraction
from view.
As shown in the test diffs, we could miscompile by
propagating flags that did not exist in the original
code.
The flags required for fmin/fmax reductions will be
fixed in a follow-up patch.
Walking the use list of a Constant (particularly, ConstantData)
is not scalable, since a given constant may be used by many
instructinos in many functions in many modules.
Differential Revision: https://reviews.llvm.org/D94713
I have removed an unnecessary assert in LoopVectorizationCostModel::getInstructionCost
that prevented a cost being calculated for select instructions when using
scalable vectors. In addition, I have changed AArch64TTIImpl::getCmpSelInstrCost
to only do special cost calculations for fixed width vectors and fall
back to the base version for scalable vectors.
I have added a simple cost model test for cmps and selects:
test/Analysis/CostModel/sve-cmpsel.ll
and some simple tests that show we vectorize loops with cmp and select:
test/Transforms/LoopVectorize/AArch64/sve-basic-vec.ll
Differential Revision: https://reviews.llvm.org/D95039
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
This is NFC-intended and removes the "OperationData"
class which had become nothing more than a recurrence
(reduction) type.
I adjusted the matching logic to distinguish
instructions from non-instructions - that's all that
the "IsLeafValue" member was keeping track of.
We were able to remove almost all of the state from
OperationData, so these don't make sense as members
of that class - just pass the RecurKind in as a param.
More streamlining is possible, but I'm trying to avoid
logic/typo bugs while fixing this. Eventually, we should
not need the `OperationData` class.
We were able to remove almost all of the state from
OperationData, so these don't make sense as members
of that class - just pass the RecurKind in as a param.
Just like llvm.assume, there are a lot of cases where we can just ignore llvm.experimental.noalias.scope.decl.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93042
A previous patch has already changed getInstructionCost to return
an InstructionCost type. This patch changes the other various
getXXXCost functions to return an InstructionCost too. This is a
non-functional change - I've added a few asserts that the costs
are valid in places where we're selecting between vector call
and intrinsic costs. However, since we don't yet return invalid
costs from any of the TTI implementations these asserts should
not fire.
See this patch for the introduction of the type: https://reviews.llvm.org/D91174
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html
Differential Revision: https://reviews.llvm.org/D94065
After much refactoring over the last 2 weeks to the reduction
matching code, I think this change is finally ready.
We effectively broke fmax/fmin vector reduction optimization
when we started canonicalizing to intrinsics in instcombine,
so this should restore that functionality for SLP.
There are still FMF problems here as noted in the code comments,
but we should be avoiding miscompiles on those for fmax/fmin by
restricting to full 'fast' ops (negative tests are included).
Fixing FMF propagation is a planned follow-up.
Differential Revision: https://reviews.llvm.org/D94913
This will avoid confusion once we start matching
min/max intrinsics. All of these hacks to accomodate
cmp+sel idioms should disappear once we canonicalize
to min/max intrinsics.
The icmp opcode is now hard-coded in the cost model call.
This will make it easier to eventually remove all opcode
queries for min/max patterns as we transition to intrinsics.
This patch changes these functions:
vectorizeLoadInsert
isExtractExtractCheap
foldExtractedCmps
scalarizeBinopOrCmp
getShuffleExtract
foldBitcastShuf
to use the class InstructionCost when calling TTI.get<something>Cost().
This patch is part of a series of patches to use InstructionCost instead of
unsigned/int for the cost model functions.
See this thread for context:
http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html
See this patch for the introduction of the type:
https://reviews.llvm.org/D91174
ps.:This patch adds the test || !NewCost.isValid(), because we want to
return false when:
!NewCost.isValid && !OldCost.isValid()->the cost to transform it expensive
and
!NewCost.isValid() && OldCost.isValid()
Therefore for simplication we only add test for !NewCost.isValid()
Differential Revision: https://reviews.llvm.org/D94069
This is NFC-intended and another step towards supporting
intrinsics as reduction candidates.
The remaining bits of the OperationData class do not make
much sense as-is, so I will try to improve that, but I'm
trying to take minimal steps because it's still not clear
how this was intended to work.
This is another NFC-intended patch to allow matching
intrinsics (example: maxnum) as candidates for reductions.
It's possible that the loop/if logic can be reduced now,
but it's still difficult to understand how this all works.
To get into this block we had: !A || B || C
and we checked C in the first 'if' clause
leaving !A || B. But the 2nd 'if' is checking:
A && !B --> !(!A || B)
This is NFC-intended. I'm still trying to figure out
how the loop where this is used works. It does not
seem like we require this data at all, but it's
hard to confirm given the complicated predicates.
In the spirit of commit fc783e91e0 (llvm-svn: 248943) we
shouldn't vectorize stores of non-packed types (i.e. types that
has padding between consecutive variables in a scalar layout,
but being packed in a vector layout).
The problem was detected as a miscompile in a downstream test case.
Reviewed By: anton-afanasyev
Differential Revision: https://reviews.llvm.org/D94446
This relates to the ongoing effort to support vectorization of multiple exit loops (see D93317).
The previous code assumed that LCSSA phis were always single entry before the vectorizer ran. This was correct, but only because the vectorizer allowed only a single exiting edge. There's nothing in the definition of LCSSA which requires single entry phis.
A common case where this comes up is with a loop with multiple exiting blocks which all reach a common exit block. (e.g. see the test updates)
Differential Revision: https://reviews.llvm.org/D93725
This patch unifies the way recipes and VPValues are printed after the
transition to VPDef.
VPSlotTracker has been updated to iterate over all recipes and all
their defined values to number those. There is no need to number
values in Value2VPValue.
It also updates a few places that only used slot numbers for
VPInstruction. All recipes now can produce numbered VPValues.
This patch is part of a series of patches that migrate integer
instruction costs to use InstructionCost. In the function
selectVectorizationFactor I have simply asserted that the cost
is valid and extracted the value as is. In future we expect
to encounter invalid costs, but we should filter out those
vectorization factors that lead to such invalid costs.
See this patch for the introduction of the type: https://reviews.llvm.org/D91174
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html
Differential Revision: https://reviews.llvm.org/D92178
A severe compile-time slowdown from this call is noted in:
https://llvm.org/PR48689
My naive fix was to put it under LLVM_DEBUG ( 267ff79 ),
but that's not limiting in the way we want.
This is a quick fix (or we could just remove the call completely
and rely on some later pass to discover potentially wrong IR?).
A bigger/better fix would be to improve/limit verifyFunction()
as noted in:
https://llvm.org/PR47712
Differential Revision: https://reviews.llvm.org/D94328
Similar to D92129, update VPWidenPHIRecipe to manage the start value as
VPValue. This allows adjusting the start value as a VPlan transform,
which will be used in a follow-up patch to support reductions during
epilogue vectorization.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D93975
This was suggested to prepare for D93975.
By moving the start value creation to widenPHInstruction, we set the
stage to manage the start value directly in VPWidenPHIRecipe, which be
used subsequently to set the 'resume' value for reductions during
epilogue vectorization.
It also moves RdxDesc to the recipe, so we do not have to rely on Legal
to look it up later.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D94175
As noted in PR48689, the verifier may have some kind
of exponential behavior that should be addressed
separately. For now, only run it in debug mode to
prevent problems for release+asserts.
That limit is what we had before D80401, and I'm
not sure if there was a reason to change it in that
patch.
In the following loop:
void foo(int *a, int *b, int N) {
for (int i=0; i<N; ++i)
a[i + 4] = a[i] + b[i];
}
The loop dependence constrains the VF to a maximum of (4, fixed), which
would mean using <4 x i32> as the vector type in vectorization.
Extending this to scalable vectorization, a VF of (4, scalable) implies
a vector type of <vscale x 4 x i32>. To determine if this is legal
vscale must be taken into account. For this example, unless
max(vscale)=1, it's unsafe to vectorize.
For SVE, the number of bits in an SVE register is architecturally
defined to be a multiple of 128 bits with a maximum of 2048 bits, thus
the maximum vscale is 16. In the loop above it is therefore unfeasible
to vectorize with SVE. However, in this loop:
void foo(int *a, int *b, int N) {
#pragma clang loop vectorize_width(X, scalable)
for (int i=0; i<N; ++i)
a[i + 32] = a[i] + b[i];
}
As long as max(vscale) multiplied by the number of lanes 'X' doesn't
exceed the dependence distance, it is safe to vectorize. For SVE a VF of
(2, scalable) is within this constraint, since a vector of <16 x 2 x 32>
will have no dependencies between lanes. For any number of lanes larger
than this it would be unsafe to vectorize.
This patch extends 'computeFeasibleMaxVF' to legalize scalable VFs
specified as loop hints, implementing the following behaviour:
* If the backend does not support scalable vectors, ignore the hint.
* If scalable vectorization is unfeasible given the loop
dependence, like in the first example above for SVE, then use a
fixed VF.
* Accept scalable VFs if it's safe to do so.
* Otherwise, clamp scalable VFs that exceed the maximum safe VF.
Reviewed By: sdesmalen, fhahn, david-arm
Differential Revision: https://reviews.llvm.org/D91718
The new test case here contains a first order recurrences and an
instruction that is replicated. The first order recurrence forces an
instruction to be sunk _into_, as opposed to after the replication
region. That causes several things to go wrong including registering
vector instructions multiple times and failing to create dominance
relations correctly.
Instead we should be sinking to after the replication region, which is
what this patch makes sure happens.
Differential Revision: https://reviews.llvm.org/D93629
After merging the shuffles, we cannot rely on the previous shuffle
anymore and need to shrink the final shuffle, if it is required.
Reported in D92668
Differential Revision: https://reviews.llvm.org/D93967
Similar to 5a1d31a28 -
This should be no-functional-change because the reduction kind
opcodes are 1-for-1 mappings to the instructions we are matching
as reductions. But we want to remove the need for the
`OperationData` opcode field because that does not work when
we start matching intrinsics (eg, maxnum) as reduction candidates.
This patch updates VPWidenIntOrFpInductionRecipe to hold the start value
for the induction variable. This makes the start value explicit and
allows for adjusting the start value for a VPlan.
The flexibility will be used in further patches.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D92129
This patch adds a new getLiveInIRValue accessor to VPValue, which
returns the underlying value, if the VPValue is defined outside of
VPlan. This is required to handle scalars in VPTransformState, which
requires dealing with scalars defined outside of VPlan.
We can simply check VPValue::Def to determine if the value is defined
inside a VPlan.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D92281
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
This should be no-functional-change because the reduction kind
opcodes are 1-for-1 mappings to the instructions we are matching
as reductions. But we want to remove the need for the
`OperationData` opcode field because that does not work when
we start matching intrinsics (eg, maxnum) as reduction candidates.
SLP tries to model 2 forms of vector reductions: pairwise and splitting.
From the cost model code comments, those are defined using an example as:
/// Pairwise:
/// (v0, v1, v2, v3)
/// ((v0+v1), (v2+v3), undef, undef)
/// Split:
/// (v0, v1, v2, v3)
/// ((v0+v2), (v1+v3), undef, undef)
I don't know the full history of this functionality, but it was partly
added back in D29402. There are apparently no users at this point (no
regression tests change). X86 might have managed to work-around the need
for this through cost model and codegen improvements.
Removing this code makes it easier to continue the work that was started
in D87416 / D88193. The alternative -- if there is some target that is
silently using this option -- is to move this logic into LoopUtils. We
have related/duplicate functionality there via llvm::createTargetReduction().
Differential Revision: https://reviews.llvm.org/D93860
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.
While here, rename the inaccurate getRecurrenceBinOp()
because that was also used to get CmpInst opcodes.
The recurrence/reduction kind should always refer to the
expected opcode for a reduction. SLP appears to be the
only direct caller of createSimpleTargetReduction(), and
that calling code ideally should not be carrying around
both an opcode and a reduction kind.
This should allow us to generalize reduction matching to
use intrinsics instead of only binops.
This is almost all mechanical search-and-replace and
no-functional-change-intended (NFC). Having a single
enum makes it easier to match/reason about the
reduction cases.
The goal is to remove `Opcode` from reduction matching
code in the vectorizers because that makes it harder to
adapt the code to handle intrinsics.
The code in RecurrenceDescriptor::AddReductionVar() is
the only place that required closer inspection. It uses
a RecurrenceDescriptor and a second InstDesc to sometimes
overwrite part of the struct. It seem like we should be
able to simplify that logic, but it's not clear exactly
which cmp+sel patterns that we are trying to handle/avoid.
If DoExtraAnalysis is true (e.g. because remarks are enabled), we
continue with the analysis rather than exiting. Update code to
conditionally check if the ExitBB has phis or not a single predecessor.
Otherwise a nullptr is dereferenced with DoExtraAnalysis.
I don't know if there's some way this changes what the vectorizers
may produce for reductions, but I have added test coverage with
3567908 and 5ced712 to show that both passes already have bugs in
this area. Hopefully this does not make things worse before we can
really fix it.
I'm not sure if the SLP enum was created before the IVDescriptor
RecurrenceDescriptor / RecurrenceKind existed, but the code in
SLP is now redundant with that class, so it just makes things
more complicated to have both. We eventually call LoopUtils
createSimpleTargetReduction() to create reduction ops, so we
might as well standardize on those enum names.
There's still a question of whether we need to use TTI::ReductionFlags
vs. MinMaxRecurrenceKind, but that can be another clean-up step.
Another option would just be to flatten the enums in RecurrenceDescriptor
into a single enum. There isn't much benefit (smaller switches?) to
having a min/max subset.
This reverts commit 4ffcd4fe9a thus restoring e4df6a40da.
The only change from the original patch is to add "llvm::" before the call to empty(iterator_range). This is a speculative fix for the ambiguity reported on some builders.
This patch is a major step towards supporting multiple exit loops in the vectorizer. This patch on it's own extends the loop forms allowed in two ways:
single exit loops which are not bottom tested
multiple exit loops w/ a single exit block reached from all exits and no phis in the exit block (because of LCSSA this implies no values defined in the loop used later)
The restrictions on multiple exit loop structures will be removed in follow up patches; disallowing cases for now makes the code changes smaller and more obvious. As before, we can only handle loops with entirely analyzable exits. Removing that restriction is much harder, and is not part of currently planned efforts.
The basic idea here is that we can force the last iteration to run in the scalar epilogue loop (if we have one). From the definition of SCEV's backedge taken count, we know that no earlier iteration can exit the vector body. As such, we can leave the decision on which exit to be taken to the scalar code and generate a bottom tested vector loop which runs all but the last iteration.
The existing code already had the notion of requiring one iteration in the scalar epilogue, this patch is mainly about generalizing that support slightly, making sure we don't try to use this mechanism when tail folding, and updating the code to reflect the difference between a single exit block and a unique exit block (very mechanical).
Differential Revision: https://reviews.llvm.org/D93317
Previously the branch from the middle block to the scalar preheader & exit
was being set-up at the end of skeleton creation in completeLoopSkeleton.
Inserting SCEV or runtime checks may result in LCSSA phis being created,
if they are required. Adjusting branches afterwards may break those
PHIs.
To avoid this, we can instead create the branch from the middle block
to the exit after we created the middle block, so we have the final CFG
before potentially adjusting/creating PHIs.
This fixes a crash for the included test case. For the non-crashing
case, this is almost a NFC with respect to the generated code. The
only change is the order of the predecessors of the involved branch
targets.
Note an assertion was moved from LoopVersioning() to
LoopVersioning::versionLoop. Adjusting the branches means loop-simplify
form may be broken before constructing LoopVersioning. But LV only uses
LoopVersioning to annotate the loop instructions with !noalias metadata,
which does not require loop-simplify form.
This is a fix for an existing issue uncovered by D93317.
I am hoping to extend the reduction matching code, and it is
hard to distinguish "ReductionData" from "ReducedValueData".
So extend the tree/root metaphor to include leaves.
Another problem is that the name "OperationData" does not
provide insight into its purpose. I'm not sure if we can alter
that underlying data structure to make the code clearer.
I think this is NFC currently, but the bug would be exposed
when we allow binary intrinsics (maxnum, etc) as candidates
for reductions.
The code in matchAssociativeReduction() is using
OperationData::getNumberOfOperands() when comparing whether
the "EdgeToVisit" iterator is in-bounds, so this code must
use the same (potentially offset) operand value to set
the "EdgeToVisit".
ScalarEvolution should be able to handle both constant and variable trip
counts using getURemExpr, so we do not have to handle them separately.
This is a small simplification of a56280094e.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D93677
This patch turns updates VPInstruction to manage the value it defines
using VPDef. The VPValue is used during VPlan construction and
codegeneration instead of the plain IR reference where possible.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D90565
When the trip-count is provably divisible by the maximal/chosen VF, folding the
loop's tail during vectorization is redundant. This commit extends the existing
test for constant trip-counts to any trip-count known to be divisible by
maximal/selected VF by SCEV.
Differential Revision: https://reviews.llvm.org/D93615
This patch makes VPRecipeBase a direct subclass of VPDef, moving the
SubclassID to VPDef.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D90564
This patch turns updates VPInterleaveRecipe to manage the values it defines
using VPDef. The VPValue is used during VPlan construction and
codegeneration instead of the plain IR reference where possible.
Reviewed By: gilr
Differential Revision: https://reviews.llvm.org/D90562
Rosie Sumpter