This work follows on from D142109 and addresses a possible regression
when we know the loop iteration counter cannot overflow.
When we know the overflow-check always evaluates to false, it's better to
use the other style of tail folding where it assumes a runtime check was
added, because that avoids having to calculate a modified trip-count.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D142894
When using tail-folding and using the predicate for both data and control-flow
(the next vector iteration's predicate is generated with the llvm.active.lane.mask
intrinsic and then tested for the backedge), the LoopVectorizer still inserts a
runtime check to see if the 'i + VF' may at any point overflow for the given
trip-count. When it does, it falls back to a scalar epilogue loop.
We can get rid of that runtime check in the pre-header and therefore also
remove the scalar epilogue loop. This reduces code-size and avoids a runtime
check.
Consider the following loop:
void foo(char * __restrict__ dst, char *src, unsigned long N) {
for (unsigned long i=0; i<N; ++i)
dst[i] = src[i] + 42;
}
If 'N' is e.g. ULONG_MAX, and the VF > 1, then the loop iteration counter
will overflow when calculating the predicate for the next vector iteration
at some point, because LLVM does:
vector.ph:
%active.lane.mask.entry = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 0, i64 %N)
vector.body:
%index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
%active.lane.mask = phi <vscale x 16 x i1> [ %active.lane.mask.entry, %vector.ph ], [ %active.lane.mask.next, %vector.body ]
...
%index.next = add i64 %index, 16
; The add above may overflow, which would affect the lane mask and control flow. Hence a runtime check is needed.
%active.lane.mask.next = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 %index.next, i64 %N)
%8 = extractelement <vscale x 16 x i1> %active.lane.mask.next, i64 0
br i1 %8, label %vector.body, label %for.cond.cleanup, !llvm.loop !7
The solution:
What we can do instead is calculate the predicate before incrementing
the loop iteration counter, such that the llvm.active.lane.mask is
calculated from 'i' to 'tripcount > VF ? tripcount - VF : 0', i.e.
vector.ph:
%active.lane.mask.entry = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 0, i64 %N)
%N_minus_VF = select %N > 16 ? %N - 16 : 0
vector.body:
%index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
%active.lane.mask = phi <vscale x 16 x i1> [ %active.lane.mask.entry, %vector.ph ], [ %active.lane.mask.next, %vector.body ]
...
%active.lane.mask.next = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 %index, i64 %N_minus_VF)
%index.next = add i64 %index, %4
; The add above may still overflow, but this time the active.lane.mask is not affected
%8 = extractelement <vscale x 16 x i1> %active.lane.mask.next, i64 0
br i1 %8, label %vector.body, label %for.cond.cleanup, !llvm.loop !7
For N = 20, we'd then get:
vector.ph:
%active.lane.mask.entry = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 0, i64 %N)
; %active.lane.mask.entry = <1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1>
%N_minus_VF = select 20 > 16 ? 20 - 16 : 0
; %N_minus_VF = 4
vector.body: (1st iteration)
... ; using <1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1> as predicate in the loop
...
%active.lane.mask.next = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 0, i64 4)
; %active.lane.mask.next = <1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>
%index.next = add i64 0, 16
; %index.next = 16
%8 = extractelement <vscale x 16 x i1> %active.lane.mask.next, i64 0
; %8 = 1
br i1 %8, label %vector.body, label %for.cond.cleanup, !llvm.loop !7
; branch to %vector.body
vector.body: (2nd iteration)
... ; using <1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0> as predicate in the loop
...
%active.lane.mask.next = tail call <vscale x 16 x i1> @llvm.get.active.lane.mask.nxv16i1.i64(i64 16, i64 4)
; %active.lane.mask.next = <0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>
%index.next = add i64 16, 16
; %index.next = 32
%8 = extractelement <vscale x 16 x i1> %active.lane.mask.next, i64 0
; %8 = 0
br i1 %8, label %vector.body, label %for.cond.cleanup, !llvm.loop !7
; branch to %for.cond.cleanup
Reviewed By: fhahn, david-arm
Differential Revision: https://reviews.llvm.org/D142109
BlockFrequencyInfo should generally only be fetched in PGO builds
where a PSI profile summary is available. However, LoopVectorize
was fetching it unconditionally.
This results in a small compile-time improvement for non-PGO builds.
Differential Revision: https://reviews.llvm.org/D144953
Previously, while calculating register usage due to invariants, it was assumed that invariant would always be part of widening
instructions. This resulted in calculating vector register types for vectors which cant be legalized(check the newly added test for more details).
An invariant might not always need a vector register. For e.g., invariant might just be used for iteration check.
This patch checks if the invariant is part of any widening instruction and considers register usage accordingly. Fixes issue 60493
Differential Revision: https://reviews.llvm.org/D143422
Previously, while calculating register usage due to invariants, it was assumed that invariant would always be part of widening
instructions. This resulted in calculating vector register types for vectors which cant be legalized(check the newly added test for more details).
An invariant might not always need a vector register. For e.g., invariant might just be used for iteration check.
This patch checks if the invariant is part of any widening instruction and considers register usage accordingly. Fixes issue 60493
Differential Revision: https://reviews.llvm.org/D143422
To query the maximum value for vscale, the LV queries the vscale_range
attribute or a TTI hook. To avoid having to reimplement the same behaviour
for multiple uses (such as in D142894), it makes sense to move this code
to a separate function.
Enabling assignment tracking without this patch, a significant amount of
additional compiler run time comes from the RemoveRedundantDbgInstrs call in
InstCombine. This patch reduces compiler run time by choosing better places to
call RemoveRedundantDbgInstrs.
In non-assignment-tracking builds, RemoveRedundantDbgInstrs is called by
InstCombine if LowerDbgDeclare makes a change (i.e. it is _sometimes_
called). In assignment tracking builds LowerDbgDeclare doesn't do anything. We
still need to clean up redundant intrinsics to avoid a large performance hit
due to the number of instructions, so the current approach is to have
InstCombine _always_ call RemoveRedundantDbgInstrs.
Instrumenting the compiler to run RemoveRedundantDbgInstrs after every pass and
dump the numbers and building CTMark/tramp3d-v4 indicates that SROA and
LoopVectorize give us a bigger bang (number removed) for buck (times pass is
run).
The compile time tracker reports that this patch reduces the number of
instructions retired building CTMark projects by an average of 1.1%.
Reviewed By: scott.linder
Differential Revision: https://reviews.llvm.org/D144483
In order to allow targets to disable interleaving for scalable vectors, pass the entire VF's ElementCount to getMaxInterleaveFactor.
This is based off of the approach used here: 8d36708507
The plan would then be to disable interleaving on scalable VFs on RISC-V in a follow up patch.
See https://reviews.llvm.org/D143723#4132349
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D144474
There is no need to update the AlsoPack field when creating
VPReplicateRecipes. It can be easily computed based on the VP def-use
chains when it is needed.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D143864
The code only needs access to INvalidCosts, ORE and TheLoop, so it can
easily be moved into a helper to make selectVectorizationFactor more
compact.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D143957
Prevoius pseudo probes were dropped out of a vectorized loop body during loop vectorization. This can result in the samples of the loop entry is used for the loop body, which in turn can cause undercounting of the loop iteration count. The undercounting can further prevent the loop from being vectorized in the next build. I'm fixing this by explicting allowing pseudo probes to be kept in the vectorized loop body, and by claiming a probe instruction is not "uniform", the vectorizer will duplicate it by the number of vector lanes.
For one internal service, I'm seeing the change causes the size increase of the .pseudoprobe section by 0.7%, which should count around 0.2% of the whole binary size.
Reviewed By: wenlei
Differential Revision: https://reviews.llvm.org/D144066
When vectorizing code with function calls in it, if we encounter
a function which only has vectorized variants requiring a mask
we can synthesize an all-true mask to enable us to proceed.
Since we want the mask to be represented in vplan, the pointer
to the chosen Function is now stored as part of the
VPWidenCallRecipe, and mask arguments are added at the
appropriate index to the recipe operands.
Reviewed By: david-arm, fhahn, reames
Differential Revision: https://reviews.llvm.org/D132458
Fixed issue where 'ConstantInt::get(IndextTy, -Part)' was executed with the wrong type for Part,
e.g. IndexTy was i64, but Part was 'unsigned', which led to things like 'mul i64 .., 4294967292',
which was obviously wrong.
Also changed sve-vector-reverse.ll to be vectorized with UF>1 to test this.
This reverts commit 1f01cdda68614dba12af3cc3aff38541d0abcc6b.
This patch updates LV to sink recipes directly using the VPlan use
chains. The initial patch only moves sinking to be purely VPlan-based.
Follow-up patches will move legality checks to VPlan as well.
At the moment, there's a single test failure remaining.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D142589
This is specifically relevant for loops that vectorize using a scalable VF,
where the code results in:
%vscale = call i32 llvm.vscale.i32()
%vf.part1 = mul i32 %vscale, 4
%gep = getelementptr ..., i32 %vf.part1
Which InstCombine then changes into:
%vscale = call i32 llvm.vscale.i32()
%vf.part1 = mul i32 %vscale, 4
%vf.part1.zext = sext i32 %vf.part1 to i64
%gep = getelementptr ..., i32 %vf.part1.zext
D143016 tried to remove these extends, but that only works when
the call to llvm.vscale.i32() has a single use. After doing any
kind of CSE on these calls the combine no longer kicks in.
It seems more sensible to ask DataLayout what type to use, rather
than relying on InstCombine to insert the extend and hoping it can
fold it away.
I've only changed this for indices that are not constant, because
I vaguely remember there was a reason for sticking with i32. It
would also mean patching up loads more tests.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D143267
This NFC (intended) patch has several small changes:
* It renames PredicationStyle to TailFoldingStyle.
* It renames TTI.emitActiveLaneMask() to TTI.getPreferredTailFoldingStyle()
* Simplifies some of its uses in the LoopVectorizer
Rationale: To my surprise PredicationStyle::None did not mean 'no
predication', but rather 'no active lane mask intrinsic', such that the
predicate is created using a splat + compare with stepvector. The enum is
also highly specific to tail folding, so it seems better to name this
around that feature, i.e. 'tail folding style'.
This also makes it more amenable to extend it to other tail folding styles,
such as the one added in D142109.
Reviewed By: david-arm
Differential Revision: https://reviews.llvm.org/D142887
LoopUnroll estimates the loop size via getInstructionCost(),
but getInstructionCost() cannot pass CostKind to getVectorInstrCost().
And so does getShuffleCost() to getBroadcastShuffleOverhead(),
getPermuteShuffleOverhead(), getExtractSubvectorOverhead(),
and getInsertSubvectorOverhead().
To address this, this patch adds an argument CostKind to these
functions.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D142116
Similar to vp_depth_first_shallow (D140512) add vp_depth_first_deep to
make existing code clearer and more compact.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D142055
LAI is cached during the LoopDistribute pass, and is later re-used during LoopVectorize. The problem is that LoopVectorize changes SCEV, and the cached LAI does not get updated. Hence, when re-using the cached LAI, it references an invalid SCEV.
Fixes#59319
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D139601
This patch adds metadata to disable runtime unrolling to the vectorized
loop. If runtime unrolling/interleaving is considered profitable, LV
will interleave the loop directly. There should be no need to perform
runtime unrolling at a later stage.
Note that we already add metadata to disable runtime unrolling to the
scalar loop after vectorization.
The additional unrolling unnecessarily increases code size and compile
time. In addition to that we have several bug reports of unncessary
runtime unrolling for vectorized loops, e.g. PR40961
Compile-time improvements:
NewPM-O3: -1.04%
NewPM-ReleaseThinLTO: -0.59%
NewPM-ReleaseLTO-g: -0.97%
https://llvm-compile-time-tracker.com/compare.php?from=ce1be13a868d0f8afa367975558c1a6175cce33a&to=78bc2e67f22e9e10e61cdb6cdac4bb857d95eb1b&stat=instructions:uFixes#40306.
Reviewed By: lebedev.ri, nikic
Differential Revision: https://reviews.llvm.org/D115261
The validation of vplans could fail if an inloop reduction was created
with a block-in mask that did not dominate the reduction. This makes
sure that the insert point is set when creating the mask, to ensure it
dominates the reduction.
Differential Revision: https://reviews.llvm.org/D141003
Adjust mergeReplicateRegions to be in line with
mergeBlocksIntoPredecessors added in 36d70a6aea6b by collecting only the
valid candidates first.
Also rename to mergeReplicateRegionsIntoSuccessors and add missing
doc-comment.
This addresses post-commit suggestions by @Ayal.
The preheader was named explicitly in 256c6b0ba14e8a7ab6373b61b7193ea8c0a3651c
which makes setting the name in prior commit 95b2aa511eea1f31e183a2a3aed4d2aa852d089c
unnecessary.
Differential Revision: https://reviews.llvm.org/D140246
Merging regions can enable new sinking opportunities (e.g. if users of a
scalar value are moved from different VPBBs into the same VPBB). Sinking
in turn can also enable new merging opportunities (e.g. if a recipe
between to merge-able regions is moved.
To enable more sinking opportunities, repeat sinking & merging if
regions could be merged.
Also fix mergeReplicateRegions to return the correct Changed status.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D139788
Add and run VPlan transform to fold blocks with a single predecessor
into the predecessor. This remove redundant blocks and addresses a TODO
to replace special handling for the vector latch VPBB.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D139927
This sets the stage for D133017 by moving out the code that performs
VPlan based simplifications to a separate transform that takes the
chosen VF & UF as arguments.
The main advantage is that this transform runs before any changes to
the CFG are being made. This allows using SCEV without worrying about
making queries while the IR is in an incomplete state.
Note that this patch switches the reasoning to use SCEV, but still only
simplifies loops with constant trip counts. Using SCEV here is needed to
access the backedge taken count, because the trip count IR value has not
been created yet.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D135017
The VFs and UFs may be more constrained as the plans are transformed
(e.g. see D135017 for an example).
To make sure the VFs/UFs included in the VPlan dump are accurate,
generate them when accessing a plan's name, rather than include them in
the name string set after initial construction.
Code generation now uses the start VPValue of induction recipes.
This makes it possible to adjust the start value of the epilogue
vector loop to use the 'resume' value of the main vector loop.
Fixes#59459.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D92132
value() has undesired exception checking semantics and calls
__throw_bad_optional_access in libc++. Moreover, the API is unavailable without
_LIBCPP_NO_EXCEPTIONS on older Mach-O platforms (see
_LIBCPP_AVAILABILITY_BAD_OPTIONAL_ACCESS).
This fixes clang.
The vectorizer has code to reject scalable vectorization of loops with very short trip counts, and instead use fixed length vectors. The current code doesn't account for the minimum vscale value known, and thus under estimates the number of lanes in the scalable type for RISCV's default configuration. This results in use of predication and a trivially dead loop where a single straight line piece of code would suffice.
Note that the code quality of the original scalable vectorization could (and probably should) be improved other ways as well. This patch is solely about whether the scalable vectorization was the right choice to begin with.
This bit of code - both with and without my change - does make the unchecked assumption that the target knows how to lower fixed length vectors whose length is provably less than the vector length.
Differential Revision: https://reviews.llvm.org/D137285
This patch extends VP-based sinking to also sink VPScalarStepsRecipe.
This takes us a step closer towards retiring the IR based sinking.
The main change is extending VPScalarIVStepsRecipe::execute to support
executing in a replicate-region.
Depends on D133758.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D133760
This patch mechanically replaces None with std::nullopt where the
compiler would warn if None were deprecated. The intent is to reduce
the amount of manual work required in migrating from Optional to
std::optional.
This is part of an effort to migrate from llvm::Optional to
std::optional:
https://discourse.llvm.org/t/deprecating-llvm-optional-x-hasvalue-getvalue-getvalueor/63716
