When it comes to the scalar cost of any predicated block, the loop
vectorizer by default regards this predication as a sign that it is
looking at an if-conversion and divides the scalar cost of the block by
2, assuming it would only be executed half the time. This however makes
no sense if the predication has been introduced to tail predicate the
loop.
Original patch by Anna Welker
Differential Revision: https://reviews.llvm.org/D86452
If we have two unknown sizes and one GEP operand and one non-GEP
operand, then we currently simply return MayAlias. The comment says
we can't do anything useful ... but we can! We can still check that
the underlying objects are different (and do so for the GEP-GEP case).
To reduce the compile-time impact, this a) checks this early, before
doing the relatively expensive GEP decomposition that will not be
used and b) doesn't do the check if the other operand is a phi or
select. In that case, the phi/select will already recurse, so this
would just do two slightly different recursive walks that arrive at
the same roots.
Compile-time is still a bit of a mixed bag: https://llvm-compile-time-tracker.com/compare.php?from=624af932a808b363a888139beca49f57313d9a3b&to=845356e14adbe651a553ed11318ddb5e79a24bcd&stat=instructions
On average this is a small improvement, but sqlite with ThinLTO has
a 0.5% regression (lencod has a 1% improvement).
The BasicAA test case checks this by using two memsets with unknown
size. However, the more interesting case where this is useful is
the LoopVectorize test case, as analysis of accesses in loops tends
to always us unknown sizes.
Differential Revision: https://reviews.llvm.org/D92401
* Steps are scaled by `vscale`, a runtime value.
* Changes to circumvent the cost-model for now (temporary)
so that the cost-model can be implemented separately.
This can vectorize the following loop [1]:
void loop(int N, double *a, double *b) {
#pragma clang loop vectorize_width(4, scalable)
for (int i = 0; i < N; i++) {
a[i] = b[i] + 1.0;
}
}
[1] This source-level example is based on the pragma proposed
separately in D89031. This patch only implements the LLVM part.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D91077
The -enable-new-pm=1 translation caused loop-vectorize to run on all
functions, then instcombine, rather than all passes on one function then
the next. This caused the output of -debug-only and -print-after to be
interleaved in an unexpected way.
This change should be fairly straight forward. If we've reached a call, check to see if we can tell the result is dereferenceable from information about the minimum object size returned by the call.
To control compile time impact, I'm only adding the call for base facts in the routine. getObjectSize can also do recursive reasoning, and we don't want that general capability here.
As a follow up patch (without separate review), I will plumb through the missing TLI parameter. That will have the effect of extending this to known libcalls - malloc, new, and the like - whereas currently this only covers calls with the explicit allocsize attribute.
Differential Revision: https://reviews.llvm.org/D90341
The initial step of the uniform-after-vectorization (lane-0 demanded only) analysis was very awkwardly written. It would revisit use list of each pointer operand of a widened load/store. As a result, it was in the worst case O(N^2) where N was the number of instructions in a loop, and had restricted operand Value types to reduce the size of use lists.
This patch replaces the original algorithm with one which is at most O(2N) in the number of instructions in the loop. (The key observation is that each use of a potentially interesting pointer is visited at most twice, once on first scan, once in the use list of *it's* operand. Only instructions within the loop have their uses scanned.)
In the process, we remove a restriction which required the operand of the uniform mem op to itself be an instruction. This allows detection of uniform mem ops involving global addresses.
Differential Revision: https://reviews.llvm.org/D92056
This is yet another attempt at providing support for epilogue
vectorization following discussions raised in RFC http://llvm.1065342.n5.nabble.com/llvm-dev-Proposal-RFC-Epilog-loop-vectorization-tt106322.html#none
and reviews D30247 and D88819.
Similar to D88819, this patch achieve epilogue vectorization by
executing a single vplan twice: once on the main loop and a second
time on the epilogue loop (using a different VF). However it's able
to handle more loops, and generates more optimal control flow for
cases where the trip count is too small to execute any code in vector
form.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D89566
In this patch I have added support for a new loop hint called
vectorize.scalable.enable that says whether we should enable scalable
vectorization or not. If a user wants to instruct the compiler to
vectorize a loop with scalable vectors they can now do this as
follows:
br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !2
...
!2 = !{!2, !3, !4}
!3 = !{!"llvm.loop.vectorize.width", i32 8}
!4 = !{!"llvm.loop.vectorize.scalable.enable", i1 true}
Setting the hint to false simply reverts the behaviour back to the
default, using fixed width vectors.
Differential Revision: https://reviews.llvm.org/D88962
This is yet another attempt at providing support for epilogue
vectorization following discussions raised in RFC http://llvm.1065342.n5.nabble.com/llvm-dev-Proposal-RFC-Epilog-loop-vectorization-tt106322.html#none
and reviews D30247 and D88819.
Similar to D88819, this patch achieve epilogue vectorization by
executing a single vplan twice: once on the main loop and a second
time on the epilogue loop (using a different VF). However it's able
to handle more loops, and generates more optimal control flow for
cases where the trip count is too small to execute any code in vector
form.
Reviewed By: SjoerdMeijer
Differential Revision: https://reviews.llvm.org/D89566
In the following loop the dependence distance is 2 and can only be
vectorized if the vector length is no larger than this.
void foo(int *a, int *b, int N) {
#pragma clang loop vectorize(enable) vectorize_width(4)
for (int i=0; i<N; ++i) {
a[i + 2] = a[i] + b[i];
}
}
However, when specifying a VF of 4 via a loop hint this loop is
vectorized. According to [1][2], loop hints are ignored if the
optimization is not safe to apply.
This patch introduces a check to bail of vectorization if the user
specified VF is greater than the maximum feasible VF, unless explicitly
forced with '-force-vector-width=X'.
[1] https://llvm.org/docs/LangRef.html#llvm-loop-vectorize-and-llvm-loop-interleave
[2] https://clang.llvm.org/docs/LanguageExtensions.html#extensions-for-loop-hint-optimizations
Reviewed By: sdesmalen, fhahn, Meinersbur
Differential Revision: https://reviews.llvm.org/D90687
Instruction ExtractValue wasn't handled in
LoopVectorizationCostModel::getInstructionCost(). As a result, it was modeled
as a mul which is not really accurate. Since it is free (most of the times),
this now gets a cost of 0 using getInstructionCost.
This is a follow-up of D92208, that required changing this regression test.
In a follow up I will look at InsertValue which also isn't handled yet.
Differential Revision: https://reviews.llvm.org/D92317
VPPredInstPHIRecipe is one of the recipes that was missed during the
initial conversion. This patch adjusts the recipe to also manage its
operand using VPUser.
This was modeled to have a cost of 1, but since we do not have a MUL.2d this is
scalarized into vector inserts/extracts and scalar muls.
Motivating precommitted test is test/Transforms/SLPVectorizer/AArch64/mul.ll,
which we don't want to SLP vectorize.
Test Transforms/LoopVectorize/AArch64/extractvalue-no-scalarization-required.ll
unfortunately needed changing, but the reason is documented in
LoopVectorize.cpp:6855:
// The cost of executing VF copies of the scalar instruction. This opcode
// is unknown. Assume that it is the same as 'mul'.
which I will address next as a follow up of this.
Differential Revision: https://reviews.llvm.org/D92208
Similar to other patches, this makes VPWidenRecipe a VPValue. Because of
the way it interacts with the reduction code it also slightly alters the
way that VPValues are registered, removing the up front NeedDef and
using getOrAddVPValue to create them on-demand if needed instead.
Differential Revision: https://reviews.llvm.org/D88447
This converts the VPReductionRecipe into a VPValue, like other
VPRecipe's in preparation for traversing def-use chains. It also makes
it a VPUser, now storing the used VPValues as operands.
It doesn't yet change how the VPReductionRecipes are created. It will
need to call replaceAllUsesWith from the original recipe they replace,
but that is not done yet as VPWidenRecipe need to be created first.
Differential Revision: https://reviews.llvm.org/D88382
Fix PR47390.
The primary induction should be considered alive when folding tail by masking,
because it will be used by said masking; even when it may otherwise appear
useless: feeding only its own 'bump', which is correctly considered dead, and
as the 'bump' of another induction variable, which may wrongfully want to
consider its bump = the primary induction, dead.
Differential Revision: https://reviews.llvm.org/D92017
A uniform load is one which loads from a uniform address across all lanes. As currently implemented, we cost model such loads as if we did a single scalar load + a broadcast, but the actual lowering replicates the load once per lane.
This change tweaks the lowering to use the REPLICATE strategy by marking such loads (and the computation leading to their memory operand) as uniform after vectorization. This is a useful change in itself, but it's real purpose is to pave the way for a following change which will generalize our uniformity logic.
In review discussion, there was an issue raised with coupling cost modeling with the lowering strategy for uniform inputs. The discussion on that item remains unsettled and is pending larger architectural discussion. We decided to move forward with this patch as is, and revise as warranted once the bigger picture design questions are settled.
Differential Revision: https://reviews.llvm.org/D91398
This is re-applying a combination of f7eac51b9b and 8ec7ea3ddc as one patch
to avoid regressions now that we have better testing in place.
Those were reverted with 32dd5870ee because of crashing in experimental intrinsics.
That bug should be fixed with 7ae346434.
Paraphrased original commit messages:
This is the last step in removing cost-kind as a consideration in the
basic class model for intrinsics.
See D89461 for the start of that.
Subsequent commits dealt with each of the special-case intrinsics that
had customization here in the basic class. This should remove a barrier
to retrying D87188 (canonicalization to the abs intrinsic).
The ARM and x86 cost diffs seen here may be wrong because the
target-specific overrides have their own bugs, but we hope this is
less wrong - if something has a significant throughput cost, then it
should have a significant size / blended cost too by default.
The only behavioral diff in current regression tests is shown in the
x86 scatter-gather test (which is misplaced or broken because it runs
the entire -O3 pipeline) - we unrolled less, and we assume that is
a improvement.
Exception: in general, we want the *size* cost for a scalar call to be
cheap even if the other costs are expensive - we expect it to just be
a branch with some optional stack manipulation.
It is likely that we will want to carve out some
exceptions/overrides to this rule as follow-up patches for
calls that have some general and/or target-specific difference
to the expected lowering.
This was noticed as a regression in unrolling, so we have a test
for that now along with a couple of direct cost model tests.
If the assumed scalarization costs for the oversized vector
calls are not realistic, that would be another follow-up
refinement of the cost models.
Differential Revision: https://reviews.llvm.org/D90554
as it's causing crashes in the optimizer. A reduced testcase has been posted as a follow-up.
This reverts commit f7eac51b9b.
Temporarily Revert "[CostModel] make default size cost for libcalls small (again)" as it depends upon the primary revert.
This reverts commit 8ec7ea3ddc.
Temporarily Revert "[CostModel] add tests for math library calls; NFC" as it depends upon the primary revert.
This reverts commit df09f82599.
Temporarily Revert "[LoopUnroll] add test for full unroll that is sensitive to cost-model; NFC" as it depends upon the primary revert.
This reverts commit 618d555e8d.
I noticed an add example like the one from D91343, so here's a similar patch.
The logic is based on existing code for the single-use demanded bits fold.
But I only matched a constant instead of using compute known bits on the
operands because that was the motivating patterni that I noticed.
I think this will allow removing a special-case (but incomplete) dedicated
fold within visitAnd(), but I need to untangle the existing code to be sure.
https://rise4fun.com/Alive/V6fP
Name: add with low mask
Pre: (C1 & (-1 u>> countLeadingZeros(C2))) == 0
%a = add i8 %x, C1
%r = and i8 %a, C2
=>
%r = and i8 %x, C2
Differential Revision: https://reviews.llvm.org/D91415
This patch turns VPWidenGEPRecipe into a VPValue and uses it
during VPlan construction and codegeneration instead of the plain IR
reference where possible.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D84683
This patch turns VPWidenSelectRecipe into a VPValue and uses it
during VPlan construction and codegeneration instead of the plain IR
reference where possible.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D84682
This is the last step in removing cost-kind as a consideration in the basic class model for intrinsics.
See D89461 for the start of that.
Subsequent commits dealt with each of the special-case intrinsics that had customization here in the
basic class. This should remove a barrier to retrying
D87188 (canonicalization to the abs intrinsic).
The ARM and x86 cost diffs seen here may be wrong because the target-specific overrides have their own
bugs, but we hope this is less wrong - if something has a significant throughput cost, then it should
have a significant size / blended cost too by default.
The only behavioral diff in current regression tests is shown in the x86 scatter-gather test (which is
misplaced or broken because it runs the entire -O3 pipeline) - we unrolled less, and we assume that is
a improvement.
Differential Revision: https://reviews.llvm.org/D90554
This patch turns VPWidenCall into a VPValue and uses it
during VPlan construction and codegeneration instead of the plain IR
reference where possible.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D84681
Claim to not have any vector support to dissuade SLP, LV and friends
from generating SIMD IR for the VE target. We will take this back once
vector isel is stable.
Reviewed By: kaz7, fhahn
Differential Revision: https://reviews.llvm.org/D90462
CallInst::updateProfWeight() creates branch_weights with i64 instead of i32.
To be more consistent everywhere and remove lots of casts from uint64_t
to uint32_t, use i64 for branch_weights.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D88609
Use -0.0 instead of 0.0 as the start value. The previous use of 0.0
was fine for all existing uses of this function though, as it is
always generated with fast flags right now, and thus nsz.
When trying to prove that a memory access touches only dereferenceable memory across all iterations of a loop, use the maximum exit count rather than an exact one. In many cases we can't prove exact exit counts whereas we can prove an upper bound.
The test included is for a single exit loop with a min(C,V) exit count, but the true motivation is support for multiple exits loops. It's just really hard to write a test case for multiple exits because the vectorizer (the primary user of this API), bails far before this. For multiple exits, this allows a mix of analyzeable and unanalyzable exits when only analyzeable exits are needed to prove deref.
-Oz normally does not allow loop header duplication so this loop wouldn't be
vectorized. However the vectorization pragma should override this and allow
for loop rotation.
rdar://problem/49281061
Original patch by Adam Nemet.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D59832
CallInst::updateProfWeight() creates branch_weights with i64 instead of i32.
To be more consistent everywhere and remove lots of casts from uint64_t
to uint32_t, use i64 for branch_weights.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D88609
The warning would fire when calling isDereferenceableAndAlignedInLoop
with a scalable load. Calling isDereferenceableAndAlignedInLoop with a
scalable load would result in the use of the now deprecated implicit
cast of TypeSize to uint64_t through the overloaded operator.
This patch fixes this issue by:
- no longer considering vector loads as candidates in
canVectorizeWithIfConvert. This doesn't make sense in the context of
identifying scalar loads to vectorize.
- making use of getFixedSize inside isDereferenceableAndAlignedInLoop --
this removes the dependency on the deprecated interface, and will
trigger an assertion error if the function is ever called with a
scalable type.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D89798
We should first try to constant fold the add expression and only
strengthen nowrap flags afterwards. This allows us to determine
stronger flags if e.g. only two operands are left after constant
folding (and thus "guaranteed no wrap region" code applies) or the
resulting operands are non-negative and thus nsw->nuw strengthening
applies.
LV fails with assertion checking that UF > 0. We already set UF to 1 if it is 0 except the case when IC > MaxInterleaveCount. The fix is to set UF to 1 for that case as well.
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D87679
This expands upon the inloop reductions added in e9761688e41cb9e976,
allowing them to be inserted into tail folded loops. Reductions are
generates with the form:
x = select(mask, vecop, zero)
v = vecreduce.add(x)
c = add chain, v
Where zero here is chosen as the identity value for add reductions. The
backend is then expected to fold the select and the vecreduce into a
single predicated instruction.
Most of the code is fairly straight forward, except for the creation of
blockmasks which need to ensure they are created in dominance order. The
order they are added is altered to be after any phis, keeping the
requirements for the underlying IR.
Differential Revision: https://reviews.llvm.org/D84451
We currently collect the ICmp and Add from an induction variable,
marking them as dead so that vplan values are not created for them. This
extends that to include any single use trunk from the ICmp, which allows
the Add to more readily be removed too.
This can help with costing vplan nodes, as the ICmp and Add are more
reliably removed and are not double-counted.
Differential Revision: https://reviews.llvm.org/D88873
Currently LAA uses getScalarSizeInBits to compute the size of an element
when computing the end bound of an access.
This does not work as expected for pointers to pointers, because
getScalarSizeInBits will return 0 for pointer types.
By using DataLayout to get the size of the element we can also correctly
handle pointer element types.
Note the changes to the existing test, which seems to also use the wrong
offset for the end.
Fixes PR47751.
Reviewed By: anemet
Differential Revision: https://reviews.llvm.org/D88953
Regarding this bug I posted earlier: https://bugs.llvm.org/show_bug.cgi?id=47035
After reading through LLVM source code and getting familiar with VPlan I was able to vectorize the code using by enabling VPlan native path. After talking with @fhahn he suggested that I contribute this as a test case. So here it is. I tried to follow the available guides how to do this best I could. I modified IR code by hand to have more clear variable names instead of numbers.
One thing what I'd like to get input from someone is that is current CHECK lines sufficient enough to verify that the inner loop has been vectorized properly?
Reviewed By: fhahn
Differential Revision: https://reviews.llvm.org/D87564
We have been running tests/benchmarks downstream with tail-predication enabled
for some time now and this behaves as expected: we are not aware of any
correctness issues, and this performs better across the board than with
tail-predication disabled. Time to flip the switch!
Differential Revision: https://reviews.llvm.org/D88093
For some expressions, we can use information from loop guards when
we are looking for a maximum. This patch applies information from
loop guards to the expression used to compute the maximum backedge
taken count in howFarToZero. It currently replaces an unknown
expression X with UMin(X, Y), if the loop is guarded by
X ult Y.
This patch is minimal in what conditions it applies, and there
are a few TODOs to generalize.
This partly addresses PR40961. We will also need an update to
LV to address it completely.
Reviewed By: reames
Differential Revision: https://reviews.llvm.org/D67178
Although LLVM supports vectorization of loops containing log10/sqrt, it did not support using SVML implementation of it. Added support so that when clang is invoked with -fveclib=SVML now an appropriate SVML library log2 implementation will be invoked.
Follow up on: https://reviews.llvm.org/D77114
Tests:
Added unit tests to svml-calls.ll, svml-calls-finite.ll. Can be run with llvm-lint.
Created a simple c++ file that tests log10/sqrt, and used clang+ to build it, and output final assembly.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D87169
This allows the backend to tell the vectorizer to produce inloop
reductions through a TTI hook.
For the moment on ARM under MVE this means allowing integer add
reductions of the correct size. In the future this can include integer
min/max too, under -Os.
Differential Revision: https://reviews.llvm.org/D75512
Although LLVM supports vectorization of loops containing log2, it did not support using SVML implementation of it. Added support so that when clang is invoked with -fveclib=SVML now an appropriate SVML library log2 implementation will be invoked.
Follow up on: https://reviews.llvm.org/D77114
Tests:
Added unit tests to svml-calls.ll, svml-calls-finite.ll. Can be run with llvm-lint.
Created a simple c++ file that tests log2, and used clang+ to build it, and output final assembly.
Reviewed By: wenlei, craig.topper
Differential Revision: https://reviews.llvm.org/D86730
Interleave for small loops that have reductions inside,
which breaks dependencies and expose.
This gives very significant performance improvements for some benchmarks.
Because small loops could be in very hot functions in real applications.
Differential Revision: https://reviews.llvm.org/D81416
addRuntimeChecks uses SCEVExpander, which relies on the DT/LoopInfo to
be up-to-date. Changing the CFG afterwards may invalidate some inserted
instructions, especially LCSSA phis.
Reorder the code to first update the CFG and then create the runtime
checks. This should not have any impact on the generated code, as we
adjust the CFG and generate runtime checks together.
Fixes PR47343.
The original take 1 was 6102310d81,
which taught InstSimplify to do that, which seemed better at time,
since we got EarlyCSE support for free.
However, it was proven that we can not do that there,
the simplified-to PHI would not be reachable from the original PHI,
and that is not something InstSimplify is allowed to do,
as noted in the commit ed90f15efb
that reverted it:
> It appears to cause compilation non-determinism and caused stage3 mismatches.
Then there was take 2 3e69871ab5,
which was InstCombine-specific, but it again showed stage2-stage3 differences,
and reverted in bdaa3f86a0.
This is quite alarming.
Here, let's try to change how we find existing PHI candidate:
due to the worklist order, and the way PHI nodes are inserted
(it may be inserted as the first one, or maybe not), let's look at *all*
PHI nodes in the block.
Effects on vanilla llvm test-suite + RawSpeed:
```
| statistic name | baseline | proposed | Δ | % | \|%\| |
|----------------------------------------------------|-----------|-----------|-------:|---------:|---------:|
| asm-printer.EmittedInsts | 7942329 | 7942457 | 128 | 0.00% | 0.00% |
| assembler.ObjectBytes | 254295632 | 254312480 | 16848 | 0.01% | 0.01% |
| correlated-value-propagation.NumPhis | 18412 | 18347 | -65 | -0.35% | 0.35% |
| early-cse.NumCSE | 2183283 | 2183267 | -16 | 0.00% | 0.00% |
| early-cse.NumSimplify | 550105 | 541842 | -8263 | -1.50% | 1.50% |
| instcombine.NumAggregateReconstructionsSimplified | 73 | 4506 | 4433 | 6072.60% | 6072.60% |
| instcombine.NumCombined | 3640311 | 3644419 | 4108 | 0.11% | 0.11% |
| instcombine.NumDeadInst | 1778204 | 1783205 | 5001 | 0.28% | 0.28% |
| instcombine.NumPHICSEs | 0 | 22490 | 22490 | 0.00% | 0.00% |
| instcombine.NumWorklistIterations | 2023272 | 2024400 | 1128 | 0.06% | 0.06% |
| instcount.NumCallInst | 1758395 | 1758802 | 407 | 0.02% | 0.02% |
| instcount.NumInvokeInst | 59478 | 59502 | 24 | 0.04% | 0.04% |
| instcount.NumPHIInst | 330557 | 330545 | -12 | 0.00% | 0.00% |
| instcount.TotalBlocks | 1077138 | 1077220 | 82 | 0.01% | 0.01% |
| instcount.TotalFuncs | 101442 | 101441 | -1 | 0.00% | 0.00% |
| instcount.TotalInsts | 8831946 | 8832606 | 660 | 0.01% | 0.01% |
| simplifycfg.NumHoistCommonCode | 24186 | 24187 | 1 | 0.00% | 0.00% |
| simplifycfg.NumInvokes | 4300 | 4410 | 110 | 2.56% | 2.56% |
| simplifycfg.NumSimpl | 1019813 | 999767 | -20046 | -1.97% | 1.97% |
```
So it fires 22490 times, which is less than ~24k the take 1 did,
but more than what take 2 did (22228 times)
.
It allows foldAggregateConstructionIntoAggregateReuse() to actually work
after PHI-of-extractvalue folds did their thing. Previously SimplifyCFG
would have done this PHI CSE, of all places. Additionally, allows some
more `invoke`->`call` folds to happen (+110, +2.56%).
All in all, expectedly, this catches less things overall,
but all the motivational cases are still caught, so all good.
While the original variant with doing this in InstSimplify (rightfully)
caused questions and ultimately was detected to be a culprit
of stage2-stage3 mismatch, it was expected that
InstCombine-based implementation would be fine.
But apparently it's not, as
http://lab.llvm.org:8011/builders/clang-with-thin-lto-ubuntu/builds/24095/steps/compare-compilers/logs/stdio
suggests.
Which suggests that somewhere in InstCombine there is a loop
over nondeterministically sorted container, which causes
different worklist ordering.
This reverts commit 3e69871ab5.
The original take was 6102310d81,
which taught InstSimplify to do that, which seemed better at time,
since we got EarlyCSE support for free.
However, it was proven that we can not do that there,
the simplified-to PHI would not be reachable from the original PHI,
and that is not something InstSimplify is allowed to do,
as noted in the commit ed90f15efb
that reverted it :
> It appears to cause compilation non-determinism and caused stage3 mismatches.
However InstCombine already does many different optimizations,
so it should be a safe place to do it here.
Note that we still can't just compare incoming values ranges,
because there is no guarantee that these PHI's we'd simplify to
were already re-visited and sorted.
However coming up with a test is problematic.
Effects on vanilla llvm test-suite + RawSpeed:
```
| statistic name | baseline | proposed | Δ | % | |%| |
|----------------------------------------------------|-----------|-----------|-------:|---------:|---------:|
| instcombine.NumPHICSEs | 0 | 22228 | 22228 | 0.00% | 0.00% |
| asm-printer.EmittedInsts | 7942329 | 7942456 | 127 | 0.00% | 0.00% |
| assembler.ObjectBytes | 254295632 | 254313792 | 18160 | 0.01% | 0.01% |
| early-cse.NumCSE | 2183283 | 2183272 | -11 | 0.00% | 0.00% |
| early-cse.NumSimplify | 550105 | 541842 | -8263 | -1.50% | 1.50% |
| instcombine.NumAggregateReconstructionsSimplified | 73 | 4506 | 4433 | 6072.60% | 6072.60% |
| instcombine.NumCombined | 3640311 | 3666911 | 26600 | 0.73% | 0.73% |
| instcombine.NumDeadInst | 1778204 | 1783318 | 5114 | 0.29% | 0.29% |
| instcount.NumCallInst | 1758395 | 1758804 | 409 | 0.02% | 0.02% |
| instcount.NumInvokeInst | 59478 | 59502 | 24 | 0.04% | 0.04% |
| instcount.NumPHIInst | 330557 | 330549 | -8 | 0.00% | 0.00% |
| instcount.TotalBlocks | 1077138 | 1077221 | 83 | 0.01% | 0.01% |
| instcount.TotalFuncs | 101442 | 101441 | -1 | 0.00% | 0.00% |
| instcount.TotalInsts | 8831946 | 8832611 | 665 | 0.01% | 0.01% |
| simplifycfg.NumInvokes | 4300 | 4410 | 110 | 2.56% | 2.56% |
| simplifycfg.NumSimpl | 1019813 | 999740 | -20073 | -1.97% | 1.97% |
```
So it fires ~22k times, which is less than ~24k the take 1 did.
It allows foldAggregateConstructionIntoAggregateReuse() to actually work
after PHI-of-extractvalue folds did their thing. Previously SimplifyCFG
would have done this PHI CSE, of all places. Additionally, allows some
more `invoke`->`call` folds to happen (+110, +2.56%).
All in all, expectedly, this catches less things overall,
but all the motivational cases are still caught, so all good.
Apparently, we don't do this, neither in EarlyCSE, nor in InstSimplify,
nor in (old) GVN, but do in NewGVN and SimplifyCFG of all places..
While i could teach EarlyCSE how to hash PHI nodes,
we can't really do much (anything?) even if we find two identical
PHI nodes in different basic blocks, same-BB case is the interesting one,
and if we teach InstSimplify about it (which is what i wanted originally,
https://reviews.llvm.org/D86530), we get EarlyCSE support for free.
So i would think this is pretty uncontroversial.
On vanilla llvm test-suite + RawSpeed, this has the following effects:
```
| statistic name | baseline | proposed | Δ | % | \|%\| |
|----------------------------------------------------|-----------|-----------|-------:|---------:|---------:|
| instsimplify.NumPHICSE | 0 | 23779 | 23779 | 0.00% | 0.00% |
| asm-printer.EmittedInsts | 7942328 | 7942392 | 64 | 0.00% | 0.00% |
| assembler.ObjectBytes | 273069192 | 273084704 | 15512 | 0.01% | 0.01% |
| correlated-value-propagation.NumPhis | 18412 | 18539 | 127 | 0.69% | 0.69% |
| early-cse.NumCSE | 2183283 | 2183227 | -56 | 0.00% | 0.00% |
| early-cse.NumSimplify | 550105 | 542090 | -8015 | -1.46% | 1.46% |
| instcombine.NumAggregateReconstructionsSimplified | 73 | 4506 | 4433 | 6072.60% | 6072.60% |
| instcombine.NumCombined | 3640264 | 3664769 | 24505 | 0.67% | 0.67% |
| instcombine.NumDeadInst | 1778193 | 1783183 | 4990 | 0.28% | 0.28% |
| instcount.NumCallInst | 1758401 | 1758799 | 398 | 0.02% | 0.02% |
| instcount.NumInvokeInst | 59478 | 59502 | 24 | 0.04% | 0.04% |
| instcount.NumPHIInst | 330557 | 330533 | -24 | -0.01% | 0.01% |
| instcount.TotalInsts | 8831952 | 8832286 | 334 | 0.00% | 0.00% |
| simplifycfg.NumInvokes | 4300 | 4410 | 110 | 2.56% | 2.56% |
| simplifycfg.NumSimpl | 1019808 | 999607 | -20201 | -1.98% | 1.98% |
```
I.e. it fires ~24k times, causes +110 (+2.56%) more `invoke` -> `call`
transforms, and counter-intuitively results in *more* instructions total.
That being said, the PHI count doesn't decrease that much,
and looking at some examples, it seems at least some of them
were previously getting PHI CSE'd in SimplifyCFG of all places..
I'm adjusting `Instruction::isIdenticalToWhenDefined()` at the same time.
As a comment in `InstCombinerImpl::visitPHINode()` already stated,
there are no guarantees on the ordering of the operands of a PHI node,
so if we just naively compare them, we may false-negatively say that
the nodes are not equal when the only difference is operand order,
which is especially important since the fold is in InstSimplify,
so we can't rely on InstCombine sorting them beforehand.
Fixing this for the general case is costly (geomean +0.02%),
and does not appear to catch anything in test-suite, but for
the same-BB case, it's trivial, so let's fix at least that.
As per http://llvm-compile-time-tracker.com/compare.php?from=04879086b44348cad600a0a1ccbe1f7776cc3cf9&to=82bdedb888b945df1e9f130dd3ac4dd3c96e2925&stat=instructions
this appears to cause geomean +0.03% compile time increase (regression),
but geomean -0.01%..-0.04% code size decrease (improvement).
This implements 2 different vectorisation fallback strategies if tail-folding
fails: 1) don't vectorise at all, or 2) vectorise using a scalar epilogue. This
can be controlled with option -prefer-predicate-over-epilogue, that has been
changed to take a numeric value corresponding to the tail-folding preference
and preferred fallback.
Patch by: Pierre van Houtryve, Sjoerd Meijer.
Differential Revision: https://reviews.llvm.org/D79783
MVE Gather scatter codegeneration is looking a lot better than it used
to, but still has some issues. The instructions we currently model as 1
cycle per element, which is a bit low for some cases. Increasing the
cost by the MVECostFactor brings them in-line with our other instruction
costs. This will have the effect of only generating then when the extra
benefit is more likely to overcome some of the issues. Notably in
running out of registers and vectorizing loops that could otherwise be
SLP vectorized.
In the short-term whilst we look at other ways of dealing with those
more directly, we can increase the costs of gathers to make them more
likely to be beneficial when created.
Differential Revision: https://reviews.llvm.org/D86444
The legacy LoopVectorize has a dependency on InjectTLIMappingsLegacy.
That cannot be expressed in the new PM since they are both normal
passes. Explicitly add -inject-tli-mappings as a pass.
Follow-up to https://reviews.llvm.org/D86492.
Reviewed By: spatel
Differential Revision: https://reviews.llvm.org/D86561
This adapts LV to the new semantics of get.active.lane.mask as discussed in
D86147, which means that the LV now emits intrinsic get.active.lane.mask with
the loop tripcount instead of the backedge-taken count as its second argument.
The motivation for this is described in D86147.
Differential Revision: https://reviews.llvm.org/D86304
If gather/scatters are enabled, ARMTargetTransformInfo now allows
tail predication for loops with a much wider range of strides, up
to anything that is loop invariant.
Differential Revision: https://reviews.llvm.org/D85410
As part of D84741, this adds a target hook for the
preferPredicatedReductionSelect option and makes use
of it under MVE, allowing us to tail predicate most
reduction loops.
Differential Revision: https://reviews.llvm.org/D85980
The normal scheme for tail folding reductions is to use:
loop:
p = phi(0, a)
mask = ...
x = masked_load(..., mask)
a = add(x, p)
s = select(mask, a, p)
This means we need to keep the register p and a alive out of the loop, plus
the mask. On a target with predicated operations we can instead generate
the phi as p = phi(0, s). This ensures the select in the loop and we can
fold select(m, add(a, b), c) to something like a vaddt c, a, b using the
m predicate. This in turn allows us to tail predicate the entire loop.
Differential Revision: https://reviews.llvm.org/D84741
D81345 appears to accidentally disables vectorization when explicitly
enabled. As PGSO isn't currently accessible from LoopAccessInfo, revert back to
the vectorization with versioning-for-unit-stride for PGSO.
Differential Revision: https://reviews.llvm.org/D85784
Arm MVE has multiple instructions such as VMLAVA.s8, which (in this
case) can take two 128bit vectors, sign extend the inputs to i32,
multiplying them together and sum the result into a 32bit general
purpose register. So taking 16 i8's as inputs, they can multiply and
accumulate the result into a single i32 without any rounding/truncating
along the way. There are also reduction instructions for plain integer
add and min/max, and operations that sum into a pair of 32bit registers
together treated as a 64bit integer (even though MVE does not have a
plain 64bit addition instruction). So giving the vectorizer the ability
to use these instructions both enables us to vectorize at higher
bitwidths, and to vectorize things we previously could not.
In order to do that we need a way to represent that the reduction
operation, specified with a llvm.experimental.vector.reduce when
vectorizing for Arm, occurs inside the loop not after it like most
reductions. This patch attempts to do that, teaching the vectorizer
about in-loop reductions. It does this through a vplan recipe
representing the reductions that the original chain of reduction
operations is replaced by. Cost modelling is currently just done through
a prefersInloopReduction TTI hook (which follows in a later patch).
Differential Revision: https://reviews.llvm.org/D75069
This reverts commit e9761688e4. It breaks the build:
```
~/src/llvm-project/llvm/lib/Analysis/IVDescriptors.cpp:868:10: error: no viable conversion from returned value of type 'SmallVector<[...], 8>' to function return type 'SmallVector<[...], 4>'
return ReductionOperations;
```
Arm MVE has multiple instructions such as VMLAVA.s8, which (in this
case) can take two 128bit vectors, sign extend the inputs to i32,
multiplying them together and sum the result into a 32bit general
purpose register. So taking 16 i8's as inputs, they can multiply and
accumulate the result into a single i32 without any rounding/truncating
along the way. There are also reduction instructions for plain integer
add and min/max, and operations that sum into a pair of 32bit registers
together treated as a 64bit integer (even though MVE does not have a
plain 64bit addition instruction). So giving the vectorizer the ability
to use these instructions both enables us to vectorize at higher
bitwidths, and to vectorize things we previously could not.
In order to do that we need a way to represent that the reduction
operation, specified with a llvm.experimental.vector.reduce when
vectorizing for Arm, occurs inside the loop not after it like most
reductions. This patch attempts to do that, teaching the vectorizer
about in-loop reductions. It does this through a vplan recipe
representing the reductions that the original chain of reduction
operations is replaced by. Cost modelling is currently just done through
a prefersInloopReduction TTI hook (which follows in a later patch).
Differential Revision: https://reviews.llvm.org/D75069
No widening decisions will be computed for instructions outside the
loop. Do not try to get a widening decision. The load/store will be just
a scalar load, so treating at as normal should be fine I think.
Fixes PR46950.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D85087
If an analysis is actually invalidated, there's already a log statement
for that: 'Invalidating analysis: FooAnalysis'.
Otherwise the statement is not very useful.
Reviewed By: asbirlea, ychen
Differential Revision: https://reviews.llvm.org/D84981
This removes some unneeded block masks when we don't have any
reductions. It should not have any effect on codegen as the values
created are dead anyway.
Differential Revision: https://reviews.llvm.org/D81415
This patch uses the feature added in D79162 to fix the cost of a
sext/zext of a masked load, or a trunc for a masked store.
Previously, those were considered cheap or even free, but it's
not the case as we cannot split the load in the same way we would for
normal loads.
This updates the costs to better reflect reality, and adds a test for it
in test/Analysis/CostModel/ARM/cast.ll.
It also adds a vectorizer test that showcases the improvement: in some
cases, the vectorizer will now choose a smaller VF when
tail-predication is enabled, which results in better codegen. (Because
if it were to use a higher VF in those cases, the code we see above
would be generated, and the vmovs would block tail-predication later in
the process, resulting in very poor codegen overall)
Original Patch by Pierre van Houtryve
Differential Revision: https://reviews.llvm.org/D79163
Summary: To match NewPM name. Also the new name is clearer and more consistent.
Subscribers: jvesely, nhaehnle, hiraditya, asbirlea, kerbowa, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D84542
It was getting difficult to see which test was in which file, so this
reorganises the test files so that now all filenames start with tail-folding-*
followed by a more descriptive name what that group of tests check.
This patch enables the LoopVectorizer to build a phi of pointer
type and provide the vector loads and stores with vector type
getelementptrs built from the pointer induction variable, which
produces much less instructions than the previous approach of
creating scalar getelementpointers and glue them together to a
vector.
Differential Revision: https://reviews.llvm.org/D81267
If a vector body has live-out values, it is probably a reduction, which needs a
final reduction step after the loop. MVE has a VADDV instruction to reduce
integer vectors, but doesn't have an equivalent one for float vectors. A
live-out value that is not recognised as reduction later in the optimisation
pipeline will result in the tail-predicated loop to be reverted to a
non-predicated loop and this is very expensive, i.e. it has a significant
performance impact, which is what we hope to avoid with fine tuning the ARM TTI
hook preferPredicateOverEpilogue implementation.
Differential Revision: https://reviews.llvm.org/D82953
This refactors option -disable-mve-tail-predication to take different arguments
so that we have 1 option to control tail-predication rather than several
different ones.
This is also a prep step for D82953, in which we want to reject reductions
unless that is requested with this option.
Differential Revision: https://reviews.llvm.org/D83133
This patch fixes D81345 and PR46652.
If a loop with a small trip count is compiled w/o -Os/-Oz, Loop Access Analysis
still generates runtime checks for unit strides that will version the loop.
In such cases, the loop vectorizer should either re-run the analysis or bail-out
from vectorizing the loop, as done prior to D81345. The latter is applied for
now as the former requires refactoring.
Differential Revision: https://reviews.llvm.org/D83470
Currently the DomTree is not kept up to date for additional blocks
generated in the vector loop, for example when vectorizing with
predication. SCEVExpander relies on dominance checks when looking for
existing instructions to re-use and in some cases that can lead to the
expander picking instructions that do not actually dominate their insert
point (e.g. as in PR46525).
Unfortunately keeping the DT up-to-date is a bit tricky, because the CFG
is only patched up after generating code for a block. For now, we can
just use the vector loop header, as this ensures the inserted
instructions dominate all uses in the vector loop. There should be no
noticeable impact on the generated code, as other passes should sink
those instructions, if profitable.
Fixes PR46525.
Reviewers: Ayal, gilr, mkazantsev, dmgreen
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D83288
If a loop is in a function marked OptSize, Loop Access Analysis should refrain
from generating runtime checks for unit strides that will version the loop.
If a loop is in a function marked OptSize and its vectorization is enabled, it
should be vectorized w/o any versioning.
Fixes PR46228.
Differential Revision: https://reviews.llvm.org/D81345
This adjusts the MVE fp16 cost model, similar to how we already do for
integer casts. It uses the base cost of 1 per cvt for most fp extend /
truncates, but adjusts it for loads and stores where we know that a
extending load has been used to get the load into the correct lane, and
only an MVE VCVTB is then needed.
Differential Revision: https://reviews.llvm.org/D81813
This alters getMemoryOpCost to use the Base TargetTransformInfo version
that includes some additional checks for whether extending loads are
legal. This will generally have the effect of making <2 x ..> and some
<4 x ..> loads/stores more expensive, which in turn should help favour
larger vector factors.
Notably it alters the cost of a <4 x half>, which with the current
codegen will be expensive if it is not extended.
Differential Revision: https://reviews.llvm.org/D82456
This patch enables the LoopVectorizer to build a phi of pointer
type and provide the vector loads and stores with vector type
getelementptrs built from the pointer induction variable, which
produces much less instructions than the previous approach of
creating scalar getelementpointers and glue them together to a
vector.
Differential Revision: https://reviews.llvm.org/D81267
D79164/2596da31740f changed getCFInstrCost to return 1 per default.
AArch64 did not have its own implementation, hence the throughput cost
of CFI instructions is overestimated. On most cores, most branches should
be predicated and essentially free throughput wise.
This restores a 9% performance regression on a SPEC2006 benchmark on
AArch64 with -O3 LTO & PGO.
This patch effectively restores pre 2596da3174 behavior for AArch64
and undoes the AArch64 test changes of the patch.
Reviewers: samparker, dmgreen, anemet
Reviewed By: samparker
Differential Revision: https://reviews.llvm.org/D82755
This emits new IR intrinsic @llvm.get.active.mask for tail-folded vectorised
loops if the intrinsic is supported by the backend, which is checked by
querying TargetTransform hook emitGetActiveLaneMask.
This intrinsic creates a mask representing active and inactive vector lanes,
which is used by the masked load/store instructions that are created for
tail-folded loops. The semantics of @llvm.get.active.mask are described here in
LangRef:
https://llvm.org/docs/LangRef.html#llvm-get-active-lane-mask-intrinsics
This intrinsic is also used to provide a hint to the backend. That is, the
second argument of the intrinsic represents the back-edge taken count of the
loop. For MVE, for example, we use that to set up tail-predication, which is a
new form of predication in MVE for vector loops that implicitely predicates the
last vector loop iteration by implicitely setting active/inactive lanes, i.e.
the tail loop is predicated. In order to set up a tail-predicated vector loop,
we need to know the number of data elements processed by the vector loop, which
corresponds the the tripcount of the scalar loop, which we can now reconstruct
using @llvm.get.active.mask.
Differential Revision: https://reviews.llvm.org/D79100
Have BasicTTI call the base implementation so that both agree on the
default behaviour, which the default being a cost of '1'. This has
required an X86 specific implementation as it seems to be very
reliant on those instructions being free. Changes are also made to
AMDGPU so that their implementations distinguish between cost kinds,
so that the unrolling isn't affected. PowerPC also has its own
implementation to prevent changes to the reg-usage vectorizer test.
The cost model test changes now reflect that ret instructions are not
generally free.
Differential Revision: https://reviews.llvm.org/D79164
Alternative approach to D80570.
canCheckPtrAtRT already contains checks the figure out for which alias
sets runtime checks are needed. But it currently sets CanDoRT to false
for alias sets for which we cannot do RT checks but also do not need
any.
If we know that we do not need RT checks based on the number of
reads/writes in the alias set, we can skip processing the AS.
This patch also adds an assertion to ensure that DepCands does not
contain more than one write from the alias set.
Reviewers: Ayal, anemet, hfinkel, dmgreen
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D80622
Similar to a recent change to the X86 backend, this changes things so
that we always produce a reduction intrinsics for all reduction types,
not just the legal ones. This gives a better chance in the backend to
custom lower them to something more suitable for MVE. Especially for
something like fadd the in-order reduction produced during DAG lowering
is already better than the shuffles produced in the midend, and we can
do even better with a bit of custom lowering.
Differential Revision: https://reviews.llvm.org/D81398
getOrCreateTripCount is used to generate code for the outer loop, but it
requires a computable backedge taken counts. Check that in the VPlan
native path.
Reviewers: Ayal, gilr, rengolin, sguggill
Reviewed By: sguggill
Differential Revision: https://reviews.llvm.org/D81088
Motivating examples are seen in the PhaseOrdering tests based on:
https://bugs.llvm.org/show_bug.cgi?id=43953#c2 - if we have
intrinsics there, some pass can fold them.
The intrinsics are still named "experimental" at this point, but
if there is no fallout from this patch, that will be a good
indicator that it is safe to finalize them.
Differential Revision: https://reviews.llvm.org/D80867
LV currently only supports power of 2 vectorization factors, which has
been made explicit with the assertion added in
840450549c.
However, if the widest type is not a power-of-2 the computed MaxVF won't
be a power-of-2 either. This patch updates computeFeasibleMaxVF to
ensure the returned value is a power-of-2 by rounding down to the
nearest power-of-2.
Fixes PR46139.
Reviewers: Ayal, gilr, rengolin
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D80870
The -reassociate pass tends to transform this kind of pattern into
something that is worse for vectorization and codegen. See PR43953:
https://bugs.llvm.org/show_bug.cgi?id=43953
Follows-up the FP version of the same transform:
rGa0ce2338a083
Philip Reames