This change enables the use of RISCV's variable length vector registers for fixed length vectors in the IR, and implicitly enables various IR transforms which generate fixed length vectors if legal (e.g. LoopVectorize). Specifically, this enables fixed length vectors which are known to be inbounds of the underlying variable hardware size.
For context, remember that the +V extension provides a minimum VLEN of 128. The embedded variants provide lower minimums. The analogy here is essentially vectorizing for SSE on a machine which may or may not include AVX2/AVX512. We won't get full utilization by default, but we will get some benefit. And of course, with an explicit mcpu we can vectorize to the exact target hardware.
The LV impact is mostly related to vectorizer robustness. In cases we haven't yet fully implemented scalable vectorization support, we can fall back to fixed length vectorization.
SLP has been disabled for now, even when fixed vectors are enabled. See a310637 and associated review. There are a few addiitional code quality issues which need worked through before turning SLP on would be reasonable.
Differential Revision: https://reviews.llvm.org/D131508
The simpler diff-checks require pointers with add-recs from the same
innermost loop, but this property wasn't check completely. Add the
missing check to ensure both addrecs are in the innermost loop.
Fixes#57315.
Add a variation of @nested_loop_outer_iv_addrec_invariant_in_inner with
the dependence sink and source swapped to extend test coverage.
Also simplifies the test by removing an unneeded reduction.
When we have a dependency with a dependence distance which can only be hit on an iteration beyond the actual trip count of the loop, we can ignore that dependency when analyzing said loop. We already had this code, but had restricted it solely to unknown dependence distances. This change applies it to all dependence distances.
Without this code, we relied on the vectorizer reducing VF such that our infeasible dependence was respected. This usually worked out to about the same result, but not always. For fixed length vectorization, this could mean a smaller VF than optimal being chosen or additional runtime checks. For scalable vectorization - where the bounds on access implied by VF are broader - we could often not find a feasible VF at all.
Differential Revision: https://reviews.llvm.org/D131924
This patch adds support for vectorizing conditionally executed div/rem operations via a variant of widening. The existing support for predicated divrem in the vectorizer requires scalarization which we can't do for scalable vectors.
The basic idea is that we can always divide (take remainder) by 1 without executing UB. As such, we can use the active lane mask to conditional select either the actual divisor for active lanes, or a constant one for inactive lanes. We already account for the cost of the active lane mask, so the only additional cost is a splat of one and the vector select. This is one of several possible approaches to this problem; see the review thread for discussion on some of the others. This one was chosen mostly because it was straight forward, and none of the others seemed oviously better.
I enabled the new code only for scalable vectors. We could also legally enable it for fixed vectors as well, but I haven't thought through the cost tradeoffs between widening and scalarization enough to know if that's profitable. This will be explored in future patches.
Differential Revision: https://reviews.llvm.org/D130164
The existing cost model for fixed-order recurrences models the phi as an
extract shuffle of a v1 vector. The shuffle produced should be a splice,
as they take two vectors inputs are extracting from a subset of the
lanes. On certain architectures the existing cost model can drastically
under-estimate the correct cost for the shuffle, so this changes it to a
SK_Splice and passes a correct Mask through to the getShuffleCost call.
I believe this might be the first use of a SK_Splice shuffle cost model
outside of scalable vectors, and some targets may require additions to
the cost-model to correctly account for them. In tree targets appear to
all have been updated where needed.
Differential Revision: https://reviews.llvm.org/D132308
Don't demand low order bits from the LHS of an Add if:
- they are not demanded in the result, and
- they are known to be zero in the RHS, so they can't possibly
overflow and affect higher bit positions
This is intended to avoid a regression from a future patch to change
the order of canonicalization of ADD and AND.
Differential Revision: https://reviews.llvm.org/D130075
(X op Y) op Z --> (Y op Z) op X
This isn't a complete solution (see TODO tests for possible refinements),
but it shows some nice wins and doesn't seem to cause any harm. I think
the most potential danger is from conflicting with other folds and causing
an infinite loop - that's the reason for avoiding patterns with constant
operands.
Alternatively, we could try this in the reassociate pass, but we would not
immediately see all of the logic folds that instcombine provides. I also
looked at improving ValueTracking's isImpliedCondition() (and we should
still add some enhancements there), but that would not work in general for
bitwise logic reduction.
The tests that reduce completely to 0/-1 are motivated by issue #56653.
Differential Revision: https://reviews.llvm.org/D131356
Whilst writing a patch to add extra tail-folding RUN lines to
existing tests I noticed a few areas where they can be
cleaned up a little:
1. scalable-reductions.ll: fmin_fast does not mark fcmp as fast.
2. sve-inductions-unusual-types.ll: remove direct references to
SSA variable names.
3. sve-strict-fadd-cost.ll: don't force vector width so we see
costs for different VFs in one go. This will be important for
the follow-on patch.
4. sve-vector-reverse.ll,vector-reverse-mask4.ll: add noalias
keyword to simplify IR.
4. sve-widen-gep.ll,sve-widen-phi.ll: regenerate using script.
These changes will make the subsequent patch adding RUN lines much
easier to review!
Differential Revision: https://reviews.llvm.org/D132219
On known hardware, reductions, gather, and scatter operations have execution latencies which correlated with the vector length (VL) of the operation. Most other operations (e.g. simply arithmetic) don't correlated in this way, and instead essentially fixed cost as VL varies.
When I'd implemented initial scalable cost model support for reductions, gather, and scatter operations, I had used an upper bound on the statically unknown VL. The argument at the time was that this prevented falsely low costs, and biased the vectorizer away from generating bad (on some hardware) code. Unfortunately, practical experience shows we were a bit too effective at that goal, and the high costs defacto prevents vectorization using these constructs at all.
This patch reverses course, and ties the returned cost not to the maximum possible VL, but the VL which would correspond to VScaleForTuning. This parameter is the same one the vectorizer uses when normalizing loop costs, so the term effectively cancels out. The result is that the vectorizer now sees these constructs as comparable in cost to their fixed length variants.
This does introduce the possibility of the cost for these operations being a significant under estimate on platforms where actual VLEN is far from that implied by VScaleForTuning. On such platforms, we might make poor heuristic choices. Probably not in LV itself (due to the cancellation mentioned above), but possibly during e.g. lowering. I'm not currently aware of any concrete examples of this, but this patch does open a concern which did not previously exist.
Previously, we had the problem of overestimating costs causing the same problem on machines much closer to default values for vscale for tuning. With this patch, we still have that problem potentially if vscale for tuning is set high (manually), and then the code is run on a narrow VLEN machine.
Differential Revision: https://reviews.llvm.org/D131519
If the incoming previous value of a fixed-order recurrence is a phi in
the header, go through incoming values from the latch until we find a
non-phi value. Use this as the new Previous, all uses in the header
will be dominated by the original phi, but need to be moved after
the non-phi previous value.
At the moment, fixed-order recurrences are modeled as a chain of
first-order recurrences.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D119661
This change reorganizes the code and comments to make the expected semantics of these routines more clear. However, this is *not* an NFC change. The functional change is having isScalarWithPredication return false if the instruction does not need predicated. Specifically, for the case of a uniform memory operation we were previously considering it *not* to be a predicated instruction, but *were* considering it to be scalable with predication.
As can be seen with the test changes, this causes uniform memory ops which should have been lowered as uniform-per-parts values to instead be lowering via naive scalarization or if scalarization is infeasible (i.e. scalable vectors) aborted entirely. I also don't trust the code to bail out correctly 100% of the time, so it's possible we had a crash or miscompile from trying to scalarize something which isn't scalaralizable. I haven't found a concrete example here, but I am suspicious.
Differential Revision: https://reviews.llvm.org/D131093
After D121595 was commited, I noticed regressions assosicated with small trip
count numbersvectorisation by tail folding with scalable vectors. As a solution
for those issues I propose to introduce the minimal trip count threshold value.
Differential Revision: https://reviews.llvm.org/D130755
This extends the handling of uniform memory operations to handle the case where a store is storing a loop invariant value. Unlike the general case of a store to an invariant address where we must use the last active lane, in this case we can use any lane since all lanes must produce the same result.
For context, the basic structure of the existing code and how the change fits in:
* First, we select a widening strategy. (The result is irrelevant for this patch.)
* Then we determine if a computation is uniform within all lanes of VF. (Note this is the uniform-per-part definition, not LAI's uniform across all unrolled iterations definition.)
* If it is, we overrule the widening strategy, and unconditionally scalarize.
* VPReplicationRecipe - which is what actually does the scalarization - knows how to handle unform-per-part values including for scalable vectors. However, we do need to know that the expression is safe to execute without predication - e.g. the uniform mem op was unconditional in the original loop. (This part was split off and already landed.)
An obvious question is why not simply implement the generic case? The answer is that I'm going to, but doing so without a canonicalization towards uniform causes regressions due to bad interaction with scalarization/uniformity of values feeding the uniform mem-op. This patch is needed to avoid those regressions.
Differential Revision: https://reviews.llvm.org/D130364
If we have interleave groups in the loop we want to vectorise then
we should fall back on normal vectorisation with a scalar epilogue. In
such cases when tail-folding is enabled we'll almost certainly go on to
create vplans with very high costs for all vector VFs and fall back on
VF=1 anyway. This is likely to be worse than if we'd just used an
unpredicated vector loop in the first place.
Once the vectoriser has proper support for analysing all the costs
for each combination of VF and vectorisation style, then we should
be able to remove this.
Added an extra test here:
Transforms/LoopVectorize/AArch64/sve-tail-folding-option.ll
Differential Revision: https://reviews.llvm.org/D128342
We already had the reasoning about uniform mem op loads; if the address is accessed at least once, we know the instruction doesn't need predicated to ensure fault safety. For stores, we do need to ensure that the values visible in memory are the same with and without predication. The easiest sub-case to check for is that all the values being stored are the same. Since we know that at least one lane is active, this tells us that the value must be visible.
Warning on confusing terminology: "uniform" vs "uniform mem op" mean two different things here, and this patch is specific to the later. It would *not* be legal to make this same change for merely "uniform" operations.
Differential Revision: https://reviews.llvm.org/D130637
This change enables vectorization (using scalable vectorization only, fixed vectors are not yet enabled) for RISCV when vector instructions are available for the target configuration.
At this point, the resulting configuration should be both stable (e.g. no crashes), and profitable (i.e. few cases where scalar loops beat vector ones), but is not going to be particularly well tuned (i.e. we emit the best possible vector loop). The goal of this change is to align testing across organizations and ensure the default configuration matches what downstreams are using as closely as possible.
This exposes a large amount of code which hasn't otherwise been on by default, and thus may not have been fully exercised. Given that, having issues fall out is not unexpected. If you find issues, please make sure to include as much information as you can when reverting this change.
Differential Revision: https://reviews.llvm.org/D129013
The problem here is target independent, but particularly painful on RISCV. If we chose to vectorize such that vscale x 2 x i32 is our widest type and fits in a register, a naive expansion of i64 comparisons results in comparisons and index types at <scalabe x 2 x i64>. This requires both an LMUL of 2, and a VSETVLI toggle in the loop. Note that we could have used <vscale x 2 x i32> for the compairons legally given the range of the trip count.
All of our other tests are functionality tests constrained to some
specific configuration. This one is intended to float with the
default configuration so that changes in that default are visible
in reviews. Note that our current default does not enable
vectorization at all; thus the current output is unvectorized.
We call tail-call-elim near the beginning of the pipeline,
but that is too early to annotate calls that get added later.
In the motivating case from issue #47852, the missing 'tail'
on memset leads to sub-optimal codegen.
I experimented with removing the early instance of
tail-call-elim instead of just adding another pass, but that
appears to be slightly worse for compile-time:
+0.15% vs. +0.08% time.
"tailcall" shows adding the pass; "tailcall2" shows moving
the pass to later, then adding the original early pass back
(so 1596886802 is functionally equivalent to 180b0439dc ):
https://llvm-compile-time-tracker.com/index.php?config=NewPM-O3&stat=instructions&remote=rotateright
Note that there was an effort to split the tail call functionality
into 2 passes - that could help reduce compile-time if we find
that this change costs more in compile-time than expected based
on the preliminary testing:
D60031
Differential Revision: https://reviews.llvm.org/D130374
This code confuses LV's "Uniform" and LVL/LAI's "Uniform". Despite the
common name, these are different.
* LVs notion means that only the first lane *of each unrolled part* is
required. That is, lanes within a single unroll factor are considered
uniform. This allows e.g. widenable memory ops to be considered
uses of uniform computations.
* LVL and LAI's notion refers to all lanes across all unrollings.
IsUniformMem is in turn defined in terms of LAI's notion. Thus a
UniformMemOpmeans is a memory operation with a loop invariant address.
This means the same address is accessed in every iteration.
The tweaked piece of code was trying to match a uniform mem op (i.e.
fully loop invariant address), but instead checked for LV's notion of
uniformity. In theory, this meant with UF > 1, we could speculate
a load which wasn't safe to execute.
This ends up being mostly silent in current code as it is nearly
impossible to create the case where this difference is visible. The
closest I've come in the test case from 54cb87, but even then, the
incorrect result is only visible in the vplan debug output; before this
change we sink the unsafely speculated load back into the user's predicate
blocks before emitting IR. Both before and after IR are correct so the
differences aren't "interesting".
The other test changes are uninteresting. They're cases where LV's uniform
analysis is slightly weaker than SCEV isLoopInvariant.
This patch adds the AArch64 hook for preferPredicateOverEpilogue,
which currently returns true if SVE is enabled and one of the
following conditions (non-exhaustive) is met:
1. The "sve-tail-folding" option is set to "all", or
2. The "sve-tail-folding" option is set to "all+noreductions"
and the loop does not contain reductions,
3. The "sve-tail-folding" option is set to "all+norecurrences"
and the loop has no first-order recurrences.
Currently the default option is "disabled", but this will be
changed in a later patch.
I've added new tests to show the options behave as expected here:
Transforms/LoopVectorize/AArch64/sve-tail-folding-option.ll
Differential Revision: https://reviews.llvm.org/D129560
This patch is in preparation for enabling vectorisation with tail-folding
by default for SVE targets. Once we do that many existing tests will
break that depend upon having normal unpredicated vector loops. For
all such tests I have added the flag:
-prefer-predicate-over-epilogue=scalar-epilogue
Differential Revision: https://reviews.llvm.org/D129137
An srem or sdiv has two cases which can cause undefined behavior, not just one. The existing code did not account for this, and as a result, we miscompiled when we encountered e.g. a srem i64 %v, -1 in a conditional block.
Instead of hand rolling the logic, just use the utility function which exists exactly for this purpose.
Differential Revision: https://reviews.llvm.org/D130106
By default if SVE is enabled we want the select instruction used for
reductions to be inside the loop, rather than outside. This makes it
possible for the backend to fold the select into the operation to
produce a single predicated add, fadd, etc.
Differential Revision: https://reviews.llvm.org/D129763
In sve-tail-folding-reductions.ll I've also added an extra RUN line
to test normal reductions, i.e. not in-loop. This patch is a pre-commit
in preparation for a follow-on patch that changes how reduction selects
are generated in the vector loop.
Differential Revision: https://reviews.llvm.org/D129761
I've simplified all of the SVE vectoriser tail-folding tests to
only care about testing the flag:
-prefer-predicate-over-epiloge=predicate-else-scalar-epilogue
In practice we always want to fall back on unpredicated vector
loops if tail-folding is not possible.
Differential Revision: https://reviews.llvm.org/D129843
This builds on the previous forked pointers patch, which only accepted
a single select as the pointer to check. A recursive function to walk
through IR has been added, which searches for either a loop-invariant
or addrec SCEV.
This will only handle a single fork at present, so selects of selects
or a GEP with a select for both the base and offset will be rejected.
There is also a recursion limit with a cli option to change it.
Reviewed By: fhahn, david-arm
Differential Revision: https://reviews.llvm.org/D108699
At the moment, the VPPRedInstPHIRecipe is not used in subsequent uses of
the predicate recipe. This incorrectly models the def-use chains, as all
later uses should use the phi recipe. Fix that by delaying recording of
the recipe.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D129436
At the moment, the cost of runtime checks for scalable vectors is
overestimated due to creating separate vscale * VF expressions for each
check. Instead re-use the first expression.
For scalable vectors, it is not sufficient to only check
MinProfitableTripCount if it is >= VF.getKnownMinValue() * UF, because
this property may not holder for larger values of vscale. In those
cases, compute umax(VF * UF, MinProfTC) instead.
This should fix
https://lab.llvm.org/buildbot/#/builders/197/builds/2262
The test shows a case where the minimum trip count check incorrectly
only checks the minimum profitable trip count computed due to runtime
checks. This is incorrect for scalable VFs, because the VF * UF may
exceed the minimum profitable trip count for vscale > 1.
This is the likely reason for
https://lab.llvm.org/buildbot/#/builders/197/builds/2262 failing.
When vectorising ordered reductions we call a function
LoopVectorizationPlanner::adjustRecipesForReductions to replace the
existing VPWidenRecipe for the fadd instruction with a new
VPReductionRecipe. We attempt to insert the new recipe in the same
place, but this is wrong because createBlockInMask may have
generated new recipes that VPReductionRecipe now depends upon. I
have changed the insertion code to append the recipe to the
VPBasicBlock instead.
Added a new RUN with tail-folding enabled to the existing test:
Transforms/LoopVectorize/AArch64/scalable-strict-fadd.ll
Differential Revision: https://reviews.llvm.org/D129550
Currently we only call replaceLoopPHINodesWithPreheaderValues() if
optimizeLoopExits() replaces the exit with an unconditional exit.
However, it is very common that this already happens as part of
eliminateIVComparison(), in which case we're leaving behind the
dead header phi.
Tweak the early bailout for already-constant exits to also call
replaceLoopPHINodesWithPreheaderValues().
Differential Revision: https://reviews.llvm.org/D129214
When calculating the cost of Instruction::Br in getInstructionCost
we query PredicatedBBsAfterVectorization to see if there is a
scalar predicated block. However, this meant that the decisions
being made for a given fixed-width VF were affecting the cost for a
scalable VF. As a result we were returning InstructionCost::Invalid
pointlessly for a scalable VF that should have a low cost. I
encountered this for some loops when enabling tail-folding for
scalable VFs.
Test added here:
Transforms/LoopVectorize/AArch64/sve-tail-folding-cost.ll
Differential Revision: https://reviews.llvm.org/D128272
Currently, for vectorised loops that use the get.active.lane.mask
intrinsic we only use the mask for predicated vector operations,
such as masked loads and stores, etc. The loop itself is still
controlled by comparing the canonical induction variable with the
trip count. However, for some targets this is inefficient when it's
cheap to use the mask itself to control the loop.
This patch adds support for using the active lane mask for control
flow by:
1. Generating the active lane mask for the next iteration of the
vector loop, rather than the current one. If there are still any
remaining iterations then at least the first bit of the mask will
be set.
2. Extract the first bit of this mask and use this bit for the
conditional branch.
I did this by creating a new VPActiveLaneMaskPHIRecipe that sets
up the initial PHI values in the vector loop pre-header. I've also
made use of the new BranchOnCond VPInstruction for the final
instruction in the loop region.
Differential Revision: https://reviews.llvm.org/D125301
The motivation here is to a) bring us closer into alignment with AArch64 under the assumption that codepath is better tested, and b) simplify pattern matching in an upcoming change.
The immediate impact is a significant IR reduction but a fairly minimal change in the generated assembly. Due to a difference in expansion behavior we get a saturating add vs an unsaturating one for the old code, but that's about it. This difference comes down to different handling of overflow, which doesn't seem to be possible here anyways, so the assembly codegen is arguably a minor regression. I don't expect that to matter in practice.
Differential Revision: https://reviews.llvm.org/D129221
Now that removeDeadRecipes can remove most dead recipes across a whole
VPlan, there is no need to first collect some dead instructions.
Instead removeDeadRecipes can simply clean them up.
Depends D127580.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D128408
These three subtarget features are meant to control where MVE
instructions take 1 vs 2 vs 4 architectural beats. The mve1beat feature
is described as "Model MVE instructions as a 1 beat per tick
architecture", meaning MVE instruction will execute over 4 cycles.
mve4beat is the opposite where the entire 4 beats of the MVE instruction
execute in a single cycle. The costs for the two were backwards though,
not matching the cycle counts like they should. This patch switches the
costs on the two to bring them in-line with expectations.
Differential Revision: https://reviews.llvm.org/D129141
This can enable additional region merging, while not losing
opportunities as region merging does not produce dead recipes.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D128831
The removed CHECK configurations are tested as well below, modulo the
dce/instcombine runs. This makes them redundant, and removing them
removes a substantial amount of uneeded checks.
The tests are focused on code-gen for first-order recurrences. There are
plenty of tests specifically for runtime check generation. Using noalias
to avoid runtime checks slightly simplifies the test output and ensures
the checks focus on the relevant bits and ensures the checks focus on
the relevant bits and ensures the checks focus on the relevant bits and
ensures the checks focus on the relevant bits.
D128820 stopped creating div/rem constant expressions by default;
this patch removes support for them entirely.
The getUDiv(), getExactUDiv(), getSDiv(), getExactSDiv(), getURem()
and getSRem() on ConstantExpr are removed, and ConstantExpr::get()
now only accepts binary operators for which
ConstantExpr::isSupportedBinOp() returns true. Uses of these methods
may be replaced either by corresponding IRBuilder methods, or
ConstantFoldBinaryOpOperands (if a constant result is required).
On the C API side, LLVMConstUDiv, LLVMConstExactUDiv, LLVMConstSDiv,
LLVMConstExactSDiv, LLVMConstURem and LLVMConstSRem are removed and
corresponding LLVMBuild methods should be used.
Importantly, this also means that constant expressions can no longer
trap! This patch still keeps the canTrap() method to minimize diff --
I plan to drop it in a separate NFC patch.
Differential Revision: https://reviews.llvm.org/D129148
This patch replaces the tight hard cut-off for the number of runtime
checks with a more accurate cost-driven approach.
The new approach allows vectorization with a larger number of runtime
checks in general, but only executes the vector loop (and runtime checks) if
considered profitable at runtime. Profitable here means that the cost-model
indicates that the runtime check cost + vector loop cost < scalar loop cost.
To do that, LV computes the minimum trip count for which runtime check cost
+ vector-loop-cost < scalar loop cost.
Note that there is still a hard cut-off to avoid excessive compile-time/code-size
increases, but it is much larger than the original limit.
The performance impact on standard test-suites like SPEC2006/SPEC2006/MultiSource
is mostly neutral, but the new approach can give substantial gains in cases where
we failed to vectorize before due to the over-aggressive cut-offs.
On AArch64 with -O3, I didn't observe any regressions outside the noise level (<0.4%)
and there are the following execution time improvements. Both `IRSmk` and `srad` are relatively short running, but the changes are far above the noise level for them on my benchmark system.
```
CFP2006/447.dealII/447.dealII -1.9%
CINT2017rate/525.x264_r/525.x264_r -2.2%
ASC_Sequoia/IRSmk/IRSmk -9.2%
Rodinia/srad/srad -36.1%
```
`size` regressions on AArch64 with -O3 are
```
MultiSource/Applications/hbd/hbd 90256.00 106768.00 18.3%
MultiSourc...ks/ASCI_Purple/SMG2000/smg2000 240676.00 257268.00 6.9%
MultiSourc...enchmarks/mafft/pairlocalalign 472603.00 489131.00 3.5%
External/S...2017rate/525.x264_r/525.x264_r 613831.00 630343.00 2.7%
External/S...NT2006/464.h264ref/464.h264ref 818920.00 835448.00 2.0%
External/S...te/538.imagick_r/538.imagick_r 1994730.00 2027754.00 1.7%
MultiSourc...nchmarks/tramp3d-v4/tramp3d-v4 1236471.00 1253015.00 1.3%
MultiSource/Applications/oggenc/oggenc 2108147.00 2124675.00 0.8%
External/S.../CFP2006/447.dealII/447.dealII 4742999.00 4759559.00 0.3%
External/S...rate/510.parest_r/510.parest_r 14206377.00 14239433.00 0.2%
```
Reviewed By: lebedev.ri, ebrevnov, dmgreen
Differential Revision: https://reviews.llvm.org/D109368
At the moment, the same VPlan can be used code generation of both the
main vector and epilogue vector loop. This can lead to wrong results, if
the plan is optimized based on the VF of the main vector loop and then
re-used for the epilogue loop.
One example where this is problematic is if the scalar loops need to
execute at least one iteration, e.g. due to interleave groups.
To prevent mis-compiles in the short-term, disable optimizing exit
conditions for VPlans when using epilogue vectorization. The proper fix
is to avoid re-using the same plan for both loops, which will require
support for cloning plans first.
Fixes#56319.
I looked at canonicalizing in the other direction, but that causes
many potential regressions and infinite loops because we already
(possibly wrongly) canonicalize "trunc X to i1" into an and+icmp.
This has a data layout restriction to avoid creating illegal
mask instructions, but we could remove that if we can show
that the backend can undo this when needed.
The motivating example from issue #56119 is modeled by the
PhaseOrdering test.
At the moment LoopVersioning is only created for inner-loop
vectorization. This patch moves it to LVP::execute, which means it will
also be added for epilogue vectorization. As a consequence, the proper
noalias metadata is now also added to epilogue vector loops.
LVer will be moved to VPTransformState as follow-up.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D127966
In some cases, there may be widened users of inductions even though the
plan includes the scalar VF. In those cases, make sure we still replace
the VPWidenIntOrFpInductionRecipe with scalar steps, as otherwise we may
try to execute a VPWidenIntOrFpInductionRecipe with a scalar VF.
Alternatively the patch could also split the range if needed.
This fixes a crash exposed by D123720.
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D128755
This change is a bit subtle. If we have a type like <vscale x 1 x i64>, the vectorizer will currently reject vectorization. The reason is that a type like <1 x i64> is likely to get simply rescalarized, and the vectorizer doesn't want to be in the game of simple unrolling.
(I've given the example in terms of 1 x types which use a single register, but the same issue exists for any N x types which use N registers. e.g. RISCV LMULs.)
This change distinguishes scalable types from fixed types under the reasoning that converting to a scalable type isn't unrolling. Because the actual vscale isn't known until runtime, using a vscale type is potentially very profitable.
This makes an important, but unchecked, assumption. Specifically, the scalable type is assumed to only be legal per the cost model if there's actually a scalable register class which is distinct from the scalar domain. This is, to my knowledge, true for all targets which return non-invalid costs for scalable vector ops today, but in theory, we could have a target decide to lower scalable to fixed length vector or even scalar registers. If that ever happens, we'd need to revisit this code.
In practice, this patch unblocks scalable vectorization for ELEN types on RISCV.
Let me sketch one alternate implementation I considered. We could have restricted this to when we know a minimum value for vscale. Specifically, for the default +v extension for RISCV, we actually know that vscale >= 2 for ELEN types. However, doing it this way means we can't generate scalable vectors when using the various embedded vector extensions which have a minimum vscale of 1.
Differential Revision: https://reviews.llvm.org/D128542
LoopVectorizer uses getVScaleForTuning for deciding how to discount the cost of a potential vector factor by the amount of work performed. Without the callback implemented, the vectorizer was defaulting to an estimated vscale of 1. This results in fixed vectorization looking falsely profitable (since it used the command line VLEN).
The test change is pretty limited since a) we don't have much coverage of the vectorizer with scalable vectors at all, and b) what little coverage we have mostly uses i64 element types. There's a separate issue with <vscale x 1 x i64> which prevents us from getting to this stage of costing, and thus only the one test explicitly written to avoid that is visible in the diff. However, this is actually a very wide impact change as it changes the practical vectorization result when both fixed and scalable is enabled to scalable.
As an aside, I think the vectorizer is at little too strongly biased towards scalable when both are legal, but we can explore that separately. For now, let's just get the cost model working the way it was intended.
Differential Revision: https://reviews.llvm.org/D128547
We currently have a costing bug around the etype == ELEN case, so add otherwise duplicate tests to show test diffs as I work on other parts of costing.
If we have an unaligned uniform store, then when costing a scalable VF we can't emit code to scalarize it. (Well, we could, but we haven't implemented that case.) This change replaces an assert with a cost-model bailout such that we reject vectorization with the scalable VF instead of crashing.
Philip Reames