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
Philip Reames