This adds cost modelling for the inloop vectorization added in
745bf6cf4471. Up until now they have been modelled as the original
underlying instruction, usually an add. This happens to works OK for MVE
with instructions that are reducing into the same type as they are
working on. But MVE's instructions can perform the equivalent of an
extended MLA as a single instruction:
%sa = sext <16 x i8> A to <16 x i32>
%sb = sext <16 x i8> B to <16 x i32>
%m = mul <16 x i32> %sa, %sb
%r = vecreduce.add(%m)
->
R = VMLADAV A, B
There are other instructions for performing add reductions of
v4i32/v8i16/v16i8 into i32 (VADDV), for doing the same with v4i32->i64
(VADDLV) and for performing a v4i32/v8i16 MLA into an i64 (VMLALDAV).
The i64 are particularly interesting as there are no native i64 add/mul
instructions, leading to the i64 add and mul naturally getting very
high costs.
Also worth mentioning, under NEON there is the concept of a sdot/udot
instruction which performs a partial reduction from a v16i8 to a v4i32.
They extend and mul/sum the first four elements from the inputs into the
first element of the output, repeating for each of the four output
lanes. They could possibly be represented in the same way as above in
llvm, so long as a vecreduce.add could perform a partial reduction. The
vectorizer would then produce a combination of in and outer loop
reductions to efficiently use the sdot and udot instructions. Although
this patch does not do that yet, it does suggest that separating the
input reduction type from the produced result type is a useful concept
to model. It also shows that a MLA reduction as a single instruction is
fairly common.
This patch attempt to improve the costmodelling of in-loop reductions
by:
- Adding some pattern matching in the loop vectorizer cost model to
match extended reduction patterns that are optionally extended and/or
MLA patterns. This marks the cost of the reduction instruction correctly
and the sext/zext/mul leading up to it as free, which is otherwise
difficult to tell and may get a very high cost. (In the long run this
can hopefully be replaced by vplan producing a single node and costing
it correctly, but that is not yet something that vplan can do).
- getExtendedAddReductionCost is added to query the cost of these
extended reduction patterns.
- Expanded the ARM costs to account for these expanded sizes, which is a
fairly simple change in itself.
- Some minor alterations to allow inloop reduction larger than the highest
vector width and i64 MVE reductions.
- An extra InLoopReductionImmediateChains map was added to the vectorizer
for it to efficiently detect which instructions are reductions in the
cost model.
- The tests have some updates to show what I believe is optimal
vectorization and where we are now.
Put together this can greatly improve performance for reduction loop
under MVE.
Differential Revision: https://reviews.llvm.org/D93476
In LoopInterchange, `findInnerReductionPhi()` looks for reduction
variables, which cannot be constants. Update it to return early in that
case.
This also addresses a blocker for removing use-lists from ConstantData,
whose users could be spread across arbitrary modules in the same
LLVMContext.
Differential Revision: https://reviews.llvm.org/D94712
If a function doesn't contain loops and does not call non-willreturn
functions, then it is willreturn. Loops are detected by checking
for backedges in the function. We don't attempt to handle finite
loops at this point.
Differential Revision: https://reviews.llvm.org/D94633
It turns out the vectorizer calls the getIntrinsicInstrCost functions
with a scalar return type and vector VF. This updates the costmodel to
handle that, still producing the correct vector costs.
A vectorizer test is added to show it vectorizing at the correct factor
again.
Summary:
The custom mapper API did not previously support the mapping names added previously. This means they were not present if a user requested debugging information while using the mapper functions. This adds basic support for passing the mapped names to the runtime library.
Reviewers: jdoerfert
Differential Revision: https://reviews.llvm.org/D94806
In https://llvm.org/PR48810 , we are crashing while trying to
propagate attributes from mempcpy (returns void*) to memcpy
(returns nothing - void).
We can avoid the crash by removing known incompatible
attributes for the void return type.
I'm not sure if this goes far enough (should we just drop all
attributes since this isn't the same function?). We also need
to audit other transforms in LibCallSimplifier to make sure
there are no other cases that have the same problem.
Differential Revision: https://reviews.llvm.org/D95088
This patch applies the idea from D93734 to LoopUnswitch.
It adds support for unswitching on conditions that are only
invariant along certain paths through a loop.
In particular, it targets conditions in the loop header that
depend on values loaded from memory. If either path from
the true or false successor through the loop does not modify
memory, perform partial loop unswitching.
That is, duplicate the instructions feeding the condition in the pre-header.
Then unswitch on the duplicated condition. The condition is now known
in the unswitched version for the 'invariant' path through the original loop.
On caveat of this approach is that one of the loops created can be partially
unswitched again. To avoid this behavior, `llvm.loop.unswitch.partial.disable`
metadata is added to the unswitched loops, to avoid subsequent partial
unswitching.
If that's the approach to go, I can move the code handling the metadata kind
into separate functions.
This increases the cases we unswitch quite a bit in SPEC2006/SPEC2000 &
MultiSource. It also allows us to eliminate a dead loop in SPEC2017's omnetpp
```
Tests: 236
Same hash: 170 (filtered out)
Remaining: 66
Metric: loop-unswitch.NumBranches
Program base patch diff
test-suite...000/255.vortex/255.vortex.test 2.00 23.00 1050.0%
test-suite...T2006/401.bzip2/401.bzip2.test 7.00 55.00 685.7%
test-suite :: External/Nurbs/nurbs.test 5.00 26.00 420.0%
test-suite...s-C/unix-smail/unix-smail.test 1.00 3.00 200.0%
test-suite.../Prolangs-C++/ocean/ocean.test 1.00 3.00 200.0%
test-suite...tions/lambda-0.1.3/lambda.test 1.00 3.00 200.0%
test-suite...yApps-C++/PENNANT/PENNANT.test 2.00 5.00 150.0%
test-suite...marks/Ptrdist/yacr2/yacr2.test 1.00 2.00 100.0%
test-suite...lications/viterbi/viterbi.test 1.00 2.00 100.0%
test-suite...plications/d/make_dparser.test 12.00 24.00 100.0%
test-suite...CFP2006/433.milc/433.milc.test 14.00 27.00 92.9%
test-suite.../Applications/lemon/lemon.test 7.00 12.00 71.4%
test-suite...ce/Applications/Burg/burg.test 6.00 10.00 66.7%
test-suite...T2006/473.astar/473.astar.test 16.00 26.00 62.5%
test-suite...marks/7zip/7zip-benchmark.test 78.00 121.00 55.1%
```
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D93764
This reverts commit d97f776be5f8cd3cd446fe73827cd355f6bab4e1.
The original problem was due to build failures in shared lib builds. D95079
moved ImportedFunctionsInliningStatistics under Analysis, unblocking
this.
Branch/assume conditions in PredicateInfo are currently handled in
a rather ad-hoc manner, with some arbitrary limitations. For example,
an `and` of two `icmp`s will be handled, but an `and` of an `icmp`
and some other condition will not. That also includes the case where
more than two conditions and and'ed together.
This patch makes the handling more general by looking through and/ors
up to a limit and considering all kinds of conditions (though operands
will only be taken for cmps of course).
Differential Revision: https://reviews.llvm.org/D94447
When using 2 InlinePass instances in the same CGSCC - one for other
mandatory inlinings, the other for the heuristic-driven ones - the order
in which the ImportedFunctionStats would be output-ed would depend on
the destruction order of the inline passes, which is not deterministic.
This patch moves the ImportedFunctionStats responsibility to the
InlineAdvisor to address this problem.
Differential Revision: https://reviews.llvm.org/D94982
The loop-unswitch transform should not be performed on a loop whose
condition is divergent. For this to happen correctly, divergence
analysis must be available. The existing divergence analysis has not
been ported to the new pass manager yet. As a result, loop unswitching
on the new pass manager is currently unsafe on targets that care about
divergence.
This test is temporarily disabled to unblock work on the new pass
manager. The issue is now tracked in bug 48819.
Reviewed By: foad
Differential Revision: https://reviews.llvm.org/D95051
Loop peeling assumes that the loop's latch is a conditional branch. Add
a check to canPeel that explicitly checks for this, and testcases that
otherwise fail an assertion when trying to peel a loop whose back-edge
is a switch case or the non-unwind edge of an invoke.
Reviewed By: skatkov, fhahn
Differential Revision: https://reviews.llvm.org/D94995
Summary: This is to address bug48712.
The solution in this patch is that when we want to merge two variable a
into the storage frame of variable b only if the alignment of a is
multiple of b.
There may be other strategies. But now I think they are hard to handle
and benefit little. Or we can implement them in the future.
Test-plan: check-llvm
Reviewers: jmorse, lxfind, junparser
Differential Revision: https://reviews.llvm.org/D94891
Currently LLVM is relying on ValueTracking's `isKnownNonZero` to attach `nonnull`, which can return true when the value is poison.
To make the semantics of `nonnull` consistent with the behavior of `isKnownNonZero`, this makes the semantics of `nonnull` to accept poison, and return poison if the input pointer isn't null.
This makes many transformations like below legal:
```
%p = gep inbounds %x, 1 ; % p is non-null pointer or poison
call void @f(%p) ; instcombine converts this to call void @f(nonnull %p)
```
Instead, this semantics makes propagation of `nonnull` to caller illegal.
The reason is that, passing poison to `nonnull` does not immediately raise UB anymore, so such program is still well defined, if the callee does not use the argument.
Having `noundef` attribute there re-allows this.
```
define void @f(i8* %p) { ; functionattr cannot mark %p nonnull here anymore
call void @g(i8* nonnull %p) ; .. because @g never raises UB if it never uses %p.
ret void
}
```
Another attribute that needs to be updated is `align`. This patch updates the semantics of align to accept poison as well.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D90529
separate sections.
For ThinLTO, all the function profiles without context has been annotated to
outline functions if possible in prelink phase. In postlink phase, profile
annotation in postlink phase is only meaningful for function profile with
context. If the profile is large, it is better to split the profile into two
parts, one with context and one without, so the profile reading in postlink
phase only has to read the part with context. To have the profile splitting,
we extend the ExtBinary format to support different section arrangement. It
will be flexible to add other section layout in the future without the need
to create new class inheriting from ExtBinary class.
Differential Revision: https://reviews.llvm.org/D94435
This reverts commit 438682de6a38ac97f89fa38faf5c8dc9b09cd9ad to fix the
bug with the reducing size of the resulting vector for the entry node
with multiple users.
D75825 and D75828 modified llvm/test/Transforms/Inline/noalias2.ll to handle llvm.assume. The checking though was broken.
The NO_ASSUME has been replaced by a normal CHECK; the ASSUME rules were never triggered and have been removed.
The test checks have been regenerated.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D94978
Relative to the original change, this adds a check that the
instruction on which we're replacing operands is safe to speculatively
execute, because that's what we're effectively doing. We're executing
the instruction with the replaced operand, which is fine if it's pure,
but not fine if can cause side-effects or UB (aka is not speculatable).
Additionally, we cannot (generally) replace operands in phi nodes,
as these may refer to a different loop iteration. This is also covered
by the speculation check.
-----
InstCombine already performs a fold where X == Y ? f(X) : Z is
transformed to X == Y ? f(Y) : Z if f(Y) simplifies. However,
if f(X) only has one use, then we can always directly replace the
use inside the instruction. To actually be profitable, limit it to
the case where Y is a non-expr constant.
This could be further extended to replace uses further up a one-use
instruction chain, but for now this only looks one level up.
Among other things, this also subsumes D94860.
Differential Revision: https://reviews.llvm.org/D94862
Just like llvm.assume, there are a lot of cases where we can just ignore llvm.experimental.noalias.scope.decl.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D93042
Split impliesPoison into two recursive walks, one over V, the
other over ValAssumedPoison. This allows us to reason about poison
implications in a number of additional cases that are important
in practice. This is a generalized form of D94859, which handles
the cmp to cmp implication in particular.
Differential Revision: https://reviews.llvm.org/D94866
83daa49758a1 made loop-rotate more conservative in the presence of
function calls in the prepare-for-lto stage. The code did not properly
account for calls that are no actual function calls, like calls to
intrinsics. This patch updates the code to ensure only calls that are
lowered to actual calls are considered inline candidates.
This caused a miscompile in Chromium, see comments on the codereview for
discussion and pointer to a reproducer.
> InstCombine already performs a fold where X == Y ? f(X) : Z is
> transformed to X == Y ? f(Y) : Z if f(Y) simplifies. However,
> if f(X) only has one use, then we can always directly replace the
> use inside the instruction. To actually be profitable, limit it to
> the case where Y is a non-expr constant.
>
> This could be further extended to replace uses further up a one-use
> instruction chain, but for now this only looks one level up.
>
> Among other things, this also subsumes D94860.
>
> Differential Revision: https://reviews.llvm.org/D94862
This also reverts the follow-up
a003f26539cf4db744655e76c41f4c4a8913f116:
> [llvm] Prevent infinite loop in InstCombine of select statements
>
> This fixes an issue where the RHS and LHS the comparison operation
> creating the predicate were swapped back and forth forever.
>
> Differential Revision: https://reviews.llvm.org/D94934
D84108 exposed a bad interaction between inlining and loop-rotation
during regular LTO, which is causing notable regressions in at least
CINT2006/473.astar.
The problem boils down to: we now rotate a loop just before the vectorizer
which requires duplicating a function call in the preheader when compiling
the individual files ('prepare for LTO'). But this then prevents further
inlining of the function during LTO.
This patch tries to resolve this issue by making LoopRotate more
conservative with respect to rotating loops that have inline-able calls
during the 'prepare for LTO' stage.
I think this change intuitively improves the current situation in
general. Loop-rotate tries hard to avoid creating headers that are 'too
big'. At the moment, it assumes all inlining already happened and the
cost of duplicating a call is equal to just doing the call. But with LTO,
inlining also happens during full LTO and it is possible that a previously
duplicated call is actually a huge function which gets inlined
during LTO.
From the perspective of LV, not much should change overall. Most loops
calling user-provided functions won't get vectorized to start with
(unless we can infer that the function does not touch memory, has no
other side effects). If we do not inline the 'inline-able' call during
the LTO stage, we merely delayed loop-rotation & vectorization. If we
inline during LTO, chances should be very high that the inlined code is
itself vectorizable or the user call was not vectorizable to start with.
There could of course be scenarios where we inline a sufficiently large
function with code not profitable to vectorize, which would have be
vectorized earlier (by scalarzing the call). But even in that case,
there probably is no big performance impact, because it should be mostly
down to the cost-model to reject vectorization in that case. And then
the version with scalarized calls should also not be beneficial. In a way,
LV should have strictly more information after inlining and make more
accurate decisions (barring cost-model issues).
There is of course plenty of room for things to go wrong unexpectedly,
so we need to keep a close look at actual performance and address any
follow-up issues.
I took a look at the impact on statistics for
MultiSource/SPEC2000/SPEC2006. There are a few benchmarks with fewer
loops rotated, but no change to the number of loops vectorized.
Reviewed By: sanwou01
Differential Revision: https://reviews.llvm.org/D94232
This patch teaches SimplifyCFG::SimplifyBranchOnICmpChain to understand select form of
(x == C1 || x == C2 || ...) / (x != C1 && x != C2 && ...) and optimize them into switch if possible.
D93065 has more context about the transition, including links to the list of optimizations being updated.
Differential Revision: https://reviews.llvm.org/D93943
After much refactoring over the last 2 weeks to the reduction
matching code, I think this change is finally ready.
We effectively broke fmax/fmin vector reduction optimization
when we started canonicalizing to intrinsics in instcombine,
so this should restore that functionality for SLP.
There are still FMF problems here as noted in the code comments,
but we should be avoiding miscompiles on those for fmax/fmin by
restricting to full 'fast' ops (negative tests are included).
Fixing FMF propagation is a planned follow-up.
Differential Revision: https://reviews.llvm.org/D94913
This is to address https://bugs.llvm.org/show_bug.cgi?id=48626.
When there are musttail calls that use parameters aliasing the newly created coroutine frame, the existing implementation will fatal.
We simply cannot perform CoroElide in such cases. In theory a precise analysis can be done to check whether the parameters of the musttail call
actually alias the frame, but it's very hard to do it before the transformation happens. Also in most cases the existence of musttail call is
generated due to symmetric transfers, and in those cases alias analysis won't be able to tell that they don't alias anyway.
Differential Revision: https://reviews.llvm.org/D94834
This patch marks some library functions as willreturn. On the first pass, I
excluded most functions that interact with streams/the filesystem.
Along with willreturn, it also adds nounwind to a set of math functions.
There probably are a few additional attributes we can add for those, but
that should be done separately.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D94684
Keys matching the tombstone/empty special values cannot be inserted in a
DenseMap. Under some circumstances, LV tries to add members to an
interleave group that match the special values. Skip adding such
members. This is unlikely to have any impact in practice, because
interleave groups with such indices are very likely to not be
vectorized, due to gaps.
This issue has been surfaced by fuzzing, see
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=11638
We can fold x*C1/C2 <= x to true if C1 <= C2. This is valid even
if the multiplication is not nuw: https://alive2.llvm.org/ce/z/vULors
The multiplication or division can be replaced by shifts. We don't
handle the case where both are shifts, as that should get folded
away by InstCombine.
D91936 placed the tracking for the assumptions into BasicAA.
However, when recursing over phis, we may use fresh AAQI instances.
In this case AssumptionBasedResults from an inner AAQI can reesult
in a removal of an element from the outer AAQI.
To avoid this, move the tracking into AAQI. This generally makes
more sense, as the NoAlias assumptions themselves are also stored
in AAQI.
The test case only produces an assertion failure with D90094
reapplied. I think the issue exists independently of that change
as well, but I wasn't able to come up with a reproducer.