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
This is a straightforward port of MemCpyOpt to MemorySSA following
the approach of D26739. MemDep queries are replaced with MSSA queries
without changing the overall structure of the pass. Some care has
to be taken to account for differences between these APIs
(MemDep also returns reads, MSSA doesn't).
Differential Revision: https://reviews.llvm.org/D89207
We were not correctly splitting a blocks for chains of length 1.
Before that change, additional instructions for blocks in chains of
length 1 were not split off from the block before removing (this was
done correctly for chains of longer size).
If this first block contained an instruction referenced elsewhere,
deleting the block, would result in invalidation of the produced value.
This caused a miscompile which motivated D92297 (before D17993,
nonnull and dereferenceable attributed were not added so MergeICmps were
not triggered.) The new test gep-references-bb.ll demonstrate the issue.
The regression was introduced in
rG0efadbbcdeb82f5c14f38fbc2826107063ca48b2.
This supersedes D92364.
Test case by MaskRay (Fangrui Song).
Differential Revision: https://reviews.llvm.org/D92375
icmp is the preferred spelling in IR because icmp analysis is
expected to be better than any other analysis. This should
lead to more follow-on folding potential.
It's difficult to say exactly what we should do in codegen to
compensate. For example on AArch64, which of these is preferred:
sub w8, w0, w1
lsr w0, w8, #31
vs:
cmp w0, w1
cset w0, lt
If there are perf regressions, then we should deal with those in
codegen on a case-by-case basis.
A possible motivating example for better optimization is shown in:
https://llvm.org/PR43198 but that will require other transforms
before anything changes there.
Alive proof:
https://rise4fun.com/Alive/o4E
Name: sign-bit splat
Pre: C1 == (width(%x) - 1)
%s = sub nsw %x, %y
%r = ashr %s, C1
=>
%c = icmp slt %x, %y
%r = sext %c
Name: sign-bit LSB
Pre: C1 == (width(%x) - 1)
%s = sub nsw %x, %y
%r = lshr %s, C1
=>
%c = icmp slt %x, %y
%r = zext %c
If the shift amount was undef for some lane, the shift amount in opposite
shift is irrelevant for that lane, and the new shift amount for that lane
can be undef.
If the shift amount was undef for some lane, the shift amount in opposite
shift is irrelevant for that lane, and the new shift amount for that lane
can be undef.
There is no correctness need for that, and since we allow live-out
uses, this could theoretically happen, because currently nothing
will move the cond to right before the branch in those tests.
But regardless, lifting that restriction even makes the transform
easier to understand.
This makes the transform happen in 81 more cases (+0.55%)
)
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
This is the #2 of 2 changes that make remarks hotness threshold option
available in more tools. The changes also allow the threshold to sync with
hotness threshold from profile summary with special value 'auto'.
This change expands remarks hotness threshold option
-fdiagnostics-hotness-threshold in clang and *-remarks-hotness-threshold in
other tools to utilize hotness threshold from profile summary.
Remarks hotness filtering relies on several driver options. Table below lists
how different options are correlated and affect final remarks outputs:
| profile | hotness | threshold | remarks printed |
|---------|---------|-----------|-----------------|
| No | No | No | All |
| No | No | Yes | None |
| No | Yes | No | All |
| No | Yes | Yes | None |
| Yes | No | No | All |
| Yes | No | Yes | None |
| Yes | Yes | No | All |
| Yes | Yes | Yes | >=threshold |
In the presence of profile summary, it is often more desirable to directly use
the hotness threshold from profile summary. The new argument value 'auto'
indicates threshold will be synced with hotness threshold from profile summary
during compilation. The "auto" threshold relies on the availability of profile
summary. In case of missing such information, no remarks will be generated.
Differential Revision: https://reviews.llvm.org/D85808
Enable performing mandatory inlinings upfront, by reusing the same logic
as the full inliner, instead of the AlwaysInliner. This has the
following benefits:
- reduce code duplication - one inliner codebase
- open the opportunity to help the full inliner by performing additional
function passes after the mandatory inlinings, but before th full
inliner. Performing the mandatory inlinings first simplifies the problem
the full inliner needs to solve: less call sites, more contextualization, and,
depending on the additional function optimization passes run between the
2 inliners, higher accuracy of cost models / decision policies.
Note that this patch does not yet enable much in terms of post-always
inline function optimization.
Differential Revision: https://reviews.llvm.org/D91567
Pass through the demanded elts mask to the source operands.
The next step will be to add support for folding to add/sub if we only demand odd/even elements.
This was suggested in D92247 - I initially committed an alternate
fix ( bfd2c216ea ) to avoid the crash/assert shown in
https://llvm.org/PR48296 ,
but that was reverted because it caused msan failures on other
tests. We can try to revive that patch using the test included
here, but I do not have an immediate plan to isolate that problem.
https://llvm.org/PR48296 shows an example where we delete all of the operands
of a phi without actually deleting the phi, and that is currently considered
invalid IR. The reduced test included here would crash for that reason.
A suggested follow-up is to loosen the assert to allow 0-operand phis
in unreachable blocks.
Differential Revision: https://reviews.llvm.org/D92247
.. because it causes miscompilation when combined with select i1 -> and/or.
It is the select fold which is incorrect; but it is costly to disable the fold, so hack this one.
D92270
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
If we decided to widen IV with zext, then unsigned comparisons
should not prevent widening (same for sext/sign comparisons).
The result of comparison in wider type does not change in this case.
Differential Revision: https://reviews.llvm.org/D92207
Reviewed By: nikic
Reverting commit due to address sanitizer errors.
> Extracting the similar regions is the first step in the IROutliner.
>
> Using the IRSimilarityIdentifier, we collect the SimilarityGroups and
> sort them by how many instructions will be removed. Each
> IRSimilarityCandidate is used to define an OutlinableRegion. Each
> region is ordered by their occurrence in the Module and the regions that
> are not compatible with previously outlined regions are discarded.
>
> Each region is then extracted with the CodeExtractor into its own
> function.
>
> We test that correctly extract in:
> test/Transforms/IROutliner/extraction.ll
> test/Transforms/IROutliner/address-taken.ll
> test/Transforms/IROutliner/outlining-same-globals.ll
> test/Transforms/IROutliner/outlining-same-constants.ll
> test/Transforms/IROutliner/outlining-different-structure.ll
>
> Reviewers: paquette, jroelofs, yroux
>
> Differential Revision: https://reviews.llvm.org/D86975
This reverts commit bf899e8913.
Extracting the similar regions is the first step in the IROutliner.
Using the IRSimilarityIdentifier, we collect the SimilarityGroups and
sort them by how many instructions will be removed. Each
IRSimilarityCandidate is used to define an OutlinableRegion. Each
region is ordered by their occurrence in the Module and the regions that
are not compatible with previously outlined regions are discarded.
Each region is then extracted with the CodeExtractor into its own
function.
We test that correctly extract in:
test/Transforms/IROutliner/extraction.ll
test/Transforms/IROutliner/address-taken.ll
test/Transforms/IROutliner/outlining-same-globals.ll
test/Transforms/IROutliner/outlining-same-constants.ll
test/Transforms/IROutliner/outlining-different-structure.ll
Reviewers: paquette, jroelofs, yroux
Differential Revision: https://reviews.llvm.org/D86975
Folding a select of vector constants that include undef elements only
applies to fixed vectors, but there's no earlier check the type is not
scalable so it crashes for scalable vectors. This adds a check so this
optimization is only attempted for fixed vectors.
Reviewed By: sdesmalen
Differential Revision: https://reviews.llvm.org/D92046
This was orginally committed in 2245fb8aaa.
but was immediately reverted in f3abd54958
because of a PHI handling issue.
Original commit message:
1. It doesn't make sense to enforce that the bonus instruction
is only used once in it's basic block. What matters is
whether those user instructions fit within our budget, sure,
but that is another question.
2. It doesn't make sense to enforce that said bonus instructions
are only used within their basic block. Perhaps the branch
condition isn't using the value computed by said bonus instruction,
and said bonus instruction is simply being calculated
to be used in successors?
So iff we can clone bonus instructions, to lift these restrictions,
we just need to carefully update their external uses
to use the new cloned instructions.
Notably, this transform (even without this change) appears to be
poison-unsafe as per alive2, but is otherwise (including the patch) legal.
We don't introduce any new PHI nodes, but only "move" the instructions
around, i'm not really seeing much potential for extra cost modelling
for the transform, especially since now we allow at most one such
bonus instruction by default.
This causes the fold to fire +11.4% more (13216 -> 14725)
as of vanilla llvm test-suite + RawSpeed.
The motivational pattern is IEEE-754-2008 Binary16->Binary32
extension code:
ca57d77fb2/src/librawspeed/common/FloatingPoint.h (L115-L120)
^ that should be a switch, but it is not now: https://godbolt.org/z/bvja5v
That being said, even thought this seemed like this would fix it: https://godbolt.org/z/xGq3TM
apparently that fold is happening somewhere else afterall,
so something else also has a similar 'artificial' restriction.