The new mask represents the order, not the mask itself. At first, need
to treat as the order, convert to mask and only after that reorder
gathered scalars to build correct clustered order.
Differential Revision: https://reviews.llvm.org/D141161
Need to include the cost of the initial insertelement to the cost of the
broadcasts. Also, need to adjust the cost of the gather/buildvector if
the element is inserted into poison/undef vector.
Differential Revision: https://reviews.llvm.org/D140498
Make a separate routine for GEPs cost calculation and make
the approach uniform across load, store and GEP tree nodes.
Additional issue fixed is GEP cost savings were applied twice
for ScatterVectorize nodes (aka gather load) making them look
unrealistically profitable for vectorization.
Differential Revision: https://reviews.llvm.org/D140789
analysis.
Missed the analysis of the shuffle mask when trying to analyze the
operands of the shuffle instruction during peeking through shuffle
instructions.
We incorrectly assume intrinsic as a function call and it prevents us from
the opportunity to vectorize. On Aarch64 Cortex-A53 we think that
llvm.fmuladd.f64 is a function call which is wrong.
Differential Revision: https://reviews.llvm.org/D140392
SLP produces two gather loads (one feeds another).
For the first set of scalar loads GEP indices are all constant.
The result of the second load is then fed into reduction (as a seed).
Differential Revision: https://reviews.llvm.org/D140785
We do not need to emit many extractelements for each particular use, we
can reuse the only one, just need to adjust it to make it dominate on
all uses.
Differential Revision: https://reviews.llvm.org/D140580
We can use ShuffleInstructionBuilder now for shrinking shuffle emission.
It allows to remove extra shuffle from the emitted code and reuse
original vector.
Part of D110978
Differential Revision: https://reviews.llvm.org/D140499
of extractvalues.
No need to get the last instruction only for vectorized extractvalues,
for gathered(buildvector sequence) still need to get the insertion
point.
Added BaseShuffleAnalysis as a base class for ShuffleInstructionBuilder
and integrated shuffle logic from shuffles for externally used scalars
into this class. This class is used as the main container that
implements smart shuffle instruction builder logic.
ShuffleInstructionBuilder uses this logic.
ShuffleInstructionBuilder is also used in building of the shuffle for
the externally used scalars instead of lambdas, which are now part of BaseShuffleAnalysis class.
Differential Revision: https://reviews.llvm.org/D140100
Opaque ptr types have a size in bits of 0. The legalised type is an i64 or
vector of i64s, which do have a size. Because of this difference in size, target
hook getMemoryOpCost modelled stores of ptr types as extending/truncating
load/stores. Now we just check for opaque ptr types and return the legalised
cost. This makes stores of pointers cheaper, and as a result we now SLP
vectorise the changed test case.
Differential Revision: https://reviews.llvm.org/D140193
When we sort entries for attempting to reorder scalars, need to use
actual vectorization factor, not the number of scalars. Otherwise the
compiler crashes, if the scalars has to be reordered.
Differential Revision: https://reviews.llvm.org/D138819
If left unchecked, the SLPVecrtorizer can move loads/stores below a stackrestore. The move can cause issues if the loads/stores have pointer operands from `alloca`s that are reset by the stackrestores. This patch adds the dependency check.
The check is conservative, in that it does not check if the pointer operands of the loads/stores are actually from `alloca`s that may be reset. We did not observe any SPECCPU2017 performance degradation so this simple fix seems sufficient.
The test could have been added to `llvm/test/Transforms/SLPVectorizer/X86/stacksave-dependence.ll`, but that test has not been updated to use opaque pointers. I am not inclined to add tests that still use typed pointers, or to refactor `llvm/test/Transforms/SLPVectorizer/X86/stacksave-dependence.ll` to use opaque pointers in this patch. If desired, I will open a different patch to refactor and consolidate the tests.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D138585
extractelements.
If the resulting type is going to be scalarized, no need to adjust the
cost of removed extractelement and insert/extract subvector costs.
Otherwise, the compiler can crash because of the wrong type sizes.
Need to count the reduced values, vectorized in the tree but not in the top node. Such scalars still must be extracted out of the vector node instead of the original scalar.
If same instruction is reduced several times, but in one graph is part
of buildvector sequence and in another it is vectorized, we may loose
information that it was part of buildvector and must be extracted from
later vectorized value.
If the graph is only the buildvector node without main operation, need
to inherit insrtpoint from the redution instruction. Otherwise the
compiler crashes trying to insert instruction at the entry block.
Need to use advanced check for the same vectorized node to avoid
possible compiler crash. We may have 2 similar nodes (vector one and
gather) after graph nodes rotation, need to do extra checks for the
exact match.
All of our insert/extract ops work on 128-bit lanes.
For `Insert`, we need to extract affected 128-bit lane,
unless it's being fully overwritten (FIXME: do we need to be
careful about legalization-induced padding that we obviously don't demand?),
perform insertions, and then insert the 128-bit lane back.
But hold on. If we are operating on an 256-bit legal vector,
and thus have two 128-bit subvectors, and are fully overwriting them both,
we don't actually need to insert *both* subvectors,
only the second one, into the implicitly-widened first one.
Also, `Insert` wasn't actually querying the costs,
but just assuming them to be `1`.
`getShuffleCost(TTI::SK_ExtractSubvector)` notes:
```
// Note that in general, the insertion starting at the beginning of a vector
// isn't free, because we need to preserve the rest of the wide vector.
```
... so as far as i can tell, we didn't account for that.
I was hoping this would allow vectorization at a higher VF at one case i looked at,
but the subvector insertion cost is still dis-advising that.
The change for `Extract` is NFC, and is for consistency only,
i wanted to get rid of of that weird explicit discounting of insertion of 0'th element,
since the general code should already deal with that.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D137913
Gather nodes are vectorized as simply vector of the scalars instead of
relying on the actual node. It leads to the fact that in some cases
we may miss incorrect transformation (non-matching set of scalars is
just ended as a gather node instead of possible vector/gather node).
Better to rely on the actual nodes, it allows to improve stability and
better detect missed cases.
Differential Revision: https://reviews.llvm.org/D135174
Need to check if the insertelement mask size is reached during cost analysis to avoid compiler crash.
Differential Revision: https://reviews.llvm.org/D137639
Gather nodes are vectorized as simply vector of the scalars instead of
relying on the actual node. It leads to the fact that in some cases
we may miss incorrect transformation (non-matching set of scalars is
just ended as a gather node instead of possible vector/gather node).
Better to rely on the actual nodes, it allows to improve stability and
better detect missed cases.
Differential Revision: https://reviews.llvm.org/D135174
Generalized the cost model estimation. Improved cost model estimation
for repeated scalars (no need to count their cost anymore), improved
cost model for extractelement instructions.
cpu2017
511.povray_r 0.57
520.omnetpp_r -0.98
521.wrf_r -0.01
525.x264_r 3.59 <+
526.blender_r -0.12
531.deepsjeng_r -0.07
538.imagick_r -1.42
Geometric mean: 0.21
Differential Revision: https://reviews.llvm.org/D115757
