Currently LLVM supports vectorization of horizontal reduction
instructions with initial value set to 0. Patch supports vectorization
of reduction with non-zero initial values. Also it supports a
vectorization of instructions with some extra arguments, like:
float f(float x[], int a, int b) {
float p = a % b;
p += x[0] + 3;
for (int i = 1; i < 32; i++)
p += x[i];
return p;
}
Patch allows vectorization of this kind of horizontal reductions.
Differential Revision: https://reviews.llvm.org/D28961
llvm-svn: 293994
This patch moves some helper functions related to interleaved access
vectorization out of LoopVectorize.cpp and into VectorUtils.cpp. We would like
to use these functions in a follow-on patch that improves interleaved load and
store lowering in (ARM/AArch64)ISelLowering.cpp. One of the functions was
already duplicated there and has been removed.
Differential Revision: https://reviews.llvm.org/D29398
llvm-svn: 293788
By calling getScalarizationOverhead with the CallInst instead of the types of
its arguments, we make sure that only unique call arguments are added to the
scalarization cost.
getScalarizationOverhead() is extended to handle calls by only passing on the
actual call arguments (which is not all the operands).
This also eliminates a wrapper function with the same name.
review: Hal Finkel
llvm-svn: 293459
The jumbled scalar loads will be sorted while building the tree and these accesses will be marked to generate shufflevector after the vectorized load with proper mask.
Reviewers: hfinkel, mssimpso, mkuper
Differential Revision: https://reviews.llvm.org/D26905
Change-Id: I9c0c8e6f91a00076a7ee1465440a3f6ae092f7ad
llvm-svn: 293386
Some checks in SLP horizontal reduction analysis function are performed
several times, though it is enough to perform these checks only once
during an initial attempt at adding candidate for the reduction
instruction/reduced value.
Differential Revision: https://reviews.llvm.org/D29175
llvm-svn: 293274
change the set of uniform instructions in the loop causing an assert
failure.
The problem is that the legalization checking also builds data
structures mapping various facts about the loop body. The immediate
cause was the set of uniform instructions. If these then change when
LCSSA is formed, the data structures would already have been built and
become stale. The included test case triggered an assert in loop
vectorize that was reduced out of the new PM's pipeline.
The solution is to form LCSSA early enough that no information is cached
across the changes made. The only really obvious position is outside of
the main logic to vectorize the loop. This also has the advantage of
removing one case where forming LCSSA could mutate the loop but we
wouldn't track that as a "Changed" state.
If it is significantly advantageous to do some legalization checking
prior to this, we can do a more careful positioning but it seemed best
to just back off to a safe position first.
llvm-svn: 293168
Refactoring to remove duplications of this method.
New method getOperandsScalarizationOverhead() that looks at the present unique
operands and add extract costs for them. Old behaviour was to just add extract
costs for one operand of the type always, which still happens in
getArithmeticInstrCost() if no operands are provided by the caller.
This is a good start of improving on this, but there are more places
that can be improved by using getOperandsScalarizationOverhead().
Review: Hal Finkel
https://reviews.llvm.org/D29017
llvm-svn: 293155
instructions.
If number of instructions in horizontal reduction list is not power of 2
then only PowerOf2Floor(NumberOfInstructions) last elements are actually
vectorized, other instructions remain scalar. Patch tries to vectorize
the remaining elements either.
Differential Revision: https://reviews.llvm.org/D28959
llvm-svn: 293042
Removed data members ReduxWidth and MinVecRegSize + some C++11 stylish
improvements.
Differential Revision: https://reviews.llvm.org/D29010
llvm-svn: 292899
This changes the vectorizer to explicitly use the loopsimplify and lcssa utils,
instead of "requiring" the transformations as if they were analyses.
This is not NFC, since it changes the LCSSA behavior - we no longer run LCSSA
for all loops, but rather only for the loops we expect to modify.
Differential Revision: https://reviews.llvm.org/D28868
llvm-svn: 292456
We currently check whether a reduction has a single outside user. We don't
really need to require that - we just need to make sure a single value is
used externally. The number of external users of that value shouldn't actually
matter.
Differential Revision: https://reviews.llvm.org/D28830
llvm-svn: 292424
If a memory instruction will be vectorized, but it's pointer operand is
non-consecutive-like, the instruction is a gather or scatter operation. Its
pointer operand will be non-uniform. This should fix PR31671.
Reference: https://llvm.org/bugs/show_bug.cgi?id=31671
Differential Revision: https://reviews.llvm.org/D28819
llvm-svn: 292254
a function's CFG when that CFG is unchanged.
This allows transformation passes to simply claim they preserve the CFG
and analysis passes to check for the CFG being preserved to remove the
fanout of all analyses being listed in all passes.
I've gone through and removed or cleaned up as many of the comments
reminding us to do this as I could.
Differential Revision: https://reviews.llvm.org/D28627
llvm-svn: 292054
The removed assert seems bogus - it's perfectly legal for the roots of the
vectorized subtrees to be equal even if the original scalar values aren't,
if the original scalars happen to be equivalent.
This fixes PR31599.
Differential Revision: https://reviews.llvm.org/D28539
llvm-svn: 291692
updated instructions:
pmulld, pmullw, pmulhw, mulsd, mulps, mulpd, divss, divps, divsd, divpd, addpd and subpd.
special optimization case which replaces pmulld with pmullw\pmulhw\pshuf seq.
In case if the real operands bitwidth <= 16.
Differential Revision: https://reviews.llvm.org/D28104
llvm-svn: 291657
arguments much like the CGSCC pass manager.
This is a major redesign following the pattern establish for the CGSCC layer to
support updates to the set of loops during the traversal of the loop nest and
to support invalidation of analyses.
An additional significant burden in the loop PM is that so many passes require
access to a large number of function analyses. Manually ensuring these are
cached, available, and preserved has been a long-standing burden in LLVM even
with the help of the automatic scheduling in the old pass manager. And it made
the new pass manager extremely unweildy. With this design, we can package the
common analyses up while in a function pass and make them immediately available
to all the loop passes. While in some cases this is unnecessary, I think the
simplicity afforded is worth it.
This does not (yet) address loop simplified form or LCSSA form, but those are
the next things on my radar and I have a clear plan for them.
While the patch is very large, most of it is either mechanically updating loop
passes to the new API or the new testing for the loop PM. The code for it is
reasonably compact.
I have not yet updated all of the loop passes to correctly leverage the update
mechanisms demonstrated in the unittests. I'll do that in follow-up patches
along with improved FileCheck tests for those passes that ensure things work in
more realistic scenarios. In many cases, there isn't much we can do with these
until the loop simplified form and LCSSA form are in place.
Differential Revision: https://reviews.llvm.org/D28292
llvm-svn: 291651
This patch delays the fix-up step for external induction variable users until
after the dominator tree has been properly updated. This should fix PR30742.
The SCEVExpander in InductionDescriptor::transform can generate code in the
wrong location if the dominator tree is not up-to-date. We should work towards
keeping the dominator tree up-to-date throughout the transformation.
Reference: https://llvm.org/bugs/show_bug.cgi?id=30742
Differential Revision: https://reviews.llvm.org/D28168
llvm-svn: 291462
This code seems to be target dependent which may not be the same for all targets.
Passed the decision whether the given stride is complex or not to the target by sending stride information via SCEV to getAddressComputationCost instead of 'IsComplex'.
Specifically at X86 targets we dont see any significant address computation cost in case of the strided access in general.
Differential Revision: https://reviews.llvm.org/D27518
llvm-svn: 291106
This patch reapplies r289863. The original patch was reverted because it
exposed a bug causing the loop vectorizer to crash in the Python runtime on
PPC. The underlying issue was fixed with r289958.
llvm-svn: 289975
After r288909, instructions feeding predicated instructions may be scalarized
if profitable. Since these instructions will remain scalar, we shouldn't
attempt to type-shrink them. We should only truncate vector types to their
minimal bit widths. This bug was exposed by enabling the vectorization of loops
containing conditional stores by default.
llvm-svn: 289958
stores by default
This uncovers a crasher in the loop vectorizer on PPC when building the
Python runtime. I'll send the testcase to the review thread for the
original commit.
llvm-svn: 289934
This patch sets the default value of the "-enable-cond-stores-vec" command line
option to "true".
Differential Revision: https://reviews.llvm.org/D27814
llvm-svn: 289863
After r289755, the AssumptionCache is no longer needed. Variables affected by
assumptions are now found by using the new operand-bundle-based scheme. This
new scheme is more computationally efficient, and also we need much less
code...
llvm-svn: 289756
We currently check if the exact trip count is known and is smaller than the
"tiny loop" bound. We should be checking the maximum bound on the trip count
instead.
Differential Revision: https://reviews.llvm.org/D27690
llvm-svn: 289583
This patch ensures the correct minimum bit width during type-shrinking.
Previously when type-shrinking, we always sign-extended values back to their
original width. However, if we are going to sign-extend, and the sign bit is
unknown, we have to increase the minimum bit width by one bit so the
sign-extend will fill the upper bits correctly. If the sign bit is known to be
zero, we can perform a zero-extend instead. This should fix PR31243.
Reference: https://llvm.org/bugs/show_bug.cgi?id=31243
Differential Revision: https://reviews.llvm.org/D27466
llvm-svn: 289470
When trying to vectorize trees that start at insertelement instructions
function tryToVectorizeList() uses vectorization factor calculated as
MinVecRegSize/ScalarTypeSize. But sometimes it does not work as tree
cost for this fixed vectorization factor is too high.
Patch tries to improve the situation. It tries different vectorization
factors from max(PowerOf2Floor(NumberOfVectorizedValues),
MinVecRegSize/ScalarTypeSize) to MinVecRegSize/ScalarTypeSize and tries
to choose the best one.
Differential Revision: https://reviews.llvm.org/D27215
llvm-svn: 289043
This patch attempts to scalarize the operand expressions of predicated
instructions if they were conditionally executed in the original loop. After
scalarization, the expressions will be sunk inside the blocks created for the
predicated instructions. The transformation essentially performs
un-if-conversion on the operands.
The cost model has been updated to determine if scalarization is profitable. It
compares the cost of a vectorized instruction, assuming it will be
if-converted, to the cost of the scalarized instruction, assuming that the
instructions corresponding to each vector lane will be sunk inside a predicated
block, possibly avoiding execution. If it's more profitable to scalarize the
entire expression tree feeding the predicated instruction, the expression will
be scalarized; otherwise, it will be vectorized. We only consider the cost of
the entire expression to accurately estimate the cost of the required
insertelement and extractelement instructions.
Differential Revision: https://reviews.llvm.org/D26083
llvm-svn: 288909
This reverts commit r288497, as it broke the AArch64 build of Compiler-RT's
builtins (twice: once in r288412 and once in r288497). We should investigate
this offline.
llvm-svn: 288508
When trying to vectorize trees that start at insertelement instructions
function tryToVectorizeList() uses vectorization factor calculated as
MinVecRegSize/ScalarTypeSize. But sometimes it does not work as tree
cost for this fixed vectorization factor is too high.
Patch tries to improve the situation. It tries different vectorization
factors from max(PowerOf2Floor(NumberOfVectorizedValues),
MinVecRegSize/ScalarTypeSize) to MinVecRegSize/ScalarTypeSize and tries
to choose the best one.
Differential Revision: https://reviews.llvm.org/D27215
llvm-svn: 288497
When trying to vectorize trees that start at insertelement instructions
function tryToVectorizeList() uses vectorization factor calculated as
MinVecRegSize/ScalarTypeSize. But sometimes it does not work as tree
cost for this fixed vectorization factor is too high.
Patch tries to improve the situation. It tries different vectorization
factors from max(PowerOf2Floor(NumberOfVectorizedValues),
MinVecRegSize/ScalarTypeSize) to MinVecRegSize/ScalarTypeSize and tries
to choose the best one.
Differential Revision: https://reviews.llvm.org/D27215
llvm-svn: 288412
Currently when cost of scalar operations is evaluated the vector type is
used for scalar operations. Patch fixes this issue and fixes evaluation
of the vector operations cost.
Several test showed that vector cost model is too optimistic. It
allowed vectorization of 8 or less add/fadd operations, though scalar
code is faster. Actually, only for 16 or more operations vector code
provides better performance.
Differential Revision: https://reviews.llvm.org/D26277
llvm-svn: 288398
Currently SLP vectorizer tries to vectorize a binary operation and dies
immediately after unsuccessful the first unsuccessfull attempt. Patch
tries to improve the situation, trying to vectorize all binary
operations of all children nodes in the binop tree.
Differential Revision: https://reviews.llvm.org/D25517
llvm-svn: 288115
Summary:
The "getVectorizablePrefix" method would give up if it found an aliasing load for a store chain.
In practice, the aliasing load can be treated as a memory barrier and all stores that precede it
are a valid vectorizable prefix.
Issue found by volkan in D26962. Testcase is a pruned version of the one in the original patch.
Reviewers: jlebar, arsenm, tstellarAMD
Subscribers: mzolotukhin, wdng, nhaehnle, anna, volkan, llvm-commits
Differential Revision: https://reviews.llvm.org/D27008
llvm-svn: 287781
This patch updates a bunch of places where add_dependencies was being explicitly called to add dependencies on intrinsics_gen to instead use the DEPENDS named parameter. This cleanup is needed for a patch I'm working on to add a dependency debugging mode to the build system.
llvm-svn: 287206
The register usage algorithm incorrectly treats instructions whose value is
not used within the loop (e.g. those that do not produce a value).
The algorithm first calculates the usages within the loop. It iterates over
the instructions in order, and records at which instruction index each use
ends (in fact, they're actually recorded against the next index, as this is
when we want to delete them from the open intervals).
The algorithm then iterates over the instructions again, adding each
instruction in turn to a list of open intervals. Instructions are then
removed from the list of open intervals when they occur in the list of uses
ended at the current index.
The problem is, instructions which are not used in the loop are skipped.
However, although they aren't used, the last use of a value may have been
recorded against that instruction index. In this case, the use is not deleted
from the open intervals, which may then bump up the estimated register usage.
This patch fixes the issue by simply moving the "is used" check after the loop
which erases the uses at the current index.
Differential Revision: https://reviews.llvm.org/D26554
llvm-svn: 286969
Michael Kuperstein