For tight loops like this:
float r = 0;
for (int i = 0; i < n; i++) {
r += a[i];
}
it's better not to vectorise at -O3 using fixed-width ordered reductions
on AArch64 targets. Although the resulting number of instructions in the
generated code ends up being comparable to not vectorising at all, there
may be additional costs on some CPUs, for example perhaps the scheduling
is worse. It makes sense to deter vectorisation in tight loops.
Differential Revision: https://reviews.llvm.org/D108292
Removed AArch64 usage of the getMaxVScale interface, replacing it with
the vscale_range(min, max) IR Attribute.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D106277
Replace vector unpack operation with a scalar extend operation.
unpack(splat(X)) --> splat(extend(X))
If we have both, unpkhi and unpklo, for the same vector then we may
save a register in some cases, e.g:
Hi = unpkhi (splat(X))
Lo = unpklo(splat(X))
--> Hi = Lo = splat(extend(X))
Differential Revision: https://reviews.llvm.org/D106929
Change-Id: I77c5c201131e3a50de1cdccbdcf84420f5b2244b
Move the last{a,b} operation to the vector operand of the binary instruction if
the binop's operand is a splat value. This essentially converts the binop
to a scalar operation.
Example:
// If x and/or y is a splat value:
lastX (binop (x, y)) --> binop(lastX(x), lastX(y))
Differential Revision: https://reviews.llvm.org/D106932
Change-Id: I93ff5302f9a7972405ee0d3854cf115f072e99c0
This takes the existing SVE costing for the various min/max reduction
intrinsics and expands it to NEON, where I believe it applies equally
well.
In the process it changes the lowering to use min/max cost, as opposed
to summing up the cost of ICmp+Select.
Differential Revision: https://reviews.llvm.org/D106239
I'm not sure this is the best way to approach this,
but the situation is rather not very detectable unless we explicitly call it out when refusing to advise to unroll.
Reviewed By: efriedma
Differential Revision: https://reviews.llvm.org/D107271
I have added a new FastMathFlags parameter to getArithmeticReductionCost
to indicate what type of reduction we are performing:
1. Tree-wise. This is the typical fast-math reduction that involves
continually splitting a vector up into halves and adding each
half together until we get a scalar result. This is the default
behaviour for integers, whereas for floating point we only do this
if reassociation is allowed.
2. Ordered. This now allows us to estimate the cost of performing
a strict vector reduction by treating it as a series of scalar
operations in lane order. This is the case when FP reassociation
is not permitted. For scalable vectors this is more difficult
because at compile time we do not know how many lanes there are,
and so we use the worst case maximum vscale value.
I have also fixed getTypeBasedIntrinsicInstrCost to pass in the
FastMathFlags, which meant fixing up some X86 tests where we always
assumed the vector.reduce.fadd/mul intrinsics were 'fast'.
New tests have been added here:
Analysis/CostModel/AArch64/reduce-fadd.ll
Analysis/CostModel/AArch64/sve-intrinsics.ll
Transforms/LoopVectorize/AArch64/strict-fadd-cost.ll
Transforms/LoopVectorize/AArch64/sve-strict-fadd-cost.ll
Differential Revision: https://reviews.llvm.org/D105432
This adds some missing single source shuffle costs for AArch64, of i16
and i8 vectors. v4i16 are the same as v4i32 with a worse case cost of 3
coming from the perfect shuffle tables. The larger vector sizes expand
into a constant pool, plus a load (and adrp) and a tbl. I arbitrarily
chose 8 for the cost to be expensive but not too expensive.
Differential Revision: https://reviews.llvm.org/D106241
This changes the cost to (LT.first-1) * cost(add) + 2, where the cost of
an add is assumed to be 1. This brings it inline with the other
reductions.
Differential Revision: https://reviews.llvm.org/D106240
At the moment, <vscale x 1 x eltty> are not yet fully handled by the
code-generator, so to avoid vectorizing loops with that VF, we mark the
cost for these types as invalid.
The reason for not adding a new "TTI::getMinimumScalableVF" is because
the type is supposed to be a type that can be legalized. It partially is,
although the support for these types need some more work.
Reviewed By: paulwalker-arm, dmgreen
Differential Revision: https://reviews.llvm.org/D103882
This patch removes the IsPairwiseForm flag from the Reduction Cost TTI
hooks, along with some accompanying code for pattern matching reductions
from trees starting at extract elements. IsPairWise is now assumed to be
false, which was the predominant way that the value was used from both
the Loop and SLP vectorizers. Since the adjustments such as D93860, the
SLP vectorizer has not relied upon this distinction between paiwise and
non-pairwise reductions.
This also removes some code that was detecting reductions trees starting
from extract elements inside the costmodel. This case was
double-counting costs though, adding the individual costs on the
individual instruction _and_ the total cost of the reduction. Removing
it changes the costs in llvm/test/Analysis/CostModel/X86/reduction.ll to
not double count. The cost of reduction intrinsics is still tested
through the various tests in
llvm/test/Analysis/CostModel/X86/reduce-xyz.ll.
Differential Revision: https://reviews.llvm.org/D105484
This patch adds a TTI function, isElementTypeLegalForScalableVector, to query
whether it is possible to vectorize a given element type. This is called by
isLegalToVectorizeInstTypesForScalable to reject scalable vectorization if
any of the instruction types in the loop are unsupported, e.g:
int foo(__int128_t* ptr, int N)
#pragma clang loop vectorize_width(4, scalable)
for (int i=0; i<N; ++i)
ptr[i] = ptr[i] + 42;
This example currently crashes if we attempt to vectorize since i128 is not a
supported type for scalable vectorization.
Reviewed By: sdesmalen, david-arm
Differential Revision: https://reviews.llvm.org/D102253
This patch adds a new ShuffleKind SK_Splice and then handle the cost in
getShuffleCost, as in experimental.vector.reverse.
Differential Revision: https://reviews.llvm.org/D104630
Loads of <4 x i8> vectors were modeled as extremely expensive. And while we
don't have a load instruction that supports this, it isn't that expensive to
create a vector of i8 elements. The codegen for this was fixed/optimised in
D105110. This now tweaks the cost model and enables SLP vectorisation of my
motivating case loadi8.ll.
Differential Revision: https://reviews.llvm.org/D103629
OR, XOR and AND entries are added to the cost table. An extra cost
is added when vector splitting occurs.
This is done to address the issue of a missed SLP vectorization
opportunity due to unreasonably high costs being attributed to the vector
Or reduction (see: https://bugs.llvm.org/show_bug.cgi?id=44593).
Differential Revision: https://reviews.llvm.org/D104538
Added a case for CTPOP to AArch64TTIImpl::getIntrinsicInstrCost so that
the cost estimate matches the codegen in
test/CodeGen/AArch64/arm64-vpopcnt.ll
Differential Revision: https://reviews.llvm.org/D103952
We were passing the RecurrenceDescriptor by value to most of the reduction analysis methods, despite it being rather bulky with TrackingVH members (that can be costly to copy). In all these cases we're only using the RecurrenceDescriptor for rather basic purposes (access to types/kinds etc.).
Differential Revision: https://reviews.llvm.org/D104029
Fixes getTypeConversion to return `TypeScalarizeScalableVector` when a scalable vector
type cannot be legalized by widening/splitting. When this is the method of legalization
found, getTypeLegalizationCost will return an Invalid cost.
The getMemoryOpCost, getMaskedMemoryOpCost & getGatherScatterOpCost functions already call
getTypeLegalizationCost and will now also return an Invalid cost for unsupported types.
Reviewed By: sdesmalen, david-arm
Differential Revision: https://reviews.llvm.org/D102515
Use llvm.experimental.vector.insert instead of storing into an alloca
when generating code for these intrinsics. This defers the codegen of
the generated vector to instruction selection, allowing existing
shufflevector style optimizations to apply.
Additionally, introduce a new target transform that can recognise fixed
predicate patterns in the svbool variants of these intrinsics.
Differential Revision: https://reviews.llvm.org/D103082
Due to the dependency on runtime unrolling, UnJ is only
enabled by default on in-order scheduling models,
and if a cpu is specified through -mcpu.
Differential Revision: https://reviews.llvm.org/D103604
Follow up to D101357 / 3fa6510f6.
Supersedes D102330.
Goal: Use flags setting rdffrs instead of rdffr + ptest.
Problem: RDFFR_P doesn't have have a flags setting equivalent.
Solution: in instcombine, canonicalize to RDFFR_PP at the IR level, and
rely on RDFFR_PP+PTEST => RDFFRS_PP optimization in
AArch64InstrInfo::optimizePTestInstr.
While here:
* Test that rdffr.z+ptest generates a rdffrs.
* Use update_{test,llc}_checks.py on the tests.
* Use sve attribute on functions.
Differential Revision: https://reviews.llvm.org/D102623
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
By converting the SVE intrinsic to a normal LLVM insertelement we give
the code generator a better chance to remove transitions between GPRs
and VPRs
Co-authored-by: Paul Walker <paul.walker@arm.com>
Depends on D101302
Differential Revision: https://reviews.llvm.org/D101167
As part of this the ptrue coalescing done in SVEIntrinsicOpts has been
modified to not introduce redundant converts, since the convert removal
will no longer run after that optimisation to clean up.
Differential Revision: https://reviews.llvm.org/D101302
When vectorising for AArch64 targets if you specify the SVE attribute
we automatically then treat masked loads and stores as legal. Also,
since we have no cost model for masked memory ops we believe it's
cheap to use the masked load/store intrinsics even for fixed width
vectors. This can lead to poor code quality as the intrinsics will
currently be scalarised in the backend. This patch adds a basic
cost model that marks fixed-width masked memory ops as significantly
more expensive than for scalable vectors.
Tests for the cost model are added here:
Transforms/LoopVectorize/AArch64/masked-op-cost.ll
Differential Revision: https://reviews.llvm.org/D100745
We were missing some instruction costs when converting vectors of
floating point half types into integers, so I've added those here.
I also manually generated assembly code for each FP->int case and
looked at the number of instructions generated, which meant
adjusting some of the existing costs too.
I've updated an existing test to reflect the new costs:
Analysis/CostModel/AArch64/sve-fptoi.ll
Differential Revision: https://reviews.llvm.org/D99935
Introduced the cost of thre reverse shuffles for AArch64, currently just
copied the costs for PermuteSingleSrc.
Differential Revision: https://reviews.llvm.org/D100871
when the predicate used by last{a,b} specifies a known vector length.
For example:
aarch64_sve_lasta(VL1, D) -> extractelement(D, #1)
aarch64_sve_lastb(VL1, D) -> extractelement(D, #0)
Co-authored-by: Paul Walker <paul.walker@arm.com>
Differential Revision: https://reviews.llvm.org/D100476
At the moment, getMemoryOpCost returns 1 for all inputs if CostKind is
CodeSize or SizeAndLatency. This fools LoopUnroll into thinking memory
operations on large vectors have a cost of one, even if they will get
expanded to a large number of memory operations in the backend.
This patch updates getMemoryOpCost to return the cost for the type
legalization for both CodeSize and SizeAndLatency. This should more
accurately reflect the number of memory operations required.
I am not sure how latency should properly be included in SizeAndLatency
from the description, but returning the size cost should be clearly more
accurate.
This does not cause any binary changes when building
MultiSource/SPEC2000/SPEC2006 with -O3 -flto for AArch64, likely because
large vector memops are not really formed by code emitted from Clang.
But using the C/C++ matrix extension can easily result in code with very
large vector operations directly from Clang, e.g.
https://clang.godbolt.org/z/6xzxcTGvb
Reviewed By: samparker
Differential Revision: https://reviews.llvm.org/D100291
Matthew Devereau