The min/max intrinsics are not yet canonical, but when they are the tail
predications analysis will change from treating them like icmp to
treating them like intrinsics. Unfortunately, they can currently produce
better code by not being tail predicated thanks to the vectorizer picking
higher VF's and the backend folding to better instructions (especially
for saturate patterns). In the long run we will need to improve the
vectorizers cost modelling, recognizing the instruction directly, but in
the meantime this treats min/max as before to prevent performance
regressions.
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
This makes a couple of changes to the costing of MLA reduction patterns,
to more accurately cost various patterns that can come up from
vectorization.
- The Arm implementation of getExtendedAddReductionCost is altered to
only provide costs for legal or smaller types. Larger than legal types
need to be split, which currently does not work very well, especially
for predicated reductions where the predicate may be legal but needs to
be split. Currently we limit it to legal or smaller input types.
- The getReductionPatternCost has learnt that reduce(ext(mul(ext, ext))
is a pattern that can come up, and can be treated the same as
reduce(mul(ext, ext)) providing the extension types match.
- And it has been adjusted to not count the ext in reduce(mul(ext, ext))
as part of a reduce(mul) pattern.
Together these changes help to more accurately cost the mla reductions
in cases such as where the extend types don't match or the extend
opcodes are different, picking better vector factors that don't result
in expanded reductions.
Differential Revision: https://reviews.llvm.org/D106166
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 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 prevents constant gep operands from being hoisted by the Constant
Hoisting pass, leaving them to CodegenPrepare which can usually do a
better job at splitting large offsets. This can, in general, improve
performance and decrease codesize, especially for v6m where many
constants have a high cost.
Differential Revision: https://reviews.llvm.org/D104877
v6m cores only have a limited number of registers available. Unrolling
can mean we spend more on stack spills and reloads than we save from the
unrolling. This patch adds an extra heuristic to put a limit on the
unroll count for loops with multiple live out values, as measured from
the LCSSA phi nodes.
Differential Revision: https://reviews.llvm.org/D104659
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
Added cost estimation for switch instruction, updated costs of branches, fixed
phi cost.
Had to increase `-amdgpu-unroll-threshold-if` default value since conditional
branch cost (size) was corrected to higher value.
Test renamed to "control-flow.ll".
Removed redundant code in `X86TTIImpl::getCFInstrCost()` and
`PPCTTIImpl::getCFInstrCost()`.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D96805
This removes the restriction that only Thumb2 targets enable runtime
loop unrolling, allowing it for Thumb1 only cores as well. The existing
T2 heuristics are used (for the time being) to control when and how
unrolling is performed.
Differential Revision: https://reviews.llvm.org/D99588
This UpperBound unrolling was already enabled so long as a series of
conditions in ARMTTIImpl::getUnrollingPreferences pass. This just always
enables it as it can help fully unroll loops that would not otherwise
pass those tests.
Differential Revision: https://reviews.llvm.org/D99174
The scalarization overhead was set deliberately high for MVE, whilst the
codegen was new. It helps protect us against the negative ramifications
of mixing scalar and vector instructions. This decreases that,
especially for floating point where the cost of extracting/inserting
lane elements can be low. For integer the cost is still fairly high due
to the cross-register-bank copy, but is no longer n^2 in the length of
the vector.
In general, this will decrease the cost of scalarizing floats and long
integer vectors. i64 increase in cost, having a high cost before and
after this patch. For floats this allows up to start doing things like
vectorizing fdiv instructions, even if they are scalarized.
Differential Revision: https://reviews.llvm.org/D98245
This uses the shuffle mask cost from D98206 to give a better cost of MVE
VREV instructions. This helps especially in VectorCombine where the cost
of shuffles is used to reorder bitcasts, which this helps keep the phase
ordering test for fp16 reductions producing optimal code. The isVREVMask
has been moved to a header file to allow it to be used across target
transform and isel lowering.
Differential Revision: https://reviews.llvm.org/D98210
This adds an Mask ArrayRef to getShuffleCost, so that if an exact mask
can be provided a more accurate cost can be provided by the backend.
For example VREV costs could be returned by the ARM backend. This should
be an NFC until then, laying the groundwork for that to be added.
Differential Revision: https://reviews.llvm.org/D98206
Recently we improved the lowering of low overhead loops and tail
predicated loops, but concentrated first on the DLS do style loops. This
extends those improvements over to the WLS while loops, improving the
chance of lowering them successfully. To do this the lowering has to
change a little as the instructions are terminators that produce a value
- something that needs to be treated carefully.
Lowering starts at the Hardware Loop pass, inserting a new
llvm.test.start.loop.iterations that produces both an i1 to control the
loop entry and an i32 similar to the llvm.start.loop.iterations
intrinsic added for do loops. This feeds into the loop phi, properly
gluing the values together:
%wls = call { i32, i1 } @llvm.test.start.loop.iterations.i32(i32 %div)
%wls0 = extractvalue { i32, i1 } %wls, 0
%wls1 = extractvalue { i32, i1 } %wls, 1
br i1 %wls1, label %loop.ph, label %loop.exit
...
loop:
%lsr.iv = phi i32 [ %wls0, %loop.ph ], [ %iv.next, %loop ]
..
%iv.next = call i32 @llvm.loop.decrement.reg.i32(i32 %lsr.iv, i32 1)
%cmp = icmp ne i32 %iv.next, 0
br i1 %cmp, label %loop, label %loop.exit
The llvm.test.start.loop.iterations need to be lowered through ISel
lowering as a pair of WLS and WLSSETUP nodes, which each get converted
to t2WhileLoopSetup and t2WhileLoopStart Pseudos. This helps prevent
t2WhileLoopStart from being a terminator that produces a value,
something difficult to control at that stage in the pipeline. Instead
the t2WhileLoopSetup produces the value of LR (essentially acting as a
lr = subs rn, 0), t2WhileLoopStart consumes that lr value (the Bcc).
These are then converted into a single t2WhileLoopStartLR at the same
point as t2DoLoopStartTP and t2LoopEndDec. Otherwise we revert the loop
to prevent them from progressing further in the pipeline. The
t2WhileLoopStartLR is a single instruction that takes a GPR and produces
LR, similar to the WLS instruction.
%1:gprlr = t2WhileLoopStartLR %0:rgpr, %bb.3
t2B %bb.1
...
bb.2.loop:
%2:gprlr = PHI %1:gprlr, %bb.1, %3:gprlr, %bb.2
...
%3:gprlr = t2LoopEndDec %2:gprlr, %bb.2
t2B %bb.3
The t2WhileLoopStartLR can then be treated similar to the other low
overhead loop pseudos, eventually being lowered to a WLS providing the
branches are within range.
Differential Revision: https://reviews.llvm.org/D97729
As a followup to D95291, getOperandsScalarizationOverhead was still
using a VF as a vector factor if the arguments were scalar, and would
assert on certain matrix intrinsics with differently sized vector
arguments. This patch removes the VF arg, instead passing the Types
through directly. This should allow it to more accurately compute the
cost without having to guess at which operands will be vectorized,
something difficult with more complex intrinsics.
This adjusts one SVE test as it is now calling the wrong intrinsic vs
veccall. Without invalid InstructCosts the cost of the scalarized
intrinsic is too low. This should get fixed when the cost of
scalarization is accounted for with scalable types.
Differential Revision: https://reviews.llvm.org/D96287
getIntrinsicInstrCost takes a IntrinsicCostAttributes holding various
parameters of the intrinsic being costed. It can either be called with a
scalar intrinsic (RetTy==Scalar, VF==1), with a vector instruction
(RetTy==Vector, VF==1) or from the vectorizer with a scalar type and
vector width (RetTy==Scalar, VF>1). A RetTy==Vector, VF>1 is considered
an error. Both of the vector modes are expected to be treated the same,
but because this is confusing many backends end up getting it wrong.
Instead of trying work with those two values separately this removes the
VF parameter, widening the RetTy/ArgTys by VF used called from the
vectorizer. This keeps things simpler, but does require some other
modifications to keep things consistent.
Most backends look like this will be an improvement (or were not using
getIntrinsicInstrCost). AMDGPU needed the most changes to keep the code
from c230965ccf working. ARM removed the fix in
dfac521da1, webassembly happens to get a fixup for an SLP cost
issue and both X86 and AArch64 seem to now be using better costs from
the vectorizer.
Differential Revision: https://reviews.llvm.org/D95291
It appears that pointer types were causing issues for the min/max cost
code in getIntrinsicInstrCost. This makes sure that when matching
icmp/select to a min/max, we only do that for normal int or float types.
A v8i32 compare will produce a v8i1 predicate, but during codegen the
v8i32 will be split into two v4i32, potentially requiring two v4i1
predicates to be merged into a single v8i1. Because this merging of two
v4i1's into a v8i1 is very expensive, we need to make the cost of the
compare equally high.
This patch adds the cost of that to ARMTTIImpl::getCmpSelInstrCost.
Because we don't know whether the user of the predicate can be split,
and the cost model is mostly pre-instruction, we may be pessimistic but
that should only be for larger and legal types. This also adds min/max
detection to the costmodel where it can be detected, to keep those in
line with the cost of simple min/max instructions. Otherwise for the
most part, costs that were already expensive have become more expensive.
Differential Revision: https://reviews.llvm.org/D96692
This adds basic MVE costs for SMIN/SMAX/UMIN/UMAX, as well as MINNUM and
MAXNUM representing fmin and fmax. It tightens up the costs, not using a
ICmp+Select cost.
Differential Revision: https://reviews.llvm.org/D96603
This refactors shouldFavorPostInc() and shouldFavorBackedgeIndex() into
getPreferredAddressingMode() so that we have one interface to steer LSR in
generating the preferred addressing mode.
Differential Revision: https://reviews.llvm.org/D96600
The vector reduction intrinsics started life as experimental ops, so backend support
was lacking. As part of promoting them to 1st-class intrinsics, however, codegen
support was added/improved:
D58015
D90247
So I think it is safe to now remove this complication from IR.
Note that we still have an IR-level codegen expansion pass for these as discussed
in D95690. Removing that is another step in simplifying the logic. Also note that
x86 was already unconditionally forming reductions in IR, so there should be no
difference for x86.
I spot checked a couple of the tests here by running them through opt+llc and did
not see any asm diffs.
If we do find functional differences for other targets, it should be possible
to (at least temporarily) restore the shuffle IR with the ExpandReductions IR
pass.
Differential Revision: https://reviews.llvm.org/D96552
This adds the CostKind to getMVEVectorCostFactor, so that it can
automatically account for CodeSize costs, where it returns a cost of 1
not the MVEFactor used for Throughput/Latency. This helps simplify the
caller code and allows us to get the codesize cost more correct in more
cases.
David Green