This enabled the select optimize patch for ARM Out of order AArch64
cores. It is trying to solve a problem that is difficult for the
compiler to fix. The criteria for when a csel is better or worse than a
branch depends heavily on whether the branch is well predicted and the
amount of ILP in the loop (as well as other criteria like the core in
question and the relative performance of the branch predictor). The
pass seems to do a decent job though, with the inner loop heuristics
being well implemented and doing a better job than I had expected in
general, even without PGO information.
I've been doing quite a bit of benchmarking. The headline numbers are
these for SPEC2017 on a Neoverse N1:
500.perlbench_r -0.12%
502.gcc_r 0.02%
505.mcf_r 6.02%
520.omnetpp_r 0.32%
523.xalancbmk_r 0.20%
525.x264_r 0.02%
531.deepsjeng_r 0.00%
541.leela_r -0.09%
548.exchange2_r 0.00%
557.xz_r -0.20%
Running benchmarks with a combination of the llvm-test-suite plus
several versions of SPEC gave between a 0.2% and 0.4% geomean
improvement depending on the core/run. The instruction count went down
by 0.1% too, which is a good sign, but the results can be a little
noisy. Some issues from other benchmarks I had ran were improved in
rGca78b5601466f8515f5f958ef8e63d787d9d812e. In summary well predicted
branches will see in improvement, badly predicted branches may get
worse, and on average performance seems to be a little better overall.
This patch enables the pass for AArch64 under -O3 for cores that will
benefit for it. i.e. not in-order cores that do not fit into the "Assume
infinite resources that allow to fully exploit the available
instruction-level parallelism" cost model. It uses a subtarget feature
for specifying when the pass will be enabled, which I have enabled under
cpu=generic as the performance increases for out of order cores seems
larger than any decreases for inorder, which were minor.
Differential Revision: https://reviews.llvm.org/D138990
A target can return if a misaligned access is 'fast' as defined
by the target or not. In reality there can be different levels
of 'fast' and 'slow'. This patch changes the boolean 'Fast'
argument of the allowsMisalignedMemoryAccesses family of functions
to an unsigned representing its speed.
A target can still define it as it wants and the direct translation
of the current code uses 0 and 1 for current false and true. This
makes the change an NFC.
Subsequent patch will start using an actual value of speed in
the load/store vectorizer to compare if a vectorized access going
to be not just fast, but not slower than before.
Differential Revision: https://reviews.llvm.org/D124217
Epilogue loop vectorization is a feature in the vectorize intended to avoid running fully scalar code when the vector length of the main loop turns out to be either longer than the trip count of the actual loop, or with a huge remainder.
In practice, this feature appears to not have been well tuned. I honestly don't think it should be on by default at all, but it definitely shouldn't be on for RISCV. Note that other targets have also disabled it, but they've done so via disabling interleaving - which is, well, completely unrelated - and we don't want to do that for RISCV.
In the near term, many examples I'm seeing have terrible codegen for epilogue vectorization. We are greatly increasing code size for little value at reasonable VLEN values for small types. In the long term, the cases that epilogue vectorization are intended to handle are likely better handled via tail folding on RISCV.
As an aside, I also don't really trust the correctness of epilogue vectorization. The code structure is such that otherwise straight forward changes sometimes break only epilogue vectorization. The reuse of an existing vplan without careful validation opens significant room for nasty bugs. Given how rarely the code is exercised, that is not a good combination.
As such, this patch introduces a TTI hook, and completely disables epilogue vectorization on RISCV.
Differential Revision: https://reviews.llvm.org/D136695
If the single-thread model is used, or the
-licm-force-thread-model-single flag is specified, skip checks
related to thread-safety. This means that store promotion for
conditionally executed stores only requires proof of
dereferenceability and writability, but not of thread-safety. For
example, this enables promotion of stores to (non-constant) globals,
as well as captured allocas.
Fixes https://github.com/llvm/llvm-project/issues/50537.
Differential Revision: https://reviews.llvm.org/D130466
The mul by constant costmodels handle power-of-2 constants, but not negated-power-of-2, despite the backends handling both.
This patch adds the OperandValueProperties::OP_NegatedPowerOf2 enum and wires it for use for basic mul cost analysis and SLP handling.
Fixes#50778
Differential Revision: https://reviews.llvm.org/D111968
Also remove new-pass-manager version of ExpandLargeDivRem because there is no way
yet to access TargetLowering in the new pass manager.
Differential Revision: https://reviews.llvm.org/D133691
This adds the ExpandLargeDivRem to the default pass pipeline.
The limit at which it expands div/rem instructions is configured
via a new TargetTransformInfo hook (default: no expansion)
X86, Arm and AArch64 backends implement this hook to expand div/rem
instructions with more than 128 bits.
Differential Revision: https://reviews.llvm.org/D130076
CGP uses a raw `getInstructionCost(I, TargetTransformInfo::TCK_SizeAndLatency) >= TCC_Expensive` check to see if its better to move an expensive instruction used in a select behind a branch instead.
This is causing issues with upcoming improvements to TCK_SizeAndLatency costs on X86 as we need to use TCK_SizeAndLatency as an uop count (so its compatible with various target-specific buffer sizes - see D132288), but we can have instructions that have a low TCK_SizeAndLatency value but should still be treated as 'expensive' (FDIV for example) - by adding a isExpensiveToSpeculativelyExecute wrapper we can keep the current behaviour but still add an x86 override in a future patch when the cost tables are updated to compensate.
This has the effect of exposing the power-of-two property for use in memory op costing, but no target actually uses it yet. The main point of this change is simple consistency with the recently changes getArithmeticInstrCost, and to remove the last (interface) use of OperandValueKind.
This change completes the process of replacing OperandValueKind and OperandValueProperties which were previously passed independently in this API with a single container class which contains both.
This is the change which motivated the whole sequence which preceeded it. In an original spike version of this change, I'd noticed a nasty bug: I'd changed the signature without changing names, and as result, we silently passed additional information through a callsite which previously dropped the power-of-two fact. This might be harmless in most cases, but at least a couple clearly dependend for correctness on not passing that property through.
I did my best to split off prior changes which reduced the scope of this one, and which made it possible to use compiler assistance. For instance, every parameter which changes type in this change also changes name. This was intentional to make sure that every call site possible effected must show up in the diff. This let me audit each one closely.
This removes the last use of OperandValueKind from the client side API, and (once this is fully plumbed through TTI implementation) allow use of the same properties in store costing as arithmetic costing.
This completes the client side transition to the OperandValueInfo version of this routine. Backend TTI implementations still use the prior versions for now.
OperandValueKind and OperandValueProperties both provide facts about the operands of an instruction for purposes of cost modeling. We've discussed merging them several times; before I plumb through more flags, let's go ahead and do so.
This change only adds the client side interface for getArithmeticInstrCost and makes a couple of minor changes in client code to prove that it works. Target TTI implementations still use the split flags. I'm deliberately splitting what could be one big change into a series of smaller ones so that I can lean on the compiler to catch errors along the way.
Defaults to TCK_RecipThroughput - as most explicit calls were assuming TCK_RecipThroughput (vectorizers) or was just doing a before-vs-after comparison (vectorcombiner). Calls via getInstructionCost were just dropping the CostKind, so again there should be no change at this time (as getShuffleCost and its expansions don't use CostKind yet) - but it will make it easier for us to better account for size/latency shuffle costs in inline/unroll passes in the future.
Differential Revision: https://reviews.llvm.org/D132287
* Replace getUserCost with getInstructionCost, covering all cost kinds.
* Remove getInstructionLatency, it's not implemented by any backends, and we should fold the functionality into getUserCost (now getInstructionCost) to make it easier for targets to handle the cost kinds with their existing cost callbacks.
Original Patch by @samparker (Sam Parker)
Differential Revision: https://reviews.llvm.org/D79483
After D121595 was commited, I noticed regressions assosicated with small trip
count numbersvectorisation by tail folding with scalable vectors. As a solution
for those issues I propose to introduce the minimal trip count threshold value.
Differential Revision: https://reviews.llvm.org/D130755
A const reference is preferred over a non-null const pointer.
`Type *` is kept as is to match the other overload.
Reviewed By: davidxl
Differential Revision: https://reviews.llvm.org/D131197
1) Overloaded (instruction-based) method is a wrapper around the current (opcode-based) method.
2) This patch also changes a few callsites (VectorCombine.cpp,
SLPVectorizer.cpp, CodeGenPrepare.cpp) to call the overloaded method.
3) This is a split of D128302.
Differential Revision: https://reviews.llvm.org/D131114
If we have interleave groups in the loop we want to vectorise then
we should fall back on normal vectorisation with a scalar epilogue. In
such cases when tail-folding is enabled we'll almost certainly go on to
create vplans with very high costs for all vector VFs and fall back on
VF=1 anyway. This is likely to be worse than if we'd just used an
unpredicated vector loop in the first place.
Once the vectoriser has proper support for analysing all the costs
for each combination of VF and vectorisation style, then we should
be able to remove this.
Added an extra test here:
Transforms/LoopVectorize/AArch64/sve-tail-folding-option.ll
Differential Revision: https://reviews.llvm.org/D128342
Now the API getExtendedAddReductionCost is used to determine the cost of extended Add reduction with optional Mul. For Arm, it could cover the cases. But for other target, for example: RISCV, they support other kinds of extended recution, such as FAdd.
This patch does the following changes:
1, Split getExtendedAddReductionCost into 2 new API: getExtendedReductionCost which handles the extended reduction with addtional input of Opcode; getMulAccReductionCost which handle the MLA cases the getExtendedAddReductionCost.
2, Refactor getReductionPatternCost, add some contraint condition to make sure the getMulAccReductionCost should only handle the reuction of Add + Mul.
Differential Revision: https://reviews.llvm.org/D130868
I am playing with the LoopDataPrefetch pass and found out that it
bails to work with a pointer in a non-zero address space. This
patch adds the target callback to check if an address space is to
be considered for prefetching. Default implementation still only
allows address space 0, so this is NFCI.
This does not currently affect any known targets, but seems to be
generally useful for the future.
Differential Revision: https://reviews.llvm.org/D129795
This patch adds a command line flag to be able to test
the type based cost-model analysis for Intrinsics.
Differential Revision: https://reviews.llvm.org/D129109
This patch adds the AArch64 hook for preferPredicateOverEpilogue,
which currently returns true if SVE is enabled and one of the
following conditions (non-exhaustive) is met:
1. The "sve-tail-folding" option is set to "all", or
2. The "sve-tail-folding" option is set to "all+noreductions"
and the loop does not contain reductions,
3. The "sve-tail-folding" option is set to "all+norecurrences"
and the loop has no first-order recurrences.
Currently the default option is "disabled", but this will be
changed in a later patch.
I've added new tests to show the options behave as expected here:
Transforms/LoopVectorize/AArch64/sve-tail-folding-option.ll
Differential Revision: https://reviews.llvm.org/D129560
Currently, for vectorised loops that use the get.active.lane.mask
intrinsic we only use the mask for predicated vector operations,
such as masked loads and stores, etc. The loop itself is still
controlled by comparing the canonical induction variable with the
trip count. However, for some targets this is inefficient when it's
cheap to use the mask itself to control the loop.
This patch adds support for using the active lane mask for control
flow by:
1. Generating the active lane mask for the next iteration of the
vector loop, rather than the current one. If there are still any
remaining iterations then at least the first bit of the mask will
be set.
2. Extract the first bit of this mask and use this bit for the
conditional branch.
I did this by creating a new VPActiveLaneMaskPHIRecipe that sets
up the initial PHI values in the vector loop pre-header. I've also
made use of the new BranchOnCond VPInstruction for the final
instruction in the loop region.
Differential Revision: https://reviews.llvm.org/D125301
enabled
The C++20 Coroutines couldn't be compiled to WebAssembly due to an
optimization named symmetric transfer requires the support for musttail
calls but WebAssembly doesn't support it yet.
This patch tries to fix the problem by adding a supportsTailCalls
method to TargetTransformImpl to skip the symmetric transfer when
tail-call feature is not supported.
Reviewed By: tlively
Differential Revision: https://reviews.llvm.org/D128794
During the reordering transformation we should try to avoid reordering bundles
like fadd,fsub because this may block them being matched into a single vector
instruction in x86.
We do this by checking if a TreeEntry is such a pattern and adding it to the
list of TreeEntries with orders that need to be considered.
Differential Revision: https://reviews.llvm.org/D125712
In some passes we need a valid number of cache line size to do analysis or
transformation, e.g., loop cache analysis and loop date prefetch. However,
for some backend targets, `TTIImpl->getCacheLineSize()` is not implemented
and hence 'TTI.getCacheLineSize()' would just return 0 which eventually might
produce invalid result.
In this patch we add a user-specified opt/llc option for cache line size.
If the option is specified by users we use the value supplied, otherwise we
fall-back to the default value obtained from `TTIImpl->->getCacheLineSize()`.
The powerpc target already has such an option, this patch generalizes
this option to TargetTransformInfo.cpp.
Reviewed By: bmahjour, #loopoptwg
Differential Revision: https://reviews.llvm.org/D127342
Previously, `getRegUsageForType` was implemented using
`getTypeLegalizationCost`. `getRegUsageForType` is used by the loop
vectorizer to estimate the register pressure caused by using a vector
type. However, `getTypeLegalizationCost` currently only appears to
understand splitting and not scalarization, so significantly
underestimates the register requirements.
Instead, use `getNumRegisters`, which understands when scalarization
can occur (via computeRegisterProperties).
This was discovered while investigating D118979 (Set maximum VF with
shouldMaximizeVectorBandwidth), where under fixed-length 512-bit SVE the
loop vectorizer previously ends up costing an v128i1 as 2 v64i*
registers where it actually occupies 128 i32 registers.
I'm sending this patch early for comment, I'm still doing some sanity checking
with LNT. I note that getRegisterClassForType appears to return VectorRC even
though the type in question (large vNi1 types) end up occupying scalar
registers. That might be worth fixing too.
Differential Revision: https://reviews.llvm.org/D125918
We can try to vectorize number of stores less than MinVecRegSize
/ scalar_value_size, if it is allowed by target. Gives an extra
opportunity for the vectorization.
Fixes PR54985.
Differential Revision: https://reviews.llvm.org/D124284
Before this patch `Args` was used to pass a broadcat's arguments by SLP.
This patch changes this. `Args` is now used for passing the operands of
the shuffle.
Differential Revision: https://reviews.llvm.org/D124202
Currently, the utility supports lowering of non atomic memory transfer routines only. This patch adds support for atomic version of memcopy. This may be useful for targets not supporting atomic memcopy.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D118443
David Green