Motivating examples are seen in the PhaseOrdering tests based on:
https://bugs.llvm.org/show_bug.cgi?id=43953#c2 - if we have
intrinsics there, some pass can fold them.
The intrinsics are still named "experimental" at this point, but
if there is no fallout from this patch, that will be a good
indicator that it is safe to finalize them.
Differential Revision: https://reviews.llvm.org/D80867
Use getMemoryOpCost from the generic implementation of getUserCost
and have getInstructionThroughput return the result of that for loads
and stores.
This also means that the X86 implementation of getUserCost can be
removed with the functionality folded into its getMemoryOpCost.
Differential Revision: https://reviews.llvm.org/D80984
Combine the two API calls into one by introducing a structure to hold
the relevant data. This has the added benefit of moving the boiler
plate code for arguments and flags, into the constructors. This is
intended to be a non-functional change, but the complicated web of
logic involved here makes it very hard to guarantee.
Differential Revision: https://reviews.llvm.org/D79941
getScalarizationOverhead is only ever called with vectors (and we already had a load of cast<VectorType> calls immediately inside the functions).
Followup to D78357
Reviewed By: @samparker
Differential Revision: https://reviews.llvm.org/D79341
Make the kind of cost explicit throughout the cost model which,
apart from making the cost clear, will allow the generic parts to
calculate better costs. It will also allow some backends to
approximate and correlate the different costs if they wish. Another
benefit is that it will also help simplify the cost model around
immediate and intrinsic costs, where we currently have multiple APIs.
RFC thread:
http://lists.llvm.org/pipermail/llvm-dev/2020-April/141263.html
Differential Revision: https://reviews.llvm.org/D79002
The improvements to the x86 vector insert/extract element costs in D74976 resulted in the estimated costs for vector initialization and scalarization increasing higher than should be expected. This is particularly noticeable on pre-SSE4 targets where the available of legal INSERT_VECTOR_ELT ops is more limited.
This patch does 2 things:
1 - it implements X86TTIImpl::getScalarizationOverhead to more accurately represent the typical costs of a ISD::BUILD_VECTOR pattern.
2 - it adds a DemandedElts mask to getScalarizationOverhead to permit the SLP's BoUpSLP::getGatherCost to be rewritten to use it directly instead of accumulating raw vector insertion costs.
This fixes PR45418 where a v4i8 (zext'd to v4i32) was no longer vectorizing.
A future patch should extend X86TTIImpl::getScalarizationOverhead to tweak the EXTRACT_VECTOR_ELT scalarization costs as well.
Reviewed By: @craig.topper
Differential Revision: https://reviews.llvm.org/D78216
There are several different types of cost that TTI tries to provide
explicit information for: throughput, latency, code size along with
a vague 'intersection of code-size cost and execution cost'.
The vectorizer is a keen user of RecipThroughput and there's at least
'getInstructionThroughput' and 'getArithmeticInstrCost' designed to
help with this cost. The latency cost has a single use and a single
implementation. The intersection cost appears to cover most of the
rest of the API.
getUserCost is explicitly called from within TTI when the user has
been explicit in wanting the code size (also only one use) as well
as a few passes which are concerned with a mixture of size and/or
a relative cost. In many cases these costs are closely related, such
as when multiple instructions are required, but one evident diverging
cost in this function is for div/rem.
This patch adds an argument so that the cost required is explicit,
so that we can make the important distinction when necessary.
Differential Revision: https://reviews.llvm.org/D78635
The API for shuffles and reductions uses generic Type parameters,
instead of VectorType, and so assertions and casts are used a lot.
This patch makes those types explicit, which means that the clients
can't be lazy, but results in less ambiguity, and that can only be a
good thing.
Bugzilla: https://bugs.llvm.org/show_bug.cgi?id=45562
Differential Revision: https://reviews.llvm.org/D78357
This is similar to what I recently did for getArithmeticReductionCost.
I'm trying to account for the narrowing from 512->256->128 as we go.
I've also added a new helper method getMinMaxCost that tries to
handle the cases where we have native min/max instructions and
fall back to cmp+select when we don't.
Differential Revision: https://reviews.llvm.org/D76634
Refines the gather/scatter cost model, but also changes the TTI
function getIntrinsicInstrCost to accept an additional parameter
which is needed for the gather/scatter cost evaluation.
This did require trivial changes in some non-ARM backends to
adopt the new parameter.
Extending gathers and truncating scatters are now priced cheaper.
Differential Revision: https://reviews.llvm.org/D75525
Add an extra parameter so alignment can be taken under
consideration in gather/scatter legalization.
Differential Revision: https://reviews.llvm.org/D71610
Soon Intrinsic::ID will be a plain integer, so this overload will not be
possible.
Rename both overloads to ensure that downstream targets observe this as
a build failure instead of a runtime failure.
Split off from D71320
Reviewers: efriedma
Differential Revision: https://reviews.llvm.org/D71381
This attempts to teach the cost model in Arm that code such as:
%s = shl i32 %a, 3
%a = and i32 %s, %b
Can under Arm or Thumb2 become:
and r0, r1, r2, lsl #3
So the cost of the shift can essentially be free. To do this without
trying to artificially adjust the cost of the "and" instruction, it
needs to get the users of the shl and check if they are a type of
instruction that the shift can be folded into. And so it needs to have
access to the actual instruction in getArithmeticInstrCost, which if
available is added as an extra parameter much like getCastInstrCost.
We otherwise limit it to shifts with a single user, which should
hopefully handle most of the cases. The list of instruction that the
shift can be folded into include ADC, ADD, AND, BIC, CMP, EOR, MVN, ORR,
ORN, RSB, SBC and SUB. This translates to Add, Sub, And, Or, Xor and
ICmp.
Differential Revision: https://reviews.llvm.org/D70966
-mvzeroupper will force the vzeroupper insertion pass to run on
CPUs that normally wouldn't. -mno-vzeroupper disables it on CPUs
where it normally runs.
To support this with the default feature handling in clang, we
need a vzeroupper feature flag in X86.td. Since this flag has
the opposite polarity of the fast-partial-ymm-or-zmm-write we
used to use to disable the pass, we now need to add this new
flag to every CPU except KNL/KNM and BTVER2 to keep identical
behavior.
Remove -fast-partial-ymm-or-zmm-write which is no longer used.
Differential Revision: https://reviews.llvm.org/D69786
PTEST and especially the MOVMSK instructions are slow on Knights Landing
or later. As a bonus, this patch increases instruction parallelism by
emitting:
KORTEST(PCMPNEQ(a, b), PCMPNEQ(c, d)) == 0
Instead of:
KORTEST(AND(PCMPEQ(a, b), PCMPEQ(c, d))) == ~0
https://reviews.llvm.org/D69157
Add an extra parameter so the backend can take the alignment into
consideration.
Differential Revision: https://reviews.llvm.org/D68400
llvm-svn: 374763
In loop-vectorize, interleave count and vector factor depend on target register number. Currently, it does not
estimate different register pressure for different register class separately(especially for scalar type,
float type should not be on the same position with int type), so it's not accurate. Specifically,
it causes too many times interleaving/unrolling, result in too many register spills in loop body and hurting performance.
So we need classify the register classes in IR level, and importantly these are abstract register classes,
and are not the target register class of backend provided in td file. It's used to establish the mapping between
the types of IR values and the number of simultaneous live ranges to which we'd like to limit for some set of those types.
For example, POWER target, register num is special when VSX is enabled. When VSX is enabled, the number of int scalar register is 32(GPR),
float is 64(VSR), but for int and float vector register both are 64(VSR). So there should be 2 kinds of register class when vsx is enabled,
and 3 kinds of register class when VSX is NOT enabled.
It runs on POWER target, it makes big(+~30%) performance improvement in one specific bmk(503.bwaves_r) of spec2017 and no other obvious degressions.
Differential revision: https://reviews.llvm.org/D67148
llvm-svn: 374634
Also Revert "[LoopVectorize] Fix non-debug builds after rL374017"
This reverts commit 9f41deccc0.
This reverts commit 18b6fe07bc.
The patch is breaking PowerPC internal build, checked with author, reverting
on behalf of him for now due to timezone.
llvm-svn: 374091
In loop-vectorize, interleave count and vector factor depend on target register number. Currently, it does not
estimate different register pressure for different register class separately(especially for scalar type,
float type should not be on the same position with int type), so it's not accurate. Specifically,
it causes too many times interleaving/unrolling, result in too many register spills in loop body and hurting performance.
So we need classify the register classes in IR level, and importantly these are abstract register classes,
and are not the target register class of backend provided in td file. It's used to establish the mapping between
the types of IR values and the number of simultaneous live ranges to which we'd like to limit for some set of those types.
For example, POWER target, register num is special when VSX is enabled. When VSX is enabled, the number of int scalar register is 32(GPR),
float is 64(VSR), but for int and float vector register both are 64(VSR). So there should be 2 kinds of register class when vsx is enabled,
and 3 kinds of register class when VSX is NOT enabled.
It runs on POWER target, it makes big(+~30%) performance improvement in one specific bmk(503.bwaves_r) of spec2017 and no other obvious degressions.
Differential revision: https://reviews.llvm.org/D67148
llvm-svn: 374017
Summary:
Similar to the previous prefer-256-bit flag. We might want to
enable this by default some CPUs. This just starts the initial
work to implement and prove that it effects TTI's vector width.
Reviewers: RKSimon, echristo, spatel, atdt
Reviewed By: RKSimon
Subscribers: lebedev.ri, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67311
llvm-svn: 371319
When considering a loop containing nontemporal stores or loads for
vectorization, suppress the vectorization if the corresponding
vectorized store or load with the aligment of the original scaler
memory op is not supported with the nontemporal hint on the target.
This adds two new functions:
bool isLegalNTStore(Type *DataType, unsigned Alignment) const;
bool isLegalNTLoad(Type *DataType, unsigned Alignment) const;
to TTI, leaving the target independent default implementation as
returning true, but with overriding implementations for X86 that
check the legality based on available Subtarget features.
This fixes https://llvm.org/PR40759
Differential Revision: https://reviews.llvm.org/D61764
llvm-svn: 363581
Summary:
This adds a BranchFusion feature to replace the usage of the MacroFusion
for AMD CPUs.
See D59688 for context.
Reviewers: andreadb, lebedev.ri
Subscribers: hiraditya, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59872
llvm-svn: 357171
This adds support for scalarizing these intrinsics as well the X86TargetTransformInfo support to avoid scalarizing them in the cases X86 can handle.
I've omitted handling special cases for constant masks for this first pass. Though CodeGenPrepare can constant fold the branch conditions and remove some of the control flow anyway.
Fixes PR40994 and is covers most of PR3666. Might want to implement constant masks to close that.
Differential Revision: https://reviews.llvm.org/D59180
llvm-svn: 356687
As this has broken the lto bootstrap build for 3 days and is
showing a significant regression on the Dither_benchmark results (from
the LLVM benchmark suite) -- specifically, on the
BENCHMARK_FLOYD_DITHER_128, BENCHMARK_FLOYD_DITHER_256, and
BENCHMARK_FLOYD_DITHER_512; the others are unchanged. These have
regressed by about 28% on Skylake, 34% on Haswell, and over 40% on
Sandybridge.
This reverts commit r353923.
llvm-svn: 354434
The use of the -mprefer-vector-width=256 command line option mixed with functions
using vector intrinsics can create situations where one function thinks 512 vectors
are legal, but another fucntion does not.
If a 512 bit vector is passed between them via a pointer, its possible ArgumentPromotion
might try to pass by value instead. This will result in type legalization for the two
functions handling the 512 bit vector differently leading to runtime failures.
Had the 512 bit vector been passed by value from clang codegen, both functions would
have been tagged with a min-legal-vector-width=512 function attribute. That would
make them be legalized the same way.
I observed this issue in 32-bit mode where a union containing a 512 bit vector was
being passed by a function that used intrinsics to one that did not. The caller
ended up passing in zmm0 and the callee tried to read it from ymm0 and ymm1.
The fix implemented here is just to consider it a mismatch if two functions
would handle 512 bit differently without looking at the types that are being
considered. This is the easist and safest fix, but it can be improved in the future.
Differential Revision: https://reviews.llvm.org/D58390
llvm-svn: 354376
Tuning flags don't have any effect on the available instructions so aren't a good reason to prevent inlining.
There are also some ISA flags that don't have any intrinsics our ABI requirements that we can exclude. I've put only the most basic ones like cmpxchg16b and lahfsahf. These are interesting because they aren't present in all 64-bit CPUs, but we have codegen workarounds when they aren't present.
Loosening these checks can help with scenarios where a caller has a more specific CPU than a callee. The default tuning flags on our generic 'x86-64' CPU can currently make it inline compatible with other CPUs. I've also added an example test for 'nocona' and 'prescott' where 'nocona' is just a 64-bit capable version of 'prescott' but in 32-bit mode they should be completely compatible.
I've based the implementation here of the similar code in AMDGPU.
Differential Revision: https://reviews.llvm.org/D58371
llvm-svn: 354355
Try to use 64-bit SLP vectorization. In addition to horizontal instrs
this change triggers optimizations for partial vector operations (for instance,
using low halfs of 128-bit registers xmm0 and xmm1 to multiply <2 x float> by
<2 x float>).
Fixes llvm.org/PR32433
llvm-svn: 353923
to reflect the new license.
We understand that people may be surprised that we're moving the header
entirely to discuss the new license. We checked this carefully with the
Foundation's lawyer and we believe this is the correct approach.
Essentially, all code in the project is now made available by the LLVM
project under our new license, so you will see that the license headers
include that license only. Some of our contributors have contributed
code under our old license, and accordingly, we have retained a copy of
our old license notice in the top-level files in each project and
repository.
llvm-svn: 351636
optsize using masked wide loads
Under Opt for Size, the vectorizer does not vectorize interleave-groups that
have gaps at the end of the group (such as a loop that reads only the even
elements: a[2*i]) because that implies that we'll require a scalar epilogue
(which is not allowed under Opt for Size). This patch extends the support for
masked-interleave-groups (introduced by D53011 for conditional accesses) to
also cover the case of gaps in a group of loads; Targets that enable the
masked-interleave-group feature don't have to invalidate interleave-groups of
loads with gaps; they could now use masked wide-loads and shuffles (if that's
what the cost model selects).
Reviewers: Ayal, hsaito, dcaballe, fhahn
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D53668
llvm-svn: 345705
interleave-group
The vectorizer currently does not attempt to create interleave-groups that
contain predicated loads/stores; predicated strided accesses can currently be
vectorized only using masked gather/scatter or scalarization. This patch makes
predicated loads/stores candidates for forming interleave-groups during the
Loop-Vectorizer's analysis, and adds the proper support for masked-interleave-
groups to the Loop-Vectorizer's planning and transformation stages. The patch
also extends the TTI API to allow querying the cost of masked interleave groups
(which each target can control); Targets that support masked vector loads/
stores may choose to enable this feature and allow vectorizing predicated
strided loads/stores using masked wide loads/stores and shuffles.
Reviewers: Ayal, hsaito, dcaballe, fhahn, javed.absar
Reviewed By: Ayal
Differential Revision: https://reviews.llvm.org/D53011
llvm-svn: 344472
In the motivating case from PR35681 and represented by the macro-fuse-cmp test:
https://bugs.llvm.org/show_bug.cgi?id=35681
...there's a 37 -> 31 byte size win for the loop because we eliminate the big base
address offsets.
SPEC2017 on Ryzen shows no significant perf difference.
Differential Revision: https://reviews.llvm.org/D42607
llvm-svn: 324289
All these headers already depend on CodeGen headers so moving them into
CodeGen fixes the layering (since CodeGen depends on Target, not the
other way around).
llvm-svn: 318490
- Targets that want to support memcmp expansions now return the list of
supported load sizes.
- Expansion codegen does not assume that all power-of-two load sizes
smaller than the max load size are valid. For examples, this is not the
case for x86(32bit)+sse2.
Fixes PR34887.
llvm-svn: 316905
Summary:
Right now there are two functions with the same name, one does the work
and the other one returns true if expansion is needed. Rename
TargetTransformInfo::expandMemCmp to make it more consistent with other
members of TargetTransformInfo.
Remove the unused Instruction* parameter.
Differential Revision: https://reviews.llvm.org/D38165
llvm-svn: 314096
This is intended to be a superset of the functionality from D31037 (EarlyCSE) but implemented
as an independent pass, so there's no stretching of scope and feature creep for an existing pass.
I also proposed a weaker version of this for SimplifyCFG in D30910. And I initially had almost
this same functionality as an addition to CGP in the motivating example of PR31028:
https://bugs.llvm.org/show_bug.cgi?id=31028
The advantage of positioning this ahead of SimplifyCFG in the pass pipeline is that it can allow
more flattening. But it needs to be after passes (InstCombine) that could sink a div/rem and
undo the hoisting that is done here.
Decomposing remainder may allow removing some code from the backend (PPC and possibly others).
Differential Revision: https://reviews.llvm.org/D37121
llvm-svn: 312862
SLP vectorizer supports horizontal reductions for Add/FAdd binary
operations. Patch adds support for horizontal min/max reductions.
Function getReductionCost() is split to getArithmeticReductionCost() for
binary operation reductions and getMinMaxReductionCost() for min/max
reductions.
Patch fixes PR26956.
Differential revision: https://reviews.llvm.org/D27846
llvm-svn: 312791
Summary:
We add the precise cache sizes and associativity for the following Intel
architectures:
- Penry
- Nehalem
- Westmere
- Sandy Bridge
- Ivy Bridge
- Haswell
- Broadwell
- Skylake
- Kabylake
Polly uses since several months a performance model for BLAS computations that
derives optimal cache and register tile sizes from cache and latency
information (based on ideas from "Analytical Modeling Is Enough for High-Performance BLIS", by Tze Meng Low published at TOMS 2016).
While bootstrapping this model, these target values have been kept in Polly.
However, as our implementation is now rather mature, it seems time to teach
LLVM itself about cache sizes.
Interestingly, L1 and L2 cache sizes are pretty constant across
micro-architectures, hence a set of architecture specific default values
seems like a good start. They can be expanded to more target specific values,
in case certain newer architectures require different values. For now a set
of Intel architectures are provided.
Just as a little teaser, for a simple gemm kernel this model allows us to
improve performance from 1.2s to 0.27s. For gemm kernels with less optimal
memory layouts even larger speedups can be reported.
Reviewers: Meinersbur, bollu, singam-sanjay, hfinkel, gareevroman, fhahn, sebpop, efriedma, asb
Reviewed By: fhahn, asb
Subscribers: lsaba, asb, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D37051
llvm-svn: 311647
Store operation takes 2 UOps on X86 processors. The exact cost calculation affects several optimization passes including loop unroling.
This change compensates performance degradation caused by https://reviews.llvm.org/D34458 and shows improvements on some benchmarks.
Differential Revision: https://reviews.llvm.org/D35888
llvm-svn: 311285
Sanjay Patel