This is the first patch adding an initial set of matrix intrinsics and a
corresponding lowering pass. This has been discussed on llvm-dev:
http://lists.llvm.org/pipermail/llvm-dev/2019-October/136240.html
The first patch introduces four new intrinsics (transpose, multiply,
columnwise load and store) and a LowerMatrixIntrinsics pass, that
lowers those intrinsics to vector operations.
Matrixes are embedded in a 'flat' vector (e.g. a 4 x 4 float matrix
embedded in a <16 x float> vector) and the intrinsics take the dimension
information as parameters. Those parameters need to be ConstantInt.
For the memory layout, we initially assume column-major, but in the RFC
we also described how to extend the intrinsics to support row-major as
well.
For the initial lowering, we split the input of the intrinsics into a
set of column vectors, transform those column vectors and concatenate
the result columns to a flat result vector.
This allows us to lower the intrinsics without any shape propagation, as
mentioned in the RFC. In follow-up patches, we plan to submit the
following improvements:
* Shape propagation to eliminate the embedding/splitting for each
intrinsic.
* Fused & tiled lowering of multiply and other operations.
* Optimization remarks highlighting matrix expressions and costs.
* Generate loops for operations on large matrixes.
* More general block processing for operation on large vectors,
exploiting shape information.
We would like to add dedicated transpose, columnwise load and store
intrinsics, even though they are not strictly necessary. For example, we
could instead emit a large shufflevector instruction instead of the
transpose. But we expect that to
(1) become unwieldy for larger matrixes (even for 16x16 matrixes,
the resulting shufflevector masks would be huge),
(2) risk instcombine making small changes, causing us to fail to
detect the transpose, preventing better lowerings
For the load/store, we are additionally planning on exploiting the
intrinsics for better alias analysis.
Reviewers: anemet, Gerolf, reames, hfinkel, andrew.w.kaylor, efriedma, rengolin
Reviewed By: anemet
Differential Revision: https://reviews.llvm.org/D70456
This reapplies: 8ff85ed905
Original commit message:
As a follow-up to my initial mail to llvm-dev here's a first pass at the O1 described there.
This change doesn't include any change to move from selection dag to fast isel
and that will come with other numbers that should help inform that decision.
There also haven't been any real debuggability studies with this pipeline yet,
this is just the initial start done so that people could see it and we could start
tweaking after.
Test updates: Outside of the newpm tests most of the updates are coming from either
optimization passes not run anymore (and without a compelling argument at the moment)
that were largely used for canonicalization in clang.
Original post:
http://lists.llvm.org/pipermail/llvm-dev/2019-April/131494.html
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65410
This reverts commit c9ddb02659.
This change doesn't include any change to move from selection dag to fast isel
and that will come with other numbers that should help inform that decision.
There also haven't been any real debuggability studies with this pipeline yet,
this is just the initial start done so that people could see it and we could start
tweaking after.
Test updates: Outside of the newpm tests most of the updates are coming from either
optimization passes not run anymore (and without a compelling argument at the moment)
that were largely used for canonicalization in clang.
Original post:
http://lists.llvm.org/pipermail/llvm-dev/2019-April/131494.html
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D65410
Add a pass to lower is.constant and objectsize intrinsics
This pass lowers is.constant and objectsize intrinsics not simplified by
earlier constant folding, i.e. if the object given is not constant or if
not using the optimized pass chain. The result is recursively simplified
and constant conditionals are pruned, so that dead blocks are removed
even for -O0. This allows inline asm blocks with operand constraints to
work all the time.
The new pass replaces the existing lowering in the codegen-prepare pass
and fallbacks in SDAG/GlobalISEL and FastISel. The latter now assert
on the intrinsics.
Differential Revision: https://reviews.llvm.org/D65280
llvm-svn: 374784
This pass lowers is.constant and objectsize intrinsics not simplified by
earlier constant folding, i.e. if the object given is not constant or if
not using the optimized pass chain. The result is recursively simplified
and constant conditionals are pruned, so that dead blocks are removed
even for -O0. This allows inline asm blocks with operand constraints to
work all the time.
The new pass replaces the existing lowering in the codegen-prepare pass
and fallbacks in SDAG/GlobalISEL and FastISel. The latter now assert
on the intrinsics.
Differential Revision: https://reviews.llvm.org/D65280
llvm-svn: 374743
This allows later passes (in particular InstCombine) to optimize more
cases.
One that's important to us is `memcmp(p, q, constant) < 0` and memcmp(p, q, constant) > 0.
llvm-svn: 364412
NOTE: Note that no attributes are derived yet. This patch will not go in
alone but only with others that derive attributes. The framework is
split for review purposes.
This commit introduces the Attributor pass infrastructure and fixpoint
iteration framework. Further patches will introduce abstract attributes
into this framework.
In a nutshell, the Attributor will update instances of abstract
arguments until a fixpoint, or a "timeout", is reached. Communication
between the Attributor and the abstract attributes that are derived is
restricted to the AbstractState and AbstractAttribute interfaces.
Please see the file comment in Attributor.h for detailed information
including design decisions and typical use case. Also consider the class
documentation for Attributor, AbstractState, and AbstractAttribute.
Reviewers: chandlerc, homerdin, hfinkel, fedor.sergeev, sanjoy, spatel, nlopes, nicholas, reames
Subscribers: mehdi_amini, mgorny, hiraditya, bollu, steven_wu, dexonsmith, dang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59918
llvm-svn: 362578
Summary:
Match NewPassManager behavior: add option for interleaved loops in the
old pass manager, and use that instead of the flag used to disable loop unroll.
No changes in the defaults.
Reviewers: chandlerc
Subscribers: mehdi_amini, jlebar, dmgreen, hsaito, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61030
llvm-svn: 359615
Summary:
Make the flags in LICM + MemorySSA tuning options in the old and new
pass managers.
Subscribers: mehdi_amini, jlebar, Prazek, george.burgess.iv, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60490
llvm-svn: 358772
Summary:
Trying to add the plumbing necessary to add tuning options to the new pass manager.
Testing with the flags for loop vectorize.
Reviewers: chandlerc
Subscribers: sanjoy, mehdi_amini, jlebar, steven_wu, dexonsmith, dang, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59723
llvm-svn: 358763
Summary:
Create a method to forget everything in SCEV.
Add a cl::opt and PassManagerBuilder option to use this in LoopUnroll.
Motivation: Certain Halide applications spend a very long time compiling in forgetLoop, and prefer to forget everything and rebuild SCEV from scratch.
Sample difference in compile time reduction: 21.04 to 14.78 using current ToT release build.
Testcase showcasing this cannot be opensourced and is fairly large.
The option disabled by default, but it may be desirable to enable by
default. Evidence in favor (two difference runs on different days/ToT state):
File Before (s) After (s)
clang-9.bc 7267.91 6639.14
llvm-as.bc 194.12 194.12
llvm-dis.bc 62.50 62.50
opt.bc 1855.85 1857.53
File Before (s) After (s)
clang-9.bc 8588.70 7812.83
llvm-as.bc 196.20 194.78
llvm-dis.bc 61.55 61.97
opt.bc 1739.78 1886.26
Reviewers: sanjoy
Subscribers: mehdi_amini, jlebar, zzheng, javed.absar, dmgreen, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60144
llvm-svn: 358304
LTO provides additional opportunities for tailcall elimination due to
link-time inlining and visibility of nocapture attribute. Testing showed
negligible impact on compilation times.
Differential Revision: https://reviews.llvm.org/D58391
llvm-svn: 356511
The basic idea of the pass is to use a circular buffer to log the execution ordering of the functions. We only log the function when it is first executed. We use a 8-byte hash to log the function symbol name.
In this pass, we add three global variables:
(1) an order file buffer: a circular buffer at its own llvm section.
(2) a bitmap for each module: one byte for each function to say if the function is already executed.
(3) a global index to the order file buffer.
At the function prologue, if the function has not been executed (by checking the bitmap), log the function hash, then atomically increase the index.
Differential Revision: https://reviews.llvm.org/D57463
llvm-svn: 355133
With or without PGO data applied, splitting early in the pipeline
(either before the inliner or shortly after it) regresses performance
across SPEC variants. The cause appears to be that splitting hides
context for subsequent optimizations.
Schedule splitting late again, in effect reversing r352080, which
scheduled the splitting pass early for code size benefits (documented in
https://reviews.llvm.org/D57082).
Differential Revision: https://reviews.llvm.org/D58258
llvm-svn: 354158
Summary:
Follow up to D57082 which moved splitting earlier in the pipeline, in
order to perform it before inlining. However, it was moved too early,
before the IR is annotated with instrumented PGO data. This caused the
splitting to incorrectly determine cold functions.
Move it to just after PGO annotation (still before inlining), in both
pass managers.
Reviewers: vsk, hiraditya, sebpop
Subscribers: mehdi_amini, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D57805
llvm-svn: 353270
Performing splitting early has several advantages:
- Inhibiting inlining of cold code early improves code size. Compared
to scheduling splitting at the end of the pipeline, this cuts code
size growth in half within the iOS shared cache (0.69% to 0.34%).
- Inhibiting inlining of cold code improves compile time. There's no
need to inline split cold functions, or to inline as much *within*
those split functions as they are marked `minsize`.
- During LTO, extra work is only done in the pre-link step. Less code
must be inlined during cross-module inlining.
An additional motivation here is that the most common cold regions
identified by the static/conservative splitting heuristic can (a) be
found before inlining and (b) do not grow after inlining. E.g.
__assert_fail, os_log_error.
The disadvantages are:
- Some opportunities for splitting out cold code may be missed. This
gap can potentially be narrowed by adding a worklist algorithm to the
splitting pass.
- Some opportunities to reduce code size may be lost (e.g. store
sinking, when one side of the CFG diamond is split). This does not
outweigh the code size benefits of splitting earlier.
On net, splitting early in the pipeline has substantial code size
benefits, and no major effects on memory locality or performance. We
measured memory locality using ktrace data, and consistently found that
10% fewer pages were needed to capture 95% of text page faults in key
iOS benchmarks. We measured performance on frequency-stabilized iOS
devices using LNT+externals.
This reverses course on the decision made to schedule splitting late in
r344869 (D53437).
Differential Revision: https://reviews.llvm.org/D57082
llvm-svn: 352080
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
If the sample profile has no inlining hierachy information included, we call
the sample profile is flattened. For flattened profile, in ThinLTO postlink
phase, SampleProfileLoader's hot function inlining and profile annotation will
do nothing, so it is better to save the effort to read in the profile and run
the sample profile loader pass. It is helpful for reducing compile time when
the flattened profile is huge.
Differential Revision: https://reviews.llvm.org/D54819
llvm-svn: 351476
Currently we have pgo options defined in PassManagerBuilder.cpp only for
instrument pgo, but not for sample pgo. We also have pgo options defined
in NewPMDriver.cpp in opt only for new pass manager and for all kinds of
pgo. They have some inconsistency.
To make the options more consistent and make tests writing easier, the
patch let old pass manager to share the same pgo options with new pass
manager in opt, and removes the options in PassManagerBuilder.cpp.
Differential Revision: https://reviews.llvm.org/D56749
llvm-svn: 351392
At -O0, globalopt is not run during the compile step, and we can have a
chain of an alias having an immediate aliasee of another alias. The
summaries are constructed assuming aliases in a canonical form
(flattened chains), and as a result only the base object but no
intermediate aliases were preserved.
Fix by adding a pass that canonicalize aliases, which ensures each
alias is a direct alias of the base object.
Reviewers: pcc, davidxl
Subscribers: mehdi_amini, inglorion, eraman, steven_wu, dexonsmith, arphaman, llvm-commits
Differential Revision: https://reviews.llvm.org/D54507
llvm-svn: 350423
Rename:
NoUnrolling to InterleaveOnlyWhenForced
and
AlwaysVectorize to !VectorizeOnlyWhenForced
Contrary to what the name 'AlwaysVectorize' suggests, it does not
unconditionally vectorize all loops, but applies a cost model to
determine whether vectorization is profitable to all loops. Hence,
passing false will disable the cost model, except when a loop is marked
with llvm.loop.vectorize.enable. The 'OnlyWhenForced' suffix (suggested
by @hfinkel in D55716) better matches this behavior.
Similarly, 'NoUnrolling' disables the profitability cost model for
interleaving (a term to distinguish it from unrolling by the
LoopUnrollPass); rename it for consistency.
Differential Revision: https://reviews.llvm.org/D55785
llvm-svn: 349513
When using clang with `-fno-unroll-loops` (implicitly added with `-O1`),
the LoopUnrollPass is not not added to the (legacy) pass pipeline. This
also means that it will not process any loop metadata such as
llvm.loop.unroll.enable (which is generated by #pragma unroll or
WarnMissedTransformationsPass emits a warning that a forced
transformation has not been applied (see
https://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20181210/610833.html).
Such explicit transformations should take precedence over disabling
heuristics.
This patch unconditionally adds LoopUnrollPass to the optimizing
pipeline (that is, it is still not added with `-O0`), but passes a flag
indicating whether automatic unrolling is dis-/enabled. This is the same
approach as LoopVectorize uses.
The new pass manager's pipeline builder has no option to disable
unrolling, hence the problem does not apply.
Differential Revision: https://reviews.llvm.org/D55716
llvm-svn: 349509
When multiple loop transformation are defined in a loop's metadata, their order of execution is defined by the order of their respective passes in the pass pipeline. For instance, e.g.
#pragma clang loop unroll_and_jam(enable)
#pragma clang loop distribute(enable)
is the same as
#pragma clang loop distribute(enable)
#pragma clang loop unroll_and_jam(enable)
and will try to loop-distribute before Unroll-And-Jam because the LoopDistribute pass is scheduled after UnrollAndJam pass. UnrollAndJamPass only supports one inner loop, i.e. it will necessarily fail after loop distribution. It is not possible to specify another execution order. Also,t the order of passes in the pipeline is subject to change between versions of LLVM, optimization options and which pass manager is used.
This patch adds 'followup' attributes to various loop transformation passes. These attributes define which attributes the resulting loop of a transformation should have. For instance,
!0 = !{!0, !1, !2}
!1 = !{!"llvm.loop.unroll_and_jam.enable"}
!2 = !{!"llvm.loop.unroll_and_jam.followup_inner", !3}
!3 = !{!"llvm.loop.distribute.enable"}
defines a loop ID (!0) to be unrolled-and-jammed (!1) and then the attribute !3 to be added to the jammed inner loop, which contains the instruction to distribute the inner loop.
Currently, in both pass managers, pass execution is in a fixed order and UnrollAndJamPass will not execute again after LoopDistribute. We hope to fix this in the future by allowing pass managers to run passes until a fixpoint is reached, use Polly to perform these transformations, or add a loop transformation pass which takes the order issue into account.
For mandatory/forced transformations (e.g. by having been declared by #pragma omp simd), the user must be notified when a transformation could not be performed. It is not possible that the responsible pass emits such a warning because the transformation might be 'hidden' in a followup attribute when it is executed, or it is not present in the pipeline at all. For this reason, this patche introduces a WarnMissedTransformations pass, to warn about orphaned transformations.
Since this changes the user-visible diagnostic message when a transformation is applied, two test cases in the clang repository need to be updated.
To ensure that no other transformation is executed before the intended one, the attribute `llvm.loop.disable_nonforced` can be added which should disable transformation heuristics before the intended transformation is applied. E.g. it would be surprising if a loop is distributed before a #pragma unroll_and_jam is applied.
With more supported code transformations (loop fusion, interchange, stripmining, offloading, etc.), transformations can be used as building blocks for more complex transformations (e.g. stripmining+stripmining+interchange -> tiling).
Reviewed By: hfinkel, dmgreen
Differential Revision: https://reviews.llvm.org/D49281
Differential Revision: https://reviews.llvm.org/D55288
llvm-svn: 348944
Summary:
In the new+old pass manager, hot cold splitting was schedule too early.
Thanks to Vedant for pointing this out.
Reviewers: sebpop, vsk
Reviewed By: sebpop, vsk
Subscribers: mehdi_amini, llvm-commits
Differential Revision: https://reviews.llvm.org/D53437
llvm-svn: 344869
This reverts commit r342387 as it's showing significant performance
regressions in a number of benchmarks. Followed up with the
committer and original thread with an example and will get performance
numbers before recommitting.
llvm-svn: 343522
Summary: This patch adds bindings to C and Go for addCoroutinePassesToExtensionPoints, which is used to add coroutine passes to the correct locations in PassManagerBuilder.
Reviewers: whitequark, deadalnix
Reviewed By: whitequark
Subscribers: mehdi_amini, modocache, llvm-commits
Differential Revision: https://reviews.llvm.org/D51642
llvm-svn: 343336
This patch turns LoopInterchange into a loop pass. It now only
considers top-level loops and tries to move the innermost loop to the
optimal position within the loop nest. By only looking at top-level
loops, we might miss a few opportunities the function pass would get
(e.g. if we have a loop nest of 3 loops, in the function pass
we might process loops at level 1 and 2 and move the inner most loop to
level 1, and then we process loops at levels 0, 1, 2 and interchange
again, because we now have a different inner loop). But I think it would
be better to handle such cases by picking the best inner loop from the
start and avoid re-visiting the same loops again.
The biggest advantage of it being a function pass is that it interacts
nicely with the other loop passes. Without this patch, there are some
performance regressions on AArch64 with loop interchanging enabled,
where no loops were interchanged, but we missed out on some other loop
optimizations.
It also removes the SimplifyCFG run. We are just changing branches, so
the CFG should not be more complicated, besides the additional 'unique'
preheaders this pass might create.
Reviewers: chandlerc, efriedma, mcrosier, javed.absar, xbolva00
Reviewed By: xbolva00
Differential Revision: https://reviews.llvm.org/D51702
llvm-svn: 343308
Rebase rL341954 since https://bugs.llvm.org/show_bug.cgi?id=38912
has been fixed by rL342055.
Precommit testing performed:
* Overnight runs of csmith comparing the output between programs
compiled with gvn-hoist enabled/disabled.
* Bootstrap builds of clang with UbSan/ASan configurations.
llvm-svn: 342387
This reverts rL341954.
The builder `sanitizer-x86_64-linux-bootstrap-ubsan` has been
failing with timeouts at stage2 clang/ubsan:
[3065/3073] Linking CXX executable bin/lld
command timed out: 1200 seconds without output running python
../sanitizer_buildbot/sanitizers/buildbot_selector.py,
attempting to kill
llvm-svn: 342001
