This produced a compile error with GCC:
llvm-project/polly/lib/Transform/ScheduleOptimizer.cpp:1220:49: error: cannot convert ‘bool’ to ‘llvm::TargetTransformInfo::RegisterKind’
1220 | RegisterBitwidth = TTI->getRegisterBitWidth(true);
BandAttr markers are added as parents of schedule tree bands. These also
appear as markers its equivalent AST, but a band does not necessarily
corresponds to a loop in this. Iterations may be peeled or the loop
being unrolled (e.g. if it has just one iteration). In such cases it may
happend that there is not loop between a BandAttr marker and the marker
for a loop nested in the former parent band/loop.
Handle the situation by giving priority to the inner marker over the
outer.
Fixes the polly-x86_64-linux-test-suite buildbot.
We enumerated the cross product Domain x Scatter, but sorted only be the
scatter key. In case there are are multiple statement instances per
scatter value, the order between statement instances of the same loop
iteration was undefined.
Propertly enumerate and sort only by the scatter value, and group the
domains using the scatter dimension again.
Thanks to Leonard Chan for the report.
Make Polly look for unrolling metadata (https://llvm.org/docs/TransformMetadata.html#loop-unrolling) that is usually only interpreted by the LoopUnroll pass and apply it to the SCoP's schedule.
While not that useful by itself (there already is an unroll pass), it introduces mechanism to apply arbitrary loop transformation directives in arbitrary order to the schedule. Transformations are applied until no more directives are found. Since ISL's rescheduling would discard the manual transformations and it is assumed that when the user specifies the sequence of transformations, they do not want any other transformations to apply. Applying user-directed transformations can be controlled using the `-polly-pragma-based-opts` switch and is enabled by default.
This does not influence the SCoP detection heuristic. As a consequence, loop that do not fulfill SCoP requirements or the initial profitability heuristic will be ignored. `-polly-process-unprofitable` can be used to disable the latter.
Other than manually editing the IR, there is currently no way for the user to add loop transformations in an order other than the order in the default pipeline, or transformations other than the one supported by clang's LoopHint. See the `unroll_double.ll` test as example that clang currently is unable to emit. My own extension of `#pragma clang loop` allowing an arbitrary order and additional transformations is available here: https://github.com/meinersbur/llvm-project/tree/pragma-clang-loop. An effort to upstream this functionality as `#pragma clang transform` (because `#pragma clang loop` has an implicit transformation order defined by the loop pipeline) is D69088.
Additional transformations from my downstream pragma-clang-loop branch are tiling, interchange, reversal, unroll-and-jam, thread-parallelization and array packing. Unroll was chosen because it uses already-defined metadata and does not require correctness checks.
Reviewed By: sebastiankreutzer
Differential Revision: https://reviews.llvm.org/D97977
Polly currently needs to be slowly refactor to use the C++ wrapper objects to handle the reference counters automatically.
I took the function of astScheduleDimIsParallel and refactored it so that it uses the C++ wrapper function as much as possible.
There are some problems with the IsParallel since it expects the C objects, so the C++ wrapper functions must be .release() and .get() first before they are able to be used with IsParallel.
When checking the ReductionDependencies Parallelism with the Build's Schedule, I opted to keep the union map as a C object rather than a C++ object. Eventually, changes will need to be made to IsParallel to refactor it to the C++ wrappers. When this is done, this function will also need to be slightly refactored to not use the C object.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D98455
This reverts commit 329aeb5db4,
and relands commit 61f006ac65.
This is a continuation of D89456.
As it was suggested there, now that SCEV models `PtrToInt`,
we can try to improve SCEV's pointer handling.
In particular, i believe, i will need this in the future
to further fix `SCEVAddExpr`operation type handling.
This removes special handling of `ConstantPointerNull`
from `ScalarEvolution::createSCEV()`, and add constant folding
into `ScalarEvolution::getPtrToIntExpr()`.
This way, `null` constants stay as such in SCEV's,
but gracefully become zero integers when asked.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D98147
This removes some (but not all) uses of type-less CreateGEP()
and CreateInBoundsGEP() APIs, which are incompatible with opaque
pointers.
There are a still a number of tricky uses left, as well as many
more variation APIs for CreateGEP.
This is a continuation of D89456.
As it was suggested there, now that SCEV models `PtrToInt`,
we can try to improve SCEV's pointer handling.
In particular, i believe, i will need this in the future
to further fix `SCEVAddExpr`operation type handling.
This removes special handling of `ConstantPointerNull`
from `ScalarEvolution::createSCEV()`, and add constant folding
into `ScalarEvolution::getPtrToIntExpr()`.
This way, `null` constants stay as such in SCEV's,
but gracefully become zero integers when asked.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D98147
Emit llvm.loop.parallel_accesses metadata instead of
llvm.mem.parallel_loop_access. The latter is deprecated because it
assumes that LoopIDs are persistent, which they are not.
We also emit parallel access metadata for all surrounding parallel
loops, not just the innermost parallel.
Polly use algorithms from the Integer Set Library (isl), which is a library written in C and which is incompatible with the rest of the LLVM as it is written in C++.
Changes made:
* Refabricating IsOutermostParallel() to take C++ bindings instead of reference-counting in C isl lib.
* Addition of manage_copy() to be used as reference for C objects instead of IsOutermostParallel()
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D97751
Currently, the IslAst library is a C library that would be incompatible with the rest of the LLVM because LLVM is written in C++.
I took one function, IsInnermostParallel(), and refactored it so that it would take the C++ wrapper object instead of using reference counters with the C ISL library. As well, all the references that use IsInnermostParallel() will use manage_copy() since they are still expecting the C object.
Reviewed By: Meinersbur
Differential Revision: https://reviews.llvm.org/D97425
Allow users to use a non-system version of perl, python and awk, which is useful
in certain package managers.
Reviewed By: JDevlieghere, MaskRay
Differential Revision: https://reviews.llvm.org/D95119
Regenerate the C++ wrapper header from the current isl version's
headers.
The most notable change is that some dimension sizes are represented by
an isl_size (instead of unsigned), which is a signed int. Additionally,
some function may return -1 in case of an error which already had been
fixed in the past. The C++ may no return -1 instead of UINT_MAX which
caused the problems.
Some types in Polly had been changed from unsigned to isl_size
(that were not already auto) and some loops/comparision had to be
changed to avoid unsigned/signed comparison warnings.
ScopDetection's DetectionContext holds AssertionVH for
RequiredInvariantLoads. An assertion is thrown if the handle's value is
erased and the ScopDetection is not yet invalidated. The ScopDetection
must remain valid durting the ScopPassManager. Enusure that all Scop
analyses are free'd when the ScopPass manager is done.
If IR generation has happened, also invalidate all other passes to avoid
possible issues because, like for the legacy pass manager, Polly does not
yet perfectly preserve them.
DetectionContext objects are stored as values in a DenseMap. When the
DenseMap reaches its maximum load factor, it is resized and all its
objects moved to a new memory allocation. Unfortunately Scop object have
a reference to its DetectionContext. When the DenseMap resizes, all the
DetectionContexts reference now point to invalid memory, even if caused
by an unrelated DetectionContext.
Even worse, NewPM's ScopPassManager called isMaxRegionInScop with the
Verify=true parameter before each pass. This caused the old
DetectionContext to be removed an a new on created and re-verified.
Of course, the Scop object was already created pointing to the old
DetectionContext. Because the new DetectionContext would
usually be stored at the same position in the DenseMap, the reference
would usually reference the new DetectionContext of the same Region.
Usually.
If not, the old position still points to memory in the DenseMap
allocation (unless also a resizing occurs) such that tools like Valgrind
and AddressSanitizer would not be able to diagnose this.
Instead of storing the DetectionContext inside the DenseMap, use a
std::unique_ptr to a DetectionContext allocation, i.e. it will not move
around anymore. This also allows use to remove the very strange
DetectionContext(const DetectionContext &&)
copy/move(?) constructor. DetectionContext objects now are neither
copied nor moved.
As a result, every re-verification of a DetectionContext will use a new
allocation. Therefore, once a Scop object has been created using a
DetectionContext, it must not be re-verified (the Scop data structure
requires its underlying Region to not change before code generation
anyway). The NewPM may call isMaxRegionInScop only with
Validate=false parameter.
The description of the -polly switch stated that it was only enabled
with -O3. This was a lie, the optimization level was ignored. Only at
-O0 Polly was not added to the pass pipeline because the pass builder,
but only because the extension points were not triggered.
In the NewPM, the VectorizerStart extensions point is actually trigger
even with -O0 which leads to the following crash:
Assertion `Level != OptimizationLevel::O0 && "Must request optimizations!"' failed.
We sanitize the optimization levels using the following rules for both
pass mangers:
1. Only enable Polly if optimizing at all (-O1, -O2 or -O3).
2. Do not enable Polly when optimizing for size.
3. Ignore the optimization level for diagnostic passes (printer, viewer
or JScop-exporter).
4. If only diagnostic passes enabled, skip the code-generation.
5. Fix the description of the -polly command line option.
These are implementation details of the IslScheduleOptimizer pass
implementation and not use anywhere else. Hence, we can move them to the
cpp file and into an anonymous namespace.
Only getPartialTilePrefixes is, aside from the pass itself, used
externally (by the ScheduleOptimizerTest) and moved into the polly
namespace.
This reverts commit b7d870eae7 and the
subsequent fix "[Polly] Fix build after AssumptionCache change (D96168)"
(commit e6810cab09).
It caused indeterminism in the output, such that e.g. the
polly-x86_64-linux buildbot failed accasionally.
Move SimplifiyVisitor from Simplify.h to Simplify.cpp. It is not
relevant for applying the pass in either the NewPM or the legacyPM.
Rename it to SimplifyImpl to account for that.
This is possible due its state not being necessary to be preserved
between runs and thefore SimplifyImpl not needed to be held in the
pass object. Instead, SimplifyImpl is only instatiated for the
current Scop. In the NewPM as a function-local variable, and in the
legacy PM inside a llvm::Optional object because the state must be
preserved between the printScop (invoked by opt -analyze) and the most
recent runOnScop calls.
"using namespace" pollutes the namespace of every file that includes
such a header and universally considered a bad thing. Even the variant
namespace polly {
using namespace llvm;
}
(previously used by LoopGenerators.h) imports more symbols than the file
is in control of. The header may include a fixed set of files from LLVM,
but the header itself may by be included together with other headers
from LLVM. For instance, LLVM's MemorySSA.h and Polly's ScopInfo.h both
declare a class 'MemoryAccess' which may conflict.
Instead of prefixing everything in Polly's header files, this patch adds
'using' statements to import only the symbols that are actually
referenced in Polly. This approach is also used by MLIR to import
commonly used symbols into the mlir namespace.
This patch also puts the symbols declared in IslNodeBuilder.h into the
Polly namespace to also be able to use the imported symbols.
Even though it has some oddities, both pipelines should be as similar as
possible. Also use report_fatal_error instead of assertions to ensure a
proper failure in release builds for unsupported options.
This finalizes the patch serious to make Polly run in the default
configuration when using the NewPM by default.
In particular, print the ast with -debug-only=polly-ast, print a
per-scop header with print<polly-ast> and force-add the analysis with
-polly-code-generation=ast.
TargetTransformInfo is required by IslScheduleOptimizer, as ScopPass.
Unfortunately it is not possible to get arbitrary larger-unit analyses
in for as ScopPass. Loop passes also already use TargetTransformInfo as
LoopStandardAnalysisResults, hence wei might expect it to be available
to Scop passes as well.
The pass-instrumentation pass is implicitly execute by the NewPM
whenever a new analysis runs. Not registering it will cause the crash
whenever a scop pass requests an analysis.
For instance this is the case for the IstAstAnalysis requesting the
DependenceAnalsis result.
In addition to that regression tests should not test the intire pass
pipeline (unless they are testing the pipeline itself), the Polly-ACC
currently does not support the new pass manager. If enabled by default,
such tests will therefore fail.
Use the -polly-gpu-runtime and -polly-gpu-arch options also as default
values for the PPCGCodeGeneration pass. This requires to move the option
to be moved from the pipeline-building Register passes to the
PPCGCodeGeneration implementation.
Fixes the spir-typesize.ll buildbot fail.
Fangrui Song