The DEBUG() macro is very generic so it might clash with other projects.
The renaming was done as follows:
- git grep -l 'DEBUG' | xargs sed -i 's/\bDEBUG\s\?(/LLVM_DEBUG(/g'
- git diff -U0 master | ../clang/tools/clang-format/clang-format-diff.py -i -p1 -style LLVM
Differential Revision: https://reviews.llvm.org/D44978
llvm-svn: 332352
Add the options -polly-codegen-trace-stmts and
-polly-codegen-trace-scalars. When enabled, adds a call to the
beginning of every generated statement that prints the executed
statement instance. With -polly-codegen-trace-scalars, it also prints
the value of all scalars that are used in the statement, and PHIs
defined in the beginning of the statement.
Differential Revision: https://reviews.llvm.org/D45743
llvm-svn: 330864
In r330292 this assert was turned incorrectly into an unreachable, but
the correct behavior (thanks Michael) is to assert for anything that is
not 64 bit, but falltrough for 64 bit. I document this in the source
code.
llvm-svn: 330309
A check in assert-builds was meant to verify that a load provides a
value in all statement instances (i.e. its domain). The domain is
commonly gist'ed within the parameter context to contain fewer
constraints. However, statement instances outside the context are
no valid executions, hence the value provided can be undefined.
Refine the check for valid loads to only needed to be defined within
the SCoP context.
In addition, the JSONImporter had to be changed to allow importing
access relations that are broader than the current access relation,
but still defined over all statement instances.
This should fix the compiler crash in test-suite's oggenc of the
-polly-process-unprofitable buildbot.
llvm-svn: 329655
Summary:
When checking the parallelism of a scheduling dimension, we first check if excluding reduction dependences the loop is parallel or not.
If the loop is not parallel, then we need to return the minimal dependence distance of all data dependences, including the previously subtracted reduction dependences.
Reviewers: grosser, Meinersbur, efriedma, eli.friedman, jdoerfert, bollu
Reviewed By: Meinersbur
Subscribers: llvm-commits, pollydev
Tags: #polly
Differential Revision: https://reviews.llvm.org/D45236
llvm-svn: 329214
During codegen, Polly attempts to clear all loops from ScalarEvolution
and LoopInfo, and it does so one block at a time. This causes undefined
behaviour, since this way a loop header might be removed from a loop
before the entire loop is erased, causing ScalarEvolution to run into an
error.
Instead, just delete the entire loop atomically. This fixes currently
failing testcases.
llvm-svn: 326643
As part of this cleanup a couple of unnecessary isl::manage(obj.copy()) pattern
are eliminated as well.
We checked for all potential cleanups by scanning for:
"grep -R isl::manage\( lib/ | grep copy"
llvm-svn: 325558
Memory transfer instructions take two pointers. It is not defined to
which of those a noalias annotation applies. To ensure correctness,
do not add noalias annotations to memcpy/memmove instructions anymore.
The caused a miscompile with test-suite's MultiSource/Applications/obsequi.
Since r321138, the MemCpyOpt pass would remove memcpy/memmove calls if
known to copy uninitialized memory. In that case, it was initialized
by another memcpy, but the annotation for the target pointer said
it would not alias. The annotation was actually meant for the source
pointer, which was was an alloca and could not alias with the target
pointer.
llvm-svn: 321371
Isl does not allow generating isl_ast_expr from an isl_pw_aff that has an
empty domain (i.e. has no pieces). We already detected the case if the
isl_pw_aff comes with an empty domain.
isl_ast_build also considers the domain empty if it is disjoint with the
parameter context (e.g. parameters values that we exclude by runtime
versioning).
Intersect the access relation domain with the parameter context to
also detect such practically empty access domains. The effective
pointer used in the generated code is unimportand because it will never
be executed.
This fixes llvm.org/PR35362
llvm-svn: 318806
Summary:
Most changes are mechanical, but in one place I changed the program semantics
by fixing a likely bug:
In `Scop::hasFeasibleRuntimeContext()`, I'm now explicitely handling the
error-case. Before, when the call to `addNonEmptyDomainConstraints()`
returned a null set, this (probably) accidentally worked because
isl_bool_error converts to true. I'm checking for nullptr now.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: Meinersbur
Subscribers: nemanjai, kbarton, pollydev, llvm-commits
Differential Revision: https://reviews.llvm.org/D39971
llvm-svn: 318632
When collecting base pointers that need to be made available in parallel
subfunctions, use the base pointer associated with the latest
ScopArrayInfo, instead of the original one.
llvm-svn: 316983
Summary:
When GPUNodeBuilder creates loops inside the kernel, it dispatches to
IslNodeBuilder. This however is surprisingly dangerous, since it accesses the
AST Node's user through the wrong type. This patch fixes this problem by
overriding createFor correctly.
This fixes PR35010.
Reviewers: grosser, bollu, Meinersbur
Reviewed By: Meinersbur
Subscribers: Meinersbur, nemanjai, pollydev, llvm-commits, kbarton
Differential Revision: https://reviews.llvm.org/D39364
llvm-svn: 316872
We make sure that the final reload of an invariant scalar memory access uses the
same stack slot into which the invariant memory access was stored originally.
Earlier, this was broken as we introduce a new stack slot aside of the preload
stack slot, which remained uninitialized and caused our escaping loads to
contain garbage. This happened due to us clearing the pre-populated values
in EscapeMap after kernel code generation. We address this issue by preserving
the original host values and restoring them after kernel code generation.
EscapeMap is not expected to be used during kernel code generation, hence we
clear it during kernel generation to make sure that any unintended uses are
noticed.
llvm-svn: 314894
This matches the behavior we already have in lib/Codegen/CodeGeneration.cpp and
makes sure that we fall back to the original code. It seems when invariant load
hoisting was introduced to the GPGPU backend we missed to reset the RTC flag,
such that kernels where invariant load hoisting failed executed the 'optimized'
SCoP, which however is set to a simple 'unreachable'. Unsurprisingly, this
results in hard to debug issues that are a lot of fun to debug.
llvm-svn: 314624
Such RTCs may introduce integer wrapping intrinsics with more than 64 bit,
which are translated to library calls on AOSP that are not part of the
runtime and will consequently cause linker errors.
Thanks to Eli Friedman for reporting this issue and reducing the test case.
llvm-svn: 314065
Since -polly-codegen reports itself to preserve DependenceInfo and IslAstInfo,
we might get those analysis that were computed by a different ScopInfo for a
different Scop structure. This would be unfortunate because DependenceInfo and
IslAstInfo hold references to resources allocated by
ScopInfo/ScopBuilder/Scop (e.g. isl_id). If -polly-codegen and
DependenceInfo/IslAstInfo do not agree on which Scop to use, unpredictable
things can happen.
When the ScopInfo/Scop object is freed, there is a high probability that the
new ScopInfo/Scop object will be created at the same heap position with the
same address. Comparing whether the Scop or ScopInfo address is the expected
therefore is unreliable.
Instead, we compare the address of the isl_ctx object. Both, DependenceInfo
and IslAstInfo must hold a reference to the isl_ctx object to ensure it is
not freed before the destruction of those analyses which might happen after
the destruction of the Scop/ScopInfo they refer to. Hence, the isl_ctx
will not be freed and its address not reused as long there is a
DependenceInfo or IslAstInfo around.
This fixes llvm.org/PR34441
llvm-svn: 313842
Update CodegenCleanup using the function-level passes added by
populatePassManager that run between EP_EarlyAsPossible and
EP_VectorizerStart in -O3.
The changes in particular are:
- Added pass create arguments, e.g. ExpensiveCombines for InstCombine.
- Remove reroll pass. The option -reroll-loops is disabled by default.
- Add passes run with UnitAtATime, which is the default.
- Add instances of LibCallsShrinkWrap, TailCallElimination, SCCP
(sparse conditional constant propagation), Float2Int
that did not run before.
- Add instances of GVN as in the default pipeline.
Notes:
- GVNHoist, GVNSink, NewGVN are still disabled in the -O3 pipeline.
- The optimization level and other optimization parameters are not
accessible outside of PassManagerBuilder, hence we cannot add passes
depending on these.
Differential Revision: https://reviews.llvm.org/D37571
llvm-svn: 312875
The type of NewValue might change due to ScalarEvolution
looking though bitcasts. The synthesized NewValue therefore
becomes the type before the bitcast.
llvm-svn: 312718
In certain situations, the context in the isl_ast_build could result for the
min/max locations of our alias sets to become empty, which would cause an
internal error in isl, which is then unable to derive a value for these
expressions. Check these conditions before code generating expressions and
instead assume that alias check succeeded. This is valid, as the corresponding
memory accesses will not be executed under any valid context.
This fixed llvm.org/PR34432. Thanks to Qirun Zhang for reporting.
llvm-svn: 312455
In Polly, we specifically add a paramter to represent the outermost dimension
size of fortran arrays. We do this because this information is statically
available from the fortran metadata generated by dragonegg.
However, we were only materializing these parameters (meaning, creating an
llvm::Value to back the isl_id) from *memory accesses*. This is wrong,
we should materialize parameters from *scop array info*.
It is wrong because if there is a case where we detect 2 fortran arrays,
but only one of them is accessed, we may not materialize the other array's
dimensions at all.
This is incorrect. We fix this by looping over all
`polly::ScopArrayInfo` in a scop, rather that just all `polly::MemoryAccess`.
Differential Revision: https://reviews.llvm.org/D37379
llvm-svn: 312350
Currently, GVN can be necessary to eliminate redundant instructions in case
of, for instance, GEMM and float type. This patch makes GVN be run during
the cleanup.
Reviewed-by: Tobias Grosser <tobias@grosser.es>,
Michael Kruse <llvm@meinersbur.de>
Differential Revision: https://reviews.llvm.org/D37340
llvm-svn: 312307
This is useful when we face certain intrinsics such as `llvm.exp.*`
which cannot be lowered by the NVPTX backend while other intrinsics can.
So, we would need to keep blacklists of intrinsics that cannot be
handled by the NVPTX backend. It is much simpler to try and promote
all intrinsics to libdevice versions.
This patch makes function/intrinsic very uniform, and will always try to use
a libdevice version if it exists.
Differential Revision: https://reviews.llvm.org/D37056
llvm-svn: 312239
The adds code generation support for the previous commit.
This patch has been re-applied, after the memory issue in the previous patch
has been fixed.
llvm-svn: 312211
Properly require and preserve the OptimizationRemarkEmitter for use in
ScopPass. Previously one had to get the ORE from ScopDetection because
CodeGeneration did not mark it as preserved. It would need to be
recomputed which results in the legacy PM to throw away all previous
SCoP analysis.
This also changes the implementation of ScopPass::getAnalysisUsage to
not unconditionally preserve all passes, but only those needed to be
preserved by any SCoP pass (at least when using the legacy PM). This
allows invalidating DependenceInfo (and IslAstInfo) in case the pass
would cause them to change (e.g. OpTree, DeLICM, MaximalArrayExpansion)
JSONImporter should also invalidate the DependenceInfo. In this patch
it marks DependenceInfo as preserved anyway because some regression
tests depend on it.
Differential Revision: https://reviews.llvm.org/D37010
llvm-svn: 311888
Whether a partial write is tautological/unsatisfiable not only
depends on the access domain, but also on the domain covered
by its node in the AST.
In the example below, there are two instances of Stmt_cond_false. It may have a partial write access that is not executed in instance Stmt_cond_false(0).
for (int c0 = 0; c0 < tmp5; c0 += 1) {
Stmt_for_body344(c0);
if (tmp5 >= c0 + 2)
Stmt_cond_false(c0);
Stmt_cond_end(c0);
}
if (tmp5 <= 0) {
Stmt_for_body344(0);
Stmt_cond_false(0);
Stmt_cond_end(0);
}
Isl cannot derive a subscript for an array element that is never accessed.
This caused an error in that no subscript expression has been generated
in IslNodeBuilder::createNewAccesses, but BlockGenerator expected one
to exist because there is an execution of that write, just not in that
ast node.
Fixed by instead of determining whether the access domain is empty,
inspect whether isl generated a constant "false" ast expression in
the current ast node.
This should fix a compiler crash of the aosp buildbot.
llvm-svn: 311663
This is a stylistic change to make the function a little more readable.
Also add a debug print to show what instruction contains a use of a
function we don't understand in the kernel.
Differential Revision: https://reviews.llvm.org/D37058
llvm-svn: 311648
Add statistics about
- Which optimizations are applied
- Number of loops in Scops at various stages
- Number of scalar/singleton writes at various stages representative
for scalar false dependencies
- Number of parallel loops
These will be useful to find regressions due to moving Polly further
down of LLVM's pass pipeline.
Differential Revision: https://reviews.llvm.org/D37049
llvm-svn: 311553
MSVC warns about comparison between a signed and unsigned integer.
The rules of C(++) define that an unsigned comparison has to be
carried-out in this case. This is unlikely to be intended.
Fix by assigning the loop's upper bound to a signed integer first.
This also avoids repeated evaluation of the invariant upper bound.
llvm-svn: 311548
Summary:
There is no need to emit alias metadata for scalars, as basicaa will easily
distinguish them from arrays. This reduces the size of the metadata we generate.
This is especially useful after we moved to -polly-position=before-vectorizer,
where a lot more scalar dependences are introduced, which increased the size of
the alias analysis metadata and made us commonly reach the limits after which
we do not emit alias metadata that have been introduced to prevent quadratic
growth of this alias metadata.
This improves 2mm performance from 1.5 seconds to 0.17 seconds.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: Meinersbur
Subscribers: pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D37028
llvm-svn: 311498
Currently, in case of GEMM and the pattern matching based optimizations, we
use only the SLP Vectorizer out of two LLVM vectorizers. Since the Loop
Vectorizer can get in the way of optimal code generation, we disable the Loop
Vectorizer for the innermost loop using mark nodes and emitting the
corresponding metadata.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D36928
llvm-svn: 311473
This feature was not enabled for `PPCGCodeGeneration`. Now that this is
enabled, we can benchmark Scops that have been optimised with
`-polly-codegen-ppcg` with the `-polly-codegen-perf-monitoring` option.
Differential Revision: https://reviews.llvm.org/D36934
llvm-svn: 311328
We still see some issues with parameter space mismatches. Revert this to get
a clean baseline. We will recommit after these issues have been resolved.
This reverts commit 0e360a14194f722ded7aa2bc9d4be2ed2efeeb49.
llvm-svn: 311268
Instead of using Twines and temporary expressions, we do string manipulation
through a std::string. This resolves a memory corruption issue, which likely
was caused by twines loosing their underlying string too soon.
llvm-svn: 311264
- We should iterate over `I`, which is `Cur` expanded out to an
instruction, and not `Cur` itself.
- This is a bugfix.
Differential Revision: https://reviews.llvm.org/D36923
llvm-svn: 311261
Summary:
This information is necessary for PPCG to perform correct life range reordering.
With these changes applied we can live-range reorder some of the important
kernels in COSMO.
We also update and rename one test case, which previously could not be optimized
and now is optimized thanks to live-range reordering. To preserve test coverage
we add a new test case scalar-writes-in-scop-requires-abort.ll, which exercises
our automatic abort in case of scalar writes in the kernel.
Reviewers: Meinersbur, bollu, singam-sanjay
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36929
llvm-svn: 311259
Kernel argument sizes now only get appended to the kernel launch parameter list if the OpenCL runtime is selected, not if CUDA runtime is chosen.
Differential revision: D36925
llvm-svn: 311248
When using -polly-ignore-integer-wrapping and -polly-acc-codegen-managed-memory
we add parameter dimensions lazily to the domains, which results in PPCG not
including parameter dimensions that are only used in memory accesses in the
kernel space. To make sure these parameters are still passed to the kernel, we
collect these parameter dimensions and align the kernel's parameter space
before code-generating it.
llvm-svn: 311239
Summary:
Drop unused parameter dimensions to reduce the size of the sets we are working
with. Especially the computed dependences tend to accumulate a lot of parameters
that are present in the input memory accesses, but often not necessary to
express the actual dependences. As isl represents maps and sets with dense
matrices, reducing the dimensionality of isl sets commonly reduces code
generation performance.
This reduces compile time from 17 to 11 seconds for our test case. While this is
not impressive, this patch helped me to identify the previous two performance
improvements and additionally also increases readability of the isl data
structures we use.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36869
llvm-svn: 311161
Summary:
They are not used and consequently do not even need to be computed. This reduces
the overall compile time for our kernel from 1m33s to 17s.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36868
llvm-svn: 311157
Summary:
This change reduces the overall number of synchronize calls for kernels with
a lot of output data at the cost of additional synchronize calls for kernels
launched in sequence without any device to host transfers in between. As the
latter pattern is a lot less frequent, this seems a better tradeoff.
Even though the above motivation would be motivation enough, this is just
a step towards enabling ppcg to not compute to and from device copy calls
at all, which would be incorrect in case we still relied on these calls to
place our synchronization statements.
Reviewers: Meinersbur, bollu, singam-sanjay
Reviewed By: bollu
Subscribers: nemanjai, kbarton, pollydev, llvm-commits
Tags: #polly
Differential Revision: https://reviews.llvm.org/D36867
llvm-svn: 311155
This avoid the construction of very large sets and in many cases also keeps the
number of parameters low. As a result, we see a compile time reduction from 5
minutes to only slightly above 1 minute for one of our larger test cases.
llvm-svn: 311127
Reuse the machinery built for replacing global arrays to replace malloc/free as
well. Example replacement that was missed earlier:
```
call void \
bitcast (void (i8*)* @free to void (%custom_type*)*) (%custom_type* %13)
```
- Since the `bitcast` is a `ConstantExpr`, `replaceAllUsesWith` would miss
this. We don't miss this anymore.
Differential Revision: https://reviews.llvm.org/D36825
llvm-svn: 311121
- If we have global arrays, we would like to rewrite them to global
pointers which are allocated using `cudaMallocManaged`.
- If we have allocas in a function, we would like to rewrite them to
heap-allocations with `cudaMallocManaged` and `cudaFree`.
- With these rewrite mechanisms, we can offload _any_ function to the
GPU with no code rewrite whatsover.
Differential Revision: https://reviews.llvm.org/D36516
llvm-svn: 311080
Summary:
During code generation for a Scop we modify the IR of a function.
While this shouldn't affect a Scop in the formal sense, the implementation
caches various information about the IR such as SCEV expressions for bounds or
parameters. This cached information needs to be updated or invalidated. To this
end, SPMUpdater allows passes to report when they've invalidated a Scop to the
PassManager, which will then flush and recompute all Scops. This in turn
invalidates all iterators, so references to Scops shouldn't be held.
Reviewers: grosser, Meinersbur, bollu
Reviewed By: grosser
Subscribers: llvm-commits, pollydev
Differential Revision: https://reviews.llvm.org/D36524
llvm-svn: 310551
This pass is useful to automatically convert a codebase that uses malloc/free
to use their managed memory counterparts.
Currently, rewrite malloc and free to the `polly_{malloc,free}Managed` variants.
A future patch will teach ManagedMemoryRewrite to rewrite global arrays
as pointers to globally allocated managed memory.
Differential Revision: https://reviews.llvm.org/D36513
llvm-svn: 310471
Codegen with -polly-parallel queried the unmapped MemoryAccess, but only
the MemoryKind after mapping is relevant for codegen.
This should fix various fails of the
perf-x86_64-penryn-O3-polly-parallel-fast buildbot.
llvm-svn: 310466
Previously, we used to compute this with `elementSizeInBits / 8`. This
would yield an element size of 0 when the array had element size < 8 in
bits.
To fix this, ask data layout what the size in bytes should be.
Differential Revision: https://reviews.llvm.org/D36459
llvm-svn: 310448
We introduce another level of alias metadata to distinguish the individual
non-aliasing accesses that have inter iteration alias-free base pointers
marked with "Inter iteration alias-free" mark nodes. To distinguish two
accesses, the comparison of raw pointers representing base pointers is used.
In case of, for example, ublas's prod function that implements GEMM, and
DeLiCM we can get accesses to same location represented by different raw
pointers. Consequently, we create different alias sets that can prevent
accesses from, for example, being sinked or hoisted.
To avoid the issue, we compare the corresponding SCEV information instead
of the corresponding raw pointers.
Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D35761
llvm-svn: 310380
To do this, we replicate what `CodeGeneration` does. We expose
`markNodeUnreachable` from `CodeGeneration` to `PPCGCodeGeneration`.
Differential Revision: https://reviews.llvm.org/D36457
llvm-svn: 310350
Summary:
This resolves some "instruction does not dominate use" errors, as we used to
prepare the arrays at the location of the first kernel, which not necessarily
dominated all other kernel calls.
Reviewers: Meinersbur, bollu, singam-sanjay
Subscribers: nemanjai, pollydev, llvm-commits, kbarton
Differential Revision: https://reviews.llvm.org/D36372
llvm-svn: 310196
Philip Pfaffe