This revision allows for attaching "debug labels" to patterns, and provides to FrozenRewritePatternSet for filtering patterns based on these labels (in addition to the debug name of the pattern). This will greatly simplify the ability to write tests targeted towards specific patterns (in cases where many patterns may interact), will also simplify debugging pattern application by observing how application changes when enabling/disabling specific patterns.
To enable better reuse of pattern rewrite options between passes, this revision also adds a new PassUtil.td file to the Rewrite/ library that will allow for passes to easily hook into a common interface for pattern debugging. Two options are used to seed this utility, `disable-patterns` and `enable-patterns`, which are used to enable the filtering behavior indicated above.
Differential Revision: https://reviews.llvm.org/D102441
I backed this off to make the previous patch easier to wrangle, but now
this is an efficient query and it is better to not replace it in CSE.
Differential Revision: https://reviews.llvm.org/D103494
The previous impl densely scanned the entire region starting with an op
when dominators were created, creating a DominatorTree for every region.
This is extremely expensive up front -- particularly for clients like
Linalg/Transforms/Fusion.cpp that construct DominanceInfo for a single
query. It is also extremely memory wasteful for IRs that use single
block regions commonly (e.g. affine.for) because it's making a
dominator tree for a region that has trivial dominance. The
implementation also had numerous unnecessary minor efficiencies, e.g.
doing multiple walks of the region tree or tryGetBlocksInSameRegion
building a DenseMap that it didn't need.
This patch switches to an approach where [Post]DominanceInfo is free
to construct, and which lazily constructs DominatorTree's for any
multiblock regions that it needs. This avoids the up-front cost
entirely, making its runtime proportional to the complexity of the
region tree instead of # ops in a region. This also avoids the memory
and time cost of creating DominatorTree's for single block regions.
Finally this rewrites the implementation for simplicity and to avoids
the constant factor problems the old implementation had.
Differential Revision: https://reviews.llvm.org/D103384
CSE is the only client of this API, refactor it a bit to pull the query
internally to make changes to DominanceInfo a bit easier. This commit
also improves comments a bit.
* Add `hasCanonicalizer` option to Dialect.
* Initialize canonicalizer with dialect-wide canonicalization patterns.
* Add test case to TestDialect.
Dialect-wide canonicalization patterns are useful if a canonicalization pattern does not conceptually associate with any single operation, i.e., it should not be registered as part of an operation's `getCanonicalizationPatterns` function. E.g., this is the case for canonicalization patterns that match an op interface.
Differential Revision: https://reviews.llvm.org/D103226
This allows C++ clients of the Canonicalize pass to specify their own
Config option struct to control how Canonicalize works, increasing reusability.
This also allows controlling these settings for the default Canonicalize pass
using command line options. This is useful for testing and for playing with
things on the command line.
Differential Revision: https://reviews.llvm.org/D103069
This exposes the iterations and top-down processing as flags, and also
allows controlling whether region simplification is desirable for a client.
This allows deleting some duplicated entrypoints to
applyPatternsAndFoldGreedily.
This also deletes the Constant Preprocessing pass, which isn't worth it
on balance.
All defaults are all kept the same, so no one should see a behavior change.
Differential Revision: https://reviews.llvm.org/D102988
Steps for normalizing dynamic memrefs for tiled layout map
1. Check if original map is tiled layout. Only tiled layout is supported.
2. Create normalized memrefType. Dimensions that include dynamic dimensions
in the map output will be dynamic dimensions.
3. Create new maps to calculate each dimension size of new memref.
In tiled layout, the dimension size can be calculated by replacing
"floordiv <tile size>" with "ceildiv <tile size>" and
"mod <tile size>" with "<tile size>".
4. Create AffineApplyOp to apply the new maps. The output of AffineApplyOp is
dynamicSizes for new AllocOp.
5. Add the new dynamic sizes in new AllocOp.
This patch also set MemRefsNormalizable trant in CastOp and DimOp since
they used with dynamic memrefs.
Reviewed By: bondhugula
Differential Revision: https://reviews.llvm.org/D97655
The previous implementation did not handle casting behavior properly and
did not consider aliases.
Differential Revision: https://reviews.llvm.org/D102785
During affine loop fusion, create private memrefs for escaping memrefs
too under the conditions that:
-- the source is not removed after fusion, and
-- the destination does not write to the memref.
This creates more fusion opportunities as illustrated in the test case.
Reviewed By: bondhugula, ayzhuang
Differential Revision: https://reviews.llvm.org/D102604
Splitting the memref dialect lead to an introduction of several dependencies
to avoid compilation issues. The canonicalize pass also depends on the
memref dialect, but it shouldn't. This patch resolves the dependencies
and the unintuitive includes are removed. However, the dependency moves
to the constructor of the std dialect.
Differential Revision: https://reviews.llvm.org/D102060
The current design uses a unique entry for each argument/result attribute, with the name of the entry being something like "arg0". This provides for a somewhat sparse design, but ends up being much more expensive (from a runtime perspective) in-practice. The design requires building a string every time we lookup the dictionary for a specific arg/result, and also requires N attribute lookups when collecting all of the arg/result attribute dictionaries.
This revision restructures the design to instead have an ArrayAttr that contains all of the attribute dictionaries for arguments and another for results. This design reduces the number of attribute name lookups to 1, and allows for O(1) lookup for individual element dictionaries. The major downside is that we can end up with larger memory usage, as the ArrayAttr contains an entry for each element even if that element has no attributes. If the memory usage becomes too problematic, we can experiment with a more sparse structure that still provides a lot of the wins in this revision.
This dropped the compilation time of a somewhat large TensorFlow model from ~650 seconds to ~400 seconds.
Differential Revision: https://reviews.llvm.org/D102035
This it to make more clear the difference between this and
an AliasAnalysis.
For example, given a sequence of subviews that create values
A -> B -> C -> d:
BufferViewFlowAnalysis::resolve(B) => {B, C, D}
AliasAnalysis::resolve(B) => {A, B, C, D}
Differential Revision: https://reviews.llvm.org/D100838
This revision takes the forward value propagation engine in SCCP and refactors it into a more generalized forward dataflow analysis framework. This framework allows for propagating information about values across the various control flow constructs in MLIR, and removes the need for users to reinvent the traversal (often not as completely). There are a few aspects of the traversal, that were conservative for SCCP, that should be relaxed to support the needs of different value analyses. To keep this revision simple, these conservative behaviors will be left in (Note that this won't produce an incorrect result, but may produce more conservative results than necessary in certain edge cases. e.g. region entry arguments for non-region branch interface operations). The framework also only focuses on computing lattices for values, given the SCCP origins, but this is something to relax as needed in the future.
Given that this logic is already in SCCP, a majority of this commit is NFC. The more interesting parts are the interface glue that clients interact with.
Differential Revision: https://reviews.llvm.org/D100915
This patch introduces the neccessary infrastructure changes to implement
cost-modelling for detensoring. In particular, it introduces the
following changes:
- An extension to the dialect conversion framework to selectively
convert sub-set of non-entry BB arguments.
- An extension to branch conversion pattern to selectively convert
sub-set of a branche's operands.
- An interface for detensoring cost-modelling.
- 2 simple implementations of 2 different cost models.
This sets the stage to explose cost-modelling for detessoring in an
easier way. We still need to come up with better cost models.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D99945
`ConversionPatternRewriter::applySignatureConversion` did not have a way
to apply a signature conversion that involved materializations.
Differential Revision: https://reviews.llvm.org/D99782
Setting the builder from a block is looking up for a parent operation
to get a context, instead by setting up the builder with an explicit
context we can support invoking this helper in absence of a parent
operation.
Fixes a bug in affine fusion pipeline where an incorrect slice is computed.
After the slice computation is done, original domain of the the source is
compared with the new domain that will result if the fusion succeeds. If the
new domain must be a subset of the original domain for the slice to be
valid. If the slice computed is incorrect, fusion based on such a slice is
avoided.
Relevant test cases are added/edited.
Fixes https://bugs.llvm.org/show_bug.cgi?id=49203
Differential Revision: https://reviews.llvm.org/D98239
If an operation has been inserted as a key in to the known values
hashtable, then it can not be modified in a way which changes its hash.
This change avoids modifying the operands of any previously recorded
operation, which prevents their hash from changing.
In an SSACFG region, it is impossible to visit an operation before
visiting its operands, so this is not a problem. This situation can only
happen in regions without strict dominance, such as graph regions.
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D99486
This allows for the conversion to match `A(B()) -> C()` with a pattern matching
`A` and marking `B` for deletion.
Also add better assertions when an operation is erased while still having uses.
Differential Revision: https://reviews.llvm.org/D99442
Add a new clone operation to the memref dialect. This operation implicitly
copies data from a source buffer to a new buffer. In contrast to the linalg.copy
operation, this operation does not accept a target buffer as an argument.
Instead, this operation performs a conceptual allocation which does not need to
be performed manually.
Furthermore, this operation resolves the dependency from the linalg-dialect
in the BufferDeallocation pass. In addition, we also extended the canonicalization
patterns to fold clone operations. The copy removal pass has been removed.
Differential Revision: https://reviews.llvm.org/D99172
This reverts commit 361b7d125b by Chris
Lattner <clattner@nondot.org> dated Fri Mar 19 21:22:15 2021 -0700.
The change to the greedy rewriter driver picking a different order was
made without adequate analysis of the trade-offs and experimentation. A
change like this has far reaching consequences on transformation
pipelines, and a major impact upstream and downstream. For eg., one
can’t be sure that it doesn’t slow down a large number of cases by small
amounts or create other issues. More discussion here:
https://llvm.discourse.group/t/speeding-up-canonicalize/3015/25
Reverting this so that improvements to the traversal order can be made
on a clean slate, in bigger steps, and higher bar.
Differential Revision: https://reviews.llvm.org/D99329
In particular for Graph Regions, the terminator needs is just a
historical artifact of the generalization of MLIR from CFG region.
Operations like Module don't need a terminator, and before Module
migrated to be an operation with region there wasn't any needed.
To validate the feature, the ModuleOp is migrated to use this trait and
the ModuleTerminator operation is deleted.
This patch is likely to break clients, if you're in this case:
- you may iterate on a ModuleOp with `getBody()->without_terminator()`,
the solution is simple: just remove the ->without_terminator!
- you created a builder with `Builder::atBlockTerminator(module_body)`,
just use `Builder::atBlockEnd(module_body)` instead.
- you were handling ModuleTerminator: it isn't needed anymore.
- for generic code, a `Block::mayNotHaveTerminator()` may be used.
Differential Revision: https://reviews.llvm.org/D98468
To match an interface or trait, users currently have to use the `MatchAny` tag. This tag can be quite problematic for compile time for things like the canonicalizer, as the `MatchAny` patterns may get applied to *every* operation. This revision adds better support by bucketing interface/trait patterns based on which registered operations have them registered. This means that moving forward we will only attempt to match these patterns to operations that have this interface registered. Two simplify defining patterns that match traits and interfaces, two new utility classes have been added: OpTraitRewritePattern and OpInterfaceRewritePattern.
Differential Revision: https://reviews.llvm.org/D98986
This nicely aligns the naming with RewritePatternSet. This type isn't
as widely used, but we keep a using declaration in to help with
downstream consumption of this change.
Differential Revision: https://reviews.llvm.org/D99131
This doesn't change APIs, this just cleans up the many in-tree uses of these
names to use the new preferred names. We'll keep the old names around for a
couple weeks to help transitions.
Differential Revision: https://reviews.llvm.org/D99127
GreedyPatternRewriteDriver was changed from bottom-up traversal to top-down traversal. Not all passes work yet with that change for traversal order. To give some time for fixing, add an option to allow to switch back to bottom-up traversal. Use this option in FusionOfTensorOpsPass which fails otherwise.
Differential Revision: https://reviews.llvm.org/D99059
This updates the codebase to pass the context when creating an instance of
OwningRewritePatternList, and starts removing extraneous MLIRContext
parameters. There are many many more to be removed.
Differential Revision: https://reviews.llvm.org/D99028
This reapplies b5d9a3c / https://reviews.llvm.org/D98609 with a one line fix in
processExistingConstants to skip() when erasing a constant we've already seen.
Original commit message:
1) Change the canonicalizer to walk the function in top-down order instead of
bottom-up order. This composes well with the "top down" nature of constant
folding and simplification, reducing iterations and re-evaluation of ops in
simple cases.
2) Explicitly enter existing constants into the OperationFolder table before
canonicalizing. Previously we would "constant fold" them and rematerialize
them, wastefully recreating a bunch fo constants, which lead to pointless
memory traffic.
Both changes together provide a 33% speedup for canonicalize on some mid-size
CIRCT examples.
One artifact of this change is that the constants generated in normal pattern
application get inserted at the top of the function as the patterns are applied.
Because of this, we get "inverted" constants more often, which is an aethetic
change to the IR but does permute some testcases.
Differential Revision: https://reviews.llvm.org/D99006
This allows for notifying callers when operations/blocks get erased, which is especially useful for the greedy pattern driver. The current greedy pattern driver "throws away" all information on constants in the operation folder because it doesn't know if they get erased or not. By passing in RewriterBase, we can directly track this and prevent the need for the pattern driver to rediscover all of the existing constants. In some situations this cuts the compile time of the canonicalizer in half.
Differential Revision: https://reviews.llvm.org/D98755
When deleting operations in DCE, the algorithm uses a post-order walk of
the IR to ensure that value uses were erased before value defs. Graph
regions do not have the same structural invariants as SSA CFG, and this
post order walk could delete value defs before uses. This problem is
guaranteed to occur when there is a cycle in the use-def graph.
This change stops DCE from visiting the operations and blocks in any
meaningful order. Instead, we rely on explicitly dropping all uses of a
value before deleting it.
Reviewed By: mehdi_amini, rriddle
Differential Revision: https://reviews.llvm.org/D98919
This reverts commit b5d9a3c923.
The commit introduced a memory error in canonicalization/operation
walking that is exposed when compiled with ASAN. It leads to crashes in
some "release" configurations.
We know that all ConstantLike operations have one result and no operands,
so check this first before doing the trait check. This change speeds up
Canonicalize on a CIRCT testcase by ~5%.
Differential Revision: https://reviews.llvm.org/D98615
Two changes:
1) Change the canonicalizer to walk the function in top-down order instead of
bottom-up order. This composes well with the "top down" nature of constant
folding and simplification, reducing iterations and re-evaluation of ops in
simple cases.
2) Explicitly enter existing constants into the OperationFolder table before
canonicalizing. Previously we would "constant fold" them and rematerialize
them, wastefully recreating a bunch fo constants, which lead to pointless
memory traffic.
Both changes together provide a 33% speedup for canonicalize on some mid-size
CIRCT examples.
One artifact of this change is that the constants generated in normal pattern
application get inserted at the top of the function as the patterns are applied.
Because of this, we get "inverted" constants more often, which is an aethetic
change to the IR but does permute some testcases.
Differential Revision: https://reviews.llvm.org/D98609
The current implementation has some inefficiencies that become noticeable when running on large modules. This revision optimizes the code, and updates some out-dated idioms with newer utilities. The main components of this optimization include:
* Add an overload of Block::eraseArguments that allows for O(N) erasure of disjoint arguments.
* Don't process entry block arguments given that we don't erase them at this point.
* Don't track individual operation results, given that we don't erase them. We can just track the parent operation.
Differential Revision: https://reviews.llvm.org/D98309
This makes it easy to compose the distribution computation with
other affine computations.
Reviewed By: mravishankar
Differential Revision: https://reviews.llvm.org/D98171
The current implementation of Value involves a pointer int pair with several different kinds of owners, i.e. BlockArgumentImpl*, Operation *, TrailingOpResult*. This design arose from the desire to save memory overhead for operations that have a very small number of results (generally 0-2). There are, unfortunately, many problematic aspects of the current implementation that make Values difficult to work with or just inefficient.
Operation result types are stored as a separate array on the Operation. This is very inefficient for many reasons: we use TupleType for multiple results, which can lead to huge amounts of memory usage if multi-result operations change types frequently(they do). It also means that simple methods like Value::getType/Value::setType now require complex logic to get to the desired type.
Value only has one pointer bit free, severely limiting the ability to use it in things like PointerUnion/PointerIntPair. Given that we store the kind of a Value along with the "owner" pointer, we only leave one bit free for users of Value. This creates situations where we end up nesting PointerUnions to be able to use Value in one.
As noted above, most of the methods in Value need to branch on at least 3 different cases which is both inefficient, possibly error prone, and verbose. The current storage of results also creates problems for utilities like ValueRange/TypeRange, which want to efficiently store base pointers to ranges (of which Operation* isn't really useful as one).
This revision greatly simplifies the implementation of Value by the introduction of a new ValueImpl class. This class contains all of the state shared between all of the various derived value classes; i.e. the use list, the type, and the kind. This shared implementation class provides several large benefits:
* Most of the methods on value are now branchless, and often one-liners.
* The "kind" of the value is now stored in ValueImpl instead of Value
This frees up all of Value's pointer bits, allowing for users to take full advantage of PointerUnion/PointerIntPair/etc. It also allows for storing more operation results as "inline", 6 now instead of 2, freeing up 1 word per new inline result.
* Operation result types are now stored in the result, instead of a side array
This drops the size of zero-result operations by 1 word. It also removes the memory crushing use of TupleType for operations results (which could lead up to hundreds of megabytes of "dead" TupleTypes in the context). This also allowed restructured ValueRange, making it simpler and one word smaller.
This revision does come with two conceptual downsides:
* Operation::getResultTypes no longer returns an ArrayRef<Type>
This conceptually makes some usages slower, as the iterator increment is slightly more complex.
* OpResult::getOwner is slightly more expensive, as it now requires a little bit of arithmetic
From profiling, neither of the conceptual downsides have resulted in any perceivable hit to performance. Given the advantages of the new design, most compiles are slightly faster.
Differential Revision: https://reviews.llvm.org/D97804
This patch continues detensorizing implementation by detensoring
internal control flow in functions.
In order to detensorize functions, all the non-entry block's arguments
are detensored and branches between such blocks are properly updated to
reflect the detensored types as well. Function entry block (signature)
is left intact.
This continues work towards handling github/google/iree#1159.
Reviewed By: silvas
Differential Revision: https://reviews.llvm.org/D97148
Just a pure method renaming.
It is a preparation step for replacing "memory space as raw integer"
with more generic "memory space as attribute", which will be done in
separate commit.
The `MemRefType::getMemorySpace` method will return `Attribute` and
become the main API, while `getMemorySpaceAsInt` will be declared as
deprecated and will be replaced in all in-tree dialects (also in separate
commits).
Reviewed By: mehdi_amini, rriddle
Differential Revision: https://reviews.llvm.org/D97476
Amy Zhuang