[Mlir-commits] [mlir] [mlir][transform] Handle multiple library preloading passes. (PR #69705)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Fri Oct 20 03:28:14 PDT 2023
Ingo =?utf-8?q?Müller?= <ingomueller at google.com>
Message-ID:
In-Reply-To: <llvm/llvm-project/pull/69705/mlir at github.com>
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
Author: Ingo Müller (ingomueller-net)
<details>
<summary>Changes</summary>
This is a new attempt at #<!-- -->69320.
The transform dialect stores a "library module" that the preload pass
can populate. Until now, each pass registered an additional module by
simply pushing it to a vector; however, the interpreter only used the
first of them. This commit turns the registration into "loading", i.e.,
each newly added module gets merged into the existing one. This allows
the loading to be split into several passes, and using the library in
the interpreter now takes all of them into account. While this design
avoids repeated merging every time the library is accessed, it requires
that the implementation of merging modules lives in the
TransformDialect target (since it at the dialect depend on each
other).
This resolves https://github.com/llvm/llvm-project/issues/69111.
---
Patch is 26.80 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/69705.diff
6 Files Affected:
- (added) mlir/include/mlir/Dialect/Transform/IR/Utils.h (+40)
- (modified) mlir/include/mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h (+1-16)
- (modified) mlir/lib/Dialect/Transform/IR/CMakeLists.txt (+1)
- (added) mlir/lib/Dialect/Transform/IR/Utils.cpp (+240)
- (modified) mlir/lib/Dialect/Transform/Transforms/TransformInterpreterPassBase.cpp (+1)
- (modified) mlir/lib/Dialect/Transform/Transforms/TransformInterpreterUtils.cpp (+2-230)
``````````diff
diff --git a/mlir/include/mlir/Dialect/Transform/IR/Utils.h b/mlir/include/mlir/Dialect/Transform/IR/Utils.h
new file mode 100644
index 000000000000000..30eb9276604e721
--- /dev/null
+++ b/mlir/include/mlir/Dialect/Transform/IR/Utils.h
@@ -0,0 +1,40 @@
+//===- Utils.h - Utils related to the transform dialect ---------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_DIALECT_TRANSFORM_IR_UTILS_H
+#define MLIR_DIALECT_TRANSFORM_IR_UTILS_H
+
+namespace mlir {
+struct LogicalResult;
+class Operation;
+template <typename>
+class OwningOpRef;
+
+namespace transform {
+namespace detail {
+
+/// Merge all symbols from `other` into `target`. Both ops need to implement the
+/// `SymbolTable` trait. Operations are moved from `other`, i.e., `other` may be
+/// modified by this function and might not verify after the function returns.
+/// Upon merging, private symbols may be renamed in order to avoid collisions in
+/// the result. Public symbols may not collide, with the exception of
+/// instances of `SymbolOpInterface`, where collisions are allowed if at least
+/// one of the two is external, in which case the other op preserved (or any one
+/// of the two if both are external).
+// TODO: Reconsider cloning individual ops rather than forcing users of the
+// function to clone (or move) `other` in order to improve efficiency.
+// This might primarily make sense if we can also prune the symbols that
+// are merged to a subset (such as those that are actually used).
+LogicalResult mergeSymbolsInto(Operation *target,
+ OwningOpRef<Operation *> other);
+
+} // namespace detail
+} // namespace transform
+} // namespace mlir
+
+#endif // MLIR_DIALECT_TRANSFORM_IR_UTILS_H
\ No newline at end of file
diff --git a/mlir/include/mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h b/mlir/include/mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h
index 3fc02267f26e9da..faebe3304ea0489 100644
--- a/mlir/include/mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h
+++ b/mlir/include/mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h
@@ -32,7 +32,7 @@ namespace detail {
/// Each entry in `paths` may either be a regular file, in which case it ends up
/// in the result list, or a directory, in which case all (regular) `.mlir`
/// files in that directory are added. Any other file types lead to a failure.
-LogicalResult expandPathsToMLIRFiles(ArrayRef<std::string> &paths,
+LogicalResult expandPathsToMLIRFiles(ArrayRef<std::string> paths,
MLIRContext *context,
SmallVectorImpl<std::string> &fileNames);
@@ -68,21 +68,6 @@ ModuleOp getPreloadedTransformModule(MLIRContext *context);
TransformOpInterface findTransformEntryPoint(
Operation *root, ModuleOp module,
StringRef entryPoint = TransformDialect::kTransformEntryPointSymbolName);
-
-/// Merge all symbols from `other` into `target`. Both ops need to implement the
-/// `SymbolTable` trait. Operations are moved from `other`, i.e., `other` may be
-/// modified by this function and might not verify after the function returns.
-/// Upon merging, private symbols may be renamed in order to avoid collisions in
-/// the result. Public symbols may not collide, with the exception of
-/// instances of `SymbolOpInterface`, where collisions are allowed if at least
-/// one of the two is external, in which case the other op preserved (or any one
-/// of the two if both are external).
-// TODO: Reconsider cloning individual ops rather than forcing users of the
-// function to clone (or move) `other` in order to improve efficiency.
-// This might primarily make sense if we can also prune the symbols that
-// are merged to a subset (such as those that are actually used).
-LogicalResult mergeSymbolsInto(Operation *target,
- OwningOpRef<Operation *> other);
} // namespace detail
/// Standalone util to apply the named sequence `entryPoint` to the payload.
diff --git a/mlir/lib/Dialect/Transform/IR/CMakeLists.txt b/mlir/lib/Dialect/Transform/IR/CMakeLists.txt
index 7e8802bc755eedf..34083b2fd7aab37 100644
--- a/mlir/lib/Dialect/Transform/IR/CMakeLists.txt
+++ b/mlir/lib/Dialect/Transform/IR/CMakeLists.txt
@@ -5,6 +5,7 @@ add_mlir_dialect_library(MLIRTransformDialect
TransformInterfaces.cpp
TransformOps.cpp
TransformTypes.cpp
+ Utils.cpp
DEPENDS
MLIRMatchInterfacesIncGen
diff --git a/mlir/lib/Dialect/Transform/IR/Utils.cpp b/mlir/lib/Dialect/Transform/IR/Utils.cpp
new file mode 100644
index 000000000000000..0bfaf5ffade6da4
--- /dev/null
+++ b/mlir/lib/Dialect/Transform/IR/Utils.cpp
@@ -0,0 +1,240 @@
+//===- Utils.cpp - Utils related to the transform dialect -------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Transform/IR/Utils.h"
+#include "mlir/Dialect/Transform/IR/TransformDialect.h"
+#include "mlir/IR/Verifier.h"
+#include "mlir/Interfaces/FunctionInterfaces.h"
+#include "llvm/Support/Debug.h"
+
+using namespace mlir;
+
+#define DEBUG_TYPE "transform-dialect-utils"
+#define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ")
+
+/// Return whether `func1` can be merged into `func2`. For that to work
+/// `func1` has to be a declaration (aka has to be external) and `func2`
+/// either has to be a declaration as well, or it has to be public (otherwise,
+/// it wouldn't be visible by `func1`).
+static bool canMergeInto(FunctionOpInterface func1, FunctionOpInterface func2) {
+ return func1.isExternal() && (func2.isPublic() || func2.isExternal());
+}
+
+/// Merge `func1` into `func2`. The two ops must be inside the same parent op
+/// and mergable according to `canMergeInto`. The function erases `func1` such
+/// that only `func2` exists when the function returns.
+static LogicalResult mergeInto(FunctionOpInterface func1,
+ FunctionOpInterface func2) {
+ assert(canMergeInto(func1, func2));
+ assert(func1->getParentOp() == func2->getParentOp() &&
+ "expected func1 and func2 to be in the same parent op");
+
+ // Check that function signatures match.
+ if (func1.getFunctionType() != func2.getFunctionType()) {
+ return func1.emitError()
+ << "external definition has a mismatching signature ("
+ << func2.getFunctionType() << ")";
+ }
+
+ // Check and merge argument attributes.
+ MLIRContext *context = func1->getContext();
+ auto *td = context->getLoadedDialect<transform::TransformDialect>();
+ StringAttr consumedName = td->getConsumedAttrName();
+ StringAttr readOnlyName = td->getReadOnlyAttrName();
+ for (unsigned i = 0, e = func1.getNumArguments(); i < e; ++i) {
+ bool isExternalConsumed = func2.getArgAttr(i, consumedName) != nullptr;
+ bool isExternalReadonly = func2.getArgAttr(i, readOnlyName) != nullptr;
+ bool isConsumed = func1.getArgAttr(i, consumedName) != nullptr;
+ bool isReadonly = func1.getArgAttr(i, readOnlyName) != nullptr;
+ if (!isExternalConsumed && !isExternalReadonly) {
+ if (isConsumed)
+ func2.setArgAttr(i, consumedName, UnitAttr::get(context));
+ else if (isReadonly)
+ func2.setArgAttr(i, readOnlyName, UnitAttr::get(context));
+ continue;
+ }
+
+ if ((isExternalConsumed && !isConsumed) ||
+ (isExternalReadonly && !isReadonly)) {
+ return func1.emitError()
+ << "external definition has mismatching consumption "
+ "annotations for argument #"
+ << i;
+ }
+ }
+
+ // `func1` is the external one, so we can remove it.
+ assert(func1.isExternal());
+ func1->erase();
+
+ return success();
+}
+
+LogicalResult
+transform::detail::mergeSymbolsInto(Operation *target,
+ OwningOpRef<Operation *> other) {
+ assert(target->hasTrait<OpTrait::SymbolTable>() &&
+ "requires target to implement the 'SymbolTable' trait");
+ assert(other->hasTrait<OpTrait::SymbolTable>() &&
+ "requires target to implement the 'SymbolTable' trait");
+
+ SymbolTable targetSymbolTable(target);
+ SymbolTable otherSymbolTable(*other);
+
+ // Step 1:
+ //
+ // Rename private symbols in both ops in order to resolve conflicts that can
+ // be resolved that way.
+ LLVM_DEBUG(DBGS() << "renaming private symbols to resolve conflicts:\n");
+ // TODO: Do we *actually* need to test in both directions?
+ for (auto &&[symbolTable, otherSymbolTable] : llvm::zip(
+ SmallVector<SymbolTable *, 2>{&targetSymbolTable, &otherSymbolTable},
+ SmallVector<SymbolTable *, 2>{&otherSymbolTable,
+ &targetSymbolTable})) {
+ Operation *symbolTableOp = symbolTable->getOp();
+ for (Operation &op : symbolTableOp->getRegion(0).front()) {
+ auto symbolOp = dyn_cast<SymbolOpInterface>(op);
+ if (!symbolOp)
+ continue;
+ StringAttr name = symbolOp.getNameAttr();
+ LLVM_DEBUG(DBGS() << " found @" << name.getValue() << "\n");
+
+ // Check if there is a colliding op in the other module.
+ auto collidingOp =
+ cast_or_null<SymbolOpInterface>(otherSymbolTable->lookup(name));
+ if (!collidingOp)
+ continue;
+
+ LLVM_DEBUG(DBGS() << " collision found for @" << name.getValue());
+
+ // Collisions are fine if both opt are functions and can be merged.
+ if (auto funcOp = dyn_cast<FunctionOpInterface>(op),
+ collidingFuncOp =
+ dyn_cast<FunctionOpInterface>(collidingOp.getOperation());
+ funcOp && collidingFuncOp) {
+ if (canMergeInto(funcOp, collidingFuncOp) ||
+ canMergeInto(collidingFuncOp, funcOp)) {
+ LLVM_DEBUG(llvm::dbgs() << " but both ops are functions and "
+ "will be merged\n");
+ continue;
+ }
+
+ // If they can't be merged, proceed like any other collision.
+ LLVM_DEBUG(llvm::dbgs() << " and both ops are function definitions");
+ }
+
+ // Collision can be resolved by renaming if one of the ops is private.
+ auto renameToUnique =
+ [&](SymbolOpInterface op, SymbolOpInterface otherOp,
+ SymbolTable &symbolTable,
+ SymbolTable &otherSymbolTable) -> LogicalResult {
+ LLVM_DEBUG(llvm::dbgs() << ", renaming\n");
+ FailureOr<StringAttr> maybeNewName =
+ symbolTable.renameToUnique(op, {&otherSymbolTable});
+ if (failed(maybeNewName)) {
+ InFlightDiagnostic diag = op->emitError("failed to rename symbol");
+ diag.attachNote(otherOp->getLoc())
+ << "attempted renaming due to collision with this op";
+ return diag;
+ }
+ LLVM_DEBUG(DBGS() << " renamed to @" << maybeNewName->getValue()
+ << "\n");
+ return success();
+ };
+
+ if (symbolOp.isPrivate()) {
+ if (failed(renameToUnique(symbolOp, collidingOp, *symbolTable,
+ *otherSymbolTable)))
+ return failure();
+ continue;
+ }
+ if (collidingOp.isPrivate()) {
+ if (failed(renameToUnique(collidingOp, symbolOp, *otherSymbolTable,
+ *symbolTable)))
+ return failure();
+ continue;
+ }
+ LLVM_DEBUG(llvm::dbgs() << ", emitting error\n");
+ InFlightDiagnostic diag = symbolOp.emitError()
+ << "doubly defined symbol @" << name.getValue();
+ diag.attachNote(collidingOp->getLoc()) << "previously defined here";
+ return diag;
+ }
+ }
+
+ // TODO: This duplicates pass infrastructure. We should split this pass into
+ // several and let the pass infrastructure do the verification.
+ for (auto *op : SmallVector<Operation *>{target, *other}) {
+ if (failed(mlir::verify(op)))
+ return op->emitError() << "failed to verify input op after renaming";
+ }
+
+ // Step 2:
+ //
+ // Move all ops from `other` into target and merge public symbols.
+ LLVM_DEBUG(DBGS() << "moving all symbols into target\n");
+ {
+ SmallVector<SymbolOpInterface> opsToMove;
+ for (Operation &op : other->getRegion(0).front()) {
+ if (auto symbol = dyn_cast<SymbolOpInterface>(op))
+ opsToMove.push_back(symbol);
+ }
+
+ for (SymbolOpInterface op : opsToMove) {
+ // Remember potentially colliding op in the target module.
+ auto collidingOp = cast_or_null<SymbolOpInterface>(
+ targetSymbolTable.lookup(op.getNameAttr()));
+
+ // Move op even if we get a collision.
+ LLVM_DEBUG(DBGS() << " moving @" << op.getName());
+ op->moveBefore(&target->getRegion(0).front(),
+ target->getRegion(0).front().end());
+
+ // If there is no collision, we are done.
+ if (!collidingOp) {
+ LLVM_DEBUG(llvm::dbgs() << " without collision\n");
+ continue;
+ }
+
+ // The two colliding ops must both be functions because we have already
+ // emitted errors otherwise earlier.
+ auto funcOp = cast<FunctionOpInterface>(op.getOperation());
+ auto collidingFuncOp =
+ cast<FunctionOpInterface>(collidingOp.getOperation());
+
+ // Both ops are in the target module now and can be treated
+ // symmetrically, so w.l.o.g. we can reduce to merging `funcOp` into
+ // `collidingFuncOp`.
+ if (!canMergeInto(funcOp, collidingFuncOp)) {
+ std::swap(funcOp, collidingFuncOp);
+ }
+ assert(canMergeInto(funcOp, collidingFuncOp));
+
+ LLVM_DEBUG(llvm::dbgs() << " with collision, trying to keep op at "
+ << collidingFuncOp.getLoc() << ":\n"
+ << collidingFuncOp << "\n");
+
+ // Update symbol table. This works with or without the previous `swap`.
+ targetSymbolTable.remove(funcOp);
+ targetSymbolTable.insert(collidingFuncOp);
+ assert(targetSymbolTable.lookup(funcOp.getName()) == collidingFuncOp);
+
+ // Do the actual merging.
+ if (failed(mergeInto(funcOp, collidingFuncOp))) {
+ return failure();
+ }
+ }
+ }
+
+ if (failed(mlir::verify(target)))
+ return target->emitError()
+ << "failed to verify target op after merging symbols";
+
+ LLVM_DEBUG(DBGS() << "done merging ops\n");
+ return success();
+}
diff --git a/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterPassBase.cpp b/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterPassBase.cpp
index 538c81fe39fddb2..4d9d1461cd759ea 100644
--- a/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterPassBase.cpp
+++ b/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterPassBase.cpp
@@ -15,6 +15,7 @@
#include "mlir/Dialect/Transform/IR/TransformDialect.h"
#include "mlir/Dialect/Transform/IR/TransformInterfaces.h"
#include "mlir/Dialect/Transform/IR/TransformOps.h"
+#include "mlir/Dialect/Transform/IR/Utils.h"
#include "mlir/Dialect/Transform/Transforms/TransformInterpreterUtils.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/Verifier.h"
diff --git a/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterUtils.cpp b/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterUtils.cpp
index e6d692072267c1f..983057c30a80102 100644
--- a/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterUtils.cpp
+++ b/mlir/lib/Dialect/Transform/Transforms/TransformInterpreterUtils.cpp
@@ -14,6 +14,7 @@
#include "mlir/Dialect/Transform/IR/TransformDialect.h"
#include "mlir/Dialect/Transform/IR/TransformInterfaces.h"
#include "mlir/Dialect/Transform/IR/TransformOps.h"
+#include "mlir/Dialect/Transform/IR/Utils.h"
#include "mlir/IR/BuiltinOps.h"
#include "mlir/IR/Verifier.h"
#include "mlir/IR/Visitors.h"
@@ -38,7 +39,7 @@ using namespace mlir;
/// in the result list, or a directory, in which case all (regular) `.mlir`
/// files in that directory are added. Any other file types lead to a failure.
LogicalResult transform::detail::expandPathsToMLIRFiles(
- ArrayRef<std::string> &paths, MLIRContext *context,
+ ArrayRef<std::string> paths, MLIRContext *context,
SmallVectorImpl<std::string> &fileNames) {
for (const std::string &path : paths) {
auto loc = FileLineColLoc::get(context, path, 0, 0);
@@ -89,12 +90,6 @@ LogicalResult transform::detail::expandPathsToMLIRFiles(
LogicalResult transform::detail::parseTransformModuleFromFile(
MLIRContext *context, llvm::StringRef transformFileName,
OwningOpRef<ModuleOp> &transformModule) {
- if (transformFileName.empty()) {
- LLVM_DEBUG(
- DBGS() << "no transform file name specified, assuming the transform "
- "module is embedded in the IR next to the top-level\n");
- return success();
- }
// Parse transformFileName content into a ModuleOp.
std::string errorMessage;
auto memoryBuffer = mlir::openInputFile(transformFileName, &errorMessage);
@@ -172,7 +167,6 @@ LogicalResult transform::detail::assembleTransformLibraryFromPaths(
{
mergedParsedLibraries.get()->setAttr("transform.with_named_sequence",
UnitAttr::get(context));
- IRRewriter rewriter(context);
// TODO: extend `mergeSymbolsInto` to support multiple `other` modules.
for (OwningOpRef<ModuleOp> &parsedLibrary : parsedLibraries) {
if (failed(transform::detail::mergeSymbolsInto(
@@ -186,228 +180,6 @@ LogicalResult transform::detail::assembleTransformLibraryFromPaths(
return success();
}
-/// Return whether `func1` can be merged into `func2`. For that to work
-/// `func1` has to be a declaration (aka has to be external) and `func2`
-/// either has to be a declaration as well, or it has to be public (otherwise,
-/// it wouldn't be visible by `func1`).
-static bool canMergeInto(FunctionOpInterface func1, FunctionOpInterface func2) {
- return func1.isExternal() && (func2.isPublic() || func2.isExternal());
-}
-
-/// Merge `func1` into `func2`. The two ops must be inside the same parent op
-/// and mergable according to `canMergeInto`. The function erases `func1` such
-/// that only `func2` exists when the function returns.
-static LogicalResult mergeInto(FunctionOpInterface func1,
- FunctionOpInterface func2) {
- assert(canMergeInto(func1, func2));
- assert(func1->getParentOp() == func2->getParentOp() &&
- "expected func1 and func2 to be in the same parent op");
-
- // Check that function signatures match.
- if (func1.getFunctionType() != func2.getFunctionType()) {
- return func1.emitError()
- << "external definition has a mismatching signature ("
- << func2.getFunctionType() << ")";
- }
-
- // Check and merge argument attributes.
- MLIRContext *context = func1->getContext();
- auto *td = context->getLoadedDialect<transform::TransformDialect>();
- StringAttr consumedName = td->getConsumedAttrName();
- StringAttr readOnlyName = td->getReadOnlyAttrName();
- for (unsigned i = 0, e = func1.getNumArguments(); i < e; ++i) {
- bool isExternalConsumed = func2.getArgAttr(i, consumedName) != nullptr;
- bool isExternalReadonly = func2.getArgAttr(i, readOnlyName) != nullptr;
- bool isConsumed = func1.getArgAttr(i, consumedName) != nullptr;
- bool isReadonly = func1.getArgAttr(i, readOnlyName) != nullptr;
- if (!isExternalConsumed && !isExternalReadonly) {
- if (isConsumed)
- func2.setArgAttr(i, consumedName, UnitAttr::get(context));
- else if (isReadonly)
- func2.setArgAttr(i, readOnlyName, UnitAttr::get(context));
- continue;
- }
-
- if ((isExternalConsumed && !isConsumed) ||
- (isExternalReadonly && !isReadonly)) {
- return func1.emitError()
- << "external definition has mismatching consumption "
- "annotations for argument #"
- << i;
- }
- }
-
- // `func1` is the external one, s...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/69705
More information about the Mlir-commits
mailing list