Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[mlir][Transforms] Dialect conversion: Make materializations optional #107109

Merged

Conversation

matthias-springer
Copy link
Member

This commit makes source/target/argument materializations (via the TypeConverter API) optional.

By default (ConversionConfig::buildMaterializations = true), the dialect conversion infrastructure tries to legalize all unresolved materializations right after the main transformation process has succeeded. If at least one unresolved materialization fails to resolve, the dialect conversion fails. (With an error message such as failed to legalize unresolved materialization ....) Automatic materializations through the TypeConverter API can now be deactivated. In that case, every unresolved materialization will show up as a builtin.unrealized_conversion_cast op in the output IR.

There used to be a complex and error-prone analysis in the dialect conversion that predicted the future uses of unresolved materializations. Based on that logic, some casts (that were deemed to unnecessary) were folded. This analysis was needed because folding happened at a point of time when some IR changes (e.g., op replacements) had not materialized yet.

This commit removes that analysis. Any folding of cast ops now happens after all other IR changes have been materialized and the uses can directly be queried from the IR. This simplifies the analysis significantly. And certain helper data structures such as inverseMapping are no longer needed for the analysis. The folding itself is done by reconcileUnrealizedCasts (which also exists as a standalone pass).

After casts have been folded, the remaining casts are materialized through the TypeConverter, as usual. This last step can be deactivated in the ConversionConfig.

ConversionConfig::buildMaterializations = false can be used to debug error messages such as failed to legalize unresolved materialization .... (It is also useful in case automatic materializations are not needed.) The materializations that failed to resolve can then be seen as builtin.unrealized_conversion_cast ops in the resulting IR. (This is better than running with -debug, because -debug shows IR where some IR changes have not been materialized yet.)

Note: This is a reupload of #104668, but with correct handling of cyclic unrealized_conversion_casts that may be generated by the dialect conversion.

@llvmbot llvmbot added mlir:core MLIR Core Infrastructure mlir:gpu mlir mlir:bufferization Bufferization infrastructure labels Sep 3, 2024
@llvmbot
Copy link
Member

llvmbot commented Sep 3, 2024

@llvm/pr-subscribers-mlir-gpu
@llvm/pr-subscribers-mlir-core
@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-bufferization

Author: Matthias Springer (matthias-springer)

Changes

This commit makes source/target/argument materializations (via the TypeConverter API) optional.

By default (ConversionConfig::buildMaterializations = true), the dialect conversion infrastructure tries to legalize all unresolved materializations right after the main transformation process has succeeded. If at least one unresolved materialization fails to resolve, the dialect conversion fails. (With an error message such as failed to legalize unresolved materialization ....) Automatic materializations through the TypeConverter API can now be deactivated. In that case, every unresolved materialization will show up as a builtin.unrealized_conversion_cast op in the output IR.

There used to be a complex and error-prone analysis in the dialect conversion that predicted the future uses of unresolved materializations. Based on that logic, some casts (that were deemed to unnecessary) were folded. This analysis was needed because folding happened at a point of time when some IR changes (e.g., op replacements) had not materialized yet.

This commit removes that analysis. Any folding of cast ops now happens after all other IR changes have been materialized and the uses can directly be queried from the IR. This simplifies the analysis significantly. And certain helper data structures such as inverseMapping are no longer needed for the analysis. The folding itself is done by reconcileUnrealizedCasts (which also exists as a standalone pass).

After casts have been folded, the remaining casts are materialized through the TypeConverter, as usual. This last step can be deactivated in the ConversionConfig.

ConversionConfig::buildMaterializations = false can be used to debug error messages such as failed to legalize unresolved materialization .... (It is also useful in case automatic materializations are not needed.) The materializations that failed to resolve can then be seen as builtin.unrealized_conversion_cast ops in the resulting IR. (This is better than running with -debug, because -debug shows IR where some IR changes have not been materialized yet.)

Note: This is a reupload of #104668, but with correct handling of cyclic unrealized_conversion_casts that may be generated by the dialect conversion.


Patch is 31.92 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/107109.diff

5 Files Affected:

  • (modified) mlir/include/mlir/Transforms/DialectConversion.h (+11)
  • (modified) mlir/lib/Transforms/Utils/DialectConversion.cpp (+99-294)
  • (modified) mlir/test/Conversion/NVGPUToNVVM/nvgpu-to-nvvm.mlir (+2-3)
  • (modified) mlir/test/Dialect/Bufferization/Transforms/finalizing-bufferize.mlir (+1)
  • (modified) mlir/test/Transforms/test-legalize-type-conversion.mlir (+5-1)
diff --git a/mlir/include/mlir/Transforms/DialectConversion.h b/mlir/include/mlir/Transforms/DialectConversion.h
index 60113bdef16a23..5f680e8eca7559 100644
--- a/mlir/include/mlir/Transforms/DialectConversion.h
+++ b/mlir/include/mlir/Transforms/DialectConversion.h
@@ -1124,6 +1124,17 @@ struct ConversionConfig {
   // already been modified) and iterators into past IR state cannot be
   // represented at the moment.
   RewriterBase::Listener *listener = nullptr;
+
+  /// If set to "true", the dialect conversion attempts to build source/target/
+  /// argument materializations through the type converter API in lieu of
+  /// builtin.unrealized_conversion_cast ops. The conversion process fails if
+  /// at least one materialization could not be built.
+  ///
+  /// If set to "false", the dialect conversion does not does not build any
+  /// custom materializations and instead inserts
+  /// builtin.unrealized_conversion_cast ops to ensure that the resulting IR
+  /// is valid.
+  bool buildMaterializations = true;
 };
 
 //===----------------------------------------------------------------------===//
diff --git a/mlir/lib/Transforms/Utils/DialectConversion.cpp b/mlir/lib/Transforms/Utils/DialectConversion.cpp
index b23fb97959ed67..cc9c9495e5155c 100644
--- a/mlir/lib/Transforms/Utils/DialectConversion.cpp
+++ b/mlir/lib/Transforms/Utils/DialectConversion.cpp
@@ -702,14 +702,12 @@ class UnresolvedMaterializationRewrite : public OperationRewrite {
     return rewrite->getKind() == Kind::UnresolvedMaterialization;
   }
 
+  void rollback() override;
+
   UnrealizedConversionCastOp getOperation() const {
     return cast<UnrealizedConversionCastOp>(op);
   }
 
-  void rollback() override;
-
-  void cleanup(RewriterBase &rewriter) override;
-
   /// Return the type converter of this materialization (which may be null).
   const TypeConverter *getConverter() const {
     return converterAndKind.getPointer();
@@ -766,7 +764,7 @@ namespace detail {
 struct ConversionPatternRewriterImpl : public RewriterBase::Listener {
   explicit ConversionPatternRewriterImpl(MLIRContext *ctx,
                                          const ConversionConfig &config)
-      : context(ctx), config(config) {}
+      : context(ctx), eraseRewriter(ctx), config(config) {}
 
   //===--------------------------------------------------------------------===//
   // State Management
@@ -834,6 +832,7 @@ struct ConversionPatternRewriterImpl : public RewriterBase::Listener {
   //===--------------------------------------------------------------------===//
   // Materializations
   //===--------------------------------------------------------------------===//
+
   /// Build an unresolved materialization operation given an output type and set
   /// of input operands.
   Value buildUnresolvedMaterialization(MaterializationKind kind,
@@ -882,7 +881,7 @@ struct ConversionPatternRewriterImpl : public RewriterBase::Listener {
 
     /// Erase the given op (unless it was already erased).
     void eraseOp(Operation *op) override {
-      if (erased.contains(op))
+      if (wasErased(op))
         return;
       op->dropAllUses();
       RewriterBase::eraseOp(op);
@@ -890,17 +889,24 @@ struct ConversionPatternRewriterImpl : public RewriterBase::Listener {
 
     /// Erase the given block (unless it was already erased).
     void eraseBlock(Block *block) override {
-      if (erased.contains(block))
+      if (wasErased(block))
         return;
       assert(block->empty() && "expected empty block");
       block->dropAllDefinedValueUses();
       RewriterBase::eraseBlock(block);
     }
 
+    bool wasErased(void *ptr) const { return erased.contains(ptr); }
+
+    bool wasErased(OperationRewrite *rewrite) const {
+      return wasErased(rewrite->getOperation());
+    }
+
     void notifyOperationErased(Operation *op) override { erased.insert(op); }
 
     void notifyBlockErased(Block *block) override { erased.insert(block); }
 
+  private:
     /// Pointers to all erased operations and blocks.
     DenseSet<void *> erased;
   };
@@ -912,6 +918,11 @@ struct ConversionPatternRewriterImpl : public RewriterBase::Listener {
   /// MLIR context.
   MLIRContext *context;
 
+  /// A rewriter that keeps track of ops/block that were already erased and
+  /// skips duplicate op/block erasures. This rewriter is used during the
+  /// "cleanup" phase.
+  SingleEraseRewriter eraseRewriter;
+
   // Mapping between replaced values that differ in type. This happens when
   // replacing a value with one of a different type.
   ConversionValueMapping mapping;
@@ -1058,10 +1069,6 @@ void UnresolvedMaterializationRewrite::rollback() {
   op->erase();
 }
 
-void UnresolvedMaterializationRewrite::cleanup(RewriterBase &rewriter) {
-  rewriter.eraseOp(op);
-}
-
 void ConversionPatternRewriterImpl::applyRewrites() {
   // Commit all rewrites.
   IRRewriter rewriter(context, config.listener);
@@ -1069,7 +1076,6 @@ void ConversionPatternRewriterImpl::applyRewrites() {
     rewrite->commit(rewriter);
 
   // Clean up all rewrites.
-  SingleEraseRewriter eraseRewriter(context);
   for (auto &rewrite : rewrites)
     rewrite->cleanup(eraseRewriter);
 }
@@ -2353,12 +2359,6 @@ struct OperationConverter {
       ConversionPatternRewriterImpl &rewriterImpl,
       DenseMap<Value, SmallVector<Value>> &inverseMapping);
 
-  /// Legalize any unresolved type materializations.
-  LogicalResult legalizeUnresolvedMaterializations(
-      ConversionPatternRewriter &rewriter,
-      ConversionPatternRewriterImpl &rewriterImpl,
-      DenseMap<Value, SmallVector<Value>> &inverseMapping);
-
   /// Legalize an operation result that was marked as "erased".
   LogicalResult
   legalizeErasedResult(Operation *op, OpResult result,
@@ -2405,6 +2405,56 @@ LogicalResult OperationConverter::convert(ConversionPatternRewriter &rewriter,
   return success();
 }
 
+static LogicalResult
+legalizeUnresolvedMaterialization(RewriterBase &rewriter,
+                                  UnresolvedMaterializationRewrite *rewrite) {
+  UnrealizedConversionCastOp op = rewrite->getOperation();
+  assert(!op.use_empty() &&
+         "expected that dead materializations have already been DCE'd");
+  Operation::operand_range inputOperands = op.getOperands();
+  Type outputType = op.getResultTypes()[0];
+
+  // Try to materialize the conversion.
+  if (const TypeConverter *converter = rewrite->getConverter()) {
+    rewriter.setInsertionPoint(op);
+    Value newMaterialization;
+    switch (rewrite->getMaterializationKind()) {
+    case MaterializationKind::Argument:
+      // Try to materialize an argument conversion.
+      newMaterialization = converter->materializeArgumentConversion(
+          rewriter, op->getLoc(), outputType, inputOperands);
+      if (newMaterialization)
+        break;
+      // If an argument materialization failed, fallback to trying a target
+      // materialization.
+      [[fallthrough]];
+    case MaterializationKind::Target:
+      newMaterialization = converter->materializeTargetConversion(
+          rewriter, op->getLoc(), outputType, inputOperands);
+      break;
+    case MaterializationKind::Source:
+      newMaterialization = converter->materializeSourceConversion(
+          rewriter, op->getLoc(), outputType, inputOperands);
+      break;
+    }
+    if (newMaterialization) {
+      assert(newMaterialization.getType() == outputType &&
+             "materialization callback produced value of incorrect type");
+      rewriter.replaceOp(op, newMaterialization);
+      return success();
+    }
+  }
+
+  InFlightDiagnostic diag = op->emitError()
+                            << "failed to legalize unresolved materialization "
+                               "from ("
+                            << inputOperands.getTypes() << ") to " << outputType
+                            << " that remained live after conversion";
+  diag.attachNote(op->getUsers().begin()->getLoc())
+      << "see existing live user here: " << *op->getUsers().begin();
+  return failure();
+}
+
 LogicalResult OperationConverter::convertOperations(ArrayRef<Operation *> ops) {
   if (ops.empty())
     return success();
@@ -2446,6 +2496,37 @@ LogicalResult OperationConverter::convertOperations(ArrayRef<Operation *> ops) {
   } else {
     rewriterImpl.applyRewrites();
   }
+
+  // Gather all unresolved materializations.
+  SmallVector<UnrealizedConversionCastOp> allCastOps;
+  DenseMap<Operation *, UnresolvedMaterializationRewrite *> rewriteMap;
+  for (std::unique_ptr<IRRewrite> &rewrite : rewriterImpl.rewrites) {
+    auto *mat = dyn_cast<UnresolvedMaterializationRewrite>(rewrite.get());
+    if (!mat)
+      continue;
+    if (rewriterImpl.eraseRewriter.wasErased(mat))
+      continue;
+    allCastOps.push_back(mat->getOperation());
+    rewriteMap[mat->getOperation()] = mat;
+  }
+
+  // Reconcile all UnrealizedConversionCastOps that were inserted by the
+  // dialect conversion frameworks. (Not the one that were inserted by
+  // patterns.)
+  SmallVector<UnrealizedConversionCastOp> remainingCastOps;
+  reconcileUnrealizedCasts(allCastOps, &remainingCastOps);
+
+  // Try to legalize all unresolved materializations.
+  if (config.buildMaterializations) {
+    IRRewriter rewriter(rewriterImpl.context, config.listener);
+    for (UnrealizedConversionCastOp castOp : remainingCastOps) {
+      auto it = rewriteMap.find(castOp.getOperation());
+      assert(it != rewriteMap.end() && "inconsistent state");
+      if (failed(legalizeUnresolvedMaterialization(rewriter, it->second)))
+        return failure();
+    }
+  }
+
   return success();
 }
 
@@ -2459,9 +2540,6 @@ OperationConverter::finalize(ConversionPatternRewriter &rewriter) {
   if (failed(legalizeConvertedOpResultTypes(rewriter, rewriterImpl,
                                             inverseMapping)))
     return failure();
-  if (failed(legalizeUnresolvedMaterializations(rewriter, rewriterImpl,
-                                                inverseMapping)))
-    return failure();
   return success();
 }
 
@@ -2577,279 +2655,6 @@ LogicalResult OperationConverter::legalizeConvertedArgumentTypes(
   return success();
 }
 
-/// Replace the results of a materialization operation with the given values.
-static void
-replaceMaterialization(ConversionPatternRewriterImpl &rewriterImpl,
-                       ResultRange matResults, ValueRange values,
-                       DenseMap<Value, SmallVector<Value>> &inverseMapping) {
-  matResults.replaceAllUsesWith(values);
-
-  // For each of the materialization results, update the inverse mappings to
-  // point to the replacement values.
-  for (auto [matResult, newValue] : llvm::zip(matResults, values)) {
-    auto inverseMapIt = inverseMapping.find(matResult);
-    if (inverseMapIt == inverseMapping.end())
-      continue;
-
-    // Update the reverse mapping, or remove the mapping if we couldn't update
-    // it. Not being able to update signals that the mapping would have become
-    // circular (i.e. %foo -> newValue -> %foo), which may occur as values are
-    // propagated through temporary materializations. We simply drop the
-    // mapping, and let the post-conversion replacement logic handle updating
-    // uses.
-    for (Value inverseMapVal : inverseMapIt->second)
-      if (!rewriterImpl.mapping.tryMap(inverseMapVal, newValue))
-        rewriterImpl.mapping.erase(inverseMapVal);
-  }
-}
-
-/// Compute all of the unresolved materializations that will persist beyond the
-/// conversion process, and require inserting a proper user materialization for.
-static void computeNecessaryMaterializations(
-    DenseMap<Operation *, UnresolvedMaterializationRewrite *>
-        &materializationOps,
-    ConversionPatternRewriter &rewriter,
-    ConversionPatternRewriterImpl &rewriterImpl,
-    DenseMap<Value, SmallVector<Value>> &inverseMapping,
-    SetVector<UnresolvedMaterializationRewrite *> &necessaryMaterializations) {
-  // Helper function to check if the given value or a not yet materialized
-  // replacement of the given value is live.
-  // Note: `inverseMapping` maps from replaced values to original values.
-  auto isLive = [&](Value value) {
-    auto findFn = [&](Operation *user) {
-      auto matIt = materializationOps.find(user);
-      if (matIt != materializationOps.end())
-        return !necessaryMaterializations.count(matIt->second);
-      return rewriterImpl.isOpIgnored(user);
-    };
-    // A worklist is needed because a value may have gone through a chain of
-    // replacements and each of the replaced values may have live users.
-    SmallVector<Value> worklist;
-    worklist.push_back(value);
-    while (!worklist.empty()) {
-      Value next = worklist.pop_back_val();
-      if (llvm::find_if_not(next.getUsers(), findFn) != next.user_end())
-        return true;
-      // This value may be replacing another value that has a live user.
-      llvm::append_range(worklist, inverseMapping.lookup(next));
-    }
-    return false;
-  };
-
-  llvm::unique_function<Value(Value, Value, Type)> lookupRemappedValue =
-      [&](Value invalidRoot, Value value, Type type) {
-        // Check to see if the input operation was remapped to a variant of the
-        // output.
-        Value remappedValue = rewriterImpl.mapping.lookupOrDefault(value, type);
-        if (remappedValue.getType() == type && remappedValue != invalidRoot)
-          return remappedValue;
-
-        // Check to see if the input is a materialization operation that
-        // provides an inverse conversion. We just check blindly for
-        // UnrealizedConversionCastOp here, but it has no effect on correctness.
-        auto inputCastOp = value.getDefiningOp<UnrealizedConversionCastOp>();
-        if (inputCastOp && inputCastOp->getNumOperands() == 1)
-          return lookupRemappedValue(invalidRoot, inputCastOp->getOperand(0),
-                                     type);
-
-        return Value();
-      };
-
-  SetVector<UnresolvedMaterializationRewrite *> worklist;
-  for (auto &rewrite : rewriterImpl.rewrites) {
-    auto *mat = dyn_cast<UnresolvedMaterializationRewrite>(rewrite.get());
-    if (!mat)
-      continue;
-    materializationOps.try_emplace(mat->getOperation(), mat);
-    worklist.insert(mat);
-  }
-  while (!worklist.empty()) {
-    UnresolvedMaterializationRewrite *mat = worklist.pop_back_val();
-    UnrealizedConversionCastOp op = mat->getOperation();
-
-    // We currently only handle target materializations here.
-    assert(op->getNumResults() == 1 && "unexpected materialization type");
-    OpResult opResult = op->getOpResult(0);
-    Type outputType = opResult.getType();
-    Operation::operand_range inputOperands = op.getOperands();
-
-    // Try to forward propagate operands for user conversion casts that result
-    // in the input types of the current cast.
-    for (Operation *user : llvm::make_early_inc_range(opResult.getUsers())) {
-      auto castOp = dyn_cast<UnrealizedConversionCastOp>(user);
-      if (!castOp)
-        continue;
-      if (castOp->getResultTypes() == inputOperands.getTypes()) {
-        replaceMaterialization(rewriterImpl, user->getResults(), inputOperands,
-                               inverseMapping);
-        necessaryMaterializations.remove(materializationOps.lookup(user));
-      }
-    }
-
-    // Try to avoid materializing a resolved materialization if possible.
-    // Handle the case of a 1-1 materialization.
-    if (inputOperands.size() == 1) {
-      // Check to see if the input operation was remapped to a variant of the
-      // output.
-      Value remappedValue =
-          lookupRemappedValue(opResult, inputOperands[0], outputType);
-      if (remappedValue && remappedValue != opResult) {
-        replaceMaterialization(rewriterImpl, opResult, remappedValue,
-                               inverseMapping);
-        necessaryMaterializations.remove(mat);
-        continue;
-      }
-    } else {
-      // TODO: Avoid materializing other types of conversions here.
-    }
-
-    // If the materialization does not have any live users, we don't need to
-    // generate a user materialization for it.
-    bool isMaterializationLive = isLive(opResult);
-    if (!isMaterializationLive)
-      continue;
-    if (!necessaryMaterializations.insert(mat))
-      continue;
-
-    // Reprocess input materializations to see if they have an updated status.
-    for (Value input : inputOperands) {
-      if (auto parentOp = input.getDefiningOp<UnrealizedConversionCastOp>()) {
-        if (auto *mat = materializationOps.lookup(parentOp))
-          worklist.insert(mat);
-      }
-    }
-  }
-}
-
-/// Legalize the given unresolved materialization. Returns success if the
-/// materialization was legalized, failure otherise.
-static LogicalResult legalizeUnresolvedMaterialization(
-    UnresolvedMaterializationRewrite &mat,
-    DenseMap<Operation *, UnresolvedMaterializationRewrite *>
-        &materializationOps,
-    ConversionPatternRewriter &rewriter,
-    ConversionPatternRewriterImpl &rewriterImpl,
-    DenseMap<Value, SmallVector<Value>> &inverseMapping) {
-  auto findLiveUser = [&](auto &&users) {
-    auto liveUserIt = llvm::find_if_not(
-        users, [&](Operation *user) { return rewriterImpl.isOpIgnored(user); });
-    return liveUserIt == users.end() ? nullptr : *liveUserIt;
-  };
-
-  llvm::unique_function<Value(Value, Type)> lookupRemappedValue =
-      [&](Value value, Type type) {
-        // Check to see if the input operation was remapped to a variant of the
-        // output.
-        Value remappedValue = rewriterImpl.mapping.lookupOrDefault(value, type);
-        if (remappedValue.getType() == type)
-          return remappedValue;
-        return Value();
-      };
-
-  UnrealizedConversionCastOp op = mat.getOperation();
-  if (!rewriterImpl.ignoredOps.insert(op))
-    return success();
-
-  // We currently only handle target materializations here.
-  OpResult opResult = op->getOpResult(0);
-  Operation::operand_range inputOperands = op.getOperands();
-  Type outputType = opResult.getType();
-
-  // If any input to this materialization is another materialization, resolve
-  // the input first.
-  for (Value value : op->getOperands()) {
-    auto valueCast = value.getDefiningOp<UnrealizedConversionCastOp>();
-    if (!valueCast)
-      continue;
-
-    auto matIt = materializationOps.find(valueCast);
-    if (matIt != materializationOps.end())
-      if (failed(legalizeUnresolvedMaterialization(
-              *matIt->second, materializationOps, rewriter, rewriterImpl,
-              inverseMapping)))
-        return failure();
-  }
-
-  // Perform a last ditch attempt to avoid materializing a resolved
-  // materialization if possible.
-  // Handle the case of a 1-1 materialization.
-  if (inputOperands.size() == 1) {
-    // Check to see if the input operation was remapped to a variant of the
-    // output.
-    Value remappedValue = lookupRemappedValue(inputOperands[0], outputType);
-    if (remappedValue && remappedValue != opResult) {
-      replaceMaterialization(rewriterImpl, opResult, remappedValue,
-                             inverseMapping);
-      return success();
-    }
-  } else {
-    // TODO: Avoid materializing other types of conversions here.
-  }
-
-  // Try to materialize the conversion.
-  if (const TypeConverter *converter = mat.getConverter()) {
-    rewriter.setInsertionPoint(op);
-    Value newMaterialization;
-    switch (mat.getMaterializationKind()) {
-    case MaterializationKind::Argument:
-      // Try to materialize an argument conversion.
-      newMaterialization = converter->materializeArgumentConversion(
-          rewriter, op->getLoc(), outputType, inputOperands);
-      if (newMaterialization)
-        break;
-      // If an argument materialization failed, fallback to trying a target
-      // materialization.
-      [[fallthrough]];
-    case MaterializationKind::Target:
-      newMaterialization = converter->materializeTargetConversion(
-          rewriter, op->getLoc(), outputType, inputOperands);
-      break;
-    case MaterializationKind::Source:
-      newMaterialization = converter->materializeSourceConversion(
-          rewriter, op->getLoc(), outputType, inputOperands);
-      break;
-    }
-    if (newMaterialization) {
-      assert(newMaterialization.getType() == outputType &&
-             "materialization callback produced value of incorrect type");
-      replaceMaterial...
[truncated]

…#104668)

This commit makes source/target/argument materializations (via the
`TypeConverter` API) optional.

By default (`ConversionConfig::buildMaterializations = true`), the
dialect conversion infrastructure tries to legalize all unresolved
materializations right after the main transformation process has
succeeded. If at least one unresolved materialization fails to resolve,
the dialect conversion fails. (With an error message such as `failed to
legalize unresolved materialization ...`.) Automatic materializations
through the `TypeConverter` API can now be deactivated. In that case,
every unresolved materialization will show up as a
`builtin.unrealized_conversion_cast` op in the output IR.

There used to be a complex and error-prone analysis in the dialect
conversion that predicted the future uses of unresolved
materializations. Based on that logic, some casts (that were deemed to
unnecessary) were folded. This analysis was needed because folding
happened at a point of time when some IR changes (e.g., op replacements)
had not materialized yet.

This commit removes that analysis. Any folding of cast ops now happens
after all other IR changes have been materialized and the uses can
directly be queried from the IR. This simplifies the analysis
significantly. And certain helper data structures such as
`inverseMapping` are no longer needed for the analysis. The folding
itself is done by `reconcileUnrealizedCasts` (which also exists as a
standalone pass).

After casts have been folded, the remaining casts are materialized
through the `TypeConverter`, as usual. This last step can be deactivated
in the `ConversionConfig`.

`ConversionConfig::buildMaterializations = false` can be used to debug
error messages such as `failed to legalize unresolved materialization
...`. (It is also useful in case automatic materializations are not
needed.) The materializations that failed to resolve can then be seen as
`builtin.unrealized_conversion_cast` ops in the resulting IR. (This is
better than running with `-debug`, because `-debug` shows IR where some
IR changes have not been materialized yet.)

Note: This is a reupload of #104668, but with correct handling of cyclic unrealized_conversion_casts that may be generated by the dialect conversion.
@matthias-springer matthias-springer force-pushed the users/matthias-springer/optional_materializations2 branch from 6fd9571 to f5190d9 Compare September 5, 2024 15:32
@matthias-springer
Copy link
Member Author

I checked with IREE and Google folks to make sure that this works with their code base and does not contain any further unknown regressions.

Copy link
Contributor

@mikeurbach mikeurbach left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Makes sense to me, thanks for confirming with Google/IREE and adding the handling of cyclic casts.

@matthias-springer matthias-springer merged commit 3815f47 into main Sep 5, 2024
8 checks passed
@matthias-springer matthias-springer deleted the users/matthias-springer/optional_materializations2 branch September 5, 2024 17:41
@maerhart
Copy link
Member

maerhart commented Sep 6, 2024

Really cool! Thanks @matthias-springer

/// builtin.unrealized_conversion_cast ops. The conversion process fails if
/// at least one materialization could not be built.
///
/// If set to "false", the dialect conversion does not does not build any
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

does not does not -> does not

VitaNuo pushed a commit to VitaNuo/llvm-project that referenced this pull request Sep 12, 2024
…llvm#107109)

This commit makes source/target/argument materializations (via the
`TypeConverter` API) optional.

By default (`ConversionConfig::buildMaterializations = true`), the
dialect conversion infrastructure tries to legalize all unresolved
materializations right after the main transformation process has
succeeded. If at least one unresolved materialization fails to resolve,
the dialect conversion fails. (With an error message such as `failed to
legalize unresolved materialization ...`.) Automatic materializations
through the `TypeConverter` API can now be deactivated. In that case,
every unresolved materialization will show up as a
`builtin.unrealized_conversion_cast` op in the output IR.

There used to be a complex and error-prone analysis in the dialect
conversion that predicted the future uses of unresolved
materializations. Based on that logic, some casts (that were deemed to
unnecessary) were folded. This analysis was needed because folding
happened at a point of time when some IR changes (e.g., op replacements)
had not materialized yet.

This commit removes that analysis. Any folding of cast ops now happens
after all other IR changes have been materialized and the uses can
directly be queried from the IR. This simplifies the analysis
significantly. And certain helper data structures such as
`inverseMapping` are no longer needed for the analysis. The folding
itself is done by `reconcileUnrealizedCasts` (which also exists as a
standalone pass).

After casts have been folded, the remaining casts are materialized
through the `TypeConverter`, as usual. This last step can be deactivated
in the `ConversionConfig`.

`ConversionConfig::buildMaterializations = false` can be used to debug
error messages such as `failed to legalize unresolved materialization
...`. (It is also useful in case automatic materializations are not
needed.) The materializations that failed to resolve can then be seen as
`builtin.unrealized_conversion_cast` ops in the resulting IR. (This is
better than running with `-debug`, because `-debug` shows IR where some
IR changes have not been materialized yet.)

Note: This is a reupload of llvm#104668, but with correct handling of cyclic
unrealized_conversion_casts that may be generated by the dialect
conversion.
matthias-springer added a commit that referenced this pull request Sep 12, 2024
…es (#106760)

Handle dropped block arguments and dropped op results in the same way:
build a source materialization (that may fold away if unused). This
simplifies the code base a bit and makes it possible to merge
`legalizeConvertedArgumentTypes` and `legalizeConvertedOpResultTypes` in
a future commit. These two functions are almost doing the same thing
now.

As a side effect, this commit also changes the dialect conversion such
that temporary circular cast ops are no longer generated. (There was a
workaround in #107109 that can now be removed again.) Example:
```
%0 = "builtin.unrealized_conversion_cast"(%1) : (!a) -> !b
%1 = "builtin.unrealized_conversion_cast"(%0) : (!b) -> !a
// No further uses of %0, %1.
```

This happened when:
1. An op was erased. (No replacement values provided.)
2. A conversion pattern for another op builds a replacement value for
the erased op's results (first cast op) during `remapValues`, but that
SSA value is not used during the pattern application.
3. During the finalization phase, `legalizeConvertedOpResultTypes`
thinks that the erased op is alive because of the cast op that was built
in Step 2. It builds a cast from that replacement value to the original
type.
4. During the commit phase, all uses of the original op are replaced
with the casted value produced in Step 3. We have generated circular IR.

This problem can be avoided by making sure that source materializations
are generated for all dropped results. This ensures that we always have
some replacement SSA value in the mapping. Previously, we sometimes had
a value mapped and sometimes not. (No more special casing is needed
anymore to distinguish between "value dropped" or "value replaced with
SSA value".)
matthias-springer added a commit that referenced this pull request Oct 23, 2024
…sult type (#113031)

This commit simplifies the result type of materialization functions.

Previously: `std::optional<Value>`
Now: `Value`

The previous implementation allowed 3 possible return values:
- Non-null value: The materialization function produced a valid
materialization.
- `std::nullopt`: The materialization function failed, but another
materialization can be attempted.
- `Value()`: The materialization failed and so should the dialect
conversion. (Previously: Dialect conversion can roll back.)

This commit removes the last variant. It is not particularly useful
because the dialect conversion will fail anyway if all other
materialization functions produced `std::nullopt`.

Furthermore, in contrast to type conversions, at least one
materialization callback is expected to succeed. In case of a failing
type conversion, the current dialect conversion can roll back and try a
different pattern. This also used to be the case for materializations,
but that functionality was removed with #107109: failed materializations
can no longer trigger a rollback. (They can just make the entire dialect
conversion fail without rollback.) With this in mind, it is even less
useful to have an additional error state for materialization functions.

This commit is in preparation of merging the 1:1 and 1:N type
converters. Target materializations will have to return multiple values
instead of a single one. With this commit, we can keep the API simple:
`SmallVector<Value>` instead of `std::optional<SmallVector<Value>>`.

Note for LLVM integration: All 1:1 materializations should return
`Value` instead of `std::optional<Value>`. Instead of `std::nullopt`
return `Value()`.
NoumanAmir657 pushed a commit to NoumanAmir657/llvm-project that referenced this pull request Nov 4, 2024
…sult type (llvm#113031)

This commit simplifies the result type of materialization functions.

Previously: `std::optional<Value>`
Now: `Value`

The previous implementation allowed 3 possible return values:
- Non-null value: The materialization function produced a valid
materialization.
- `std::nullopt`: The materialization function failed, but another
materialization can be attempted.
- `Value()`: The materialization failed and so should the dialect
conversion. (Previously: Dialect conversion can roll back.)

This commit removes the last variant. It is not particularly useful
because the dialect conversion will fail anyway if all other
materialization functions produced `std::nullopt`.

Furthermore, in contrast to type conversions, at least one
materialization callback is expected to succeed. In case of a failing
type conversion, the current dialect conversion can roll back and try a
different pattern. This also used to be the case for materializations,
but that functionality was removed with llvm#107109: failed materializations
can no longer trigger a rollback. (They can just make the entire dialect
conversion fail without rollback.) With this in mind, it is even less
useful to have an additional error state for materialization functions.

This commit is in preparation of merging the 1:1 and 1:N type
converters. Target materializations will have to return multiple values
instead of a single one. With this commit, we can keep the API simple:
`SmallVector<Value>` instead of `std::optional<SmallVector<Value>>`.

Note for LLVM integration: All 1:1 materializations should return
`Value` instead of `std::optional<Value>`. Instead of `std::nullopt`
return `Value()`.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
mlir:bufferization Bufferization infrastructure mlir:core MLIR Core Infrastructure mlir:gpu mlir
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants