[llvm-branch-commits] [mlir] [draft] Dialect Conversion without Rollback (PR #93412)

Matthias Springer via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Tue May 28 12:12:26 PDT 2024


================
@@ -1053,3 +1055,241 @@ LogicalResult mlir::applyOpPatternsAndFold(
   });
   return converged;
 }
+
+//===----------------------------------------------------------------------===//
+// One-Shot Dialect Conversion Infrastructure
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// A conversion rewriter for the One-Shot Dialect Conversion. This rewriter
+/// immediately materializes all IR changes. It derives from
+/// `ConversionPatternRewriter` so that the existing conversion patterns can
+/// be used with the One-Shot Dialect Conversion.
+class OneShotConversionPatternRewriter : public ConversionPatternRewriter {
+public:
+  OneShotConversionPatternRewriter(MLIRContext *ctx)
+      : ConversionPatternRewriter(ctx) {}
+
+  bool canRecoverFromRewriteFailure() const override { return false; }
+
+  void replaceOp(Operation *op, ValueRange newValues) override;
+
+  void replaceOp(Operation *op, Operation *newOp) override {
+    replaceOp(op, newOp->getResults());
+  }
+
+  void eraseOp(Operation *op) override { PatternRewriter::eraseOp(op); }
+
+  void eraseBlock(Block *block) override { PatternRewriter::eraseBlock(block); }
+
+  void inlineBlockBefore(Block *source, Block *dest, Block::iterator before,
+                         ValueRange argValues = std::nullopt) override {
+    PatternRewriter::inlineBlockBefore(source, dest, before, argValues);
+  }
+  using PatternRewriter::inlineBlockBefore;
+
+  void startOpModification(Operation *op) override {
+    PatternRewriter::startOpModification(op);
+  }
+
+  void finalizeOpModification(Operation *op) override {
+    PatternRewriter::finalizeOpModification(op);
+  }
+
+  void cancelOpModification(Operation *op) override {
+    PatternRewriter::cancelOpModification(op);
+  }
+
+  void setCurrentTypeConverter(const TypeConverter *converter) override {
+    typeConverter = converter;
+  }
+
+  const TypeConverter *getCurrentTypeConverter() const override {
+    return typeConverter;
+  }
+
+  LogicalResult getAdapterOperands(StringRef valueDiagTag,
+                                   std::optional<Location> inputLoc,
+                                   ValueRange values,
+                                   SmallVector<Value> &remapped) override;
+
+private:
+  /// Build an unrealized_conversion_cast op or look it up in the cache.
+  Value buildUnrealizedConversionCast(Location loc, Type type, Value value);
+
+  /// The current type converter.
+  const TypeConverter *typeConverter;
+
+  /// A cache for unrealized_conversion_casts. To ensure that identical casts
+  /// are not built multiple times.
+  DenseMap<std::pair<Value, Type>, Value> castCache;
----------------
matthias-springer wrote:

I think this can happen during `getAdapterOperands`. Each conversion pattern can have its own type converter. (That's why there is `currentTypeConverter` etc.)

```
func.func @foo(%arg0: i1) {
  "op1"(%arg0)
  "op2"(%arg0)
}
```

Let's assume that there is no conversion pattern for `func.func`. I.e., the block argument is not converted. We could have two different patterns for "op1" and "op2" with different type converters that convert `i1` to different types. Now we have to insert two unrealized_conversion_casts with different result types.

(This probably does not happen very often.)


https://github.com/llvm/llvm-project/pull/93412


More information about the llvm-branch-commits mailing list