[flang] Remove materialization workaround in type converter

[matthias-springer]

This change is in preparation of #97903, which adds extra checks for materializations: it is now enforced that they produce an SSA value of the correct type, so the current workaround no longer works.

The original workaround avoided target materializations by directly returning the to-be-converted SSA value from the materialization callback. This can be avoided by initializing the lowering patterns that insert the materializations without a type converter. For cg::XEmboxOp, the existing workaround that skips unrealized_conversion_cast ops is still in place.

Also remove the lowering pattern for unrealized_conversion_cast. This pattern has no effect because unrealized_conversion_cast ops that are inserted by the dialect conversion framework are never matched by the pattern driver.

[llvmbot]

@llvm/pr-subscribers-flang-fir-hlfir

@llvm/pr-subscribers-flang-codegen

Author: Matthias Springer (matthias-springer)

Changes

This change is in preparation of #97903, which adds extra checks for materializations: it is now enforced that they produce an SSA value of the correct type, so the current workaround no longer works.

For fir.has_value the fix is simple: no target materializations on the operands are performed if the lowering patterns is initialized without a type converter. For cg::XEmboxOp, the existing workaround that skips unrealized_conversion_cast ops can be generalized. (This is still a workaround.)

Also remove the lowering pattern for unrealized_conversion_cast. This pattern has no effect because unrealized_conversion_cast ops that are inserted by the dialect conversion framework are never matched by the pattern driver.

---

Full diff: https://github.com/llvm/llvm-project/pull/98743.diff

3 Files Affected:

• (modified) flang/include/flang/Tools/CLOptions.inc (+5)
• (modified) flang/lib/Optimizer/CodeGen/CodeGen.cpp (+19-41)
• (modified) flang/test/Fir/basic-program.fir (+1)

diff --git a/flang/include/flang/Tools/CLOptions.inc b/flang/include/flang/Tools/CLOptions.inc
index 7f2910c5cfd3c..7df5044949463 100644
--- a/flang/include/flang/Tools/CLOptions.inc
+++ b/flang/include/flang/Tools/CLOptions.inc
@@ -9,6 +9,7 @@
 /// This file defines some shared command-line options that can be used when
 /// debugging the test tools. This file must be included into the tool.
 
+#include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h"
 #include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
 #include "mlir/Dialect/LLVMIR/LLVMAttrs.h"
 #include "mlir/Pass/PassManager.h"
@@ -223,6 +224,10 @@ inline void addFIRToLLVMPass(
   options.forceUnifiedTBAATree = useOldAliasTags;
   addPassConditionally(pm, disableFirToLlvmIr,
       [&]() { return fir::createFIRToLLVMPass(options); });
+  // The dialect conversion framework may leave dead unrealized_conversion_cast
+  // ops behind, so run reconcile-unrealized-casts to clean them up.
+  addPassConditionally(pm, disableFirToLlvmIr,
+      [&]() { return mlir::createReconcileUnrealizedCastsPass(); });
 }
 
 inline void addLLVMDialectToLLVMPass(
diff --git a/flang/lib/Optimizer/CodeGen/CodeGen.cpp b/flang/lib/Optimizer/CodeGen/CodeGen.cpp
index 7483acfcd1ca7..bcd8a8c37c2a5 100644
--- a/flang/lib/Optimizer/CodeGen/CodeGen.cpp
+++ b/flang/lib/Optimizer/CodeGen/CodeGen.cpp
@@ -35,7 +35,6 @@
 #include "mlir/Conversion/MathToLLVM/MathToLLVM.h"
 #include "mlir/Conversion/MathToLibm/MathToLibm.h"
 #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h"
-#include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h"
 #include "mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/DLTI/DLTI.h"
@@ -1726,10 +1725,9 @@ struct XReboxOpConversion : public EmboxCommonConversion<fir::cg::XReboxOp> {
     // fir.box was translated to an llvm.ptr<llvm.struct<>> and the MLIR pass
     // manager inserted a builtin.unrealized_conversion_cast that was inserted
     // and needs to be removed here.
-    if (isInGlobalOp(rewriter))
-      if (auto unrealizedCast =
-              loweredBox.getDefiningOp<mlir::UnrealizedConversionCastOp>())
-        loweredBox = unrealizedCast.getInputs()[0];
+    if (auto unrealizedCast =
+            loweredBox.getDefiningOp<mlir::UnrealizedConversionCastOp>())
+      loweredBox = unrealizedCast.getInputs()[0];
 
     TypePair inputBoxTyPair = getBoxTypePair(rebox.getBox().getType());
 
@@ -2669,8 +2667,9 @@ struct TypeDescOpConversion : public fir::FIROpConversion<fir::TypeDescOp> {
 };
 
 /// Lower `fir.has_value` operation to `llvm.return` operation.
-struct HasValueOpConversion : public fir::FIROpConversion<fir::HasValueOp> {
-  using FIROpConversion::FIROpConversion;
+struct HasValueOpConversion
+    : public mlir::OpConversionPattern<fir::HasValueOp> {
+  using OpConversionPattern::OpConversionPattern;
 
   llvm::LogicalResult
   matchAndRewrite(fir::HasValueOp op, OpAdaptor adaptor,
@@ -3515,29 +3514,6 @@ struct MustBeDeadConversion : public fir::FIROpConversion<FromOp> {
   }
 };
 
-struct UnrealizedConversionCastOpConversion
-    : public fir::FIROpConversion<mlir::UnrealizedConversionCastOp> {
-  using FIROpConversion::FIROpConversion;
-
-  llvm::LogicalResult
-  matchAndRewrite(mlir::UnrealizedConversionCastOp op, OpAdaptor adaptor,
-                  mlir::ConversionPatternRewriter &rewriter) const override {
-    assert(op.getOutputs().getTypes().size() == 1 && "expect a single type");
-    mlir::Type convertedType = convertType(op.getOutputs().getTypes()[0]);
-    if (convertedType == adaptor.getInputs().getTypes()[0]) {
-      rewriter.replaceOp(op, adaptor.getInputs());
-      return mlir::success();
-    }
-
-    convertedType = adaptor.getInputs().getTypes()[0];
-    if (convertedType == op.getOutputs().getType()[0]) {
-      rewriter.replaceOp(op, adaptor.getInputs());
-      return mlir::success();
-    }
-    return mlir::failure();
-  }
-};
-
 struct ShapeOpConversion : public MustBeDeadConversion<fir::ShapeOp> {
   using MustBeDeadConversion::MustBeDeadConversion;
 };
@@ -3714,7 +3690,8 @@ class FIRToLLVMLowering
       signalPassFailure();
     }
 
-    // Run pass to add comdats to functions that have weak linkage on relevant platforms
+    // Run pass to add comdats to functions that have weak linkage on relevant
+    // platforms
     if (fir::getTargetTriple(mod).supportsCOMDAT()) {
       mlir::OpPassManager comdatPM("builtin.module");
       comdatPM.addPass(mlir::LLVM::createLLVMAddComdats());
@@ -3789,16 +3766,17 @@ void fir::populateFIRToLLVMConversionPatterns(
       DivcOpConversion, EmboxOpConversion, EmboxCharOpConversion,
       EmboxProcOpConversion, ExtractValueOpConversion, FieldIndexOpConversion,
       FirEndOpConversion, FreeMemOpConversion, GlobalLenOpConversion,
-      GlobalOpConversion, HasValueOpConversion, InsertOnRangeOpConversion,
-      InsertValueOpConversion, IsPresentOpConversion, LenParamIndexOpConversion,
-      LoadOpConversion, MulcOpConversion, NegcOpConversion,
-      NoReassocOpConversion, SelectCaseOpConversion, SelectOpConversion,
-      SelectRankOpConversion, SelectTypeOpConversion, ShapeOpConversion,
-      ShapeShiftOpConversion, ShiftOpConversion, SliceOpConversion,
-      StoreOpConversion, StringLitOpConversion, SubcOpConversion,
-      TypeDescOpConversion, TypeInfoOpConversion, UnboxCharOpConversion,
-      UnboxProcOpConversion, UndefOpConversion, UnreachableOpConversion,
-      UnrealizedConversionCastOpConversion, XArrayCoorOpConversion,
+      GlobalOpConversion, InsertOnRangeOpConversion, InsertValueOpConversion,
+      IsPresentOpConversion, LenParamIndexOpConversion, LoadOpConversion,
+      MulcOpConversion, NegcOpConversion, NoReassocOpConversion,
+      SelectCaseOpConversion, SelectOpConversion, SelectRankOpConversion,
+      SelectTypeOpConversion, ShapeOpConversion, ShapeShiftOpConversion,
+      ShiftOpConversion, SliceOpConversion, StoreOpConversion,
+      StringLitOpConversion, SubcOpConversion, TypeDescOpConversion,
+      TypeInfoOpConversion, UnboxCharOpConversion, UnboxProcOpConversion,
+      UndefOpConversion, UnreachableOpConversion, XArrayCoorOpConversion,
       XEmboxOpConversion, XReboxOpConversion, ZeroOpConversion>(converter,
                                                                 options);
+
+  patterns.insert<HasValueOpConversion>(patterns.getContext());
 }
diff --git a/flang/test/Fir/basic-program.fir b/flang/test/Fir/basic-program.fir
index 7bbfd709b0aaf..dda4f32872fef 100644
--- a/flang/test/Fir/basic-program.fir
+++ b/flang/test/Fir/basic-program.fir
@@ -119,4 +119,5 @@ func.func @_QQmain() {
 // PASSES-NEXT:   (S) 0 num-dce'd - Number of operations eliminated
 // PASSES-NEXT: TargetRewrite
 // PASSES-NEXT: FIRToLLVMLowering
+// PASSES-NEXT: ReconcileUnrealizedCasts
 // PASSES-NEXT: LLVMIRLoweringPass

[matthias-springer]

Note: Type conversions do not know the context in which they are converting a type. But each pattern can be initialized with a different type converter. That's the basic idea behind this PR.

I tried the same for the workaround in XReboxOpConversion, but without success:

    // Inside a fir.global, the input box was produced as an llvm.struct<>
    // because objects cannot be handled in memory inside a fir.global body that
    // must be constant foldable. However, the type translation are not
    // contextual, so the fir.box<T> type of the operation that produced the
    // fir.box was translated to an llvm.ptr<llvm.struct<>> and the MLIR pass
    // manager inserted a builtin.unrealized_conversion_cast that was inserted
    // and needs to be removed here.
    if (isInGlobalOp(rewriter))
      if (auto unrealizedCast =
              loweredBox.getDefiningOp<mlir::UnrealizedConversionCastOp>())
        loweredBox = unrealizedCast.getInputs()[0];

Maybe the type conversion in that pattern is truly context-specific.

[jeanPerier]

Note: Type conversions do not know the context in which they are converting a type. But each pattern can be initialized with a different type converter. That's the basic idea behind this PR.

I tried the same for the workaround in XReboxOpConversion, but without success:

Yes, the handling of the fir.box type (Fotran descriptor) is context specific. It is lowered to an "unique pointer" to a struct type, except inside globals, where it is directly lowered to struct because memory cannot be manipulated in a static initializer and what we want to produce is the struct value itself.
I am not very happy with this context specific interpretation. The solution is probably to introduce new FIR types for the global contexts (which would push the contextual problem to lowering, so the end-to-end problem would not be much simpler).

[jeanPerier]

Thank you @matthias-springer, looks good to me.