func.func fails to lower to llvm.func when arguments or return types contain dynamic or boolean tensors

[Subway2023]

Description
While testing lowering MLIR func.func to LLVM IR, we found that functions with dynamic tensors, boolean tensors, or other special types in arguments or return values often fail to lower.

Case 1

module {
  func.func nested @func1() -> vector<30xi64> {
    %cst = arith.constant 3.472000e+04 : f16
    ...
  }
  func.func @func2() -> tensor<?xi64> {
    %cst = arith.constant 3.472000e+04 : f16
    ...
  }
}

Lowering command:

/llvm-project-llvmorg-21.1.2/build/bin/mlir-opt \
  --arith-expand \
  --arith-unsigned-when-equivalent \
  --convert-vector-to-scf \
  --convert-scf-to-cf \
  --convert-vector-to-llvm \
  --one-shot-bufferize="dialect-filter=tensor,linalg,bufferization" \
  --finalize-memref-to-llvm \
  --convert-arith-to-llvm \
  --convert-cf-to-llvm \
  --convert-func-to-llvm \
  --reconcile-unrealized-casts \
  program_352287.mlir

Result:

• @func1 is successfully lowered to llvm.func because it returns a vector<30xi64>.
• @func2 is not lowered, remains as func.func because it returns a dynamic tensor<?xi64>.

---

Case 2

module {
  func.func private @func1() { ... }
  func.func private @func2(%arg0: vector<10xf32>, %arg1: memref<?xi32>, %arg2: tensor<?x21xi32>) -> tensor<21x21xi1> {
    ...
  }
}

Lowering command:

/llvm-project-llvmorg-21.1.2/build/bin/mlir-opt \
  --one-shot-bufferize="dialect-filter=tensor,linalg,bufferization" \
  --canonicalize \
  --buffer-deallocation-pipeline \
  --convert-vector-to-scf \
  --convert-scf-to-cf \
  --convert-vector-to-llvm \
  --convert-arith-to-llvm \
  --finalize-memref-to-llvm \
  --convert-cf-to-llvm \
  --convert-arith-to-llvm \
  --finalize-memref-to-llvm \
  --convert-func-to-llvm \
  --reconcile-unrealized-casts \
  program_3279.mlir

Result:

• @func1 is successfully lowered.
• @func2 remains as func.func, because it has arguments and return values with vector, memref, and tensor<?xiN> / tensor<...xi1> types, which are not fully supported for lowering.

---

Expected Behavior

• All func.func should lower to llvm.func, even if arguments or return types include dynamic tensors, boolean tensors, or vectors.
• Alternatively, documentation should clearly indicate which types cannot be lowered and suggest proper handling (e.g., bufferization).

---

Actual Behavior

• Functions with dynamic tensor or boolean tensor types are not lowered.
• Functions with vector or memref types can be lowered correctly.

---

Environment

• MLIR version: LLVM 21.1.2
• OS: Linux
• Command-line flags: as above

---

program_352287.txt
program_3279.txt

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (Subway2023)

**Description** While testing lowering MLIR `func.func` to LLVM IR, we found that functions with **dynamic tensors**, **boolean tensors**, or other special types in arguments or return values often fail to lower.

Case 1

module {
  func.func nested @<!-- -->func1() -&gt; vector&lt;30xi64&gt; {
    %cst = arith.constant 3.472000e+04 : f16
    ...
  }
  func.func @<!-- -->func2() -&gt; tensor&lt;?xi64&gt; {
    %cst = arith.constant 3.472000e+04 : f16
    ...
  }
}

Lowering command:

/llvm-project-llvmorg-21.1.2/build/bin/mlir-opt \
  --arith-expand \
  --arith-unsigned-when-equivalent \
  --convert-vector-to-scf \
  --convert-scf-to-cf \
  --convert-vector-to-llvm \
  --one-shot-bufferize="dialect-filter=tensor,linalg,bufferization" \
  --finalize-memref-to-llvm \
  --convert-arith-to-llvm \
  --convert-cf-to-llvm \
  --convert-func-to-llvm \
  --reconcile-unrealized-casts \
  program_352287.mlir

Result:

• @<!-- -->func1 is successfully lowered to llvm.func because it returns a vector<30xi64>.
• @<!-- -->func2 is not lowered, remains as func.func because it returns a dynamic tensor<?xi64>.

---

Case 2

module {
  func.func private @<!-- -->func1() { ... }
  func.func private @<!-- -->func2(%arg0: vector&lt;10xf32&gt;, %arg1: memref&lt;?xi32&gt;, %arg2: tensor&lt;?x21xi32&gt;) -&gt; tensor&lt;21x21xi1&gt; {
    ...
  }
}

Lowering command:

/llvm-project-llvmorg-21.1.2/build/bin/mlir-opt \
  --one-shot-bufferize="dialect-filter=tensor,linalg,bufferization" \
  --canonicalize \
  --buffer-deallocation-pipeline \
  --convert-vector-to-scf \
  --convert-scf-to-cf \
  --convert-vector-to-llvm \
  --convert-arith-to-llvm \
  --finalize-memref-to-llvm \
  --convert-cf-to-llvm \
  --convert-arith-to-llvm \
  --finalize-memref-to-llvm \
  --convert-func-to-llvm \
  --reconcile-unrealized-casts \
  program_3279.mlir

Result:

• @<!-- -->func1 is successfully lowered.
• @<!-- -->func2 remains as func.func, because it has arguments and return values with vector, memref, and tensor<?xiN> / tensor<...xi1> types, which are not fully supported for lowering.

---

Expected Behavior

• All func.func should lower to llvm.func, even if arguments or return types include dynamic tensors, boolean tensors, or vectors.
• Alternatively, documentation should clearly indicate which types cannot be lowered and suggest proper handling (e.g., bufferization).

---

Actual Behavior

• Functions with dynamic tensor or boolean tensor types are not lowered.
• Functions with vector or memref types can be lowered correctly.

---

Environment

• MLIR version: LLVM 21.1.2
• OS: Linux
• Command-line flags: as above

---

program_352287.txt
program_3279.txt

[linuxlonelyeagle]

you should use -one-shot-bufferize="bufferize-function-boundaries" to bufferize func dialect, then it will work well.

mlir-opt test.mlir --arith-expand -arith-unsigned-when-equivalent -convert-vector-to-scf -convert-scf-to-cf  -convert-vector-to-llvm -one-shot-bufferize="bufferize-function-boundaries"  -finalize-memref-to-llvm -convert-arith-to-llvm -convert-cf-to-llvm -convert-func-to-llvm