[MLIR] Invalid LLVM dialect generated after lowering

[RJerrica]

Hi, I used MLIR to lower a program to LLVM IR. However, the process generated an invalid LLVM dialect program that cannot be translated to LLVM IR. Here is the code example.

Input Program

// vector_add.mlir
module {
  func.func @vector_add(%A: memref<100xf32>, %B: memref<100xf32>, %C: memref<100xf32>) {
    %c0 = index.constant 0
    %c1 = index.constant 1
    %c100 = index.constant 100
    %c256 = index.constant 256
    
    gpu.launch blocks(%block_x, %block_y, %block_z) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1)
                 threads(%thread_x, %thread_y, %thread_z) in (%block_size_x = %c256, %block_size_y = %c1, %block_size_z = %c1) {
      
      %temp_global_id = index.mul %block_x, %block_size_x
      %global_id = index.add %temp_global_id, %thread_x
      
      %within_bounds = index.cmp slt(%global_id, %c100)
      
      scf.if %within_bounds {
        %a_val = memref.load %A[%global_id] : memref<100xf32>
        %b_val = memref.load %B[%global_id] : memref<100xf32>
        
        %sum = arith.addf %a_val, %b_val : f32
        
        memref.store %sum, %C[%global_id] : memref<100xf32>
      }
      
      gpu.terminator
    }
    
    func.return
  }
}

MLIR Command

I used the following command to lower the program to​ the LLVM dialect and then to​ LLVM IR.

mlir-opt vector_add.mlir \
  --convert-scf-to-cf \
  --convert-gpu-to-nvvm \
  --convert-cf-to-llvm \
  --convert-index-to-llvm \
  --convert-arith-to-llvm \
  --convert-math-to-llvm \
  --convert-func-to-llvm \
  --finalize-memref-to-llvm \
  -convert-nvvm-to-llvm \
  --reconcile-unrealized-casts \
  -o vector_add_llvm.mlir

Then, I use mlir-translate to convert the LLVM dialect module to LLVM IR. However, the generated LLVM dialect program is invalid because it contains numerous unrealized_conversion_cast operations, such as %20 = builtin.unrealized_conversion_cast %19 : i64 to index (See below for the logs).

mlir-translate --mlir-to-llvmir vector_add_llvm.mlir -o vector_add.ll
vector_add_llvm.mlir:6:10: error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast
    %3 = builtin.unrealized_conversion_cast %2 : i64 to index
         ^
vector_add_llvm.mlir:6:10: note: see current operation: %3 = "builtin.unrealized_conversion_cast"(%2) : (i64) -> index

Version

mlir-opt --version
LLVM (http://llvm.org/):
  LLVM version 22.0.0git
  Optimized build with assertions.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: None (RJerrica)

Hi, I used MLIR to lower a program to LLVM IR. However, the process generated an invalid LLVM dialect program that cannot be translated to LLVM IR. Here is the code example.

Input Program

// vector_add.mlir
module {
  func.func @<!-- -->vector_add(%A: memref&lt;100xf32&gt;, %B: memref&lt;100xf32&gt;, %C: memref&lt;100xf32&gt;) {
    %c0 = index.constant 0
    %c1 = index.constant 1
    %c100 = index.constant 100
    %c256 = index.constant 256
    
    gpu.launch blocks(%block_x, %block_y, %block_z) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1)
                 threads(%thread_x, %thread_y, %thread_z) in (%block_size_x = %c256, %block_size_y = %c1, %block_size_z = %c1) {
      
      %temp_global_id = index.mul %block_x, %block_size_x
      %global_id = index.add %temp_global_id, %thread_x
      
      %within_bounds = index.cmp slt(%global_id, %c100)
      
      scf.if %within_bounds {
        %a_val = memref.load %A[%global_id] : memref&lt;100xf32&gt;
        %b_val = memref.load %B[%global_id] : memref&lt;100xf32&gt;
        
        %sum = arith.addf %a_val, %b_val : f32
        
        memref.store %sum, %C[%global_id] : memref&lt;100xf32&gt;
      }
      
      gpu.terminator
    }
    
    func.return
  }
}

MLIR Command

I used the following command to lower the program to​ the LLVM dialect and then to​ LLVM IR.

mlir-opt vector_add.mlir \
  --convert-scf-to-cf \
  --convert-gpu-to-nvvm \
  --convert-cf-to-llvm \
  --convert-index-to-llvm \
  --convert-arith-to-llvm \
  --convert-math-to-llvm \
  --convert-func-to-llvm \
  --finalize-memref-to-llvm \
  -convert-nvvm-to-llvm \
  --reconcile-unrealized-casts \
  -o vector_add_llvm.mlir

Then, I use mlir-translate to convert the LLVM dialect module to LLVM IR. However, the generated LLVM dialect program is invalid because it contains numerous unrealized_conversion_cast operations, such as %20 = builtin.unrealized_conversion_cast %19 : i64 to index (See below for the logs).

mlir-translate --mlir-to-llvmir vector_add_llvm.mlir -o vector_add.ll
vector_add_llvm.mlir:6:10: error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast
    %3 = builtin.unrealized_conversion_cast %2 : i64 to index
         ^
vector_add_llvm.mlir:6:10: note: see current operation: %3 = "builtin.unrealized_conversion_cast"(%2) : (i64) -> index

Version

mlir-opt --version
LLVM (http://llvm.org/):
  LLVM version 22.0.0git
  Optimized build with assertions.

[GrumpyPigSkin]

I managed to get this to compile using the following args for mlir-opt, it looks the the args you were using only partially lowered the code:

{
      "type": "lldb",
      "request": "launch",
      "name": "Launch MLIR",
      "program": "${workspaceFolder}/build_debug/bin/mlir-opt",
      "args": [
        "vector_add.mlir",
        // MOve the inlined gpu kernel into its own function
        "-gpu-kernel-outlining",
        "--convert-scf-to-cf",
        // Lower device kernel to NVVM
        "--convert-gpu-to-nvvm",
        // Clean up types (index -> i64) inside the kernel BEFORE compiling it
        "-reconcile-unrealized-casts",
        // Compile the kernel to a binary string (PTX)
        "--nvvm-attach-target=chip=sm_75",
        "--gpu-module-to-binary",
        // Lower Host GPU Launch to calls (now it can see the binary)
        "-gpu-to-llvm",
        // Standard Host Lowering
        "--convert-cf-to-llvm",
        "--convert-index-to-llvm",
        "--convert-arith-to-llvm",
        "--convert-math-to-llvm",
        "--convert-func-to-llvm",
        "--finalize-memref-to-llvm",
        "-convert-nvvm-to-llvm",
        // Final cleanup for Host types
        "-reconcile-unrealized-casts",
        "-o",
        "vector_add_llvm.mlir"
      ],
      "cwd": "${workspaceFolder}"
    }