[mlir] `memref.store` and `vector.store` not working in `scf.for` with `mlir-vulkan-runner`

[angelz913]

I was writing an integration test with mlir-vulkan-runner. When I was initializing an input buffer using memref.store in an scf.for loop in the host function, I got an error that possibly resulted from finalize-memref-to-llvm in the pass pipeline of mlir-vulkan-runner (https://github.com/llvm/llvm-project/blob/main/mlir/tools/mlir-vulkan-runner/mlir-vulkan-runner.cpp#L78). I also tried using vector.store instead but still got the same error.

error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast
Error: could not convert to LLVM IR

The host function is:

func.func @main() {
  // Allocate 3 buffers.
  %in_buf = memref.alloc() : memref<128xi32>
  %out_buf = memref.alloc() : memref<1xi32>
  %total_count_buf = memref.alloc() : memref<1xi32>

  // Constants.
  %cst0 = arith.constant 0 : i32
  %idx0 = arith.constant 0 : index
  %idx1 = arith.constant 1 : index
  %idx32 = arith.constant 32 : index
  %idx64 = arith.constant 64 : index
  %idx65 = arith.constant 65 : index
  %idx128 = arith.constant 128 : index

  // Initialize input buffer.
  // The input will be [0, 1, ..., 62, 63, 63, 62, ..., 1, 0]
  scf.for %iter = %idx0 to %idx64 step %idx1 {
    %iter_i32 = index.castu %iter : index to i32
    memref.store %iter_i32, %in_buf[%iter] : memref<128xi32>
  }
  %cst127 = arith.constant 127 : i32
  scf.for %iter = %idx65 to %idx128 step %idx1 {
    %iter_i32 = index.castu %iter : index to i32
    %elem = arith.subi %cst127, %iter_i32 : i32
    memref.store %elem, %in_buf[%iter] : memref<128xi32>
  }

  // More code here...
}

[llvmbot]

@llvm/issue-subscribers-mlir-spirv

Author: Angel Zhang (angelz913)

I was writing an integration test with `mlir-vulkan-runner`. When I was initializing an input buffer using `memref.store` in an `scf.for` loop in the host function, I got an error that possibly resulted from `finalize-memref-to-llvm` in the pass pipeline of `mlir-vulkan-runner` (https://github.com/llvm/llvm-project/blob/main/mlir/tools/mlir-vulkan-runner/mlir-vulkan-runner.cpp#L78). I also tried using `vector.store` instead but still got the same error. ``` error: LLVM Translation failed for operation: builtin.unrealized_conversion_cast Error: could not convert to LLVM IR ``` The host function is: ``` func.func @main() { // Allocate 3 buffers. %in_buf = memref.alloc() : memref<128xi32> %out_buf = memref.alloc() : memref<1xi32> %total_count_buf = memref.alloc() : memref<1xi32> // Constants. %cst0 = arith.constant 0 : i32 %idx0 = arith.constant 0 : index %idx1 = arith.constant 1 : index %idx32 = arith.constant 32 : index %idx64 = arith.constant 64 : index %idx65 = arith.constant 65 : index %idx128 = arith.constant 128 : index // Initialize input buffer. // The input will be [0, 1, ..., 62, 63, 63, 62, ..., 1, 0] scf.for %iter = %idx0 to %idx64 step %idx1 { %iter_i32 = index.castu %iter : index to i32 memref.store %iter_i32, %in_buf[%iter] : memref<128xi32> } %cst127 = arith.constant 127 : i32 scf.for %iter = %idx65 to %idx128 step %idx1 { %iter_i32 = index.castu %iter : index to i32 %elem = arith.subi %cst127, %iter_i32 : i32 memref.store %elem, %in_buf[%iter] : memref<128xi32> } // More code here... } ```