[Bug] Cannot lower `addptr` for ` gm_slot_buffer` in `aix_initialize_pipe`

[fiskrt]

Component

PTO Dialect / ODS (include/PTO/IR)

Description

Cannot use pointer arithmetic using pto.addptr when passing gm_slot_buffer to aix_initialize_pipe

error: addptr must feed make_tensor_view or load/store_scalar for lowering

Reproduction (minimal)

module {
  func.func @cube_kernel(%arg0: !pto.ptr<f32>, %arg1: !pto.ptr<f32>) attributes {pto.kernel_kind = #pto.kernel_kind<cube>} {
    %c2048 = arith.constant 2048 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = pto.get_block_idx
    %1 = arith.index_cast %0 : i64 to index
    %2 = arith.muli %1, %c2048 : index
    %3 = pto.addptr %arg0, %2 : <f32> -> <f32>
    %4 = pto.import_reserved_buffer{name = "c2v_fifo", peer_func = @vector_kernel} -> i32
    pto.aic_initialize_pipe{dir_mask = 1, slot_size = 1024} (gm_slot_buffer = %3 : !pto.ptr<f32>, c2v_consumer_buf = %4 : i32, v2c_consumer_buf = %c0_i32 : i32)
    return
  }
  func.func @vector_kernel(%arg0: !pto.ptr<f32>, %arg1: !pto.ptr<f32>) attributes {pto.kernel_kind = #pto.kernel_kind<vector>} {
    %c2048 = arith.constant 2048 : index
    %c0_i32 = arith.constant 0 : i32
    %0 = pto.get_block_idx
    %1 = arith.index_cast %0 : i64 to index
    %2 = arith.muli %1, %c2048 : index
    %3 = pto.addptr %arg0, %2 : <f32> -> <f32>
    %4 = pto.reserve_buffer{name = "c2v_fifo", size = 8192, location = <vec>, auto = true} -> i32
    pto.aiv_initialize_pipe{dir_mask = 1, slot_size = 1024} (gm_slot_buffer = %3 : !pto.ptr<f32>, c2v_consumer_buf = %4 : i32, v2c_consumer_buf = %c0_i32 : i32)
    pto.tfree_from_aic{split = 1}
    return
  }
  func.func @call_both(%arg0: memref<256xi64>, %arg1: !pto.ptr<f32>, %arg2: !pto.ptr<f32>, %arg3: !pto.ptr<f32>) attributes {pto.entry} {
    pto.set_ffts %arg0 : memref<256xi64>
    call @cube_kernel(%arg1, %arg2) : (!pto.ptr<f32>, !pto.ptr<f32>) -> ()
    call @vector_kernel(%arg1, %arg3) : (!pto.ptr<f32>, !pto.ptr<f32>) -> ()
    return
  }
}

Run with

ptoas --pto-arch a3 repro.mlir

Expected behavior

Generate a TPipe object with offset

 auto pipe = TPipe<0, Direction::DIR_C2V, 1024, 8, 8>(v1+2048*get_block_idx(), v4, v9);

Actual behavior / error logs

loc("repro.mlir":8:10): error: addptr must feed make_tensor_view or load/store_scalar for lowering

Git commit

dd58f5b

Host platform

Linux (aarch64)

Target Ascend arch (if relevant)

a3

PTOAS build level (if relevant)

None