[PATCH] D144086: [AArch64] Load into zero vector patterns

[Benoit Jacob via Phabricator via llvm-commits]

Benoit added a comment.

Thank you very much for the quick fix. I confirm that https://reviews.llvm.org/rG1c6ea961938488997712763762079e535b8b704e fixes the regression.

You probably won't need this anymore since you were able to fix this without it, but just for completeness, here was how to reproduce:

1. Build https://github.com/openxla/iree - following normal build instructions <https://openxla.github.io/iree/building-from-source/getting-started/> - note that IREE uses its own submodule `third_party/llvm-project`.
2. Run the IREE compiler from the build directory with these flags:

  tools/iree-compile --iree-llvm-target-triple=aarch64-none-linux-android29 --iree-hal-target-backends=llvm-cpu ~/pack_testcase.mlir -o /tmp/a.vmfb --iree-llvm-keep-linker-artifacts

Where the input file `pack_testcase.mlir` is:

  func.func @pack_pad_transpose_1x9xi8_into_2x4x8x4xi8(%arg0 : tensor<1x9xi8>) -> tensor<2x4x8x4xi8> {
    %empty = tensor.empty() : tensor<2x4x8x4xi8>
    %c0_i8 = arith.constant 0 : i8
    %pack = tensor.pack %arg0 padding_value(%c0_i8 : i8) outer_dims_perm = [1, 0] inner_dims_pos = [1, 0] inner_tiles = [8, 4] into %empty {encoding = #iree_linalg_ext.encoding<MATMUL_I8I8I32_RHS>} : tensor<1x9xi8> -> tensor<2x4x8x4xi8>
    return %pack : tensor<2x4x8x4xi8>
  }

Thanks to the `--iree-llvm-keep-linker-artifacts` flag, it will print the path to the generated `.so`, like this

  /usr/local/google/home/benoitjacob/pack_testcase.mlir:4:11: remark: linker artifacts for embedded_elf_arm_64 preserved:
      /tmp/pack_pad_transpose_1x9xi8_into_2x4x8x4xi8_dispatch_0-9c98ea.so

So you can then objdump that as usual,

  $ANDROID_NDK/toolchains/llvm/prebuilt/linux-x86_64/bin/llvm-objdump -d /tmp/pack_pad_transpose_1x9xi8_into_2x4x8x4xi8_dispatch_0-9c98ea.so




Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D144086/new/

https://reviews.llvm.org/D144086