Google Heir Heir

Dump intermediate IR during compilation

Overview

This skill guides the agent in adding flags to the heir-opt tool to dump the compiled IR after each compiler pass to a series of files. These files can then be inspected to identify the failing pass as well as the IR input to the failing pass.

Usage

Converting a bazel run command

To convert a bazel run command, add the flags --mlir-print-ir-before-all and --mlir-print-ir-tree-dir=<PATH>, where <PATH> is a path to a temporary directory.

For example, given the following input command (which can be produced by the lit-to-bazel skill),

bazel run //tools:heir-opt -- --mlir-to-ckks='ciphertext-degree=8' --scheme-to-openfhe='entry-function=dot_product' /path/to/file

The agent should produce the following (again where <PATH> is chosen appropriately)

bazel run //tools:heir-opt -- --mlir-to-ckks='ciphertext-degree=8' --scheme-to-openfhe='entry-function=dot_product' --mlir-print-ir-before-all --mlir-print-ir-tree-dir=<PATH> /path/to/file

Tree structure

--mlir-print-ir-tree-dir nests the dumped IR in a tree structure that matches the nesting structure of the pass pipeline. In particular, by default all passes will be nested under a builtin_module_no-symbol-name/ folder, and any passes in the pipeline that are anchored on nested operations like func.func will be placed in subdirectories based on the name, such as func_func_matvec__decrypt__result0 for a pass anchored on func.func when the IR contains a function called matvec__decrypt__result0.

For example, if the flag --mlir-print-ir-tree-dir=/tmp/mlir is used, then the dumped files might look like:

/tmp/mlir/
└── builtin_module_no-symbol-name
    ├── 0_annotate-module.mlir
    ├── 10_compare-to-sign-rewrite.mlir
    ├── 11_canonicalize.mlir
    ├── 12_cse.mlir
    ├── 13_apply-folders.mlir
    ├── 14_canonicalize.mlir
    ├── 15_insert-rotate.mlir
    ├── 16_cse.mlir
    ├── 17_canonicalize.mlir
    ├── 18_cse.mlir
    ├── 19_collapse-insertion-chains.mlir
    ├── 1_debug-validate-names.mlir
    ├── 20_sccp.mlir
    ├── 21_canonicalize.mlir
    ...
    ├── func_func__assign_layout_9911323688340753831
    │   ├── 47_0_affine-expand-index-ops.mlir
    │   ├── 47_1_affine-simplify-structures.mlir
    │   ├── 47_2_affine-loop-normalize.mlir
    │   └── 47_3_forward-insert-slice-to-extract-slice.mlir
    ├── func_func_matvec
    │   ├── 47_0_affine-expand-index-ops.mlir
    │   ├── 47_1_affine-simplify-structures.mlir
    │   ├── 47_2_affine-loop-normalize.mlir
    │   └── 47_3_forward-insert-slice-to-extract-slice.mlir
    ├── func_func_matvec__decrypt__result0
    │   ├── 47_0_affine-expand-index-ops.mlir
    │   ├── 47_1_affine-simplify-structures.mlir
    │   ├── 47_2_affine-loop-normalize.mlir
    │   └── 47_3_forward-insert-slice-to-extract-slice.mlir
    └── func_func_matvec__encrypt__arg0
        ├── 47_0_affine-expand-index-ops.mlir
        ├── 47_1_affine-simplify-structures.mlir
        ├── 47_2_affine-loop-normalize.mlir
        └── 47_3_forward-insert-slice-to-extract-slice.mlir

Dumped file contents

At the top of each file, a comment is printed indicating how the file was generated, containing e.g., IR Dump Before CSEPass (cse). This will say whether the IR was dumped before or after a pass, as well as whether (in the case of --mlir-print-ir-after-failure) the pass failed.

Below the comment, the MLIR is printed in textual format.

Gotchas

• Stale files: --mlir-print-ir-tree-dir does not delete existing files, so if the command is re-run, be sure to remove the existing temporary files first.
• Dump before vs after: --mlir-print-ir-before-all prints the IR before a pass runs, which is required to extract the IR if a pass has a hard error like a segfault. To dump the IR after each pass, use --mlir-print-ir-after-all.
• Dumping after failure: For some errors without stack traces (like a dialect conversion failure where the compiler fails gracefully), you can use --mlir-print-ir-after-failure to dump the IR at failure time, which can provide additional debugging information. One notable example is when dialect conversion fails, the IR dumped by --mlir-print-ir-after-failure includes the partially-converted IR including any unrealized_conversion_cast ops that failed to be removed.
• Dumping after change: The number of files can be reduced by using --mlir-print-ir-after-change, which only prints the IR if the compiler pass changed it. There is no "before change" analogue.
• Dumping after change: The number of files can be reduced by using --mlir-print-ir-before=<pass_name>, which only prints the IR before the target pass. Since a pass can be run multiple times in a pipeline, this may still produce many files if, for example, the pass is a common one like cse.

Copy this checklist and track progress:

- [ ] Step 1: Identify the command to convert, perhaps with `lit-to-bazel`
- [ ] Step 2: Add the relevant flags, defaulting to `--mlir-print-ir-before-all` and `--mlir-print-ir-tree-dir` if there is no specific reason to use the more specific flags.
- [ ] Step 3: Inspect the generated command and run it.
- [ ] Step 4: Inspect the dumped files for information.

Source: google/heir — distributed by TomeVault.