| Issue |
170255
|
| Summary |
[MLIR] `-split-input-file` support for `mlir-runner`
|
| Labels |
mlir
|
| Assignees |
|
| Reporter |
lemoncmd
|
### Summary
`mlir-opt` supports the `-split-input-file` option, which makes it convenient to keep multiple similar tests/benchmarks in a single `.mlir` file. However, `mlir-runner` does not support processing such split-input-file style inputs, which makes it hard to write integration tests that mirror how we usually structure MLIR tests.
In particular, when writing integration tests that run something like:
```sh
mlir-opt pipeline.mlir | mlir-runner ...
```
there is currently no “normal” way to handle multiple `// -----`-separated modules in a single file. We can hack around it with POSIX tools, but it would be much nicer if `mlir-runner` could consume split inputs directly.
### Current behavior
* For unit tests, it’s common to have many similar tests in one file, e.g.:
```sh
mlir-opt %s -split-input-file -some-pass ...
```
`mlir-opt` automatically splits on `// -----` markers and runs the pass pipeline per chunk.
* For integration tests, we typically do something like:
```sh
mlir-opt %s -some-pass... | \
mlir-runner -e main -entry-point-result=void ...
```
* If `input.mlir` contains multiple `// -----`-separated modules (for benchmarking multiple similar cases in one file), there is no built-in way to have `mlir-runner` execute each split chunk. `mlir-runner` expects a single module on stdin.
* The only way today is to:
* either manually split the file,
* or use shell tricks / POSIX tools to emulate `-split-input-file`, and then call `mlir-runner` per split file.
This works but is awkward and makes integration tests less straightforward and less portable.
### Expected / desired behavior
It would be very helpful if `mlir-runner` could consume split-input-file style inputs directly, similar to `mlir-opt`. For example, one of the following would work:
1. **`-split-input-file` support in `mlir-runner` itself**, so we could write:
```sh
mlir-opt -pass-pipeline='...' my_benchmarks.mlir \
| mlir-runner -split-input-file -e main -entry-point-result=void ...
```
and have `mlir-runner` execute each chunk independently.
2. Or, some other option / mechanism that allows `mlir-runner` to:
* accept multiple modules separated by `// -----` on stdin, and
* either run them sequentially, or fail with a clear error if that’s not supported.
The key point is: **I want to keep multiple similar benchmarks in a single `.mlir` file and still be able to run them via `mlir-runner` in a “standard” way, without resorting to custom splitting scripts.**
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
