visibility of `unnamed_addr` and `visibility_` when using `mlir-opt`.

experiment.py

# /// script
# requires-python = ">=3.11"
# dependencies = [
#     "polars",
# ]
# ///

import itertools
import re
import subprocess
import polars as pl


def run_mlir_opt(mlir_input: str) -> str:
    # Runs mlir-opt and returns generic output
    process = subprocess.run(
        ["mlir-opt", "--mlir-print-op-generic"],
        input=mlir_input,
        capture_output=True,
        text=True,
        check=True
    )
    return process.stdout


def extract_attr(output: str, attr_name: str) -> str:
    # Extracts the value of an attribute from generic MLIR output
    if output.startswith("ERROR:"):
        return "ERR"
    match = re.search(f"{attr_name} = (\\d+) : i64", output)
    return match.group(1) if match else "-"


# Templates
generic_template = '''"builtin.module"() ({{
"llvm.func"() ({{
"llvm.return"() : () -> ()
}}) {{
sym_name = "f",
function_type = !llvm.func<void ()>,
linkage = #llvm.linkage<external>,
CConv = #llvm.cconv<ccc>{extra_attrs}
}} : () -> ()
}}) : () -> ()
'''

# Custom format uses keywords, not i64 attributes:
#   visibility_: 0=default (no keyword), 1=hidden, 2=protected
#   unnamed_addr: 0=default (no keyword), 1=local_unnamed_addr, 2=unnamed_addr
VIS_KEYWORDS = {None: "", 0: "", 1: "hidden ", 2: "protected "}
UNNAMED_KEYWORDS = {None: "", 0: "", 1: "local_unnamed_addr ", 2: "unnamed_addr "}

custom_template = '''llvm.func {vis}{unnamed}@f() {{
  llvm.return
}}
'''

results = []
for fmt in ["generic", "custom"]:
    for vis, unnamed in itertools.product([None, 0, 1], repeat=2):
        # Build input based on format
        if fmt == "generic":
            attrs = []
            if vis is not None:
                attrs.append(f"visibility_ = {vis} : i64")
            if unnamed is not None:
                attrs.append(f"unnamed_addr = {unnamed} : i64")
            extra = (", " + ", ".join(attrs)) if attrs else ""
            mlir_input = generic_template.format(extra_attrs=extra)
        else:
            vis_kw = VIS_KEYWORDS[vis]
            unnamed_kw = UNNAMED_KEYWORDS[unnamed]
            mlir_input = custom_template.format(vis=vis_kw, unnamed=unnamed_kw)
        
        output = run_mlir_opt(mlir_input)
        results.append({
            "format": fmt,
            "in_vis": str(vis) if vis is not None else "-",
            "in_unnamed": str(unnamed) if unnamed is not None else "-",
            "out_vis": extract_attr(output, "visibility_"),
            "out_unnamed": extract_attr(output, "unnamed_addr"),
        })

df = pl.DataFrame(results)
with pl.Config(tbl_hide_column_data_types=True, tbl_rows=-1):
    print(df)

run.txt

$ uv run experiment.py
  shape: (18, 5)
  ┌─────────┬────────┬────────────┬─────────┬─────────────┐
  │ format  ┆ in_vis ┆ in_unnamed ┆ out_vis ┆ out_unnamed │
  ╞═════════╪════════╪════════════╪═════════╪═════════════╡
  │ generic ┆ -      ┆ -          ┆ 0       ┆ -           │
  │ generic ┆ -      ┆ 0          ┆ 0       ┆ 0           │
  │ generic ┆ -      ┆ 1          ┆ 0       ┆ 1           │
  │ generic ┆ 0      ┆ -          ┆ 0       ┆ -           │
  │ generic ┆ 0      ┆ 0          ┆ 0       ┆ 0           │
  │ generic ┆ 0      ┆ 1          ┆ 0       ┆ 1           │
  │ generic ┆ 1      ┆ -          ┆ 1       ┆ -           │
  │ generic ┆ 1      ┆ 0          ┆ 1       ┆ 0           │
  │ generic ┆ 1      ┆ 1          ┆ 1       ┆ 1           │
  │ custom  ┆ -      ┆ -          ┆ 0       ┆ 0           │
  │ custom  ┆ -      ┆ 0          ┆ 0       ┆ 0           │
  │ custom  ┆ -      ┆ 1          ┆ 0       ┆ 1           │
  │ custom  ┆ 0      ┆ -          ┆ 0       ┆ 0           │
  │ custom  ┆ 0      ┆ 0          ┆ 0       ┆ 0           │
  │ custom  ┆ 0      ┆ 1          ┆ 0       ┆ 1           │
  │ custom  ┆ 1      ┆ -          ┆ 1       ┆ 0           │
  │ custom  ┆ 1      ┆ 0          ┆ 1       ┆ 0           │
  │ custom  ┆ 1      ┆ 1          ┆ 1       ┆ 1           │
  └─────────┴────────┴────────────┴─────────┴─────────────┘


Conclusions:

- both the generic and custom format behave the same (but place the `visibility_` and `unnamed_addr` attributes differently.
- `visibility_`: always defaults to 0.
- `unnamed_addr`: always defaults to 0 in custom format. only printed if explicitly set in generic format.