Add pattern catalog page to MLIR website (#230)

This PR adds a basic pattern-search index to the MLIR website.

## Features:

- Autocomplete operation name on search input
- Links to GitHub search for the class name in `path:mlir`
- Clicking on a search result shows a popup with the class and all the
operations it touches (each of which is a link to the search)
- Filters by C++ namespace of the pattern and the "modification type"
(insert, remove, replace) the pattern applies to that op
- Support regex inputs for operation name (e.g., `.*\.for`)

## Screenshots

<img width="1580" height="1308" alt="image"
src="https://github.com/user-attachments/assets/d70f60cc-de86-4d3d-ad13-e0b43e8d57a3"
/>


<img width="1598" height="1010" alt="image"
src="https://github.com/user-attachments/assets/eaf25e57-720b-4022-9a3a-d4aa114f969b"
/>

---------

Co-authored-by: Jeremy Kun <[email protected]>
Co-authored-by: Mehdi Amini <[email protected]>

Author: j2kun

.github/workflows/main.yml

@@ -31,7 +31,7 @@ jobs:
       run: ./mlir-www-helper.sh --help
 
     - name: Install dependencies
-      run: sudo apt-get install doxygen graphviz
+      run: sudo apt-get install doxygen graphviz python3
 
     - name: Clone LLVM repo
       uses: actions/checkout@v3
@@ -44,8 +44,10 @@ jobs:
       run: |
         mkdir llvm_src/build
         cd llvm_src/build
+        # Assert mode is needed for the pattern search index
         cmake ../llvm \
           -DCMAKE_BUILD_TYPE=Release \
+          -DLLVM_ENABLE_ASSERTIONS=ON \
           -DLLVM_ENABLE_PROJECTS=mlir \
           -DLLVM_TARGETS_TO_BUILD="host" \
           -DMLIR_INCLUDE_DOCS="true" \
@@ -55,6 +57,25 @@ jobs:
     - name: Install docs
       run: ./mlir-www-helper.sh --install-docs "llvm_src" "website"
 
+    - name: Build pattern search index
+      run: |
+        cd llvm_src/build
+        # Run the tests once to ensure they are passing.
+        # The next run used to generate the patterns log is disabling error checking. This first
+        # run will ensure we don't generate the website when tests are broken.
+        make -j$(nproc) check-mlir
+        echo "finished building tests"
+        # ignore nonzero exit codes in this command, since MLIR_GENERATE_PATTERN_CATALOG
+        # implicitly enables`--mlir-disable-threading` which some tests are incompatible with.
+        { MLIR_GENERATE_PATTERN_CATALOG=1 bin/llvm-lit -j$(nproc) -v -a tools/mlir/test || true ; } \
+          | grep '# | \[pattern-logging-listener' \
+          | sed 's/# | \[pattern-logging-listener:[0-9]*\] PatternLoggingListener.cpp:[0-9]* //g' \
+          | sort | uniq > $HOME/pattern_catalog.txt
+        # back to mlir-www root
+        cd ../..
+        head $HOME/pattern_catalog.txt
+        python3 build_pattern_catalog_index.py $HOME/pattern_catalog.txt website/static/pattern-search/pattern-index.json
+
     - name: Setup Hugo
       uses: peaceiris/actions-hugo@v2
       with:

build_pattern_catalog_index.py

@@ -0,0 +1,147 @@
+#!/usr/bin/env python3
+"""
+Convert pipe-delimited text file to JSON format for MLIR operation search.
+Usage: python convert_to_json.py input.txt output.json
+"""
+
+from collections import defaultdict
+import json
+import os
+import re
+import sys
+
+
+def extract_namespace(class_name):
+    """Extract namespace from qualified class name."""
+    # Remove template parameters and anonymous namespace markers
+    cleaned = re.sub(r"<[^>]*>", "", class_name)
+    cleaned = re.sub(r"\(anonymous namespace\)::", "", cleaned)
+
+    # Split by :: and take all but the last part
+    parts = cleaned.split("::")
+    if len(parts) > 1:
+        return "::".join(parts[:-1])
+    return ""
+
+
+def fix_method_name(method):
+    """Convert from the logger's method name to a UI-friendly name."""
+    if method == "notifyOperationReplaced (with op)":
+        return "replace op with new op"
+    if method == "notifyOperationReplaced (with values)":
+        return "replace op with values"
+    if method == "notifyOperationErased":
+        return "erase op"
+    if method == "notifyOperationInserted":
+        return "insert op"
+    if method == "notifyOperationModified":
+        return "modify op"
+
+
+def parse_line(line):
+    """Parse a single line from the input file."""
+    parts = [part.strip() for part in line.strip().split("|")]
+
+    if len(parts) < 3 or parts[0].strip() == "":
+        return None
+
+    class_name = parts[0]
+    method = parts[1]
+    operations = parts[2:]
+
+    namespace = extract_namespace(class_name)
+    method = fix_method_name(method)
+
+    return {
+        "className": class_name,
+        "namespace": namespace,
+        "method": method,
+        "operations": operations,
+    }
+
+
+def build_reverse_index(entries):
+    """Build reverse index from operation names to class entries."""
+    index = defaultdict(list)
+
+    for entry in entries:
+        for op in entry["operations"]:
+            index[op].append(
+                {
+                    "className": entry["className"],
+                    "namespace": entry["namespace"],
+                    "method": entry["method"],
+                    "operations": entry["operations"],
+                }
+            )
+
+    return dict(index)
+
+
+def extract_metadata(entries):
+    """Extract unique namespaces and methods for filtering."""
+    namespaces = set()
+    methods = set()
+
+    for entry in entries:
+        if entry["namespace"]:
+            namespaces.add(entry["namespace"])
+        methods.add(entry["method"])
+
+    return {"namespaces": sorted(list(namespaces)), "methods": sorted(list(methods))}
+
+
+def main():
+    if len(sys.argv) != 3:
+        print("Usage: python convert_to_json.py input.txt output.json")
+        sys.exit(1)
+
+    input_file = sys.argv[1]
+    output_file = sys.argv[2]
+
+    entries = []
+    lines = []
+
+    try:
+        with open(input_file, "r", encoding="utf-8") as f:
+            for line_num, line in enumerate(f, 1):
+                if not line.strip():
+                    continue
+                lines.append((line_num, line))
+    except FileNotFoundError:
+        print(f"Error: Input file '{input_file}' not found")
+        sys.exit(1)
+
+    for line_num, line in lines:
+        entry = parse_line(line)
+        if entry:
+            entries.append(entry)
+        else:
+            print(f"Warning: Skipped malformed line {line_num}: {line.strip()}")
+
+    print(f"Parsed {len(entries)} entries")
+
+    # Build reverse index
+    operations_index = build_reverse_index(entries)
+    print(f"Created index for {len(operations_index)} operations")
+
+    # Extract metadata
+    metadata = extract_metadata(entries)
+    print(
+        f"Found {len(metadata['namespaces'])} namespaces and {len(metadata['methods'])} methods"
+    )
+
+    # Create final JSON structure
+    result = {"operations": operations_index, "metadata": metadata}
+
+    directory = os.path.dirname(output_file)
+    os.makedirs(directory, exist_ok=True)
+    # Write to output file
+    with open(output_file, "w", encoding="utf-8") as f:
+        json.dump(result, f, indent=2, ensure_ascii=False)
+
+    print(f"Successfully wrote JSON to {output_file}")
+
+
+if __name__ == "__main__":
+    main()

website/content/docs/PatternSearch.md

@@ -0,0 +1,22 @@
+---
+title: "Pattern Search"
+date: 1970-01-01T00:00:00Z
+draft: false
+---
+
+Search for an operation name below (e.g., `linalg.matvec`) to find a list of
+all Pattern classes that insert, erase, or modify that operation.
+
+The search index used here is generated by running all `lit` tests in MLIR with
+the additional flag `--pattern-logging-listener`, which produces a record of
+each rewrite operation and the names of the operations it affected.
+
+Caveats:
+
+ - Some records may be missing due to incomplete test coverage.
+ - Patterns are only indexed if they define a `debugName`, and while most
+   patterns have one, they are not required to.
+ - Patterns defined in PDLL are not supported.
+ - Dialect Conversion patterns are only partially supported (only insertions).
+
+{{< pattern-search >}}