[mlir][core] Add an MLIR "pattern catalog" generator #146228
Conversation
This PR adds a new cmake build option MLIR_ENABLE_CATALOG_GENERATOR. When enabled, it attaches a listener to all RewritePatterns that emits the name of the operation being modified to a special file. When the MLIR test suite is run, these files can be combined into an index linking operations to the patterns that insert, modify, or replace them. This index is intended to be used to create a website that allows one to look up patterns from an operation name.
llvmbot
added
mlir:core
|
@llvm/pr-subscribers-mlir-core Author: Jeremy Kun (j2kun) ChangesThis PR adds a new cmake build option MLIR_ENABLE_CATALOG_GENERATOR. When enabled, it attaches a listener to all RewritePatterns that emits the name of the operation being modified to a special file. When the MLIR test suite is run, these files can be combined into an index linking operations to the patterns that insert, modify, or replace them. This index is intended to be used to create a website that allows one to look up patterns from an operation name. Full diff: https://github.com/llvm/llvm-project/pull/146228.diff 5 Files Affected:
diff --git a/mlir/CMakeLists.txt b/mlir/CMakeLists.txt
index 44493b75b8a8c..bd30d94e1ccb4 100644
--- a/mlir/CMakeLists.txt
+++ b/mlir/CMakeLists.txt
@@ -202,6 +202,24 @@ if(MLIR_ENABLE_BINDINGS_PYTHON)
mlir_configure_python_dev_packages()
endif()
+#-------------------------------------------------------------------------------
+# MLIR Pattern Catalog Generator Configuration
+# Requires:
+# RTTI to be enabled (set with -DLLVM_ENABLE_RTTI=ON)
+# When enabled, causes all rewriter patterns to dump their type names and the
+# names of affected operations, which can be used to build a search index
+# mapping operations to patterns.
+#-------------------------------------------------------------------------------
+
+set(MLIR_ENABLE_CATALOG_GENERATOR 0 CACHE BOOL
+ "Enables construction of a catalog of rewrite patterns.")
+
+if (MLIR_ENABLE_CATALOG_GENERATOR)
+ message(STATUS "Enabling MLIR pattern catalog generator")
+ add_definitions(-DMLIR_ENABLE_CATALOG_GENERATOR)
+ add_definitions(-DLLVM_ENABLE_RTTI)
+endif()
+
set(CMAKE_INCLUDE_CURRENT_DIR ON)
include_directories(BEFORE
@@ -322,3 +340,4 @@ endif()
if(MLIR_STANDALONE_BUILD)
llvm_distribution_add_targets()
endif()
+
diff --git a/mlir/cmake/modules/MLIRConfig.cmake.in b/mlir/cmake/modules/MLIRConfig.cmake.in
index 71f3e028b1e88..f4ae70a22b3d2 100644
--- a/mlir/cmake/modules/MLIRConfig.cmake.in
+++ b/mlir/cmake/modules/MLIRConfig.cmake.in
@@ -16,6 +16,7 @@ set(MLIR_IRDL_TO_CPP_EXE "@MLIR_CONFIG_IRDL_TO_CPP_EXE@")
set(MLIR_INSTALL_AGGREGATE_OBJECTS "@MLIR_INSTALL_AGGREGATE_OBJECTS@")
set(MLIR_ENABLE_BINDINGS_PYTHON "@MLIR_ENABLE_BINDINGS_PYTHON@")
set(MLIR_ENABLE_EXECUTION_ENGINE "@MLIR_ENABLE_EXECUTION_ENGINE@")
+set(MLIR_ENABLE_CATALOG_GENERATOR "@MLIR_ENABLE_CATALOG_GENERATOR@")
set_property(GLOBAL PROPERTY MLIR_ALL_LIBS "@MLIR_ALL_LIBS@")
set_property(GLOBAL PROPERTY MLIR_DIALECT_LIBS "@MLIR_DIALECT_LIBS@")
diff --git a/mlir/include/mlir/IR/PatternMatch.h b/mlir/include/mlir/IR/PatternMatch.h
index 10cfe851765dc..141b3c6806ed8 100644
--- a/mlir/include/mlir/IR/PatternMatch.h
+++ b/mlir/include/mlir/IR/PatternMatch.h
@@ -475,6 +475,80 @@ class RewriterBase : public OpBuilder {
RewriterBase::Listener *rewriteListener;
};
+ struct CatalogingListener : public RewriterBase::ForwardingListener {
+ CatalogingListener(OpBuilder::Listener *listener,
+ const std::string &patternName, raw_ostream &os,
+ std::mutex &writeMutex)
+ : RewriterBase::ForwardingListener(listener), patternName(patternName),
+ os(os), writeMutex(writeMutex) {}
+
+ void notifyOperationInserted(Operation *op, InsertPoint previous) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationInserted"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationInserted(op, previous);
+ }
+
+ void notifyOperationModified(Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationModified"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationModified(op);
+ }
+
+ void notifyOperationReplaced(Operation *op, Operation *newOp) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationReplaced (with op)"
+ << " | " << op->getName() << " | " << newOp->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationReplaced(op, newOp);
+ }
+
+ void notifyOperationReplaced(Operation *op,
+ ValueRange replacement) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationReplaced (with values)"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationReplaced(op, replacement);
+ }
+
+ void notifyOperationErased(Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationErased"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationErased(op);
+ }
+
+ void notifyPatternBegin(const Pattern &pattern, Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyPatternBegin"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyPatternBegin(pattern, op);
+ }
+
+ private:
+ const std::string &patternName;
+ raw_ostream &os;
+ std::mutex &writeMutex;
+ };
+
/// Move the blocks that belong to "region" before the given position in
/// another region "parent". The two regions must be different. The caller
/// is responsible for creating or updating the operation transferring flow
diff --git a/mlir/lib/Rewrite/PatternApplicator.cpp b/mlir/lib/Rewrite/PatternApplicator.cpp
index 4a12183492fd4..c66aaf267881f 100644
--- a/mlir/lib/Rewrite/PatternApplicator.cpp
+++ b/mlir/lib/Rewrite/PatternApplicator.cpp
@@ -15,8 +15,19 @@
#include "ByteCode.h"
#include "llvm/Support/Debug.h"
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/raw_ostream.h"
+#include <cxxabi.h>
+#include <mutex>
+#endif
+
#define DEBUG_TYPE "pattern-application"
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+static std::mutex catalogWriteMutex;
+#endif
+
using namespace mlir;
using namespace mlir::detail;
@@ -152,6 +163,16 @@ LogicalResult PatternApplicator::matchAndRewrite(
unsigned anyIt = 0, anyE = anyOpPatterns.size();
unsigned pdlIt = 0, pdlE = pdlMatches.size();
LogicalResult result = failure();
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ std::error_code ec;
+ llvm::raw_fd_ostream catalogOs("pattern_catalog.txt", ec,
+ llvm::sys::fs::OF_Append);
+ if (ec) {
+ op->emitError("Failed to open pattern catalog file: " + ec.message());
+ return failure();
+ }
+#endif
+
do {
// Find the next pattern with the highest benefit.
const Pattern *bestPattern = nullptr;
@@ -206,14 +227,38 @@ LogicalResult PatternApplicator::matchAndRewrite(
} else {
LLVM_DEBUG(llvm::dbgs() << "Trying to match \""
<< bestPattern->getDebugName() << "\"\n");
-
const auto *pattern =
static_cast<const RewritePattern *>(bestPattern);
- result = pattern->matchAndRewrite(op, rewriter);
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ OpBuilder::Listener *oldListener = rewriter.getListener();
+ int status;
+ const char *mangledPatternName = typeid(*pattern).name();
+ char *demangled = abi::__cxa_demangle(mangledPatternName, nullptr,
+ nullptr, &status);
+ std::string demangledPatternName;
+ if (status == 0 && demangled) {
+ demangledPatternName = demangled;
+ free(demangled);
+ } else {
+ // Fallback in case demangling fails.
+ demangledPatternName = mangledPatternName;
+ }
+
+ RewriterBase::CatalogingListener *catalogingListener =
+ new RewriterBase::CatalogingListener(
+ oldListener, demangledPatternName, catalogOs,
+ catalogWriteMutex);
+ rewriter.setListener(catalogingListener);
+#endif
+ result = pattern->matchAndRewrite(op, rewriter);
LLVM_DEBUG(llvm::dbgs()
<< "\"" << bestPattern->getDebugName() << "\" result "
<< succeeded(result) << "\n");
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ rewriter.setListener(oldListener);
+ delete catalogingListener;
+#endif
}
// Process the result of the pattern application.
diff --git a/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel b/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
index 23d89f41a3a45..281b2566304b0 100644
--- a/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
@@ -44,6 +44,7 @@ expand_template(
"@MLIR_ENABLE_SPIRV_CPU_RUNNER@": "0",
"@MLIR_ENABLE_VULKAN_RUNNER@": "0",
"@MLIR_ENABLE_BINDINGS_PYTHON@": "0",
+ "@MLIR_ENABLE_CATALOG_GENERATOR@": "0",
"@MLIR_RUN_AMX_TESTS@": "0",
"@MLIR_RUN_ARM_SVE_TESTS@": "0",
"@MLIR_RUN_ARM_SME_TESTS@": "0",
|
|
@llvm/pr-subscribers-mlir Author: Jeremy Kun (j2kun) ChangesThis PR adds a new cmake build option MLIR_ENABLE_CATALOG_GENERATOR. When enabled, it attaches a listener to all RewritePatterns that emits the name of the operation being modified to a special file. When the MLIR test suite is run, these files can be combined into an index linking operations to the patterns that insert, modify, or replace them. This index is intended to be used to create a website that allows one to look up patterns from an operation name. Full diff: https://github.com/llvm/llvm-project/pull/146228.diff 5 Files Affected:
diff --git a/mlir/CMakeLists.txt b/mlir/CMakeLists.txt
index 44493b75b8a8c..bd30d94e1ccb4 100644
--- a/mlir/CMakeLists.txt
+++ b/mlir/CMakeLists.txt
@@ -202,6 +202,24 @@ if(MLIR_ENABLE_BINDINGS_PYTHON)
mlir_configure_python_dev_packages()
endif()
+#-------------------------------------------------------------------------------
+# MLIR Pattern Catalog Generator Configuration
+# Requires:
+# RTTI to be enabled (set with -DLLVM_ENABLE_RTTI=ON)
+# When enabled, causes all rewriter patterns to dump their type names and the
+# names of affected operations, which can be used to build a search index
+# mapping operations to patterns.
+#-------------------------------------------------------------------------------
+
+set(MLIR_ENABLE_CATALOG_GENERATOR 0 CACHE BOOL
+ "Enables construction of a catalog of rewrite patterns.")
+
+if (MLIR_ENABLE_CATALOG_GENERATOR)
+ message(STATUS "Enabling MLIR pattern catalog generator")
+ add_definitions(-DMLIR_ENABLE_CATALOG_GENERATOR)
+ add_definitions(-DLLVM_ENABLE_RTTI)
+endif()
+
set(CMAKE_INCLUDE_CURRENT_DIR ON)
include_directories(BEFORE
@@ -322,3 +340,4 @@ endif()
if(MLIR_STANDALONE_BUILD)
llvm_distribution_add_targets()
endif()
+
diff --git a/mlir/cmake/modules/MLIRConfig.cmake.in b/mlir/cmake/modules/MLIRConfig.cmake.in
index 71f3e028b1e88..f4ae70a22b3d2 100644
--- a/mlir/cmake/modules/MLIRConfig.cmake.in
+++ b/mlir/cmake/modules/MLIRConfig.cmake.in
@@ -16,6 +16,7 @@ set(MLIR_IRDL_TO_CPP_EXE "@MLIR_CONFIG_IRDL_TO_CPP_EXE@")
set(MLIR_INSTALL_AGGREGATE_OBJECTS "@MLIR_INSTALL_AGGREGATE_OBJECTS@")
set(MLIR_ENABLE_BINDINGS_PYTHON "@MLIR_ENABLE_BINDINGS_PYTHON@")
set(MLIR_ENABLE_EXECUTION_ENGINE "@MLIR_ENABLE_EXECUTION_ENGINE@")
+set(MLIR_ENABLE_CATALOG_GENERATOR "@MLIR_ENABLE_CATALOG_GENERATOR@")
set_property(GLOBAL PROPERTY MLIR_ALL_LIBS "@MLIR_ALL_LIBS@")
set_property(GLOBAL PROPERTY MLIR_DIALECT_LIBS "@MLIR_DIALECT_LIBS@")
diff --git a/mlir/include/mlir/IR/PatternMatch.h b/mlir/include/mlir/IR/PatternMatch.h
index 10cfe851765dc..141b3c6806ed8 100644
--- a/mlir/include/mlir/IR/PatternMatch.h
+++ b/mlir/include/mlir/IR/PatternMatch.h
@@ -475,6 +475,80 @@ class RewriterBase : public OpBuilder {
RewriterBase::Listener *rewriteListener;
};
+ struct CatalogingListener : public RewriterBase::ForwardingListener {
+ CatalogingListener(OpBuilder::Listener *listener,
+ const std::string &patternName, raw_ostream &os,
+ std::mutex &writeMutex)
+ : RewriterBase::ForwardingListener(listener), patternName(patternName),
+ os(os), writeMutex(writeMutex) {}
+
+ void notifyOperationInserted(Operation *op, InsertPoint previous) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationInserted"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationInserted(op, previous);
+ }
+
+ void notifyOperationModified(Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationModified"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationModified(op);
+ }
+
+ void notifyOperationReplaced(Operation *op, Operation *newOp) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationReplaced (with op)"
+ << " | " << op->getName() << " | " << newOp->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationReplaced(op, newOp);
+ }
+
+ void notifyOperationReplaced(Operation *op,
+ ValueRange replacement) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationReplaced (with values)"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationReplaced(op, replacement);
+ }
+
+ void notifyOperationErased(Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyOperationErased"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyOperationErased(op);
+ }
+
+ void notifyPatternBegin(const Pattern &pattern, Operation *op) override {
+ {
+ std::lock_guard<std::mutex> lock(writeMutex);
+ os << patternName << " | notifyPatternBegin"
+ << " | " << op->getName() << "\n";
+ os.flush();
+ }
+ ForwardingListener::notifyPatternBegin(pattern, op);
+ }
+
+ private:
+ const std::string &patternName;
+ raw_ostream &os;
+ std::mutex &writeMutex;
+ };
+
/// Move the blocks that belong to "region" before the given position in
/// another region "parent". The two regions must be different. The caller
/// is responsible for creating or updating the operation transferring flow
diff --git a/mlir/lib/Rewrite/PatternApplicator.cpp b/mlir/lib/Rewrite/PatternApplicator.cpp
index 4a12183492fd4..c66aaf267881f 100644
--- a/mlir/lib/Rewrite/PatternApplicator.cpp
+++ b/mlir/lib/Rewrite/PatternApplicator.cpp
@@ -15,8 +15,19 @@
#include "ByteCode.h"
#include "llvm/Support/Debug.h"
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/raw_ostream.h"
+#include <cxxabi.h>
+#include <mutex>
+#endif
+
#define DEBUG_TYPE "pattern-application"
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+static std::mutex catalogWriteMutex;
+#endif
+
using namespace mlir;
using namespace mlir::detail;
@@ -152,6 +163,16 @@ LogicalResult PatternApplicator::matchAndRewrite(
unsigned anyIt = 0, anyE = anyOpPatterns.size();
unsigned pdlIt = 0, pdlE = pdlMatches.size();
LogicalResult result = failure();
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ std::error_code ec;
+ llvm::raw_fd_ostream catalogOs("pattern_catalog.txt", ec,
+ llvm::sys::fs::OF_Append);
+ if (ec) {
+ op->emitError("Failed to open pattern catalog file: " + ec.message());
+ return failure();
+ }
+#endif
+
do {
// Find the next pattern with the highest benefit.
const Pattern *bestPattern = nullptr;
@@ -206,14 +227,38 @@ LogicalResult PatternApplicator::matchAndRewrite(
} else {
LLVM_DEBUG(llvm::dbgs() << "Trying to match \""
<< bestPattern->getDebugName() << "\"\n");
-
const auto *pattern =
static_cast<const RewritePattern *>(bestPattern);
- result = pattern->matchAndRewrite(op, rewriter);
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ OpBuilder::Listener *oldListener = rewriter.getListener();
+ int status;
+ const char *mangledPatternName = typeid(*pattern).name();
+ char *demangled = abi::__cxa_demangle(mangledPatternName, nullptr,
+ nullptr, &status);
+ std::string demangledPatternName;
+ if (status == 0 && demangled) {
+ demangledPatternName = demangled;
+ free(demangled);
+ } else {
+ // Fallback in case demangling fails.
+ demangledPatternName = mangledPatternName;
+ }
+
+ RewriterBase::CatalogingListener *catalogingListener =
+ new RewriterBase::CatalogingListener(
+ oldListener, demangledPatternName, catalogOs,
+ catalogWriteMutex);
+ rewriter.setListener(catalogingListener);
+#endif
+ result = pattern->matchAndRewrite(op, rewriter);
LLVM_DEBUG(llvm::dbgs()
<< "\"" << bestPattern->getDebugName() << "\" result "
<< succeeded(result) << "\n");
+#ifdef MLIR_ENABLE_CATALOG_GENERATOR
+ rewriter.setListener(oldListener);
+ delete catalogingListener;
+#endif
}
// Process the result of the pattern application.
diff --git a/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel b/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
index 23d89f41a3a45..281b2566304b0 100644
--- a/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
+++ b/utils/bazel/llvm-project-overlay/mlir/test/BUILD.bazel
@@ -44,6 +44,7 @@ expand_template(
"@MLIR_ENABLE_SPIRV_CPU_RUNNER@": "0",
"@MLIR_ENABLE_VULKAN_RUNNER@": "0",
"@MLIR_ENABLE_BINDINGS_PYTHON@": "0",
+ "@MLIR_ENABLE_CATALOG_GENERATOR@": "0",
"@MLIR_RUN_AMX_TESTS@": "0",
"@MLIR_RUN_ARM_SVE_TESTS@": "0",
"@MLIR_RUN_ARM_SME_TESTS@": "0",
|
|
|
j2kun
changed the title
| OpBuilder::Listener *oldListener = rewriter.getListener(); | ||
| int status; | ||
| const char *mangledPatternName = typeid(*pattern).name(); | ||
| char *demangled = abi::__cxa_demangle(mangledPatternName, nullptr, | ||
| nullptr, &status); | ||
| std::string demangledPatternName; | ||
| if (status == 0 && demangled) { | ||
| demangledPatternName = demangled; | ||
| free(demangled); | ||
| } else { | ||
| // Fallback in case demangling fails. | ||
| demangledPatternName = mangledPatternName; | ||
| } |
There was a problem hiding this comment.
can't you just do pattern->getDebugName()? incidentally that name is set using llvm::getTypeName (that's how it avoids demangling...).
There was a problem hiding this comment.
I can do this, but roughly 1% of the catalog (177/12632) corresponds to patterns with no debug name (checking now if this is simply the same set of patterns that fails to demangle)
There was a problem hiding this comment.
It looks like the demangling differences are pretty significant in the two implementations. For example, RTTI can get this one:
:mlir::RewritePatternSet::add<mlir::acc::AtomicUpdateOp>(llvm::LogicalResult (*)(mlir::acc::AtomicUpdateOp, mlir::PatternRewriter&), mlir::PatternBenefit, llvm::ArrayRef<llvm::StringRef>)::FnPattern | notifyOperationErased | acc.atomic.update
But that corresponds to an empty debugName. I think this corresponds to canonicalization patterns that are automatically added during mlir-tablegen via this code path
There was a problem hiding this comment.
Some of the other (minor) differences I see:
- cxx demangle includes default template attributes (e.g., mlir::AttrConvertPassThrough in VectorConvertToLLVMPattern) while debugName omits them.
- debugName gives the name of an enum, while cxx demangle only has the ordinal (e.g.,
mlir::ReshapeOpKind::kCollapsevs(mlir::ReshapeOpKind)1) - debugName properly demangles the instances mentioned in the PR description that fail under cxx demangle.
Overall debugName seems better, so the main thing I want to figure out is how to get it to handle those canonicalize methods. Maybe the solution is simply to rewrite those canonicalize methods as proper patterns?
|
I support this with every fiber of my being! Bravo. My only comment (aside from |
I had originally tried something like this, and my obstacle was that I couldn't get the test suite to produce useful outputs because stuff was piped to FileCheck (when I tried just dumping to llvm::outs()), and I didn't have a way to modify the RUN command of all the lit tests in a systematic/pragmatic way to add a debug-only flag. |
If you make |
|
err sorry - you'd have to get |
|
Is not relying on a filesystem all that critical? The mutex is mildly annoying, but it's scoped pretty tightly, and the alternative requires changes in a few other components (wouldn't I also need to plumb these environment variables through the lit config?) and additional cleanup of the output. I agree setting environment variables would be the cleanest approach. I'm just not confident in my ability to push a big change to |
I personally don't care (I think it's a reasonable use) - maybe someone else can comment on whether this kind of thing is acceptable. |
|
Spot checking a few examples, it looks like I also need to hook into the dialect conversion framework. |
This PR adds a new cmake build option MLIR_ENABLE_CATALOG_GENERATOR. When enabled, it attaches a listener to all RewritePatterns that emits the name of the operation being modified to a special file. When the MLIR test suite is run, these files can be combined into an index linking operations to the patterns that insert, modify, or replace them. This index is intended to be used to create a website that allows one to look up patterns from an operation name.
Example usage:
Sample pattern catalog output (that fits in a gist): https://gist.github.com/j2kun/02d1ab8d31c10d71027724984c89905a
[edit]: removed use of RTTI