Lowering Through LLVM (#26)

* add poly_to_llvm test file

* add starter poly-to-llvm pipeline

* add convert-func-to-llvm

* add arith-to-llvm

* add elementwise-to-linalg

* add back arith-to-llvm

* add tensor-to-linalg

* add do-nothing linalg-to-loops pass

* add scf-to-cf and cf-to-llvm passes

* update upstream MLIR commit

* fix conj pattern broken by mlir update

* migrate use of constFoldBinaryOp

* fix conflict caused by new function overload

* move arith-to-llvm to the end of the pass

* add one-shot bufferization pass

* move func-to-llvm after bufferization, enabling linalg-to-loops

* memref-expand-strided-metadata and finalize-memref-to-llvm

* move func-to-llvm even later

* add canonicalization cleanup

* encode end-to-end compilation pipeline as a test

* add a poly.eval-specific end-to-end test

* ensure llc uses the same PositionIndependentExecutable option as clang

* fix the off-by-one error in the lowering of eval

* udpate cmake build to use same LLVM commit

* update subproject commit

* add project_src_dir so tests work with both bazel and cmake

---------

Co-authored-by: Jeremy Kun <j2kun@users.noreply.github.com>

Author: j2kun

.github/workflows/build_and_test_cmake.yml

@@ -27,7 +27,7 @@ jobs:
         with:
           path: |
             ./externals/llvm-project
-          key: ${{ runner.os }}-norm-${{ hashFiles('**/CMakeLists.txt') }}
+          key: ${{ runner.os }}-cmake-${{ hashFiles('bazel/import_llvm.bzl') }}-${{ hashFiles('**/CMakeLists.txt') }}
 
       - name: Git config
         run: |
@@ -36,8 +36,10 @@ jobs:
       - name: Build LLVM
         if: steps.cache-llvm.outputs.cache-hit != 'true'
         run: |
+          LLVM_COMMIT=$(grep LLVM_COMMIT ${GITHUB_WORKSPACE}/bazel/import_llvm.bzl | head -n 1 | cut -d'"' -f 2 )
           git submodule update --init --recursive
           cd externals/llvm-project
+          git checkout ${LLVM_COMMIT}
           mkdir build && cd build
           cmake -G Ninja ../llvm -DLLVM_ENABLE_PROJECTS=mlir -DLLVM_BUILD_EXAMPLES=ON -DLLVM_ENABLE_ASSERTIONS=ON -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_RTTI=ON -DLLVM_TARGETS_TO_BUILD="host"
           cmake --build . --target check-mlir
@@ -50,4 +52,4 @@ jobs:
           cmake --build . --target MLIRMulToAddPasses
           cmake --build . --target mlir-headers
           cmake --build . --target tutorial-opt
-          cmake --build . --target check-mlir-tutorial
+          cmake --build . --target check-mlir-tutorial

bazel/import_llvm.bzl

@@ -8,8 +8,8 @@ load(
 def import_llvm(name):
     """Imports LLVM."""
 
-    # June 5, 2023
-    LLVM_COMMIT = "cd5fcea6d4c70a7328ca9538c9098d9f5af69682"
+    # 2023-10-30
+    LLVM_COMMIT = "896749aa0d420ae573255a64a349bc2a76cfed37"
 
     new_git_repository(
         name = name,

externals/llvm-project

@@ -48,7 +48,7 @@ struct ConvertAdd : public OpConversionPattern<AddOp> {
       ConversionPatternRewriter &rewriter) const override {
     arith::AddIOp addOp = rewriter.create<arith::AddIOp>(
         op.getLoc(), adaptor.getLhs(), adaptor.getRhs());
-    rewriter.replaceOp(op.getOperation(), {addOp});
+    rewriter.replaceOp(op.getOperation(), addOp);
     return success();
   }
 };
@@ -64,7 +64,7 @@ struct ConvertSub : public OpConversionPattern<SubOp> {
       ConversionPatternRewriter &rewriter) const override {
     arith::SubIOp subOp = rewriter.create<arith::SubIOp>(
         op.getLoc(), adaptor.getLhs(), adaptor.getRhs());
-    rewriter.replaceOp(op.getOperation(), {subOp});
+    rewriter.replaceOp(op.getOperation(), subOp);
     return success();
   }
 };
@@ -149,6 +149,8 @@ struct ConvertEval : public OpConversionPattern<EvalOp> {
     auto lowerBound =
         b.create<arith::ConstantOp>(b.getIndexType(), b.getIndexAttr(1));
     auto numTermsOp = b.create<arith::ConstantOp>(b.getIndexType(),
+                                                  b.getIndexAttr(numTerms));
+    auto upperBound = b.create<arith::ConstantOp>(b.getIndexType(),
                                                   b.getIndexAttr(numTerms + 1));
     auto step = lowerBound;
 
@@ -163,7 +165,7 @@ struct ConvertEval : public OpConversionPattern<EvalOp> {
     auto accum =
         b.create<arith::ConstantOp>(b.getI32Type(), b.getI32IntegerAttr(0));
     auto loop = b.create<scf::ForOp>(
-        lowerBound, numTermsOp, step, accum.getResult(),
+        lowerBound, upperBound, step, accum.getResult(),
         [&](OpBuilder &builder, Location loc, Value loopIndex,
             ValueRange loopState) {
           ImplicitLocOpBuilder b(op.getLoc(), builder);

lib/Conversion/PolyToStandard/PolyToStandard.cpp

@@ -16,12 +16,12 @@ OpFoldResult ConstantOp::fold(ConstantOp::FoldAdaptor adaptor) {
 }
 
 OpFoldResult AddOp::fold(AddOp::FoldAdaptor adaptor) {
-  return constFoldBinaryOp<IntegerAttr, APInt>(
+  return constFoldBinaryOp<IntegerAttr, APInt, void>(
       adaptor.getOperands(), [&](APInt a, APInt b) { return a + b; });
 }
 
 OpFoldResult SubOp::fold(SubOp::FoldAdaptor adaptor) {
-  return constFoldBinaryOp<IntegerAttr, APInt>(
+  return constFoldBinaryOp<IntegerAttr, APInt, void>(
       adaptor.getOperands(), [&](APInt a, APInt b) { return a - b; });
 }
 

lib/Dialect/Poly/PolyOps.cpp

@@ -6,8 +6,8 @@ include "mlir/Dialect/Complex/IR/ComplexOps.td"
 include "mlir/IR/PatternBase.td"
 
 def LiftConjThroughEval : Pat<
-  (Poly_EvalOp $f, (ConjOp $z)),
-  (ConjOp (Poly_EvalOp $f, $z))
+  (Poly_EvalOp $f, (ConjOp $z, $fastmath)),
+  (ConjOp (Poly_EvalOp $f, $z), $fastmath)
 >;
 
 def HasOneUse: Constraint<CPred<"$_self.hasOneUse()">, "has one use">;

lib/Dialect/Poly/PolyPatterns.td

@@ -46,7 +46,7 @@ struct PowerOfTwoExpand : public OpRewritePattern<MulIOp> {
     MulIOp newMul = rewriter.create<MulIOp>(op.getLoc(), lhs, newConstant);
     AddIOp newAdd = rewriter.create<AddIOp>(op.getLoc(), newMul, newMul);
 
-    rewriter.replaceOp(op, {newAdd});
+    rewriter.replaceOp(op, newAdd);
     rewriter.eraseOp(rhsDefiningOp);
 
     return success();
@@ -79,7 +79,7 @@ struct PeelFromMul : public OpRewritePattern<MulIOp> {
     MulIOp newMul = rewriter.create<MulIOp>(op.getLoc(), lhs, newConstant);
     AddIOp newAdd = rewriter.create<AddIOp>(op.getLoc(), newMul, lhs);
 
-    rewriter.replaceOp(op, {newAdd});
+    rewriter.replaceOp(op, newAdd);
     rewriter.eraseOp(rhsDefiningOp);
 
     return success();

lib/Transform/Arith/MulToAdd.cpp

@@ -6,12 +6,16 @@ filegroup(
     testonly = True,
     data = [
         "//tests:lit.cfg.py",
+        "//tests:poly_to_llvm_main.c",
         "//tools:tutorial-opt",
+        "@llvm-project//clang:clang",
         "@llvm-project//llvm:FileCheck",
         "@llvm-project//llvm:count",
+        "@llvm-project//llvm:llc",
         "@llvm-project//llvm:not",
         "@llvm-project//mlir:mlir-cpu-runner",
         "@llvm-project//mlir:mlir-opt",
+        "@llvm-project//mlir:mlir-translate",
         "@mlir_tutorial_pip_deps_lit//:pkg",
     ],
 )

tests/BUILD

@@ -40,3 +40,8 @@
     + ":"
     + os.environ["PATH"]
 )
+
+substitutions = {
+    "%project_source_dir": runfiles_dir.joinpath(Path('mlir_tutorial')),
+}
+config.substitutions.extend(substitutions.items())

tests/lit.cfg.py

@@ -25,6 +25,7 @@
 
 config.substitutions.append(("%PATH%", config.environment["PATH"]))
 config.substitutions.append(("%shlibext", config.llvm_shlib_ext))
+config.substitutions.append(("%project_source_dir", config.project_source_dir))
 
 llvm_config.with_system_environment(["HOME", "INCLUDE", "LIB", "TMP", "TEMP"])
 
@@ -49,4 +50,4 @@
     "tutorial-opt"
 ]
 
-llvm_config.add_tool_substitutions(tools, tool_dirs)
+llvm_config.add_tool_substitutions(tools, tool_dirs)