PDLL article (#43)

* update llvm commit hash to 2024-07-12

* update implementation for MLIR version bump

* add dedicated test file

* add first attempt at porting MulToAdd to PDLL

* add LHS/RHS versions of power of two pattern

* add PeelFromMul

* add Cmake build

---------

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

Author: j2kun

bazel/import_llvm.bzl

@@ -8,8 +8,8 @@ load(
 def import_llvm(name):
     """Imports LLVM."""
 
-    # 2023-11-13
-    LLVM_COMMIT = "f778eafdd878e8b11ad76f9e0a312ce7791a7481"
+    # 2024-07-12
+    LLVM_COMMIT = "0913547d0e3939cc420e88ecd037240f33736820"
 
     new_git_repository(
         name = name,

externals/llvm-project

@@ -42,7 +42,7 @@ ReduceNoiseAnalysis::ReduceNoiseAnalysis(Operation *op) {
   op->walk([&](Operation *op) {
     // FIXME: assumes all reduce_noise ops have already been removed and their
     // values forwarded.
-    if (!llvm::isa<noisy::AddOp, noisy::SubOp, noisy::MulOp>(op)) {
+    if (!isa<noisy::AddOp, noisy::SubOp, noisy::MulOp>(op)) {
       return;
     }
 
@@ -81,8 +81,8 @@ ReduceNoiseAnalysis::ReduceNoiseAnalysis(Operation *op) {
     // In the tutorial, there is no control flow, so these are the function
     // arguments of the main function being analyzed. A real compiler would
     // need to handle this more generically.
-    if (value.isa<BlockArgument>() ||
-        llvm::isa<noisy::EncodeOp>(value.getDefiningOp())) {
+    if (isa<BlockArgument>(value) ||
+        isa<noisy::EncodeOp>(value.getDefiningOp())) {
       MPConstraint *const ct =
           solver->MakeRowConstraint(INITIAL_NOISE, INITIAL_NOISE, "");
       ct->SetCoefficient(var, 1);

lib/Analysis/ReduceNoiseAnalysis/ReduceNoiseAnalysis.cpp

@@ -13,7 +13,7 @@ include "mlir/Interfaces/SideEffectInterfaces.td"
 class Noisy_BinOp<string mnemonic> : Op<Noisy_Dialect, mnemonic, [
     Pure,
     SameOperandsAndResultType,
-    DeclareOpInterfaceMethods<InferIntRangeInterface>
+    DeclareOpInterfaceMethods<InferIntRangeInterface, ["inferResultRanges"]>
 ]> {
   let arguments = (ins Noisy_I32:$lhs, Noisy_I32:$rhs);
   let results = (outs Noisy_I32:$output);
@@ -33,7 +33,7 @@ def Noisy_MulOp : Noisy_BinOp<"mul"> {
 }
 
 def Noisy_EncodeOp : Op<Noisy_Dialect, "encode", [
-    Pure, DeclareOpInterfaceMethods<InferIntRangeInterface>]> {
+    Pure, DeclareOpInterfaceMethods<InferIntRangeInterface, ["inferResultRanges"]>]> {
   let summary = "Encodes a noisy i32 from a small-width integer, injecting 12 bits of noise.";
   let arguments = (ins AnyIntOfWidths<[1, 2, 3, 4, 5]>:$input);
   let results = (outs Noisy_I32:$output);
@@ -48,7 +48,7 @@ def Noisy_DecodeOp : Op<Noisy_Dialect, "decode", [Pure]> {
 }
 
 def Noisy_ReduceNoiseOp : Op<Noisy_Dialect, "reduce_noise", [
-    Pure, SameOperandsAndResultType,  DeclareOpInterfaceMethods<InferIntRangeInterface>]> {
+    Pure, SameOperandsAndResultType,  DeclareOpInterfaceMethods<InferIntRangeInterface, ["inferResultRanges"]>]> {
   let summary = "Reduces the noise in a noisy integer to a fixed noise level. Expensive!";
   let arguments = (ins Noisy_I32:$input);
   let results = (outs Noisy_I32:$output);

lib/Dialect/Noisy/NoisyOps.td

@@ -31,7 +31,7 @@ OpFoldResult MulOp::fold(MulOp::FoldAdaptor adaptor) {
 
   if (!lhs || !rhs) return nullptr;
 
-  auto degree = getResult().getType().cast<PolynomialType>().getDegreeBound();
+  auto degree = llvm::cast<PolynomialType>(getResult().getType()).getDegreeBound();
   auto maxIndex = lhs.size() + rhs.size() - 1;
 
   SmallVector<APInt, 8> result;
@@ -68,7 +68,7 @@ OpFoldResult FromTensorOp::fold(FromTensorOp::FoldAdaptor adaptor) {
 LogicalResult EvalOp::verify() {
   auto pointTy = getPoint().getType();
   bool isSignlessInteger = pointTy.isSignlessInteger(32);
-  auto complexPt = llvm::dyn_cast<ComplexType>(pointTy);
+  auto complexPt = dyn_cast<ComplexType>(pointTy);
   return isSignlessInteger || complexPt ? success()
                                         : emitOpError(
                                               "argument point must be a 32-bit "

lib/Dialect/Poly/PolyOps.cpp

@@ -25,6 +25,8 @@ gentbl_cc_library(
     td_file = "Passes.td",
     deps = [
         "@llvm-project//mlir:OpBaseTdFiles",
+        "@llvm-project//mlir:PDLDialectTdFiles",
+        "@llvm-project//mlir:PDLInterpOpsTdFiles",
         "@llvm-project//mlir:PassBaseTdFiles",
     ],
 )
@@ -47,6 +49,39 @@ cc_library(
     hdrs = ["Passes.h"],
     deps = [
         ":MulToAdd",
+        ":MulToAddPdll",
         ":pass_inc_gen",
     ],
 )
+
+gentbl_cc_library(
+    name = "MulToAddPdllIncGen",
+    tbl_outs = [
+        (
+            ["-x=cpp"],
+            "MulToAddPdll.h.inc",
+        ),
+    ],
+    tblgen = "@llvm-project//mlir:mlir-pdll",
+    td_file = "MulToAdd.pdll",
+    deps = [
+        "@llvm-project//mlir:ArithDialect",
+        "@llvm-project//mlir:FuncDialect",
+        "@llvm-project//mlir:ArithOpsTdFiles",
+    ],
+)
+
+cc_library(
+    name = "MulToAddPdll",
+    srcs = ["MulToAddPdll.cpp"],
+    hdrs = ["MulToAddPdll.h"],
+    deps = [
+        ":pass_inc_gen",
+        ":MulToAddPdllIncGen",
+        "@llvm-project//mlir:ArithDialect",
+        "@llvm-project//mlir:FuncDialect",
+        "@llvm-project//mlir:Pass",
+        "@llvm-project//mlir:Transforms",
+    ],
+)
+

lib/Transform/Arith/BUILD

@@ -1,11 +1,18 @@
+add_mlir_pdll_library(MulToAddPdllIncGen
+  MulToAdd.pdll
+  MulToAddPdll.h.inc
+)
+
 add_mlir_library(MulToAdd
     MulToAdd.cpp
+    MulToAddPdll.cpp
 
     ${PROJECT_SOURCE_DIR}/lib/Transform/Arith/
     ADDITIONAL_HEADER_DIRS
 
     DEPENDS
     MLIRMulToAddPasses
+    MulToAddPdllIncGen
 
     LINK_LIBS PUBLIC
 )

lib/Transform/Arith/CMakeLists.txt

@@ -0,0 +1,67 @@
+#include "mlir/Dialect/Arith/IR/ArithOps.td"
+
+Constraint IsPowerOfTwo(attr: Attr) [{
+  int64_t value = cast<::mlir::IntegerAttr>(attr).getValue().getSExtValue();
+  return success((value & (value - 1)) == 0);
+}];
+
+// Currently, constraints that return values must be defined in C++
+Constraint Halve(atttr: Attr) -> Attr;
+Constraint MinusOne(attr: Attr) -> Attr;
+
+// Replace y = C*x with y = C/2*x + C/2*x, when C is a power of 2, otherwise do
+// nothing.
+Pattern PowerOfTwoExpandRhs with benefit(2) {
+  let root = op<arith.muli>(op<arith.constant> {value = const: Attr}, rhs: Value);
+  IsPowerOfTwo(const);
+  let halved: Attr = Halve(const);
+
+  rewrite root with {
+    let newConst = op<arith.constant> {value = halved};
+    let newMul = op<arith.muli>(newConst, rhs);
+    let newAdd = op<arith.addi>(newMul, newMul);
+    replace root with newAdd;
+  };
+}
+
+Pattern PowerOfTwoExpandLhs with benefit(2) {
+  let root = op<arith.muli>(lhs: Value, op<arith.constant> {value = const: Attr});
+  IsPowerOfTwo(const);
+  let halved: Attr = Halve(const);
+
+  rewrite root with {
+    let newConst = op<arith.constant> {value = halved};
+    let newMul = op<arith.muli>(lhs, newConst);
+    let newAdd = op<arith.addi>(newMul, newMul);
+    replace root with newAdd;
+  };
+}
+
+// Replace y = 9*x with y = 8*x + x
+Pattern PeelFromMulRhs with benefit(1) {
+  let root = op<arith.muli>(lhs: Value, op<arith.constant> {value = const: Attr});
+
+  // We are guaranteed `value` is not a power of two, because the greedy
+  // rewrite engine ensures the PowerOfTwoExpand pattern is run first, since
+  // it has higher benefit.
+  let minusOne: Attr = MinusOne(const);
+
+  rewrite root with {
+    let newConst = op<arith.constant> {value = minusOne};
+    let newMul = op<arith.muli>(lhs, newConst);
+    let newAdd = op<arith.addi>(newMul, lhs);
+    replace root with newAdd;
+  };
+}
+
+Pattern PeelFromMulLhs with benefit(1) {
+  let root = op<arith.muli>(op<arith.constant> {value = const: Attr}, rhs: Value);
+  let minusOne: Attr = MinusOne(const);
+
+  rewrite root with {
+    let newConst = op<arith.constant> {value = minusOne};
+    let newMul = op<arith.muli>(newConst, rhs);
+    let newAdd = op<arith.addi>(newMul, rhs);
+    replace root with newAdd;
+  };
+}

lib/Transform/Arith/MulToAdd.pdll

@@ -0,0 +1,49 @@
+#include "lib/Transform/Arith/MulToAddPdll.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/PDLInterp/IR/PDLInterp.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "mlir/include/mlir/Pass/Pass.h"
+
+namespace mlir {
+namespace tutorial {
+
+#define GEN_PASS_DEF_MULTOADDPDLL
+#include "lib/Transform/Arith/Passes.h.inc"
+
+LogicalResult halveImpl(PatternRewriter &rewriter, PDLResultList &results,
+                        ArrayRef<PDLValue> args) {
+  Attribute attr = args[0].cast<Attribute>();
+  IntegerAttr cAttr = cast<IntegerAttr>(attr);
+  int64_t value = cAttr.getValue().getSExtValue();
+  results.push_back(rewriter.getIntegerAttr(cAttr.getType(), value / 2));
+  return success();
+}
+
+LogicalResult minusOneImpl(PatternRewriter &rewriter, PDLResultList &results,
+                           ArrayRef<PDLValue> args) {
+  Attribute attr = args[0].cast<Attribute>();
+  IntegerAttr cAttr = cast<IntegerAttr>(attr);
+  int64_t value = cAttr.getValue().getSExtValue();
+  results.push_back(rewriter.getIntegerAttr(cAttr.getType(), value - 1));
+  return success();
+}
+
+void registerNativeConstraints(RewritePatternSet &patterns) {
+  patterns.getPDLPatterns().registerConstraintFunction("Halve", halveImpl);
+  patterns.getPDLPatterns().registerConstraintFunction("MinusOne", minusOneImpl);
+}
+
+struct MulToAddPdll : impl::MulToAddPdllBase<MulToAddPdll> {
+  using MulToAddPdllBase::MulToAddPdllBase;
+
+  void runOnOperation() {
+    mlir::RewritePatternSet patterns(&getContext());
+    populateGeneratedPDLLPatterns(patterns);
+    registerNativeConstraints(patterns);
+    (void)applyPatternsAndFoldGreedily(getOperation(), std::move(patterns));
+  }
+};
+
+} // namespace tutorial
+} // namespace mlir

lib/Transform/Arith/MulToAddPdll.cpp

@@ -0,0 +1,20 @@
+#ifndef LIB_TRANSFORM_ARITH_MULTOADDPDLL_H_
+#define LIB_TRANSFORM_ARITH_MULTOADDPDLL_H_
+
+#include "mlir/Pass/Pass.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Parser/Parser.h"
+
+namespace mlir {
+namespace tutorial {
+
+#define GEN_PASS_DECL_MULTOADDPDLL
+#include "lib/Transform/Arith/Passes.h.inc"
+
+#include "lib/Transform/Arith/MulToAddPdll.h.inc"
+
+}  // namespace tutorial
+}  // namespace mlir
+
+#endif  // LIB_TRANSFORM_ARITH_MULTOADDPDLL_H_

lib/Transform/Arith/MulToAddPdll.h

@@ -2,6 +2,7 @@
 #define LIB_TRANSFORM_ARITH_PASSES_H_
 
 #include "lib/Transform/Arith/MulToAdd.h"
+#include "lib/Transform/Arith/MulToAddPdll.h"
 
 namespace mlir {
 namespace tutorial {

lib/Transform/Arith/Passes.h

@@ -1,6 +1,8 @@
 #ifndef LIB_TRANSFORM_ARITH_PASSES_TD_
 #define LIB_TRANSFORM_ARITH_PASSES_TD_
 
+include "mlir/Dialect/PDL/IR/PDLDialect.td"
+include "mlir/Dialect/PDLInterp/IR/PDLInterpOps.td"
 include "mlir/Pass/PassBase.td"
 
 def MulToAdd : Pass<"mul-to-add"> {
@@ -10,4 +12,15 @@ def MulToAdd : Pass<"mul-to-add"> {
   }];
 }
 
+def MulToAddPdll : Pass<"mul-to-add-pdll"> {
+  let summary = "Convert multiplications to repeated additions using pdll";
+  let description = [{
+    Convert multiplications to repeated additions (using pdll).
+  }];
+  let dependentDialects = [
+    "mlir::pdl::PDLDialect",
+    "mlir::pdl_interp::PDLInterpDialect",
+];
+}
+
 #endif  // LIB_TRANSFORM_ARITH_PASSES_TD_

lib/Transform/Arith/Passes.td

@@ -0,0 +1,33 @@
+// RUN: tutorial-opt %s --mul-to-add-pdll | FileCheck %s
+
+func.func @just_power_of_two(%arg: i32) -> i32 {
+  %0 = arith.constant 8 : i32
+  %1 = arith.muli %arg, %0 : i32
+  func.return %1 : i32
+}
+
+// CHECK-LABEL: func.func @just_power_of_two(
+// CHECK-SAME:    %[[ARG:.*]]: i32
+// CHECK-SAME:  ) -> i32 {
+// CHECK:   %[[SUM_0:.*]] = arith.addi %[[ARG]], %[[ARG]]
+// CHECK:   %[[SUM_1:.*]] = arith.addi %[[SUM_0]], %[[SUM_0]]
+// CHECK:   %[[SUM_2:.*]] = arith.addi %[[SUM_1]], %[[SUM_1]]
+// CHECK:   return %[[SUM_2]] : i32
+// CHECK: }
+
+
+func.func @power_of_two_plus_one(%arg: i32) -> i32 {
+  %0 = arith.constant 9 : i32
+  %1 = arith.muli %arg, %0 : i32
+  func.return %1 : i32
+}
+
+// CHECK-LABEL: func.func @power_of_two_plus_one(
+// CHECK-SAME:    %[[ARG:.*]]: i32
+// CHECK-SAME:  ) -> i32 {
+// CHECK:   %[[SUM_0:.*]] = arith.addi %[[ARG]], %[[ARG]]
+// CHECK:   %[[SUM_1:.*]] = arith.addi %[[SUM_0]], %[[SUM_0]]
+// CHECK:   %[[SUM_2:.*]] = arith.addi %[[SUM_1]], %[[SUM_1]]
+// CHECK:   %[[SUM_3:.*]] = arith.addi %[[SUM_2]], %[[ARG]]
+// CHECK:   return %[[SUM_3]] : i32
+// CHECK: }