Alexseev Danila. Lab 4. Opt 2.

mlir/lib/Transforms/lab4/alexseev_danila/AlexseevMulAddMergePass.cpp

@@ -0,0 +1,50 @@
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Tools/Plugins/PassPlugin.h"
+
+class AlexseevMulAddMergePass
+    : public mlir::PassWrapper<AlexseevMulAddMergePass,
+                               mlir::OperationPass<mlir::ModuleOp>> {
+  void createFMAOperation(mlir::LLVM::FAddOp &addOp, mlir::LLVM::FMulOp &mulOp,
+                          mlir::Value &otherOperand) {
+    mlir::OpBuilder builder(addOp);
+    mlir::Value fma = builder.create<mlir::LLVM::FMAOp>(
+        addOp.getLoc(), addOp.getType(), mulOp.getOperand(0),
+        mulOp.getOperand(1), otherOperand);
+    addOp.replaceAllUsesWith(fma);
+    addOp.erase();
+    if (mulOp.use_empty())
+      mulOp.erase();
+  }
+
+public:
+  mlir::StringRef getArgument() const final { return "alexseev_mul_add_merge"; }
+
+  mlir::StringRef getDescription() const final {
+    return "Merge multiplication and addition into a single math.fma";
+  }
+
+  void runOnOperation() override {
+    getOperation().walk([&](mlir::LLVM::FAddOp addOp) {
+      mlir::Value addLHS = addOp.getOperand(0);
+      mlir::Value addRHS = addOp.getOperand(1);
+
+      if (auto mulOpLHS = addLHS.getDefiningOp<mlir::LLVM::FMulOp>()) {
+        createFMAOperation(addOp, mulOpLHS, addRHS);
+      } else if (auto mulOpRHS = addRHS.getDefiningOp<mlir::LLVM::FMulOp>()) {
+        createFMAOperation(addOp, mulOpRHS, addLHS);
+      }
+    });
+  }
+};
+
+MLIR_DECLARE_EXPLICIT_TYPE_ID(AlexseevMulAddMergePass)
+MLIR_DEFINE_EXPLICIT_TYPE_ID(AlexseevMulAddMergePass)
+
+extern "C" LLVM_ATTRIBUTE_WEAK mlir::PassPluginLibraryInfo
+mlirGetPassPluginInfo() {
+  return {MLIR_PLUGIN_API_VERSION, "alexseev_mul_add_merge",
+          LLVM_VERSION_STRING,
+          []() { mlir::PassRegistration<AlexseevMulAddMergePass>(); }};
+}

mlir/lib/Transforms/lab4/alexseev_danila/CMakeLists.txt

@@ -0,0 +1,12 @@
+set(PluginName AlexseevMulAddMergePass)
+
+file(GLOB_RECURSE ALL_SOURCE_FILES *.cpp *.h)
+add_llvm_pass_plugin(${PluginName}
+        ${ALL_SOURCE_FILES}
+        DEPENDS
+        intrinsics_gen
+        MLIRBuiltinLocationAttributesIncGen
+        BUILDTREE_ONLY
+        )
+
+set(MLIR_TEST_DEPENDS ${PluginName} ${MLIR_TEST_DEPENDS} PARENT_SCOPE)

mlir/test/Transforms/lab4/alexseev_danila/test.mlir

@@ -0,0 +1,40 @@
+// RUN: mlir-opt -load-pass-plugin=%mlir_lib_dir/AlexseevMulAddMergePass%shlibext --pass-pipeline="builtin.module(alexseev_mul_add_merge)" %s | FileCheck %s
+
+module attributes {dlti.dl_spec = #dlti.dl_spec<#dlti.dl_entry<f80, dense<128> : vector<2xi32>>, #dlti.dl_entry<i64, dense<64> : vector<2xi32>>, #dlti.dl_entry<!llvm.ptr<272>, dense<64> : vector<4xi32>>, #dlti.dl_entry<i1, dense<8> : vector<2xi32>>, #dlti.dl_entry<!llvm.ptr, dense<64> : vector<4xi32>>, #dlti.dl_entry<i32, dense<32> : vector<2xi32>>, #dlti.dl_entry<i16, dense<16> : vector<2xi32>>, #dlti.dl_entry<i8, dense<8> : vector<2xi32>>, #dlti.dl_entry<!llvm.ptr<270>, dense<32> : vector<4xi32>>, #dlti.dl_entry<!llvm.ptr<271>, dense<32> : vector<4xi32>>, #dlti.dl_entry<f128, dense<128> : vector<2xi32>>, #dlti.dl_entry<f16, dense<16> : vector<2xi32>>, #dlti.dl_entry<f64, dense<64> : vector<2xi32>>, #dlti.dl_entry<"dlti.stack_alignment", 128 : i32>, #dlti.dl_entry<"dlti.endianness", "little">>} {
+  // double c = a * 6.0 + b;
+  llvm.func local_unnamed_addr @foo1(%arg0: f64 {llvm.noundef}, %arg1: f64 {llvm.noundef}) attributes {memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>, passthrough = ["mustprogress", "nofree", "norecurse", "nosync", "nounwind", "willreturn", ["uwtable", "2"], ["min-legal-vector-width", "0"], ["no-trapping-math", "true"], ["stack-protector-buffer-size", "8"], ["target-cpu", "x86-64"], ["target-features", "+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87"], ["tune-cpu", "generic"]]} {
+    %0 = llvm.mlir.constant(6.000000e+00 : f64) : f64
+    // CHECK-NOT:  %1 = llvm.fmul %arg0, %0  : f64
+    // CHECK-NOT:  %2 = llvm.fadd %1, %arg1  : f64
+    // CHECK:      %1 = llvm.intr.fma(%arg0, %0, %arg1)  : (f64, f64, f64) -> f64
+    %1 = llvm.fmul %arg0, %0  : f64
+    %2 = llvm.fadd %1, %arg1  : f64
+    llvm.return
+  }
+
+  // double c = a + 10.0 * b;
+  llvm.func local_unnamed_addr @foo2(%arg0: f64 {llvm.noundef}, %arg1: f64 {llvm.noundef}) attributes {memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>, passthrough = ["mustprogress", "nofree", "norecurse", "nosync", "nounwind", "willreturn", ["uwtable", "2"], ["min-legal-vector-width", "0"], ["no-trapping-math", "true"], ["stack-protector-buffer-size", "8"], ["target-cpu", "x86-64"], ["target-features", "+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87"], ["tune-cpu", "generic"]]} {
+    %0 = llvm.mlir.constant(10.000000e+00 : f64) : f64
+    // CHECK-NOT:  %1 = llvm.fmul %arg1, %0  : f64
+    // CHECK-NOT:  %2 = llvm.fadd %1, %arg0  : f64
+    // CHECK:      %1 = llvm.intr.fma(%arg1, %0, %arg0)  : (f64, f64, f64) -> f64
+    %1 = llvm.fmul %arg1, %0  : f64
+    %2 = llvm.fadd %1, %arg0  : f64
+    llvm.return
+  }
+
+  //     double c = a * b;
+  //     double d = c + e;
+  //     double f = c + x;
+  llvm.func local_unnamed_addr @foo3(%arg0: f64 {llvm.noundef}, %arg1: f64 {llvm.noundef}, %arg2: f64 {llvm.noundef}, %arg3: f64 {llvm.noundef}) attributes {memory = #llvm.memory_effects<other = none, argMem = none, inaccessibleMem = none>, passthrough = ["mustprogress", "nofree", "norecurse", "nosync", "nounwind", "willreturn", ["uwtable", "2"], ["min-legal-vector-width", "0"], ["no-trapping-math", "true"], ["stack-protector-buffer-size", "8"], ["target-cpu", "x86-64"], ["target-features", "+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87"], ["tune-cpu", "generic"]]} {
+    // CHECK-NOT: %0 = llvm.fmul %arg0, %arg1  : f64
+    // CHECK-NOT: %1 = llvm.fadd %0, %arg2  : f64
+    // CHECK:     %0 = llvm.intr.fma(%arg0, %arg1, %arg2)  : (f64, f64, f64) -> f64
+	%0 = llvm.fmul %arg0, %arg1  : f64
+    %1 = llvm.fadd %0, %arg2  : f64
+    // CHECK-NOT: %2 = llvm.fadd %0, %arg3  : f64
+    // CHECK:     %1 = llvm.intr.fma(%arg0, %arg1, %arg3)  : (f64, f64, f64) -> f64
+	%2 = llvm.fadd %0, %arg3  : f64
+    llvm.return
+  }
+}