[flang] introduce fir.copy to avoid load store of aggregates

flang/include/flang/Optimizer/Dialect/FIROps.td

@@ -342,6 +342,44 @@ def fir_StoreOp : fir_Op<"store", [FirAliasTagOpInterface]> {
   }];
 }
 
+def fir_CopyOp : fir_Op<"copy", []> {
+  let summary = "copy constant size memory";
+
+  let description = [{
+    Copy the memory from a source with compile time constant size to
+    a destination of the same type.
+
+    This is meant to be used for aggregate types where load and store
+    are not appropriate to make a copy because LLVM is not meant to
+    handle load and store of "big" aggregates.
+
+    Its "no_overlap" attribute allows indicating that the source and destination
+    are known to not overlap at compile time.
+
+    ```
+      !t =!fir.type<t{x:!fir.array<1000xi32>}>
+      fir.copy %x to %y : !fir.ref<!t>, !fir.ref<!t>
+    ```
+    TODO: add FirAliasTagOpInterface to carry TBAA.
+  }];
+
+  let arguments = (ins Arg<AnyReferenceLike, "", [MemRead]>:$source,
+                       Arg<AnyReferenceLike, "", [MemWrite]>:$destination,
+                       OptionalAttr<UnitAttr>:$no_overlap);
+
+  let builders = [OpBuilder<(ins "mlir::Value":$source,
+                                 "mlir::Value":$destination,
+                                  CArg<"bool", "false">:$no_overlap)>];
+
+  let assemblyFormat = [{
+    $source `to` $destination (`no_overlap` $no_overlap^)?
+    attr-dict `:` type(operands)
+  }];
+
+  let hasVerifier = 1;
+}
+
+
 def fir_SaveResultOp : fir_Op<"save_result", [AttrSizedOperandSegments]> {
   let summary = [{
     save an array, box, or record function result SSA-value to a memory location

flang/lib/Optimizer/CodeGen/CodeGen.cpp

@@ -3539,6 +3539,36 @@ struct StoreOpConversion : public fir::FIROpConversion<fir::StoreOp> {
   }
 };
 
+/// `fir.copy` --> `llvm.memcpy` or `llvm.memmove`
+struct CopyOpConversion : public fir::FIROpConversion<fir::CopyOp> {
+  using FIROpConversion::FIROpConversion;
+
+  llvm::LogicalResult
+  matchAndRewrite(fir::CopyOp copy, OpAdaptor adaptor,
+                  mlir::ConversionPatternRewriter &rewriter) const override {
+    mlir::Location loc = copy.getLoc();
+    mlir::Value llvmSource = adaptor.getSource();
+    mlir::Value llvmDestination = adaptor.getDestination();
+    mlir::Type i64Ty = mlir::IntegerType::get(rewriter.getContext(), 64);
+    mlir::Type copyTy = fir::unwrapRefType(copy.getSource().getType());
+    mlir::Value copySize =
+        genTypeStrideInBytes(loc, i64Ty, rewriter, convertType(copyTy));
+
+    mlir::LLVM::AliasAnalysisOpInterface newOp;
+    if (copy.getNoOverlap())
+      newOp = rewriter.create<mlir::LLVM::MemcpyOp>(
+          loc, llvmDestination, llvmSource, copySize, /*isVolatile=*/false);
+    else
+      newOp = rewriter.create<mlir::LLVM::MemmoveOp>(
+          loc, llvmDestination, llvmSource, copySize, /*isVolatile=*/false);
+
+    // TODO: propagate TBAA once FirAliasTagOpInterface added to CopyOp.
+    attachTBAATag(newOp, copyTy, copyTy, nullptr);
+    rewriter.eraseOp(copy);
+    return mlir::success();
+  }
+};
+
 namespace {
 
 /// Convert `fir.unboxchar` into two `llvm.extractvalue` instructions. One for
@@ -4142,11 +4172,11 @@ void fir::populateFIRToLLVMConversionPatterns(
       BoxOffsetOpConversion, BoxProcHostOpConversion, BoxRankOpConversion,
       BoxTypeCodeOpConversion, BoxTypeDescOpConversion, CallOpConversion,
       CmpcOpConversion, ConvertOpConversion, CoordinateOpConversion,
-      DTEntryOpConversion, DeclareOpConversion, DivcOpConversion,
-      EmboxOpConversion, EmboxCharOpConversion, EmboxProcOpConversion,
-      ExtractValueOpConversion, FieldIndexOpConversion, FirEndOpConversion,
-      FreeMemOpConversion, GlobalLenOpConversion, GlobalOpConversion,
-      InsertOnRangeOpConversion, IsPresentOpConversion,
+      CopyOpConversion, DTEntryOpConversion, DeclareOpConversion,
+      DivcOpConversion, EmboxOpConversion, EmboxCharOpConversion,
+      EmboxProcOpConversion, ExtractValueOpConversion, FieldIndexOpConversion,
+      FirEndOpConversion, FreeMemOpConversion, GlobalLenOpConversion,
+      GlobalOpConversion, InsertOnRangeOpConversion, IsPresentOpConversion,
       LenParamIndexOpConversion, LoadOpConversion, MulcOpConversion,
       NegcOpConversion, NoReassocOpConversion, SelectCaseOpConversion,
       SelectOpConversion, SelectRankOpConversion, SelectTypeOpConversion,

flang/lib/Optimizer/Dialect/FIROps.cpp

@@ -3940,6 +3940,29 @@ void fir::StoreOp::build(mlir::OpBuilder &builder, mlir::OperationState &result,
   build(builder, result, value, memref, {});
 }
 
+//===----------------------------------------------------------------------===//
+// CopyOp
+//===----------------------------------------------------------------------===//
+
+void fir::CopyOp::build(mlir::OpBuilder &builder, mlir::OperationState &result,
+                        mlir::Value source, mlir::Value destination,
+                        bool noOverlap) {
+  mlir::UnitAttr noOverlapAttr =
+      noOverlap ? builder.getUnitAttr() : mlir::UnitAttr{};
+  build(builder, result, source, destination, noOverlapAttr);
+}
+
+llvm::LogicalResult fir::CopyOp::verify() {
+  mlir::Type sourceType = fir::unwrapRefType(getSource().getType());
+  mlir::Type destinationType = fir::unwrapRefType(getDestination().getType());
+  if (sourceType != destinationType)
+    return emitOpError("source and destination must have the same value type");
+  if (fir::hasDynamicSize(sourceType))
+    return emitOpError(
+        "source value type must have a compile time constant size");
+  return mlir::success();
+}
+
 //===----------------------------------------------------------------------===//
 // StringLitOp
 //===----------------------------------------------------------------------===//

flang/test/Fir/copy-codegen.fir

@@ -0,0 +1,35 @@
+// Test fir.copy codegen.
+// RUN: fir-opt --fir-to-llvm-ir %s -o - | FileCheck %s
+
+!t=!fir.type<sometype{i:!fir.array<9xi32>}>
+
+module attributes {llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"} {
+
+func.func @test_copy_1(%arg0: !fir.ref<!t>, %arg1: !fir.ref<!t>) {
+  fir.copy %arg0 to %arg1 no_overlap : !fir.ref<!t>, !fir.ref<!t>
+  return
+}
+// CHECK-LABEL:   llvm.func @test_copy_1(
+// CHECK-SAME:                           %[[VAL_0:[0-9]+|[a-zA-Z$._-][a-zA-Z0-9$._-]*]]: !llvm.ptr,
+// CHECK-SAME:                           %[[VAL_1:[0-9]+|[a-zA-Z$._-][a-zA-Z0-9$._-]*]]: !llvm.ptr) {
+// CHECK:           %[[VAL_2:.*]] = llvm.mlir.zero : !llvm.ptr
+// CHECK:           %[[VAL_3:.*]] = llvm.getelementptr %[[VAL_2]][1] : (!llvm.ptr) -> !llvm.ptr, !llvm.struct<"sometype", (array<9 x i32>)>
+// CHECK:           %[[VAL_4:.*]] = llvm.ptrtoint %[[VAL_3]] : !llvm.ptr to i64
+// CHECK:           "llvm.intr.memcpy"(%[[VAL_1]], %[[VAL_0]], %[[VAL_4]]) <{isVolatile = false}> : (!llvm.ptr, !llvm.ptr, i64) -> ()
+// CHECK:           llvm.return
+// CHECK:         }
+
+func.func @test_copy_2(%arg0: !fir.ref<!t>, %arg1: !fir.ref<!t>) {
+  fir.copy %arg0 to %arg1 : !fir.ref<!t>, !fir.ref<!t>
+  return
+}
+// CHECK-LABEL:   llvm.func @test_copy_2(
+// CHECK-SAME:                           %[[VAL_0:[0-9]+|[a-zA-Z$._-][a-zA-Z0-9$._-]*]]: !llvm.ptr,
+// CHECK-SAME:                           %[[VAL_1:[0-9]+|[a-zA-Z$._-][a-zA-Z0-9$._-]*]]: !llvm.ptr) {
+// CHECK:           %[[VAL_2:.*]] = llvm.mlir.zero : !llvm.ptr
+// CHECK:           %[[VAL_3:.*]] = llvm.getelementptr %[[VAL_2]][1] : (!llvm.ptr) -> !llvm.ptr, !llvm.struct<"sometype", (array<9 x i32>)>
+// CHECK:           %[[VAL_4:.*]] = llvm.ptrtoint %[[VAL_3]] : !llvm.ptr to i64
+// CHECK:           "llvm.intr.memmove"(%[[VAL_1]], %[[VAL_0]], %[[VAL_4]]) <{isVolatile = false}> : (!llvm.ptr, !llvm.ptr, i64) -> ()
+// CHECK:           llvm.return
+// CHECK:         }
+}

flang/test/Fir/fir-ops.fir

@@ -933,3 +933,12 @@ func.func @test_call_arg_attrs_indirect(%arg0: i16, %arg1: (i16)-> i16) -> i16 {
   %0 = fir.call %arg1(%arg0) : (i16 {llvm.noundef, llvm.signext}) -> (i16 {llvm.signext})
   return %0 : i16
 }
+
+// CHECK-LABEL: @test_copy(
+// CHECK-SAME:    %[[VAL_0:.*]]: !fir.ref<!fir.type<sometype{i:i32}>>,
+// CHECK-SAME:    %[[VAL_1:.*]]: !fir.ptr<!fir.type<sometype{i:i32}>>
+func.func @test_copy(%arg0: !fir.ref<!fir.type<sometype{i:i32}>>, %arg1: !fir.ptr<!fir.type<sometype{i:i32}>>) {
+  // CHECK: fir.copy %[[VAL_0]] to %[[VAL_1]] no_overlap : !fir.ref<!fir.type<sometype{i:i32}>>, !fir.ptr<!fir.type<sometype{i:i32}>>
+  fir.copy %arg0 to %arg1 no_overlap : !fir.ref<!fir.type<sometype{i:i32}>>, !fir.ptr<!fir.type<sometype{i:i32}>>
+  return
+}

flang/test/Fir/invalid.fir

@@ -1018,3 +1018,31 @@ func.func @bad_is_assumed_size(%arg0: !fir.ref<!fir.array<*:none>>) {
   %1 = fir.is_assumed_size %arg0 : (!fir.ref<!fir.array<*:none>>) -> i1
   return
 }
+
+// -----
+
+!t=!fir.type<sometype{i:i32}>
+!t2=!fir.type<sometype2{j:i32}>
+func.func @bad_copy_1(%arg0: !fir.ref<!t>, %arg1: !fir.ref<!t2>) {
+  // expected-error@+1{{'fir.copy' op source and destination must have the same value type}}
+  fir.copy %arg0 to %arg1 no_overlap : !fir.ref<!t>, !fir.ref<!t2>
+  return
+}
+
+// -----
+
+!t=!fir.type<sometype{i:i32}>
+func.func @bad_copy_2(%arg0: !fir.ref<!t>, %arg1: !t) {
+  // expected-error@+1{{'fir.copy' op operand #0 must be any reference, but got '!fir.type<sometype{i:i32}>'}}
+  fir.copy %arg1 to %arg0 no_overlap : !t, !fir.ref<!t>
+  return
+}
+
+// -----
+
+!t=!fir.array<?xi32>
+func.func @test_copy(%arg0: !fir.ref<!t>, %arg1: !fir.ref<!t>) {
+  // expected-error@+1{{'fir.copy' op source value type must have a compile time constant size}}
+  fir.copy %arg0 to %arg1 no_overlap : !fir.ref<!t>, !fir.ref<!t>
+  return
+}