[flang-commits] [flang] [flang] Inline hlfir.copy_in for trivial types (PR #138718)

[Kajetan Puchalski via flang-commits]

================
@@ -0,0 +1,178 @@
+//===- InlineHLFIRCopyIn.cpp - Inline hlfir.copy_in ops -------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// Transform hlfir.copy_in array operations into loop nests performing element
+// per element assignments. For simplicity, the inlining is done for trivial
+// data types when the copy_in does not require a corresponding copy_out and
+// when the input array is not behind a pointer. This may change in the future.
+//===----------------------------------------------------------------------===//
+
+#include "flang/Optimizer/Builder/FIRBuilder.h"
+#include "flang/Optimizer/Builder/HLFIRTools.h"
+#include "flang/Optimizer/Dialect/FIRType.h"
+#include "flang/Optimizer/HLFIR/HLFIROps.h"
+#include "flang/Optimizer/OpenMP/Passes.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Support/LLVM.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+namespace hlfir {
+#define GEN_PASS_DEF_INLINEHLFIRCOPYIN
+#include "flang/Optimizer/HLFIR/Passes.h.inc"
+} // namespace hlfir
+
+#define DEBUG_TYPE "inline-hlfir-copy-in"
+
+static llvm::cl::opt<bool> noInlineHLFIRCopyIn(
+    "no-inline-hlfir-copy-in",
+    llvm::cl::desc("Do not inline hlfir.copy_in operations"),
+    llvm::cl::init(false));
+
+namespace {
+class InlineCopyInConversion : public mlir::OpRewritePattern<hlfir::CopyInOp> {
+public:
+  using mlir::OpRewritePattern<hlfir::CopyInOp>::OpRewritePattern;
+
+  llvm::LogicalResult
+  matchAndRewrite(hlfir::CopyInOp copyIn,
+                  mlir::PatternRewriter &rewriter) const override;
+};
+
+llvm::LogicalResult
+InlineCopyInConversion::matchAndRewrite(hlfir::CopyInOp copyIn,
+                                        mlir::PatternRewriter &rewriter) const {
+  fir::FirOpBuilder builder(rewriter, copyIn.getOperation());
+  mlir::Location loc = copyIn.getLoc();
+  hlfir::Entity inputVariable{copyIn.getVar()};
+  if (!fir::isa_trivial(inputVariable.getFortranElementType()))
+    return rewriter.notifyMatchFailure(copyIn,
+                                       "CopyInOp's data type is not trivial");
+
+  // There should be exactly one user of WasCopied - the corresponding
+  // CopyOutOp.
+  if (!copyIn.getWasCopied().hasOneUse())
+    return rewriter.notifyMatchFailure(
+        copyIn, "CopyInOp's WasCopied has no single user");
+  // The copy out should always be present, either to actually copy or just
+  // deallocate memory.
+  auto copyOut = mlir::dyn_cast<hlfir::CopyOutOp>(
+      copyIn.getWasCopied().user_begin().getCurrent().getUser());
+
+  if (!copyOut)
+    return rewriter.notifyMatchFailure(copyIn,
+                                       "CopyInOp has no direct CopyOut");
+
+  // Only inline the copy_in when copy_out does not need to be done, i.e. in
+  // case of intent(in).
+  if (copyOut.getVar())
+    return rewriter.notifyMatchFailure(copyIn, "CopyIn needs a copy-out");
+
+  inputVariable =
+      hlfir::derefPointersAndAllocatables(loc, builder, inputVariable);
+  mlir::Type sequenceType =
+      hlfir::getFortranElementOrSequenceType(inputVariable.getType());
+  fir::BoxType resultBoxType = fir::BoxType::get(sequenceType);
+  mlir::Value isContiguous =
+      builder.create<fir::IsContiguousBoxOp>(loc, inputVariable);
+  mlir::Operation::result_range results =
+      builder
+          .genIfOp(loc, {resultBoxType, builder.getI1Type()}, isContiguous,
+                   /*withElseRegion=*/true)
+          .genThen([&]() {
+            mlir::Value result = inputVariable;
+            if (fir::isPointerType(inputVariable.getType())) {
+              auto boxAddr = builder.create<fir::BoxAddrOp>(loc, inputVariable);
+              fir::ReferenceType refTy = fir::ReferenceType::get(sequenceType);
+              mlir::Value refVal = builder.createConvert(loc, refTy, boxAddr);
+              mlir::Value shape = hlfir::genShape(loc, builder, inputVariable);
+              result = builder.create<fir::EmboxOp>(loc, resultBoxType, refVal,
+                                                    shape);
+            }
+            builder.create<fir::ResultOp>(
+                loc, mlir::ValueRange{result, builder.createBool(loc, false)});
+          })
+          .genElse([&] {
+            mlir::Value shape = hlfir::genShape(loc, builder, inputVariable);
+            llvm::SmallVector<mlir::Value> extents =
+                hlfir::getIndexExtents(loc, builder, shape);
+            llvm::StringRef tmpName{".tmp.copy_in"};
+            llvm::SmallVector<mlir::Value> lenParams;
+            mlir::Value alloc = builder.createHeapTemporary(
+                loc, sequenceType, tmpName, extents, lenParams);
+
+            auto declareOp = builder.create<hlfir::DeclareOp>(
+                loc, alloc, tmpName, shape, lenParams,
+                /*dummy_scope=*/nullptr);
+            hlfir::Entity temp{declareOp.getBase()};
+            hlfir::LoopNest loopNest =
+                hlfir::genLoopNest(loc, builder, extents, /*isUnordered=*/true,
+                                   flangomp::shouldUseWorkshareLowering(copyIn),
+                                   /*couldVectorize=*/false);
+            builder.setInsertionPointToStart(loopNest.body);
+            hlfir::Entity elem = hlfir::getElementAt(
+                loc, builder, inputVariable, loopNest.oneBasedIndices);
+            elem = hlfir::loadTrivialScalar(loc, builder, elem);
+            hlfir::Entity tempElem = hlfir::getElementAt(
+                loc, builder, temp, loopNest.oneBasedIndices);
+            builder.create<hlfir::AssignOp>(loc, elem, tempElem);
+            builder.setInsertionPointAfter(loopNest.outerOp);
+
+            mlir::Value result;
+            // Make sure the result is always a boxed array by boxing it
+            // ourselves if need be.
+            if (mlir::isa<fir::BaseBoxType>(temp.getType())) {
+              result = temp;
+            } else {
+              fir::ReferenceType refTy =
+                  fir::ReferenceType::get(temp.getElementOrSequenceType());
+              mlir::Value refVal = builder.createConvert(loc, refTy, temp);
+              result = builder.create<fir::EmboxOp>(loc, resultBoxType, refVal,
+                                                    shape);
+            }
+
+            builder.create<fir::ResultOp>(
+                loc, mlir::ValueRange{result, builder.createBool(loc, true)});
+          })
+          .getResults();
+
+  mlir::OpResult resultBox = results[0];
+  mlir::OpResult needsCleanup = results[1];
+
+  auto alloca = builder.create<fir::AllocaOp>(loc, resultBox.getType());
----------------
mrkajetanp wrote:

That tempBox is a
```
 %3 = "fir.alloca"() <{in_type = !fir.box<!fir.heap<!fir.array<?xf64>>>, operandSegmentSizes = array<i32: 0, 0>}> : () -> !fir.ref<!fir.box<!fir.heap<!fir.array<?xf64>>>>
```
while the box resulting from the if block is just a box<array> because in one of the branches it's not actually a heap type. If I try storing it directly I get a type mismatch error on that account because the op is:
```
"fir.store"(%20#0, %5) : (!fir.box<!fir.array<?xf64>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xf64>>>>) -> ()
```
Do you mean I should force convert the resulting box? Or do some other type tricks?
The thinking here was that the existing tempBox will get cleaned up anyway so it's better to make a specific new alloca with the type I want.

https://github.com/llvm/llvm-project/pull/138718