[flang] translate derived type array init to attribute if possible #140268
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
llvmbot
added
flang
Member
|
@llvm/pr-subscribers-flang-fir-hlfir @llvm/pr-subscribers-flang-codegen Author: None (jeanPerier) ChangesThis patch relies on #140235 and #139724 to speed-up compilations of files with derived type array global with initial value. This chain was later folded in LLVM dialect to LLVM IR using LLVM IR (not dialect) folding. This insert chain generation and folding is very expensive for big arrays. For instance, this patch brings down the compilation of FM_lib fmsave.f95 from 50s to 0.5s. Instead of creating attributes directly in lowering, they are created late in codegen so that the fir.box component embox/rebox representation and code generation do not have to be duplicated some new FIR attribute based representation. In lowering, generating the insert chain in FIR is rather cheap thanks to the insert_on_range operation that avoids duplicating the insert by the dimension of the arrays when inserting the same value. The translation to an attribute is made at the LLVM IR dialect level so that all FIR abstraction are implemented. A folder is added in flang for that purpose. It is quite limited, and it is a bad idea to make it a fold method of the LLVM operation itself because trying to fold insert chain randomly in the code could be very expansive. In flang codegen, we know it is worth spending the few ms doing it because we know we are very likely to obtain a constant. That is not true in the LLVM dialect in general. Full diff: https://github.com/llvm/llvm-project/pull/140268.diff 7 Files Affected:
diff --git a/flang/include/flang/Optimizer/CodeGen/LLVMInsertChainFolder.h b/flang/include/flang/Optimizer/CodeGen/LLVMInsertChainFolder.h
new file mode 100644
index 0000000000000..d577c4c0fa70b
--- /dev/null
+++ b/flang/include/flang/Optimizer/CodeGen/LLVMInsertChainFolder.h
@@ -0,0 +1,31 @@
+//===-- LLVMInsertChainFolder.h -- insertvalue chain folder ----*- C++ -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// Helper to fold LLVM dialect llvm.insertvalue chain representing constants
+// into an Attribute representation.
+// This sits in Flang because it is incomplete and tailored for flang needs.
+//
+//===----------------------------------------------------------------------===//
+
+namespace mlir {
+class Attribute;
+class OpBuilder;
+class Value;
+} // namespace mlir
+
+namespace fir {
+
+/// Attempt to fold an llvm.insertvalue chain into an attribute representation
+/// suitable as llvm.constant operand. The returned value will be a null pointer
+/// if this is not an llvm.insertvalue result pr if the chain is not a constant,
+/// or cannot be represented as an Attribute. The operations are not deleted,
+/// but some llvm.insertvalue value operands may be folded with the builder on
+/// the way.
+mlir::Attribute tryFoldingLLVMInsertChain(mlir::Value insertChainResult,
+ mlir::OpBuilder &builder);
+} // namespace fir
diff --git a/flang/include/flang/Optimizer/Dialect/FIROps.td b/flang/include/flang/Optimizer/Dialect/FIROps.td
index 458b780806144..dc66885f776f0 100644
--- a/flang/include/flang/Optimizer/Dialect/FIROps.td
+++ b/flang/include/flang/Optimizer/Dialect/FIROps.td
@@ -2129,6 +2129,11 @@ def fir_InsertOnRangeOp : fir_OneResultOp<"insert_on_range", [NoMemoryEffect]> {
$seq `,` $val custom<CustomRangeSubscript>($coor) attr-dict `:` functional-type(operands, results)
}];
+ let extraClassDeclaration = [{
+ /// Is this insert_on_range inserting on all the values of the result type?
+ bool isFullRange();
+ }];
+
let hasVerifier = 1;
}
diff --git a/flang/lib/Optimizer/CodeGen/CMakeLists.txt b/flang/lib/Optimizer/CodeGen/CMakeLists.txt
index 04480bac552b7..980307db315d9 100644
--- a/flang/lib/Optimizer/CodeGen/CMakeLists.txt
+++ b/flang/lib/Optimizer/CodeGen/CMakeLists.txt
@@ -3,6 +3,7 @@ add_flang_library(FIRCodeGen
CodeGen.cpp
CodeGenOpenMP.cpp
FIROpPatterns.cpp
+ LLVMInsertChainFolder.cpp
LowerRepackArrays.cpp
PreCGRewrite.cpp
TBAABuilder.cpp
diff --git a/flang/lib/Optimizer/CodeGen/CodeGen.cpp b/flang/lib/Optimizer/CodeGen/CodeGen.cpp
index ad9119ba4a031..ed76a77ced047 100644
--- a/flang/lib/Optimizer/CodeGen/CodeGen.cpp
+++ b/flang/lib/Optimizer/CodeGen/CodeGen.cpp
@@ -14,6 +14,7 @@
#include "flang/Optimizer/CodeGen/CodeGenOpenMP.h"
#include "flang/Optimizer/CodeGen/FIROpPatterns.h"
+#include "flang/Optimizer/CodeGen/LLVMInsertChainFolder.h"
#include "flang/Optimizer/CodeGen/TypeConverter.h"
#include "flang/Optimizer/Dialect/FIRAttr.h"
#include "flang/Optimizer/Dialect/FIRCG/CGOps.h"
@@ -2412,15 +2413,38 @@ struct InsertOnRangeOpConversion
doRewrite(fir::InsertOnRangeOp range, mlir::Type ty, OpAdaptor adaptor,
mlir::ConversionPatternRewriter &rewriter) const override {
- llvm::SmallVector<std::int64_t> dims;
- auto type = adaptor.getOperands()[0].getType();
+ auto arrayType = adaptor.getSeq().getType();
// Iteratively extract the array dimensions from the type.
+ llvm::SmallVector<std::int64_t> dims;
+ mlir::Type type = arrayType;
while (auto t = mlir::dyn_cast<mlir::LLVM::LLVMArrayType>(type)) {
dims.push_back(t.getNumElements());
type = t.getElementType();
}
+ // Avoid generating long insert chain that are very slow to fold back
+ // (which is required in globals when later generating LLVM IR). Attempt to
+ // fold the inserted element value to an attribute and build an ArrayAttr
+ // for the resulting array.
+ if (range.isFullRange()) {
+ if (mlir::Attribute cst =
+ fir::tryFoldingLLVMInsertChain(adaptor.getVal(), rewriter)) {
+ mlir::Attribute dimVal = cst;
+ for (auto dim : llvm::reverse(dims)) {
+ // Use std::vector in case the number of elements is big.
+ std::vector<mlir::Attribute> elements(dim, dimVal);
+ dimVal = mlir::ArrayAttr::get(range.getContext(), elements);
+ }
+ // Replace insert chain with constant.
+ rewriter.replaceOpWithNewOp<mlir::LLVM::ConstantOp>(range, arrayType,
+ dimVal);
+ return mlir::success();
+ }
+ }
+
+ // The inserted value cannot be folded to an attribute, turn the
+ // insert_range into an llvm.insertvalue chain.
llvm::SmallVector<std::int64_t> lBounds;
llvm::SmallVector<std::int64_t> uBounds;
@@ -2434,8 +2458,8 @@ struct InsertOnRangeOpConversion
auto &subscripts = lBounds;
auto loc = range.getLoc();
- mlir::Value lastOp = adaptor.getOperands()[0];
- mlir::Value insertVal = adaptor.getOperands()[1];
+ mlir::Value lastOp = adaptor.getSeq();
+ mlir::Value insertVal = adaptor.getVal();
while (subscripts != uBounds) {
lastOp = rewriter.create<mlir::LLVM::InsertValueOp>(
@@ -3131,7 +3155,7 @@ struct GlobalOpConversion : public fir::FIROpConversion<fir::GlobalOp> {
// initialization is on the full range.
auto insertOnRangeOps = gr.front().getOps<fir::InsertOnRangeOp>();
for (auto insertOp : insertOnRangeOps) {
- if (isFullRange(insertOp.getCoor(), insertOp.getType())) {
+ if (insertOp.isFullRange()) {
auto seqTyAttr = convertType(insertOp.getType());
auto *op = insertOp.getVal().getDefiningOp();
auto constant = mlir::dyn_cast<mlir::arith::ConstantOp>(op);
@@ -3161,22 +3185,7 @@ struct GlobalOpConversion : public fir::FIROpConversion<fir::GlobalOp> {
return mlir::success();
}
- bool isFullRange(mlir::DenseIntElementsAttr indexes,
- fir::SequenceType seqTy) const {
- auto extents = seqTy.getShape();
- if (indexes.size() / 2 != static_cast<int64_t>(extents.size()))
- return false;
- auto cur_index = indexes.value_begin<int64_t>();
- for (unsigned i = 0; i < indexes.size(); i += 2) {
- if (*(cur_index++) != 0)
- return false;
- if (*(cur_index++) != extents[i / 2] - 1)
- return false;
- }
- return true;
- }
-
- // TODO: String comparaison should be avoided. Replace linkName with an
+ // TODO: String comparisons should be avoided. Replace linkName with an
// enumeration.
mlir::LLVM::Linkage
convertLinkage(std::optional<llvm::StringRef> optLinkage) const {
diff --git a/flang/lib/Optimizer/CodeGen/LLVMInsertChainFolder.cpp b/flang/lib/Optimizer/CodeGen/LLVMInsertChainFolder.cpp
new file mode 100644
index 0000000000000..0fc8697b735cf
--- /dev/null
+++ b/flang/lib/Optimizer/CodeGen/LLVMInsertChainFolder.cpp
@@ -0,0 +1,204 @@
+//===-- LLVMInsertChainFolder.cpp -----------------------------------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#include "flang/Optimizer/CodeGen/LLVMInsertChainFolder.h"
+#include "mlir/Dialect/LLVMIR/LLVMAttrs.h"
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/IR/Builders.h"
+#include "llvm/Support/Debug.h"
+
+#define DEBUG_TYPE "flang-insert-folder"
+
+#include <deque>
+
+namespace {
+// Helper class to construct the attribute elements of an aggregate value being
+// folded without creating a full mlir::Attribute representation for each step
+// of the insert value chain, which would both be expensive in terms of
+// compilation time and memory (since the intermediate Attribute would survive,
+// unused, inside the mlir context).
+class InsertChainBackwardFolder {
+ // Type for the current value of an element of the aggregate value being
+ // constructed by the insert chain.
+ // At any point of the insert chain, the value of an element is either:
+ // - nullptr: not yet known, the insert has not yet been seen.
+ // - an mlir::Attribute: the element is fully defined.
+ // - a nested InsertChainBackwardFolder: the element is itself an aggregate
+ // and its sub-elements have been partially defined (insert with mutliple
+ // indices have been seen).
+
+ // The insertion folder assumes backward walk of the insert chain. Once an
+ // element or sub-element has been defined, it is not overriden by new
+ // insertions (last insert wins).
+ using InFlightValue =
+ llvm::PointerUnion<mlir::Attribute, InsertChainBackwardFolder *>;
+
+public:
+ InsertChainBackwardFolder(
+ mlir::Type type, std::deque<InsertChainBackwardFolder> *folderStorage)
+ : values(getNumElements(type), mlir::Attribute{}),
+ folderStorage{folderStorage}, type{type} {}
+
+ /// Push
+ bool pushValue(mlir::Attribute val, llvm::ArrayRef<int64_t> at);
+
+ mlir::Attribute finalize(mlir::Attribute defaultFieldValue);
+
+private:
+ static int64_t getNumElements(mlir::Type type) {
+ if (auto structTy =
+ llvm::dyn_cast_if_present<mlir::LLVM::LLVMStructType>(type))
+ return structTy.getBody().size();
+ if (auto arrayTy =
+ llvm::dyn_cast_if_present<mlir::LLVM::LLVMArrayType>(type))
+ return arrayTy.getNumElements();
+ return 0;
+ }
+
+ static mlir::Type getSubElementType(mlir::Type type, int64_t field) {
+ if (auto arrayTy =
+ llvm::dyn_cast_if_present<mlir::LLVM::LLVMArrayType>(type))
+ return arrayTy.getElementType();
+ if (auto structTy =
+ llvm::dyn_cast_if_present<mlir::LLVM::LLVMStructType>(type))
+ return structTy.getBody()[field];
+ return {};
+ }
+
+ // Current element value of the aggregate value being built.
+ llvm::SmallVector<InFlightValue> values;
+ // std::deque is used to allocate storage for nested list and guarantee the
+ // stability of the InsertChainBackwardFolder* used as element value.
+ std::deque<InsertChainBackwardFolder> *folderStorage;
+ // Type of the aggregate value being built.
+ mlir::Type type;
+};
+} // namespace
+
+// Helper to fold the value being inserted by an llvm.insert_value.
+// This may call tryFoldingLLVMInsertChain if the value is an aggregate and
+// was itself constructed by a different insert chain.
+static mlir::Attribute getAttrIfConstant(mlir::Value val,
+ mlir::OpBuilder &rewriter) {
+ if (auto cst = val.getDefiningOp<mlir::LLVM::ConstantOp>())
+ return cst.getValue();
+ if (auto insert = val.getDefiningOp<mlir::LLVM::InsertValueOp>())
+ return fir::tryFoldingLLVMInsertChain(val, rewriter);
+ if (val.getDefiningOp<mlir::LLVM::ZeroOp>())
+ return mlir::LLVM::ZeroAttr::get(val.getContext());
+ if (val.getDefiningOp<mlir::LLVM::UndefOp>())
+ return mlir::LLVM::UndefAttr::get(val.getContext());
+ if (mlir::Operation *op = val.getDefiningOp()) {
+ unsigned resNum = llvm::cast<mlir::OpResult>(val).getResultNumber();
+ llvm::SmallVector<mlir::Value> results;
+ if (mlir::succeeded(rewriter.tryFold(op, results)) &&
+ results.size() > resNum) {
+ if (auto cst = results[resNum].getDefiningOp<mlir::LLVM::ConstantOp>())
+ return cst.getValue();
+ }
+ }
+ if (auto trunc = val.getDefiningOp<mlir::LLVM::TruncOp>())
+ if (auto attr = getAttrIfConstant(trunc.getArg(), rewriter))
+ if (auto intAttr = llvm::dyn_cast<mlir::IntegerAttr>(attr))
+ return mlir::IntegerAttr::get(trunc.getType(), intAttr.getInt());
+ LLVM_DEBUG(llvm::dbgs() << "cannot fold insert value operand: " << val
+ << "\n");
+ return {};
+}
+
+mlir::Attribute
+InsertChainBackwardFolder::finalize(mlir::Attribute defaultFieldValue) {
+ std::vector<mlir::Attribute> attrs;
+ attrs.reserve(values.size());
+ for (InFlightValue &inFlight : values) {
+ if (!inFlight) {
+ attrs.push_back(defaultFieldValue);
+ } else if (auto attr = llvm::dyn_cast<mlir::Attribute>(inFlight)) {
+ attrs.push_back(attr);
+ } else {
+ auto *inFlightList = llvm::cast<InsertChainBackwardFolder *>(inFlight);
+ attrs.push_back(inFlightList->finalize(defaultFieldValue));
+ }
+ }
+ return mlir::ArrayAttr::get(type.getContext(), attrs);
+}
+
+bool InsertChainBackwardFolder::pushValue(mlir::Attribute val,
+ llvm::ArrayRef<int64_t> at) {
+ if (at.size() == 0 || at[0] >= static_cast<int64_t>(values.size()))
+ return false;
+ InFlightValue &inFlight = values[at[0]];
+ if (!inFlight) {
+ if (at.size() == 1) {
+ inFlight = val;
+ return true;
+ }
+ // This is the first insert to a nested field. Create a
+ // InsertChainBackwardFolder for the current element value.
+ InsertChainBackwardFolder &inFlightList = folderStorage->emplace_back(
+ getSubElementType(type, at[0]), folderStorage);
+ inFlight = &inFlightList;
+ return inFlightList.pushValue(val, at.drop_front());
+ }
+ // Keep last inserted value if already set.
+ if (llvm::isa<mlir::Attribute>(inFlight))
+ return true;
+ auto *inFlightList = llvm::cast<InsertChainBackwardFolder *>(inFlight);
+ if (at.size() == 1) {
+ if (!llvm::isa<mlir::LLVM::ZeroAttr, mlir::LLVM::UndefAttr>(val)) {
+ LLVM_DEBUG(llvm::dbgs()
+ << "insert chain sub-element partially overwritten initial "
+ "value is not zero or undef\n");
+ return false;
+ }
+ inFlight = inFlightList->finalize(val);
+ return true;
+ }
+ return inFlightList->pushValue(val, at.drop_front());
+}
+
+mlir::Attribute fir::tryFoldingLLVMInsertChain(mlir::Value val,
+ mlir::OpBuilder &rewriter) {
+ if (auto cst = val.getDefiningOp<mlir::LLVM::ConstantOp>())
+ return cst.getValue();
+ if (auto insert = val.getDefiningOp<mlir::LLVM::InsertValueOp>()) {
+ LLVM_DEBUG(llvm::dbgs() << "trying to fold insert chain:" << val << "\n");
+ if (auto structTy =
+ llvm::dyn_cast<mlir::LLVM::LLVMStructType>(insert.getType())) {
+ mlir::LLVM::InsertValueOp currentInsert = insert;
+ mlir::LLVM::InsertValueOp lastInsert;
+ std::deque<InsertChainBackwardFolder> folderStorage;
+ InsertChainBackwardFolder inFlightList(structTy, &folderStorage);
+ while (currentInsert) {
+ mlir::Attribute attr =
+ getAttrIfConstant(currentInsert.getValue(), rewriter);
+ if (!attr)
+ return {};
+ if (!inFlightList.pushValue(attr, currentInsert.getPosition()))
+ return {};
+ lastInsert = currentInsert;
+ currentInsert = currentInsert.getContainer()
+ .getDefiningOp<mlir::LLVM::InsertValueOp>();
+ }
+ mlir::Attribute defaultVal;
+ if (lastInsert) {
+ if (lastInsert.getContainer().getDefiningOp<mlir::LLVM::ZeroOp>())
+ defaultVal = mlir::LLVM::ZeroAttr::get(val.getContext());
+ else if (lastInsert.getContainer().getDefiningOp<mlir::LLVM::UndefOp>())
+ defaultVal = mlir::LLVM::UndefAttr::get(val.getContext());
+ }
+ if (!defaultVal) {
+ LLVM_DEBUG(llvm::dbgs()
+ << "insert chain initial value is not Zero or Undef\n");
+ return {};
+ }
+ return inFlightList.finalize(defaultVal);
+ }
+ }
+ return {};
+}
diff --git a/flang/lib/Optimizer/Dialect/FIROps.cpp b/flang/lib/Optimizer/Dialect/FIROps.cpp
index d85b38c467857..e12af7782a578 100644
--- a/flang/lib/Optimizer/Dialect/FIROps.cpp
+++ b/flang/lib/Optimizer/Dialect/FIROps.cpp
@@ -2365,6 +2365,21 @@ llvm::LogicalResult fir::InsertOnRangeOp::verify() {
return mlir::success();
}
+bool fir::InsertOnRangeOp::isFullRange() {
+ auto extents = getType().getShape();
+ mlir::DenseIntElementsAttr indexes = getCoor();
+ if (indexes.size() / 2 != static_cast<int64_t>(extents.size()))
+ return false;
+ auto cur_index = indexes.value_begin<int64_t>();
+ for (unsigned i = 0; i < indexes.size(); i += 2) {
+ if (*(cur_index++) != 0)
+ return false;
+ if (*(cur_index++) != extents[i / 2] - 1)
+ return false;
+ }
+ return true;
+}
+
//===----------------------------------------------------------------------===//
// InsertValueOp
//===----------------------------------------------------------------------===//
diff --git a/flang/test/Fir/convert-and-fold-insert-on-range.fir b/flang/test/Fir/convert-and-fold-insert-on-range.fir
new file mode 100644
index 0000000000000..df18614d80b63
--- /dev/null
+++ b/flang/test/Fir/convert-and-fold-insert-on-range.fir
@@ -0,0 +1,33 @@
+// Test codegen of constant insert_on_range without symbol reference into mlir.constant.
+// RUN: fir-opt --cg-rewrite --split-input-file --fir-to-llvm-ir %s | FileCheck %s
+
+module attributes {dlti.dl_spec = #dlti.dl_spec<!llvm.ptr<270> = dense<32> : vector<4xi64>, !llvm.ptr<271> = dense<32> : vector<4xi64>, !llvm.ptr<272> = dense<64> : vector<4xi64>, i64 = dense<64> : vector<2xi64>, i128 = dense<128> : vector<2xi64>, f80 = dense<128> : vector<2xi64>, !llvm.ptr = dense<64> : vector<4xi64>, i1 = dense<8> : vector<2xi64>, i8 = dense<8> : vector<2xi64>, i16 = dense<16> : vector<2xi64>, i32 = dense<32> : vector<2xi64>, f16 = dense<16> : vector<2xi64>, f64 = dense<64> : vector<2xi64>, f128 = dense<128> : vector<2xi64>, "dlti.endianness" = "little", "dlti.mangling_mode" = "e", "dlti.stack_alignment" = 128 : i64>, fir.defaultkind = "a1c4d8i4l4r4", fir.kindmap = "", 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", llvm.target_triple = "x86_64-unknown-linux-gnu"} {
+ fir.global @derived_array : !fir.array<2x!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>> {
+ %c0 = arith.constant 0 : index
+ %0 = fir.undefined !fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>
+ %1 = fir.zero_bits !fir.heap<!fir.array<?xf64>>
+ %2 = fir.shape %c0 : (index) -> !fir.shape<1>
+ %3 = fir.embox %1(%2) : (!fir.heap<!fir.array<?xf64>>, !fir.shape<1>) -> !fir.box<!fir.heap<!fir.array<?xf64>>>
+ %4 = fir.insert_value %0, %3, ["comp", !fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>] : (!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>, !fir.box<!fir.heap<!fir.array<?xf64>>>) -> !fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>
+ %5 = fir.undefined !fir.array<2x!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>
+ %6 = fir.insert_on_range %5, %4 from (0) to (1) : (!fir.array<2x!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>, !fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>) -> !fir.array<2x!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>
+ fir.has_value %6 : !fir.array<2x!fir.type<sometype{comp:!fir.box<!fir.heap<!fir.array<?xf64>>>}>>
+ }
+}
+
+//CHECK-LABEL: llvm.mlir.global external @derived_array()
+//CHECK: %[[CST:.*]] = llvm.mlir.constant([
+//CHECK-SAME: [
+//CHECK-SAME: [#llvm.zero, 8, 20240719 : i32, 1 : i8, 28 : i8, 2 : i8, 0 : i8,
+//CHECK-SAME: [
+//CHECK-SAME: [1, 0 : index, 8]
+//CHECK-SAME: ]
+//CHECK-SAME: ],
+//CHECK-SAME: [
+//CHECK-SAME: [#llvm.zero, 8, 20240719 : i32, 1 : i8, 28 : i8, 2 : i8, 0 : i8,
+//CHECK-SAME: [
+//CHECK-SAME: [1, 0 : index, 8]
+//CHECK-SAME: ]
+//CHECK-SAME: ]) :
+//CHECK-SAME: !llvm.array<2 x struct<"sometype", (struct<(ptr, i64, i32, i8, i8, i8, i8, array<1 x array<3 x i64>>)>)>>
+//CHECK: llvm.return %[[CST]] : !llvm.array<2 x struct<"sometype", (struct<(ptr, i64, i32, i8, i8, i8, i8, array<1 x array<3 x i64>>)>)>>
|
ashermancinelli
approved these changes
May 16, 2025
Contributor
ashermancinelli
left a comment
ashermancinelli
left a comment
There was a problem hiding this comment.
Thanks, this is a really interesting fix! My review comments are just nits/questions.
Thanks Slava! Co-authored-by: Slava Zakharin <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch relies on #140235 and #139724 to speed-up compilations of files with derived type array global with initial value.
Currently, such derived type global init was lowered to an llvm.mlir.insertvalue chain in the LLVM IR dialect because there was no way to represent such value via attributes.
This chain was later folded in LLVM dialect to LLVM IR using LLVM IR (not dialect) folding. This insert chain generation and folding is very expensive for big arrays. For instance, this patch brings down the compilation of FM_lib fmsave.f95 from 50s to 0.5s.
Instead of creating attributes directly in lowering, they are created late in codegen so that the fir.box component embox/rebox representation and code generation do not have to be duplicated some new FIR attribute based representation.
In lowering, generating the insert chain in FIR is rather cheap thanks to the insert_on_range operation that avoids duplicating the insert by the dimension of the arrays when inserting the same value.
The translation to an attribute is made at the LLVM IR dialect level so that all FIR abstraction are implemented. A folder is added in flang for that purpose. It is quite limited, and it is a bad idea to make it a fold method of the LLVM operation itself because trying to fold insert chain randomly in the code could be very expansive. In flang codegen, we know it is worth spending the few ms doing it because we know we are very likely to obtain a constant. That is not true in the LLVM dialect in general.