[flang][openacc] Do not loose attributes on folding #80516
Merged
Conversation
This file contains 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
|
@llvm/pr-subscribers-openacc Author: Valentin Clement (バレンタイン クレメン) (clementval) Changeshlfir.declare introduce some boxes that can be later optimized away. The OpenACC lowering is currently setting some attributes on FIR operations to track declare variables. When the boxes are optimized away these attributes are lost. This patch propagate OpenACC attributes from box_addr op to the defining op of the folding result. Full diff: https://github.com/llvm/llvm-project/pull/80516.diff 2 Files Affected:
diff --git a/flang/lib/Optimizer/Dialect/FIROps.cpp b/flang/lib/Optimizer/Dialect/FIROps.cpp
index f826f2566b897..96a676149203d 100644
--- a/flang/lib/Optimizer/Dialect/FIROps.cpp
+++ b/flang/lib/Optimizer/Dialect/FIROps.cpp
@@ -20,6 +20,7 @@
#include "flang/Optimizer/Support/Utils.h"
#include "mlir/Dialect/CommonFolders.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/OpenACC/OpenACC.h"
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/BuiltinAttributes.h"
@@ -36,6 +37,18 @@ namespace {
#include "flang/Optimizer/Dialect/CanonicalizationPatterns.inc"
} // namespace
+static void propagateAttributes(mlir::Operation *fromOp,
+ mlir::Operation *toOp) {
+ if (!fromOp || !toOp)
+ return;
+
+ for (mlir::NamedAttribute attr : fromOp->getAttrs()) {
+ if (attr.getName().str().rfind(
+ mlir::acc::OpenACCDialect::getDialectNamespace(), 0) == 0)
+ toOp->setAttr(attr.getName().str(), attr.getValue());
+ }
+}
+
/// Return true if a sequence type is of some incomplete size or a record type
/// is malformed or contains an incomplete sequence type. An incomplete sequence
/// type is one with more unknown extents in the type than have been provided
@@ -626,8 +639,10 @@ mlir::OpFoldResult fir::BoxAddrOp::fold(FoldAdaptor adaptor) {
if (auto *v = getVal().getDefiningOp()) {
if (auto box = mlir::dyn_cast<fir::EmboxOp>(v)) {
// Fold only if not sliced
- if (!box.getSlice() && box.getMemref().getType() == getType())
+ if (!box.getSlice() && box.getMemref().getType() == getType()) {
+ propagateAttributes(getOperation(), box.getMemref().getDefiningOp());
return box.getMemref();
+ }
}
if (auto box = mlir::dyn_cast<fir::EmboxCharOp>(v))
if (box.getMemref().getType() == getType())
diff --git a/flang/test/Fir/OpenACC/propagate-attr-folding.fir b/flang/test/Fir/OpenACC/propagate-attr-folding.fir
new file mode 100644
index 0000000000000..f84269f5fef26
--- /dev/null
+++ b/flang/test/Fir/OpenACC/propagate-attr-folding.fir
@@ -0,0 +1,39 @@
+// RUN: fir-opt %s --opt-bufferization | FileCheck %s
+
+// Check that OpenACC attributes are propagated to the defining operations when
+// fir.box_addr is folded in bufferization optimization.
+
+func.func @_QPsub1(%arg0: !fir.ref<!fir.array<?x?xf32>> {fir.bindc_name = "a"}, %arg1: !fir.ref<i32> {fir.bindc_name = "n1"}, %arg2: !fir.ref<i32> {fir.bindc_name = "n2"}) {
+ %c1 = arith.constant 1 : index
+ %c0 = arith.constant 0 : index
+ %0 = fir.declare %arg1 {uniq_name = "_QFsub1En1"} : (!fir.ref<i32>) -> !fir.ref<i32>
+ %1 = fir.declare %arg2 {uniq_name = "_QFsub1En2"} : (!fir.ref<i32>) -> !fir.ref<i32>
+ %2 = fir.load %0 : !fir.ref<i32>
+ %3 = fir.convert %2 : (i32) -> index
+ %4 = arith.cmpi sgt, %3, %c0 : index
+ %5 = arith.select %4, %3, %c0 : index
+ %6 = fir.load %1 : !fir.ref<i32>
+ %7 = fir.convert %6 : (i32) -> index
+ %8 = arith.cmpi sgt, %7, %c0 : index
+ %9 = arith.select %8, %7, %c0 : index
+ %10 = fir.shape %5, %9 : (index, index) -> !fir.shape<2>
+ %11 = fir.declare %arg0(%10) {acc.declare = #acc.declare<dataClause = acc_present>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.ref<!fir.array<?x?xf32>>
+ %12 = fir.embox %11(%10) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.box<!fir.array<?x?xf32>>
+ %13:3 = fir.box_dims %12, %c0 : (!fir.box<!fir.array<?x?xf32>>, index) -> (index, index, index)
+ %14 = arith.subi %13#1, %c1 : index
+ %15 = acc.bounds lowerbound(%c0 : index) upperbound(%14 : index) extent(%13#1 : index) stride(%13#2 : index) startIdx(%c1 : index) {strideInBytes = true}
+ %16 = arith.muli %13#2, %13#1 : index
+ %17:3 = fir.box_dims %12, %c1 : (!fir.box<!fir.array<?x?xf32>>, index) -> (index, index, index)
+ %18 = arith.subi %17#1, %c1 : index
+ %19 = acc.bounds lowerbound(%c0 : index) upperbound(%18 : index) extent(%17#1 : index) stride(%16 : index) startIdx(%c1 : index) {strideInBytes = true}
+ %20 = acc.present varPtr(%11 : !fir.ref<!fir.array<?x?xf32>>) bounds(%15, %19) -> !fir.ref<!fir.array<?x?xf32>> {name = "a"}
+ %21 = acc.declare_enter dataOperands(%20 : !fir.ref<!fir.array<?x?xf32>>)
+ acc.declare_exit token(%21)
+ return
+}
+
+// CHECK-LABEL: func.func @_QPsub1(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.array<?x?xf32>> {fir.bindc_name = "a"}
+// CHECK: %[[DECL:.*]] = fir.declare %[[ARG0]](%{{.*}}) {acc.declare = #acc.declare<dataClause = acc_present>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.ref<!fir.array<?x?xf32>>
+// CHECK: %[[PRES:.*]] = acc.present varPtr(%[[DECL]] : !fir.ref<!fir.array<?x?xf32>>) bounds(%{{.*}}, %{{.*}}) -> !fir.ref<!fir.array<?x?xf32>> {name = "a"}
+// CHECK: %{{.*}} = acc.declare_enter dataOperands(%[[PRES]] : !fir.ref<!fir.array<?x?xf32>>)
|
|
@llvm/pr-subscribers-flang-fir-hlfir Author: Valentin Clement (バレンタイン クレメン) (clementval) Changeshlfir.declare introduce some boxes that can be later optimized away. The OpenACC lowering is currently setting some attributes on FIR operations to track declare variables. When the boxes are optimized away these attributes are lost. This patch propagate OpenACC attributes from box_addr op to the defining op of the folding result. Full diff: https://github.com/llvm/llvm-project/pull/80516.diff 2 Files Affected:
diff --git a/flang/lib/Optimizer/Dialect/FIROps.cpp b/flang/lib/Optimizer/Dialect/FIROps.cpp
index f826f2566b897..96a676149203d 100644
--- a/flang/lib/Optimizer/Dialect/FIROps.cpp
+++ b/flang/lib/Optimizer/Dialect/FIROps.cpp
@@ -20,6 +20,7 @@
#include "flang/Optimizer/Support/Utils.h"
#include "mlir/Dialect/CommonFolders.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/OpenACC/OpenACC.h"
#include "mlir/Dialect/OpenMP/OpenMPDialect.h"
#include "mlir/IR/Attributes.h"
#include "mlir/IR/BuiltinAttributes.h"
@@ -36,6 +37,18 @@ namespace {
#include "flang/Optimizer/Dialect/CanonicalizationPatterns.inc"
} // namespace
+static void propagateAttributes(mlir::Operation *fromOp,
+ mlir::Operation *toOp) {
+ if (!fromOp || !toOp)
+ return;
+
+ for (mlir::NamedAttribute attr : fromOp->getAttrs()) {
+ if (attr.getName().str().rfind(
+ mlir::acc::OpenACCDialect::getDialectNamespace(), 0) == 0)
+ toOp->setAttr(attr.getName().str(), attr.getValue());
+ }
+}
+
/// Return true if a sequence type is of some incomplete size or a record type
/// is malformed or contains an incomplete sequence type. An incomplete sequence
/// type is one with more unknown extents in the type than have been provided
@@ -626,8 +639,10 @@ mlir::OpFoldResult fir::BoxAddrOp::fold(FoldAdaptor adaptor) {
if (auto *v = getVal().getDefiningOp()) {
if (auto box = mlir::dyn_cast<fir::EmboxOp>(v)) {
// Fold only if not sliced
- if (!box.getSlice() && box.getMemref().getType() == getType())
+ if (!box.getSlice() && box.getMemref().getType() == getType()) {
+ propagateAttributes(getOperation(), box.getMemref().getDefiningOp());
return box.getMemref();
+ }
}
if (auto box = mlir::dyn_cast<fir::EmboxCharOp>(v))
if (box.getMemref().getType() == getType())
diff --git a/flang/test/Fir/OpenACC/propagate-attr-folding.fir b/flang/test/Fir/OpenACC/propagate-attr-folding.fir
new file mode 100644
index 0000000000000..f84269f5fef26
--- /dev/null
+++ b/flang/test/Fir/OpenACC/propagate-attr-folding.fir
@@ -0,0 +1,39 @@
+// RUN: fir-opt %s --opt-bufferization | FileCheck %s
+
+// Check that OpenACC attributes are propagated to the defining operations when
+// fir.box_addr is folded in bufferization optimization.
+
+func.func @_QPsub1(%arg0: !fir.ref<!fir.array<?x?xf32>> {fir.bindc_name = "a"}, %arg1: !fir.ref<i32> {fir.bindc_name = "n1"}, %arg2: !fir.ref<i32> {fir.bindc_name = "n2"}) {
+ %c1 = arith.constant 1 : index
+ %c0 = arith.constant 0 : index
+ %0 = fir.declare %arg1 {uniq_name = "_QFsub1En1"} : (!fir.ref<i32>) -> !fir.ref<i32>
+ %1 = fir.declare %arg2 {uniq_name = "_QFsub1En2"} : (!fir.ref<i32>) -> !fir.ref<i32>
+ %2 = fir.load %0 : !fir.ref<i32>
+ %3 = fir.convert %2 : (i32) -> index
+ %4 = arith.cmpi sgt, %3, %c0 : index
+ %5 = arith.select %4, %3, %c0 : index
+ %6 = fir.load %1 : !fir.ref<i32>
+ %7 = fir.convert %6 : (i32) -> index
+ %8 = arith.cmpi sgt, %7, %c0 : index
+ %9 = arith.select %8, %7, %c0 : index
+ %10 = fir.shape %5, %9 : (index, index) -> !fir.shape<2>
+ %11 = fir.declare %arg0(%10) {acc.declare = #acc.declare<dataClause = acc_present>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.ref<!fir.array<?x?xf32>>
+ %12 = fir.embox %11(%10) : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.box<!fir.array<?x?xf32>>
+ %13:3 = fir.box_dims %12, %c0 : (!fir.box<!fir.array<?x?xf32>>, index) -> (index, index, index)
+ %14 = arith.subi %13#1, %c1 : index
+ %15 = acc.bounds lowerbound(%c0 : index) upperbound(%14 : index) extent(%13#1 : index) stride(%13#2 : index) startIdx(%c1 : index) {strideInBytes = true}
+ %16 = arith.muli %13#2, %13#1 : index
+ %17:3 = fir.box_dims %12, %c1 : (!fir.box<!fir.array<?x?xf32>>, index) -> (index, index, index)
+ %18 = arith.subi %17#1, %c1 : index
+ %19 = acc.bounds lowerbound(%c0 : index) upperbound(%18 : index) extent(%17#1 : index) stride(%16 : index) startIdx(%c1 : index) {strideInBytes = true}
+ %20 = acc.present varPtr(%11 : !fir.ref<!fir.array<?x?xf32>>) bounds(%15, %19) -> !fir.ref<!fir.array<?x?xf32>> {name = "a"}
+ %21 = acc.declare_enter dataOperands(%20 : !fir.ref<!fir.array<?x?xf32>>)
+ acc.declare_exit token(%21)
+ return
+}
+
+// CHECK-LABEL: func.func @_QPsub1(
+// CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.array<?x?xf32>> {fir.bindc_name = "a"}
+// CHECK: %[[DECL:.*]] = fir.declare %[[ARG0]](%{{.*}}) {acc.declare = #acc.declare<dataClause = acc_present>, uniq_name = "_QFsub1Ea"} : (!fir.ref<!fir.array<?x?xf32>>, !fir.shape<2>) -> !fir.ref<!fir.array<?x?xf32>>
+// CHECK: %[[PRES:.*]] = acc.present varPtr(%[[DECL]] : !fir.ref<!fir.array<?x?xf32>>) bounds(%{{.*}}, %{{.*}}) -> !fir.ref<!fir.array<?x?xf32>> {name = "a"}
+// CHECK: %{{.*}} = acc.declare_enter dataOperands(%[[PRES]] : !fir.ref<!fir.array<?x?xf32>>)
|
vzakhari
reviewed
Feb 3, 2024
agozillon
pushed a commit
to agozillon/llvm-project
that referenced
this pull request
Feb 5, 2024
hlfir.declare introduce some boxes that can be later optimized away. The OpenACC lowering is currently setting some attributes on FIR operations to track declare variables. When the boxes are optimized away these attributes are lost. This patch propagate OpenACC attributes from box_addr op to the defining op of the folding result.
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.
hlfir.declare introduce some boxes that can be later optimized away. The OpenACC lowering is currently setting some attributes on FIR operations to track declare variables. When the boxes are optimized away these attributes are lost. This patch propagate OpenACC attributes from box_addr op to the defining op of the folding result.