[flang] Change traversal order for OptimizeArrayRepackingPass. #153136
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
A long chain of fir.pack_arrays might require multiple iterations of the greedy rewriter, if we use down-top traversal. The rewriter may not converge in 10 (default) iterations. It is not an error, but it was reported as such. This patch changes the traversal to top-down and also disabled the hard error, if the rewriter does not converge soon enough.
llvmbot
added
flang
Member
|
@llvm/pr-subscribers-flang-fir-hlfir Author: Slava Zakharin (vzakhari) ChangesA long chain of fir.pack_arrays might require multiple iterations This patch changes the traversal to top-down and also disabled Full diff: https://github.com/llvm/llvm-project/pull/153136.diff 2 Files Affected:
diff --git a/flang/lib/Optimizer/Transforms/OptimizeArrayRepacking.cpp b/flang/lib/Optimizer/Transforms/OptimizeArrayRepacking.cpp
index 1688f2887a57a..173ce1da041ae 100644
--- a/flang/lib/Optimizer/Transforms/OptimizeArrayRepacking.cpp
+++ b/flang/lib/Optimizer/Transforms/OptimizeArrayRepacking.cpp
@@ -26,6 +26,8 @@ namespace fir {
#include "flang/Optimizer/Transforms/Passes.h.inc"
} // namespace fir
+#define DEBUG_TYPE "optimize-array-repacking"
+
namespace {
class OptimizeArrayRepackingPass
: public fir::impl::OptimizeArrayRepackingBase<OptimizeArrayRepackingPass> {
@@ -78,13 +80,19 @@ void OptimizeArrayRepackingPass::runOnOperation() {
mlir::MLIRContext *context = &getContext();
mlir::RewritePatternSet patterns(context);
mlir::GreedyRewriteConfig config;
- config.setRegionSimplificationLevel(
- mlir::GreedySimplifyRegionLevel::Disabled);
+ config
+ .setRegionSimplificationLevel(mlir::GreedySimplifyRegionLevel::Disabled)
+ // Traverse the operations top-down, so that fir.pack_array
+ // operations are optimized before their using fir.pack_array
+ // operations. This way the rewrite may converge faster.
+ .setUseTopDownTraversal();
patterns.insert<PackingOfContiguous>(context);
patterns.insert<NoopUnpacking>(context);
if (mlir::failed(
mlir::applyPatternsGreedily(funcOp, std::move(patterns), config))) {
- mlir::emitError(funcOp.getLoc(), "failure in array repacking optimization");
- signalPassFailure();
+ // Failure may happen if the rewriter does not converge soon enough.
+ // That is not an error, so just report a diagnostic under debug.
+ LLVM_DEBUG(mlir::emitError(funcOp.getLoc(),
+ "failure in array repacking optimization"));
}
}
diff --git a/flang/test/Transforms/optimize-array-repacking.fir b/flang/test/Transforms/optimize-array-repacking.fir
index 6269fa441fe44..15a3e3941f44f 100644
--- a/flang/test/Transforms/optimize-array-repacking.fir
+++ b/flang/test/Transforms/optimize-array-repacking.fir
@@ -658,3 +658,136 @@ func.func @_QPneg_test_pointer(%arg0: !fir.ref<!fir.box<!fir.ptr<!fir.array<?xf3
fir.unpack_array %9 to %7 heap : !fir.box<!fir.array<?xf32>>
return
}
+
+// Test a long chain of fir.pack_array operations.
+// The rewriter used to use a down-top traversal that optimized
+// fir.pack_array operations starting from the innermost one.
+// The rewriter did not converge in 10 (default) iterations
+// causing the pass to report a failure.
+// A top-down traversal should fix this an allow optimizing
+// all the repackings.
+// CHECK-LABEL: func.func @test_long_chain(
+// CHECK-NOT: fir.pack_array
+// CHECK-NOT: fir.unpack_array
+func.func @test_long_chain(%pred: i1) {
+ %c10 = arith.constant 10 : index
+ %3 = fir.dummy_scope : !fir.dscope
+ %4 = fir.address_of(@aaa) : !fir.ref<!fir.array<10x10xi32>>
+ %5 = fir.shape %c10, %c10 : (index, index) -> !fir.shape<2>
+ %6 = fir.declare %4(%5) {uniq_name = "aaa"} : (!fir.ref<!fir.array<10x10xi32>>, !fir.shape<2>) -> !fir.ref<!fir.array<10x10xi32>>
+ %9 = fir.embox %6(%5) : (!fir.ref<!fir.array<10x10xi32>>, !fir.shape<2>) -> !fir.box<!fir.array<10x10xi32>>
+ %10 = fir.convert %9 : (!fir.box<!fir.array<10x10xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %11 = fir.dummy_scope : !fir.dscope
+ %12 = fir.pack_array %10 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %13 = fir.declare %12 dummy_scope %11 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %14 = fir.rebox %13 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb17, ^bb1
+^bb1: // pred: ^bb0
+ %20 = fir.dummy_scope : !fir.dscope
+ %21 = fir.pack_array %14 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %22 = fir.declare %21 dummy_scope %20 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %23 = fir.rebox %22 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %28 = fir.dummy_scope : !fir.dscope
+ %29 = fir.pack_array %23 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %30 = fir.declare %29 dummy_scope %28 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %31 = fir.rebox %30 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb16, ^bb2
+^bb2: // pred: ^bb1
+ %37 = fir.dummy_scope : !fir.dscope
+ %38 = fir.pack_array %31 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %39 = fir.declare %38 dummy_scope %37 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %40 = fir.rebox %39 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %45 = fir.dummy_scope : !fir.dscope
+ %46 = fir.pack_array %40 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %47 = fir.declare %46 dummy_scope %45 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %48 = fir.rebox %47 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb15, ^bb3
+^bb3: // pred: ^bb2
+ %54 = fir.dummy_scope : !fir.dscope
+ %55 = fir.pack_array %48 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %56 = fir.declare %55 dummy_scope %54 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %57 = fir.rebox %56 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %62 = fir.dummy_scope : !fir.dscope
+ %63 = fir.pack_array %57 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %64 = fir.declare %63 dummy_scope %62 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %65 = fir.rebox %64 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb14, ^bb4
+^bb4: // pred: ^bb3
+ %71 = fir.dummy_scope : !fir.dscope
+ %72 = fir.pack_array %65 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %73 = fir.declare %72 dummy_scope %71 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %74 = fir.rebox %73 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %79 = fir.dummy_scope : !fir.dscope
+ %80 = fir.pack_array %74 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %81 = fir.declare %80 dummy_scope %79 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %82 = fir.rebox %81 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb13, ^bb5
+^bb5: // pred: ^bb4
+ %88 = fir.dummy_scope : !fir.dscope
+ %89 = fir.pack_array %82 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %90 = fir.declare %89 dummy_scope %88 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %91 = fir.rebox %90 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %96 = fir.dummy_scope : !fir.dscope
+ %97 = fir.pack_array %91 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %98 = fir.declare %97 dummy_scope %96 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %99 = fir.rebox %98 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb12, ^bb6
+^bb6: // pred: ^bb5
+ %105 = fir.dummy_scope : !fir.dscope
+ %106 = fir.pack_array %99 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %107 = fir.declare %106 dummy_scope %105 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %108 = fir.rebox %107 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %113 = fir.dummy_scope : !fir.dscope
+ %114 = fir.pack_array %108 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %115 = fir.declare %114 dummy_scope %113 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %116 = fir.rebox %115 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb11, ^bb7
+^bb7: // pred: ^bb6
+ %122 = fir.dummy_scope : !fir.dscope
+ %123 = fir.pack_array %116 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %124 = fir.declare %123 dummy_scope %122 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %125 = fir.rebox %124 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %130 = fir.dummy_scope : !fir.dscope
+ %131 = fir.pack_array %125 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ %132 = fir.declare %131 dummy_scope %130 {uniq_name = "aaa"} : (!fir.box<!fir.array<?x?xi32>>, !fir.dscope) -> !fir.box<!fir.array<?x?xi32>>
+ %133 = fir.rebox %132 : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ cf.cond_br %pred, ^bb9, ^bb8
+^bb8: // pred: ^bb7
+ %139 = fir.dummy_scope : !fir.dscope
+ %140 = fir.pack_array %133 heap innermost : (!fir.box<!fir.array<?x?xi32>>) -> !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %140 to %133 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb9
+^bb9: // 2 preds: ^bb7, ^bb8
+ fir.unpack_array %131 to %125 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb10
+^bb10: // pred: ^bb9
+ fir.unpack_array %123 to %116 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb11
+^bb11: // 2 preds: ^bb6, ^bb10
+ fir.unpack_array %114 to %108 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %106 to %99 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb12
+^bb12: // 2 preds: ^bb5, ^bb11
+ fir.unpack_array %97 to %91 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %89 to %82 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb13
+^bb13: // 2 preds: ^bb4, ^bb12
+ fir.unpack_array %80 to %74 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %72 to %65 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb14
+^bb14: // 2 preds: ^bb3, ^bb13
+ fir.unpack_array %63 to %57 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %55 to %48 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb15
+^bb15: // 2 preds: ^bb2, ^bb14
+ fir.unpack_array %46 to %40 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %38 to %31 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb16
+^bb16: // 2 preds: ^bb1, ^bb15
+ fir.unpack_array %29 to %23 heap : !fir.box<!fir.array<?x?xi32>>
+ fir.unpack_array %21 to %14 heap : !fir.box<!fir.array<?x?xi32>>
+ cf.br ^bb17
+^bb17: // 2 preds: ^bb0, ^bb16
+ fir.unpack_array %12 to %10 heap : !fir.box<!fir.array<?x?xi32>>
+ return
+}
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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.
A long chain of fir.pack_arrays might require multiple iterations
of the greedy rewriter, if we use down-top traversal. The rewriter
may not converge in 10 (default) iterations. It is not an error,
but it was reported as such.
This patch changes the traversal to top-down and also disabled
the hard error, if the rewriter does not converge soon enough.