[IROutliner] Strip loop info when moving code into outlined function.

[AndrewLitteken]

Summary

As pointed out in https://github.com/llvm/llvm-project/issues/54155#issuecomment-1057465479, there was a crash when loop info was being outlined. It was not being properly stripped and adjusted, so would point to the wrong location. This uses similar logic found in the CodeExtractor to adjust the loop debug info.

[fhahn]

llvm/test/Transforms/IROutliner/outlining-strip-loop-info.ll

13

Is all the complexity needed in the test case? IIUC it would be sufficient to have a function with a simple outline-able loop + !llvm.loop metadata with debug info.

[paquette]

Just verified that this resolves the crash in the benchmark on my end.

I agree that this test case can be simplified. After that, I think this is good to go.

[AndrewLitteken]

Reducing test case

[paquette]

This testcase is better, but there are still lots of instructions which aren't necessary.

How about this:

https://godbolt.org/z/r8Gq7vdbx

?

llvm/lib/Transforms/IPO/IROutliner.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Transforms/IPO/IROutliner.h"
 #include "llvm/Analysis/IRSimilarityIdentifier.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Attributes.h"
-#include "llvm/IR/DebugInfoMetadata.h"
 #include "llvm/IR/DIBuilder.h"
+#include "llvm/IR/DebugInfo.h"
+#include "llvm/IR/DebugInfoMetadata.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Mangler.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Transforms/IPO.h"
 #include <map>
for (BasicBlock &CurrBB : llvm::make_early_inc_range(Old)) {
     std::vector<Instruction *> DebugInsts;
 
     for (Instruction &Val : CurrBB) {
       // We must handle the scoping of called functions differently than
       // other outlined instructions.
       if (!isa<CallInst>(&Val)) {
         // Remove the debug information for outlined functions.
         Val.setDebugLoc(DebugLoc());
+
+        // Loop info metadata may contain line locations. Update them to have no
+        // value in the new subprogram since the outlined code could be from
+        // several locations.
+        auto updateLoopInfoLoc = [&New](Metadata *MD) -> Metadata * {
+          if (DISubprogram *SP = New.getSubprogram())
+            if (auto *Loc = dyn_cast_or_null<DILocation>(MD))
+              return DILocation::get(New.getContext(), Loc->getLine(),
+                                     Loc->getColumn(), SP, nullptr);
+          return MD;
+        };
+        updateLoopMetadataDebugLocations(Val, updateLoopInfoLoc);
         continue;
       }
 
       // From this point we are only handling call instructions.
       CallInst *CI = cast<CallInst>(&Val);
 
       // We add any debug statements here, to be removed after.  Since the
       // instructions originate from many different locations in the program,

llvm/test/Transforms/IROutliner/outlining-strip-loop-info.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --include-generated-funcs
+; RUN: opt -S -verify -iroutliner -ir-outlining-no-cost < %s | FileCheck %s
+
+; Make sure that we strip loop debug info from instructions that are outlined as
+; well as regular debug info.
+
+target triple = "arm64-apple-unknown"
+
+%struct.hoge = type { %struct.widget, i32, i32, [4 x i8] }
+%struct.widget = type <{ i32 (...)**, %struct.quux*, i32, [4 x i8], %struct.hoge.0*, %struct.hoge.0*, i32 }>
+%struct.quux = type { i32 (...)** }
+%struct.hoge.0 = type { i32, i32, i32, i32, i8*, %struct.quux* }
+%struct.barney = type <{ i32 (...)**, %struct.quux*, i32, [4 x i8], %struct.hoge.0*, %struct.hoge.0*, i32, [4 x i8] }>
+
+define void @ham(%struct.hoge* nocapture noundef nonnull readonly align 8 dereferenceable(52) %arg, %struct.hoge.0* noundef nonnull align 8 dereferenceable(32) %arg1) unnamed_addr #0 align 2 !dbg !1 {
+bb:
+  br i1 undef, label %bb2, label %bb21
+
+bb2:                                              ; preds = %bb
+  %var = load i32, i32* undef, align 8
+  %var3 = icmp ult i32 %var, 65
+  br i1 %var3, label %bb10, label %bb4
+
+bb4:                                              ; preds = %bb2
+  %var5 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 1
+  %var6 = load i32, i32* %var5, align 4
+  %var7 = icmp eq i32 %var6, 0
+  br i1 %var7, label %bb22, label %bb8
+
+bb8:                                              ; preds = %bb4
+  %var9 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 4
+  br label %bb13
+
+bb10:                                             ; preds = %bb2
+  %var11 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 2
+  %var12 = bitcast i32* %var11 to <2 x i32>*
+  store <2 x i32> zeroinitializer, <2 x i32>* %var12, align 8
+  br label %bb22
+
+bb13:                                             ; preds = %bb13, %bb8
+  %var14 = phi i64 [ 0, %bb8 ], [ %var17, %bb13 ]
+  %var15 = load i8*, i8** %var9, align 8
+  %var16 = getelementptr inbounds i8, i8* %var15, i64 %var14
+  store i8 0, i8* %var16, align 1
+  %var17 = add nuw nsw i64 %var14, 1
+  %var18 = load i32, i32* %var5, align 4
+  %var19 = zext i32 %var18 to i64
+  %var20 = icmp ult i64 %var17, %var19
+  br i1 %var20, label %bb13, label %bb22, !dbg !7, !llvm.loop !11
+
+bb21:                                             ; preds = %bb
+  unreachable
+
+bb22:                                             ; preds = %bb13, %bb10, %bb4
+  ret void
+}
+
+; Function Attrs: optsize ssp uwtable
+define void @wombat(%struct.hoge* nocapture noundef nonnull readonly align 8 dereferenceable(52) %arg, %struct.hoge.0* noundef nonnull align 8 dereferenceable(32) %arg1) unnamed_addr #0 align 2 !dbg !13 {
+bb:
+  br i1 undef, label %bb2, label %bb21
+
+bb2:                                              ; preds = %bb
+  %var = load i32, i32* undef, align 8
+  %var3 = icmp ult i32 %var, 65
+  br i1 %var3, label %bb10, label %bb4
+
+bb4:                                              ; preds = %bb2
+  %var5 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 1
+  %var6 = load i32, i32* %var5, align 4
+  %var7 = icmp eq i32 %var6, 0
+  br i1 %var7, label %bb22, label %bb8
+
+bb8:                                              ; preds = %bb4
+  %var9 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 4
+  br label %bb13
+
+bb10:                                             ; preds = %bb2
+  %var11 = getelementptr inbounds %struct.hoge.0, %struct.hoge.0* %arg1, i64 0, i32 2
+  %var12 = bitcast i32* %var11 to <2 x i32>*
+  store <2 x i32> zeroinitializer, <2 x i32>* %var12, align 8
+  br label %bb22
+
+bb13:                                             ; preds = %bb13, %bb8
+  %var14 = phi i64 [ 0, %bb8 ], [ %var17, %bb13 ]
+  %var15 = load i8*, i8** %var9, align 8
+  %var16 = getelementptr inbounds i8, i8* %var15, i64 %var14
+  store i8 0, i8* %var16, align 1
+  %var17 = add nuw nsw i64 %var14, 1
+  %var18 = load i32, i32* %var5, align 4
+  %var19 = zext i32 %var18 to i64
+  %var20 = icmp ult i64 %var17, %var19
+  br i1 %var20, label %bb13, label %bb22
+
+bb21:                                             ; preds = %bb
+  unreachable
+
+bb22:                                             ; preds = %bb13, %bb10, %bb4
+  ret void
+}
+
+; Function Attrs: inlinehint nounwind optsize ssp uwtable
+define linkonce_odr void @foo(%struct.barney* noundef nonnull align 8 dereferenceable(44) %arg) unnamed_addr #1 align 2 !dbg !14 {
+bb:
+  unreachable
+}
+
+attributes #0 = { optsize ssp uwtable }
+attributes #1 = { inlinehint nounwind optsize ssp uwtable }
+
+!llvm.module.flags = !{!0}
+!llvm.dbg.cu = !{}
+
+!0 = !{i32 2, !"Debug Info Version", i32 3}
+!1 = distinct !DISubprogram(name: "pluto", scope: !2, file: !2, line: 160, type: !3, scopeLine: 161, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !5)
+!2 = !DIFile(filename: "file", directory: "dir")
+!3 = !DISubroutineType(types: !4)
+!4 = !{}
+!5 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus, file: !6, producer: "ver", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, splitDebugInlining: false, nameTableKind: None, sysroot: "sysroot", sdk: "sdk")
+!6 = !DIFile(filename: "file", directory: "dir")
+!7 = !DILocation(line: 354, column: 13, scope: !8, inlinedAt: !10)
+!8 = distinct !DISubprogram(name: "baz", scope: !9, file: !9, line: 345, type: !3, scopeLine: 346, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !5)
+!9 = !DIFile(filename: "file", directory: "dir")
+!10 = distinct !DILocation(line: 164, column: 15, scope: !1)
+!11 = distinct !{!11, !7, !12}
+!12 = !DILocation(line: 355, column: 37, scope: !8, inlinedAt: !10)
+!13 = distinct !DISubprogram(name: "foo", scope: !2, file: !2, line: 172, type: !3, scopeLine: 173, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !5)
+!14 = distinct !DISubprogram(name: "bar", scope: !15, file: !15, line: 219, type: !3, scopeLine: 220, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !5)
+!15 = !DIFile(filename: "file", directory: "dir")
+; CHECK-LABEL: @ham(
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    br i1 undef, label [[BB2:%.*]], label [[BB21:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    call void @outlined_ir_func_0(%struct.hoge.0* [[ARG1:%.*]]), !dbg [[DBG6:![0-9]+]]
+; CHECK-NEXT:    br label [[BB22:%.*]]
+; CHECK:       bb21:
+; CHECK-NEXT:    unreachable
+; CHECK:       bb22:
+; CHECK-NEXT:    ret void
+;
+;
+; CHECK-LABEL: @wombat(
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    br i1 undef, label [[BB2:%.*]], label [[BB21:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    call void @outlined_ir_func_0(%struct.hoge.0* [[ARG1:%.*]]), !dbg [[DBG10:![0-9]+]]
+; CHECK-NEXT:    br label [[BB22:%.*]]
+; CHECK:       bb21:
+; CHECK-NEXT:    unreachable
+; CHECK:       bb22:
+; CHECK-NEXT:    ret void
+;
+;
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    unreachable
+;
+;
+; CHECK-LABEL: define internal void @outlined_ir_func_0(
+; CHECK-NEXT:  newFuncRoot:
+; CHECK-NEXT:    br label [[BB2_TO_OUTLINE:%.*]]
+; CHECK:       bb2_to_outline:
+; CHECK-NEXT:    [[VAR:%.*]] = load i32, i32* undef, align 8
+; CHECK-NEXT:    [[VAR3:%.*]] = icmp ult i32 [[VAR]], 65
+; CHECK-NEXT:    br i1 [[VAR3]], label [[BB10:%.*]], label [[BB4:%.*]]
+; CHECK:       bb4:
+; CHECK-NEXT:    [[VAR5:%.*]] = getelementptr inbounds [[STRUCT_HOGE_0:%.*]], %struct.hoge.0* [[TMP0:%.*]], i64 0, i32 1
+; CHECK-NEXT:    [[VAR6:%.*]] = load i32, i32* [[VAR5]], align 4
+; CHECK-NEXT:    [[VAR7:%.*]] = icmp eq i32 [[VAR6]], 0
+; CHECK-NEXT:    br i1 [[VAR7]], label [[BB22_EXITSTUB:%.*]], label [[BB8:%.*]]
+; CHECK:       bb8:
+; CHECK-NEXT:    [[VAR9:%.*]] = getelementptr inbounds [[STRUCT_HOGE_0]], %struct.hoge.0* [[TMP0]], i64 0, i32 4
+; CHECK-NEXT:    br label [[BB13:%.*]]
+; CHECK:       bb10:
+; CHECK-NEXT:    [[VAR11:%.*]] = getelementptr inbounds [[STRUCT_HOGE_0]], %struct.hoge.0* [[TMP0]], i64 0, i32 2
+; CHECK-NEXT:    [[VAR12:%.*]] = bitcast i32* [[VAR11]] to <2 x i32>*
+; CHECK-NEXT:    store <2 x i32> zeroinitializer, <2 x i32>* [[VAR12]], align 8
+; CHECK-NEXT:    br label [[BB22_EXITSTUB]]
+; CHECK:       bb13:
+; CHECK-NEXT:    [[VAR14:%.*]] = phi i64 [ 0, [[BB8]] ], [ [[VAR17:%.*]], [[BB13]] ]
+; CHECK-NEXT:    [[VAR15:%.*]] = load i8*, i8** [[VAR9]], align 8
+; CHECK-NEXT:    [[VAR16:%.*]] = getelementptr inbounds i8, i8* [[VAR15]], i64 [[VAR14]]
+; CHECK-NEXT:    store i8 0, i8* [[VAR16]], align 1
+; CHECK-NEXT:    [[VAR17]] = add nuw nsw i64 [[VAR14]], 1
+; CHECK-NEXT:    [[VAR18:%.*]] = load i32, i32* [[VAR5]], align 4
+; CHECK-NEXT:    [[VAR19:%.*]] = zext i32 [[VAR18]] to i64
+; CHECK-NEXT:    [[VAR20:%.*]] = icmp ult i64 [[VAR17]], [[VAR19]]
+; CHECK-NEXT:    br i1 [[VAR20]], label [[BB13]], label [[BB22_EXITSTUB]], !llvm.loop [[LOOP13:![0-9]+]]
+; CHECK:       bb22.exitStub:
+; CHECK-NEXT:    ret void
+;