[llvm] [DebugInfo] Preserve line and column number when merging debug info. (PR #129960)
Snehasish Kumar via llvm-commits
llvm-commits at lists.llvm.org
Mon Mar 24 15:16:46 PDT 2025
https://github.com/snehasish updated https://github.com/llvm/llvm-project/pull/129960
>From cb23baa91558598ffff660bd68164a4d6ba0d3ea Mon Sep 17 00:00:00 2001
From: Snehasish Kumar <snehasishk at google.com>
Date: Wed, 5 Mar 2025 22:52:03 +0000
Subject: [PATCH 1/5] [DebugInfo] Preserve line and column number when merging
debug info.
This patch introduces a new option `-preserve-merged-debug-info` to
preserve an arbitrary version of debug information when DILocations are
merged. This is intended to be used in production environments from
which sample based profiles are derived such as AutoFDO and MemProf.
With this patch we have see a 0.2% improvement on an internal workload
at Google when generating AutoFDO profiles. It also significantly
improves the ability for MemProf by preserving debug info for merged
call instructions used in the contextual profile.
Co-authored-by: Krzysztof Pszeniczny <kpszeniczny at google.com>
---
llvm/lib/IR/DebugInfoMetadata.cpp | 15 +++++
.../DebugInfo/preserve-merged-debug-info.ll | 65 +++++++++++++++++++
2 files changed, 80 insertions(+)
create mode 100644 llvm/test/DebugInfo/preserve-merged-debug-info.ll
diff --git a/llvm/lib/IR/DebugInfoMetadata.cpp b/llvm/lib/IR/DebugInfoMetadata.cpp
index f975d4ca33ad9..b18db09820068 100644
--- a/llvm/lib/IR/DebugInfoMetadata.cpp
+++ b/llvm/lib/IR/DebugInfoMetadata.cpp
@@ -21,6 +21,7 @@
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
+#include "llvm/Support/CommandLine.h"
#include <numeric>
#include <optional>
@@ -34,6 +35,12 @@ cl::opt<bool> EnableFSDiscriminator(
cl::desc("Enable adding flow sensitive discriminators"));
} // namespace llvm
+// When true, preserves line and column number by picking one of the merged
+// location info in a deterministic manner to assist sample based PGO.
+static cl::opt<bool> PreserveMergedDebugInfo(
+ "preserve-merged-debug-info", cl::init(false), cl::Hidden,
+ cl::desc("Preserve line and column number when merging locations."));
+
uint32_t DIType::getAlignInBits() const {
return (getTag() == dwarf::DW_TAG_LLVM_ptrauth_type ? 0 : SubclassData32);
}
@@ -125,6 +132,14 @@ DILocation *DILocation::getMergedLocation(DILocation *LocA, DILocation *LocB) {
if (LocA == LocB)
return LocA;
+ if (PreserveMergedDebugInfo) {
+ auto A = std::make_tuple(LocA->getLine(), LocA->getColumn(),
+ LocA->getDiscriminator());
+ auto B = std::make_tuple(LocB->getLine(), LocB->getColumn(),
+ LocB->getDiscriminator());
+ return A < B ? LocA : LocB;
+ }
+
LLVMContext &C = LocA->getContext();
using LocVec = SmallVector<const DILocation *>;
diff --git a/llvm/test/DebugInfo/preserve-merged-debug-info.ll b/llvm/test/DebugInfo/preserve-merged-debug-info.ll
new file mode 100644
index 0000000000000..d2b5af0d2d2c3
--- /dev/null
+++ b/llvm/test/DebugInfo/preserve-merged-debug-info.ll
@@ -0,0 +1,65 @@
+; RUN: opt %s -passes=simplifycfg -hoist-common-insts -preserve-merged-debug-info -S | FileCheck %s
+; CHECK: tail call i32 @bar{{.*!dbg !}}[[TAG:[0-9]+]]
+; CHECK: ![[TAG]] = !DILocation(line: 9, column: 16, scope: !9)
+
+; ModuleID = '../llvm/test/DebugInfo/Inputs/debug-info-merge-call.c'
+source_filename = "../llvm/test/DebugInfo/Inputs/debug-info-merge-call.c"
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Function Attrs: nounwind uwtable
+define dso_local i32 @test(i32 noundef %n) local_unnamed_addr #0 !dbg !9 {
+entry:
+ %call = tail call i32 @foo(i32 noundef %n) #2, !dbg !12
+ %cmp1 = icmp sgt i32 %n, 100, !dbg !13
+ br i1 %cmp1, label %if.then, label %if.else, !dbg !13
+
+if.then: ; preds = %entry
+ %call2 = tail call i32 @bar(i32 noundef %n) #2, !dbg !14
+ %add = add nsw i32 %call2, %call, !dbg !15
+ br label %if.end, !dbg !16
+
+if.else: ; preds = %entry
+ %call4 = tail call i32 @bar(i32 noundef %n) #2, !dbg !17
+ br label %if.end
+
+if.end: ; preds = %if.else, %if.then
+ %r.0 = phi i32 [ %add, %if.then ], [ %call4, %if.else ], !dbg !18
+ ret i32 %r.0, !dbg !19
+}
+
+declare !dbg !20 i32 @foo(i32 noundef) local_unnamed_addr #1
+
+declare !dbg !21 i32 @bar(i32 noundef) local_unnamed_addr #1
+
+attributes #0 = { nounwind uwtable "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
+attributes #1 = { "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
+attributes #2 = { nounwind }
+
+!llvm.dbg.cu = !{!0}
+!llvm.module.flags = !{!2, !3, !4, !5, !6, !7}
+!llvm.ident = !{!8}
+
+!0 = distinct !DICompileUnit(language: DW_LANG_C11, file: !1, producer: "clang version 21.0.0git (git at github.com:snehasish/llvm-project.git 6ce41db6b0275d060d6e60f88b96a1657024345c)", isOptimized: true, runtimeVersion: 0, emissionKind: LineTablesOnly, splitDebugInlining: false, nameTableKind: None)
+!1 = !DIFile(filename: "../llvm/test/DebugInfo/Inputs/debug-info-merge-call.c", directory: "/usr/local/google/home/snehasishk/working/llvm-project/build-assert", checksumkind: CSK_MD5, checksum: "ac1be6c40dad11691922d600f9d55c55")
+!2 = !{i32 7, !"Dwarf Version", i32 5}
+!3 = !{i32 2, !"Debug Info Version", i32 3}
+!4 = !{i32 1, !"wchar_size", i32 4}
+!5 = !{i32 8, !"PIC Level", i32 2}
+!6 = !{i32 7, !"PIE Level", i32 2}
+!7 = !{i32 7, !"uwtable", i32 2}
+!8 = !{!"clang version 21.0.0git (git at github.com:snehasish/llvm-project.git 6ce41db6b0275d060d6e60f88b96a1657024345c)"}
+!9 = distinct !DISubprogram(name: "test", scope: !1, file: !1, line: 5, type: !10, scopeLine: 5, flags: DIFlagPrototyped | DIFlagAllCallsDescribed, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0)
+!10 = !DISubroutineType(types: !11)
+!11 = !{}
+!12 = !DILocation(line: 7, column: 13, scope: !9)
+!13 = !DILocation(line: 8, column: 8, scope: !9)
+!14 = !DILocation(line: 9, column: 16, scope: !9)
+!15 = !DILocation(line: 9, column: 14, scope: !9)
+!16 = !DILocation(line: 10, column: 3, scope: !9)
+!17 = !DILocation(line: 11, column: 10, scope: !9)
+!18 = !DILocation(line: 0, scope: !9)
+!19 = !DILocation(line: 13, column: 3, scope: !9)
+!20 = !DISubprogram(name: "foo", scope: !1, file: !1, line: 2, type: !10, flags: DIFlagPrototyped, spFlags: DISPFlagOptimized)
+!21 = !DISubprogram(name: "bar", scope: !1, file: !1, line: 1, type: !10, flags: DIFlagPrototyped, spFlags: DISPFlagOptimized)
+
>From df04e25cffc85b3a4d9721aad68e0fab2f3f6453 Mon Sep 17 00:00:00 2001
From: Snehasish Kumar <snehasishk at google.com>
Date: Wed, 5 Mar 2025 22:52:03 +0000
Subject: [PATCH 2/5] Address comments.
---
.../DebugInfo/preserve-merged-debug-info.ll | 38 ++++++++++++-------
1 file changed, 25 insertions(+), 13 deletions(-)
diff --git a/llvm/test/DebugInfo/preserve-merged-debug-info.ll b/llvm/test/DebugInfo/preserve-merged-debug-info.ll
index d2b5af0d2d2c3..a5b7bf01c6020 100644
--- a/llvm/test/DebugInfo/preserve-merged-debug-info.ll
+++ b/llvm/test/DebugInfo/preserve-merged-debug-info.ll
@@ -1,6 +1,22 @@
-; RUN: opt %s -passes=simplifycfg -hoist-common-insts -preserve-merged-debug-info -S | FileCheck %s
-; CHECK: tail call i32 @bar{{.*!dbg !}}[[TAG:[0-9]+]]
-; CHECK: ![[TAG]] = !DILocation(line: 9, column: 16, scope: !9)
+;; This test verifies that we assign a deterministic location for merged
+;; instructions when -preserve-merged-debug-info is enabled. We use the
+;; simplifycfg pass to test this behaviour since it was a common source of
+;; merged instructions, however we intend this to apply to all users of the
+;; getMergedLocation API.
+
+;; Run simplifycfg and check that only 1 call to bar remains and it's debug
+;; location has a valid line number (lexicographically smallest).
+; RUN: opt %s -passes=simplifycfg -hoist-common-insts -preserve-merged-debug-info -S | FileCheck %s --check-prefix=ENABLED
+; ENABLED: call i32 @bar{{.*!dbg !}}[[TAG:[0-9]+]]
+; ENABLED-NOT: call i32 @bar
+; ENABLED: ![[TAG]] = !DILocation(line: 9, column: 16, scope: !9)
+
+;; Run simplifycfg without the pass to ensure that we don't spuriously start
+;; passing the test if simplifycfg behaviour changes.
+; RUN: opt %s -passes=simplifycfg -hoist-common-insts -preserve-merged-debug-info=false -S | FileCheck %s --check-prefix=DISABLED
+; DISABLED: call i32 @bar{{.*!dbg !}}[[TAG:[0-9]+]]
+; DISABLED-NOT: call i32 @bar
+; DISABLED: ![[TAG]] = !DILocation(line: 0, scope: !9)
; ModuleID = '../llvm/test/DebugInfo/Inputs/debug-info-merge-call.c'
source_filename = "../llvm/test/DebugInfo/Inputs/debug-info-merge-call.c"
@@ -8,19 +24,19 @@ target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:
target triple = "x86_64-unknown-linux-gnu"
; Function Attrs: nounwind uwtable
-define dso_local i32 @test(i32 noundef %n) local_unnamed_addr #0 !dbg !9 {
+define dso_local i32 @test(i32 %n) !dbg !9 {
entry:
- %call = tail call i32 @foo(i32 noundef %n) #2, !dbg !12
+ %call = call i32 @foo(i32 %n), !dbg !12
%cmp1 = icmp sgt i32 %n, 100, !dbg !13
br i1 %cmp1, label %if.then, label %if.else, !dbg !13
if.then: ; preds = %entry
- %call2 = tail call i32 @bar(i32 noundef %n) #2, !dbg !14
+ %call2 = call i32 @bar(i32 %n), !dbg !14
%add = add nsw i32 %call2, %call, !dbg !15
br label %if.end, !dbg !16
if.else: ; preds = %entry
- %call4 = tail call i32 @bar(i32 noundef %n) #2, !dbg !17
+ %call4 = call i32 @bar(i32 %n), !dbg !17
br label %if.end
if.end: ; preds = %if.else, %if.then
@@ -28,13 +44,9 @@ if.end: ; preds = %if.else, %if.then
ret i32 %r.0, !dbg !19
}
-declare !dbg !20 i32 @foo(i32 noundef) local_unnamed_addr #1
-
-declare !dbg !21 i32 @bar(i32 noundef) local_unnamed_addr #1
+declare !dbg !20 i32 @foo(i32)
-attributes #0 = { nounwind uwtable "min-legal-vector-width"="0" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
-attributes #1 = { "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+cmov,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
-attributes #2 = { nounwind }
+declare !dbg !21 i32 @bar(i32)
!llvm.dbg.cu = !{!0}
!llvm.module.flags = !{!2, !3, !4, !5, !6, !7}
>From e43ffd1118bd3d6b974bda0c4c8d0f221ab271bf Mon Sep 17 00:00:00 2001
From: Snehasish Kumar <snehasishk at google.com>
Date: Fri, 14 Mar 2025 22:49:39 +0000
Subject: [PATCH 3/5] Update docs to mention SamplePGO debug info usage.
---
llvm/docs/HowToUpdateDebugInfo.rst | 14 ++++++++++----
llvm/docs/SourceLevelDebugging.rst | 8 ++++++++
2 files changed, 18 insertions(+), 4 deletions(-)
diff --git a/llvm/docs/HowToUpdateDebugInfo.rst b/llvm/docs/HowToUpdateDebugInfo.rst
index d8c300f2f3a70..b68f9619f9417 100644
--- a/llvm/docs/HowToUpdateDebugInfo.rst
+++ b/llvm/docs/HowToUpdateDebugInfo.rst
@@ -9,7 +9,8 @@ Introduction
============
Certain kinds of code transformations can inadvertently result in a loss of
-debug info, or worse, make debug info misrepresent the state of a program.
+debug info, or worse, make debug info misrepresent the state of a program. Debug
+info availability is also essential for SamplePGO.
This document specifies how to correctly update debug info in various kinds of
code transformations, and offers suggestions for how to create targeted debug
@@ -89,9 +90,14 @@ has a location with an accurate scope attached, and b) to prevent misleading
single-stepping (or breakpoint) behavior. Often, merged instructions are memory
accesses which can trap: having an accurate scope attached greatly assists in
crash triage by identifying the (possibly inlined) function where the bad
-memory access occurred. This rule is also meant to assist SamplePGO by banning
-scenarios in which a sample of a block containing a merged instruction is
-misattributed to a block containing one of the instructions-to-be-merged.
+memory access occurred.
+
+For SamplePGO, it is often beneficial to retain an arbitrary but deterministic
+location instead of discarding line and column information as part of merging.
+In particular, loss of location information for calls inhibit optimizations
+such as indirect call promotion. This behavior can be optionally enabled until
+support for accurately representing merged instructions in the line table is
+implemented.
Examples of transformations that should follow this rule include:
diff --git a/llvm/docs/SourceLevelDebugging.rst b/llvm/docs/SourceLevelDebugging.rst
index c1a95efd2d8bc..f0e233560ede7 100644
--- a/llvm/docs/SourceLevelDebugging.rst
+++ b/llvm/docs/SourceLevelDebugging.rst
@@ -55,6 +55,8 @@ the stored debug information into source-language specific information. As
such, a debugger must be aware of the source-language, and is thus tied to a
specific language or family of languages.
+.. _intro_consumers:
+
Debug information consumers
---------------------------
@@ -71,6 +73,12 @@ as Visual Studio and WinDBG. LLVM's debug information format is mostly derived
from and inspired by DWARF, but it is feasible to translate into other target
debug info formats such as STABS.
+SamplePGO (also known as `AutoFDO <https://gcc.gnu.org/wiki/AutoFDO>`_)
+is a variant of profile guided optimizations which uses hardware sampling based
+profilers to collect branch frequency data to with low overhead in production
+environments. It relies on debug information to associate profile information
+to LLVM IR which is then used to guide optimization heuristics.
+
It would also be reasonable to use debug information to feed profiling tools
for analysis of generated code, or, tools for reconstructing the original
source from generated code.
>From ffea5316806d11ad9adc881b820b031cd7307da6 Mon Sep 17 00:00:00 2001
From: Snehasish Kumar <snehasishk at google.com>
Date: Tue, 18 Mar 2025 03:37:59 +0000
Subject: [PATCH 4/5] Fix typo.
---
llvm/docs/SourceLevelDebugging.rst | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/llvm/docs/SourceLevelDebugging.rst b/llvm/docs/SourceLevelDebugging.rst
index f0e233560ede7..43bec9c22e65f 100644
--- a/llvm/docs/SourceLevelDebugging.rst
+++ b/llvm/docs/SourceLevelDebugging.rst
@@ -75,7 +75,7 @@ debug info formats such as STABS.
SamplePGO (also known as `AutoFDO <https://gcc.gnu.org/wiki/AutoFDO>`_)
is a variant of profile guided optimizations which uses hardware sampling based
-profilers to collect branch frequency data to with low overhead in production
+profilers to collect branch frequency data with low overhead in production
environments. It relies on debug information to associate profile information
to LLVM IR which is then used to guide optimization heuristics.
>From 713ae857aa9c49a5a86c7f05d7625f81ba7ee642 Mon Sep 17 00:00:00 2001
From: Snehasish Kumar <snehasishk at google.com>
Date: Mon, 24 Mar 2025 22:14:19 +0000
Subject: [PATCH 5/5] Enhance the documentation for SamplePGO.
---
llvm/docs/HowToUpdateDebugInfo.rst | 16 ++++++++--------
llvm/docs/SourceLevelDebugging.rst | 7 ++++++-
2 files changed, 14 insertions(+), 9 deletions(-)
diff --git a/llvm/docs/HowToUpdateDebugInfo.rst b/llvm/docs/HowToUpdateDebugInfo.rst
index b68f9619f9417..3088f59c1066a 100644
--- a/llvm/docs/HowToUpdateDebugInfo.rst
+++ b/llvm/docs/HowToUpdateDebugInfo.rst
@@ -90,14 +90,14 @@ has a location with an accurate scope attached, and b) to prevent misleading
single-stepping (or breakpoint) behavior. Often, merged instructions are memory
accesses which can trap: having an accurate scope attached greatly assists in
crash triage by identifying the (possibly inlined) function where the bad
-memory access occurred.
-
-For SamplePGO, it is often beneficial to retain an arbitrary but deterministic
-location instead of discarding line and column information as part of merging.
-In particular, loss of location information for calls inhibit optimizations
-such as indirect call promotion. This behavior can be optionally enabled until
-support for accurately representing merged instructions in the line table is
-implemented.
+memory access occurred.
+
+To maintain distinct source locations for SamplePGO, it is often beneficial to
+retain an arbitrary but deterministic location instead of discarding line and
+column information as part of merging. In particular, loss of location
+information for calls inhibit optimizations such as indirect call promotion.
+This behavior can be optionally enabled until support for accurately
+representing merged instructions in the line table is implemented.
Examples of transformations that should follow this rule include:
diff --git a/llvm/docs/SourceLevelDebugging.rst b/llvm/docs/SourceLevelDebugging.rst
index 43bec9c22e65f..ef70b86ef8685 100644
--- a/llvm/docs/SourceLevelDebugging.rst
+++ b/llvm/docs/SourceLevelDebugging.rst
@@ -77,7 +77,12 @@ SamplePGO (also known as `AutoFDO <https://gcc.gnu.org/wiki/AutoFDO>`_)
is a variant of profile guided optimizations which uses hardware sampling based
profilers to collect branch frequency data with low overhead in production
environments. It relies on debug information to associate profile information
-to LLVM IR which is then used to guide optimization heuristics.
+to LLVM IR which is then used to guide optimization heuristics. Maintaining
+deterministic and distinct source locations is necessary to maximize the
+accuracy of mapping hardware sample counts to LLVM IR. For example, DWARF
+`discriminators <https://wiki.dwarfstd.org/Path_Discriminators.md>`_ allow
+SamplePGO to distinguish between multiple paths of execution which map to the
+same source line.
It would also be reasonable to use debug information to feed profiling tools
for analysis of generated code, or, tools for reconstructing the original
More information about the llvm-commits
mailing list