[lld] [LLD][ELF] Add --why-live flag to report GC liveness reason (PR #119279)
Daniel Thornburgh via llvm-commits
llvm-commits at lists.llvm.org
Wed Jan 29 16:17:22 PST 2025
https://github.com/mysterymath updated https://github.com/llvm/llvm-project/pull/119279
>From 802fc57ddd4291404b009aa7e99d3e7bf83ecd74 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 13 Nov 2024 15:12:55 -0800
Subject: [PATCH 01/30] Pass through parent enqueue section and offset
---
lld/ELF/MarkLive.cpp | 28 ++++++++++++++++++----------
1 file changed, 18 insertions(+), 10 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index b6c22884d91769..04caf531720e4a 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -42,6 +42,11 @@ using namespace lld;
using namespace lld::elf;
namespace {
+struct LiveParent {
+ InputSectionBase *sec;
+ std::optional<uint64_t> offset;
+};
+
template <class ELFT> class MarkLive {
public:
MarkLive(Ctx &ctx, unsigned partition) : ctx(ctx), partition(partition) {}
@@ -50,7 +55,7 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, uint64_t offset);
+ void enqueue(InputSectionBase *sec, uint64_t offset, std::optional<LiveParent> parent);
void markSymbol(Symbol *sym);
void mark();
@@ -101,6 +106,8 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
Symbol &sym = sec.file->getRelocTargetSym(rel);
sym.used = true;
+ LiveParent parent = {&sec, rel.r_offset};
+
if (auto *d = dyn_cast<Defined>(&sym)) {
auto *relSec = dyn_cast_or_null<InputSectionBase>(d->section);
if (!relSec)
@@ -120,7 +127,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
// discarded, marking the LSDA will unnecessarily retain the text section.
if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
relSec->nextInSectionGroup)))
- enqueue(relSec, offset);
+ enqueue(relSec, offset, parent);
return;
}
@@ -129,7 +136,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
cast<SharedFile>(ss->file)->isNeeded = true;
for (InputSectionBase *sec : cNamedSections.lookup(sym.getName()))
- enqueue(sec, 0);
+ enqueue(sec, 0, parent);
}
// The .eh_frame section is an unfortunate special case.
@@ -187,7 +194,8 @@ static bool isReserved(InputSectionBase *sec) {
}
template <class ELFT>
-void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset) {
+void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
+ std::optional<LiveParent> parent) {
// Usually, a whole section is marked as live or dead, but in mergeable
// (splittable) sections, each piece of data has independent liveness bit.
// So we explicitly tell it which offset is in use.
@@ -209,7 +217,7 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset) {
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
if (auto *d = dyn_cast_or_null<Defined>(sym))
if (auto *isec = dyn_cast_or_null<InputSectionBase>(d->section))
- enqueue(isec, d->value);
+ enqueue(isec, d->value, std::nullopt);
}
// This is the main function of the garbage collector.
@@ -256,7 +264,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
}
for (InputSectionBase *sec : ctx.inputSections) {
if (sec->flags & SHF_GNU_RETAIN) {
- enqueue(sec, 0);
+ enqueue(sec, 0, std::nullopt);
continue;
}
if (sec->flags & SHF_LINK_ORDER)
@@ -295,7 +303,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
// Preserve special sections and those which are specified in linker
// script KEEP command.
if (isReserved(sec) || ctx.script->shouldKeep(sec)) {
- enqueue(sec, 0);
+ enqueue(sec, 0, std::nullopt);
} else if ((!ctx.arg.zStartStopGC || sec->name.starts_with("__libc_")) &&
isValidCIdentifier(sec->name)) {
// As a workaround for glibc libc.a before 2.34
@@ -323,11 +331,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, 0);
+ enqueue(isec, 0, LiveParent{&sec, std::nullopt});
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, 0);
+ enqueue(sec.nextInSectionGroup, 0, LiveParent{&sec, std::nullopt});
}
}
@@ -353,7 +361,7 @@ template <class ELFT> void MarkLive<ELFT>::moveToMain() {
continue;
if (ctx.symtab->find(("__start_" + sec->name).str()) ||
ctx.symtab->find(("__stop_" + sec->name).str()))
- enqueue(sec, 0);
+ enqueue(sec, 0, std::nullopt);
}
mark();
>From 6d376d5d528ea1151ee3235f403249f45adb0b24 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 27 Nov 2024 16:07:13 -0800
Subject: [PATCH 02/30] Track parents
---
lld/ELF/MarkLive.cpp | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 04caf531720e4a..59603e61f4f3cb 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -75,6 +75,8 @@ template <class ELFT> class MarkLive {
// There are normally few input sections whose names are valid C
// identifiers, so we just store a SmallVector instead of a multimap.
DenseMap<StringRef, SmallVector<InputSectionBase *, 0>> cNamedSections;
+
+ DenseMap<InputSectionBase*, LiveParent> parents;
};
} // namespace
@@ -209,6 +211,9 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
return;
sec->partition = sec->partition ? 1 : partition;
+ if (parent)
+ parents.try_emplace(sec, *parent);
+
// Add input section to the queue.
if (InputSection *s = dyn_cast<InputSection>(sec))
queue.push_back(s);
>From 5239cdf819ac98cbf879b1c487c2e6ec4e2e59a8 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 2 Dec 2024 15:51:10 -0800
Subject: [PATCH 03/30] Recast as LiveOffset
---
lld/ELF/MarkLive.cpp | 21 +++++++++++++--------
1 file changed, 13 insertions(+), 8 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 59603e61f4f3cb..cd2bc456ea2037 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -42,9 +42,13 @@ using namespace lld;
using namespace lld::elf;
namespace {
-struct LiveParent {
+struct LiveOffset {
InputSectionBase *sec;
std::optional<uint64_t> offset;
+
+ LiveOffset(InputSectionBase *sec,
+ std::optional<uint64_t> offset = std::nullopt)
+ : sec(sec), offset(offset) {}
};
template <class ELFT> class MarkLive {
@@ -55,7 +59,8 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, uint64_t offset, std::optional<LiveParent> parent);
+ void enqueue(InputSectionBase *sec, uint64_t offset,
+ std::optional<LiveOffset> parent);
void markSymbol(Symbol *sym);
void mark();
@@ -76,7 +81,7 @@ template <class ELFT> class MarkLive {
// identifiers, so we just store a SmallVector instead of a multimap.
DenseMap<StringRef, SmallVector<InputSectionBase *, 0>> cNamedSections;
- DenseMap<InputSectionBase*, LiveParent> parents;
+ DenseMap<LiveOffset, LiveOffset> whyLive;
};
} // namespace
@@ -108,7 +113,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
Symbol &sym = sec.file->getRelocTargetSym(rel);
sym.used = true;
- LiveParent parent = {&sec, rel.r_offset};
+ LiveOffset parent = {&sec, rel.r_offset};
if (auto *d = dyn_cast<Defined>(&sym)) {
auto *relSec = dyn_cast_or_null<InputSectionBase>(d->section);
@@ -197,7 +202,7 @@ static bool isReserved(InputSectionBase *sec) {
template <class ELFT>
void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
- std::optional<LiveParent> parent) {
+ std::optional<LiveOffset> parent) {
// Usually, a whole section is marked as live or dead, but in mergeable
// (splittable) sections, each piece of data has independent liveness bit.
// So we explicitly tell it which offset is in use.
@@ -212,7 +217,7 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
sec->partition = sec->partition ? 1 : partition;
if (parent)
- parents.try_emplace(sec, *parent);
+ whyLive.try_emplace(LiveOffset{sec, offset}, *parent);
// Add input section to the queue.
if (InputSection *s = dyn_cast<InputSection>(sec))
@@ -336,11 +341,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, 0, LiveParent{&sec, std::nullopt});
+ enqueue(isec, 0, &sec);
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, 0, LiveParent{&sec, std::nullopt});
+ enqueue(sec.nextInSectionGroup, 0, &sec);
}
}
>From d025c6976ff3d23cead2fe051770a7f9c3497258 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 2 Dec 2024 16:00:58 -0800
Subject: [PATCH 04/30] Use a pair
---
lld/ELF/MarkLive.cpp | 13 +++----------
1 file changed, 3 insertions(+), 10 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index cd2bc456ea2037..90947fc1da2223 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -42,14 +42,7 @@ using namespace lld;
using namespace lld::elf;
namespace {
-struct LiveOffset {
- InputSectionBase *sec;
- std::optional<uint64_t> offset;
-
- LiveOffset(InputSectionBase *sec,
- std::optional<uint64_t> offset = std::nullopt)
- : sec(sec), offset(offset) {}
-};
+typedef std::pair<InputSectionBase *, uint64_t> LiveOffset;
template <class ELFT> class MarkLive {
public:
@@ -341,11 +334,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, 0, &sec);
+ enqueue(isec, 0, {{&sec, 0}});
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, 0, &sec);
+ enqueue(sec.nextInSectionGroup, 0, {{&sec, 0}});
}
}
>From 329c65d1489312726bb6ba966319fe8f531a76c0 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 3 Dec 2024 11:00:50 -0800
Subject: [PATCH 05/30] Also record against the zero offset
---
lld/ELF/MarkLive.cpp | 7 ++++++-
1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 90947fc1da2223..869eb87a2c0219 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -209,8 +209,13 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
return;
sec->partition = sec->partition ? 1 : partition;
- if (parent)
+ if (parent) {
whyLive.try_emplace(LiveOffset{sec, offset}, *parent);
+ // Offset zero is treated as a stand-in for the section itself. The parent
+ // is both a specific reason that an offset within this section is alive and
+ // a generic reason the section itself is alive.
+ whyLive.try_emplace(LiveOffset{sec, 0}, *parent);
+ }
// Add input section to the queue.
if (InputSection *s = dyn_cast<InputSection>(sec))
>From 30be9ebf3ad148bc7805125f082605ba293fba9a Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 3 Dec 2024 11:10:44 -0800
Subject: [PATCH 06/30] Track live objects as either symbols or sections
---
lld/ELF/MarkLive.cpp | 26 ++++++++++++++------------
1 file changed, 14 insertions(+), 12 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 869eb87a2c0219..f8cbdd62bf1545 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -29,9 +29,11 @@
#include "Target.h"
#include "lld/Common/CommonLinkerContext.h"
#include "lld/Common/Strings.h"
+#include "llvm/ADT/DenseMapInfoVariant.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/Object/ELF.h"
#include "llvm/Support/TimeProfiler.h"
+#include <variant>
#include <vector>
using namespace llvm;
@@ -42,7 +44,7 @@ using namespace lld;
using namespace lld::elf;
namespace {
-typedef std::pair<InputSectionBase *, uint64_t> LiveOffset;
+typedef std::variant<InputSectionBase *, Defined *> LiveObject;
template <class ELFT> class MarkLive {
public:
@@ -53,7 +55,7 @@ template <class ELFT> class MarkLive {
private:
void enqueue(InputSectionBase *sec, uint64_t offset,
- std::optional<LiveOffset> parent);
+ std::optional<LiveObject> parent);
void markSymbol(Symbol *sym);
void mark();
@@ -74,7 +76,7 @@ template <class ELFT> class MarkLive {
// identifiers, so we just store a SmallVector instead of a multimap.
DenseMap<StringRef, SmallVector<InputSectionBase *, 0>> cNamedSections;
- DenseMap<LiveOffset, LiveOffset> whyLive;
+ DenseMap<LiveObject, LiveObject> whyLive;
};
} // namespace
@@ -106,7 +108,8 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
Symbol &sym = sec.file->getRelocTargetSym(rel);
sym.used = true;
- LiveOffset parent = {&sec, rel.r_offset};
+ Defined *parentSym = sec.getEnclosingSymbol(rel.r_offset);
+ auto parent = parentSym ? LiveObject(parentSym) : LiveObject(&sec);
if (auto *d = dyn_cast<Defined>(&sym)) {
auto *relSec = dyn_cast_or_null<InputSectionBase>(d->section);
@@ -195,7 +198,7 @@ static bool isReserved(InputSectionBase *sec) {
template <class ELFT>
void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
- std::optional<LiveOffset> parent) {
+ std::optional<LiveObject> parent) {
// Usually, a whole section is marked as live or dead, but in mergeable
// (splittable) sections, each piece of data has independent liveness bit.
// So we explicitly tell it which offset is in use.
@@ -210,11 +213,10 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
sec->partition = sec->partition ? 1 : partition;
if (parent) {
- whyLive.try_emplace(LiveOffset{sec, offset}, *parent);
- // Offset zero is treated as a stand-in for the section itself. The parent
- // is both a specific reason that an offset within this section is alive and
- // a generic reason the section itself is alive.
- whyLive.try_emplace(LiveOffset{sec, 0}, *parent);
+ whyLive.try_emplace(sec, *parent);
+ Defined *sym = sec->getEnclosingSymbol(offset);
+ if (sym)
+ whyLive.try_emplace(sym, *parent);
}
// Add input section to the queue.
@@ -339,11 +341,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, 0, {{&sec, 0}});
+ enqueue(isec, 0, &sec);
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, 0, {{&sec, 0}});
+ enqueue(sec.nextInSectionGroup, 0, &sec);
}
}
>From 427f6f5a0f737de6a65d44b5e9d1df44565eca5e Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 3 Dec 2024 11:26:04 -0800
Subject: [PATCH 07/30] Be clear about the zero offset vs section distinction
---
lld/ELF/MarkLive.cpp | 27 +++++++++++++++------------
1 file changed, 15 insertions(+), 12 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index f8cbdd62bf1545..4c67efe74334f5 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -54,7 +54,7 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, uint64_t offset,
+ void enqueue(InputSectionBase *sec, std::optional<uint64_t> offset,
std::optional<LiveObject> parent);
void markSymbol(Symbol *sym);
void mark();
@@ -139,7 +139,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
cast<SharedFile>(ss->file)->isNeeded = true;
for (InputSectionBase *sec : cNamedSections.lookup(sym.getName()))
- enqueue(sec, 0, parent);
+ enqueue(sec, std::nullopt, parent);
}
// The .eh_frame section is an unfortunate special case.
@@ -197,13 +197,14 @@ static bool isReserved(InputSectionBase *sec) {
}
template <class ELFT>
-void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
+void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
+ std::optional<uint64_t> offset,
std::optional<LiveObject> parent) {
// Usually, a whole section is marked as live or dead, but in mergeable
// (splittable) sections, each piece of data has independent liveness bit.
// So we explicitly tell it which offset is in use.
if (auto *ms = dyn_cast<MergeInputSection>(sec))
- ms->getSectionPiece(offset).live = true;
+ ms->getSectionPiece(offset.value_or(0)).live = true;
// Set Sec->Partition to the meet (i.e. the "minimum") of Partition and
// Sec->Partition in the following lattice: 1 < other < 0. If Sec->Partition
@@ -214,9 +215,11 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
if (parent) {
whyLive.try_emplace(sec, *parent);
- Defined *sym = sec->getEnclosingSymbol(offset);
- if (sym)
- whyLive.try_emplace(sym, *parent);
+ if (offset) {
+ Defined *sym = sec->getEnclosingSymbol(*offset);
+ if (sym)
+ whyLive.try_emplace(sym, *parent);
+ }
}
// Add input section to the queue.
@@ -274,7 +277,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
}
for (InputSectionBase *sec : ctx.inputSections) {
if (sec->flags & SHF_GNU_RETAIN) {
- enqueue(sec, 0, std::nullopt);
+ enqueue(sec, std::nullopt, std::nullopt);
continue;
}
if (sec->flags & SHF_LINK_ORDER)
@@ -313,7 +316,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
// Preserve special sections and those which are specified in linker
// script KEEP command.
if (isReserved(sec) || ctx.script->shouldKeep(sec)) {
- enqueue(sec, 0, std::nullopt);
+ enqueue(sec, std::nullopt, std::nullopt);
} else if ((!ctx.arg.zStartStopGC || sec->name.starts_with("__libc_")) &&
isValidCIdentifier(sec->name)) {
// As a workaround for glibc libc.a before 2.34
@@ -341,11 +344,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, 0, &sec);
+ enqueue(isec, std::nullopt, &sec);
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, 0, &sec);
+ enqueue(sec.nextInSectionGroup, std::nullopt, &sec);
}
}
@@ -371,7 +374,7 @@ template <class ELFT> void MarkLive<ELFT>::moveToMain() {
continue;
if (ctx.symtab->find(("__start_" + sec->name).str()) ||
ctx.symtab->find(("__stop_" + sec->name).str()))
- enqueue(sec, 0, std::nullopt);
+ enqueue(sec, std::nullopt, std::nullopt);
}
mark();
>From b9378f5a7e176926f49e44ef684a0218e77cc4e8 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 3 Dec 2024 11:37:59 -0800
Subject: [PATCH 08/30] Track shared symbols too
---
lld/ELF/MarkLive.cpp | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 4c67efe74334f5..c0c6b69aab1d3d 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -44,7 +44,7 @@ using namespace lld;
using namespace lld::elf;
namespace {
-typedef std::variant<InputSectionBase *, Defined *> LiveObject;
+typedef std::variant<InputSectionBase *, Symbol *> LiveObject;
template <class ELFT> class MarkLive {
public:
@@ -134,9 +134,12 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
return;
}
- if (auto *ss = dyn_cast<SharedSymbol>(&sym))
- if (!ss->isWeak())
+ if (auto *ss = dyn_cast<SharedSymbol>(&sym)) {
+ if (!ss->isWeak()) {
cast<SharedFile>(ss->file)->isNeeded = true;
+ whyLive.try_emplace(&sym, parent);
+ }
+ }
for (InputSectionBase *sec : cNamedSections.lookup(sym.getName()))
enqueue(sec, std::nullopt, parent);
>From 372760c321e6e751d698683dc6d09801edd86d8a Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 3 Dec 2024 14:49:45 -0800
Subject: [PATCH 09/30] Hax
---
lld/ELF/MarkLive.cpp | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index c0c6b69aab1d3d..a0f7deec45166b 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -56,6 +56,7 @@ template <class ELFT> class MarkLive {
private:
void enqueue(InputSectionBase *sec, std::optional<uint64_t> offset,
std::optional<LiveObject> parent);
+ void printWhyLive(const Symbol *s) const;
void markSymbol(Symbol *sym);
void mark();
@@ -230,6 +231,10 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
queue.push_back(s);
}
+template <class ELFT>
+void MarkLive<ELFT>::printWhyLive(const Symbol *s) const {
+}
+
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
if (auto *d = dyn_cast_or_null<Defined>(sym))
if (auto *isec = dyn_cast_or_null<InputSectionBase>(d->section))
@@ -353,6 +358,8 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
if (sec.nextInSectionGroup)
enqueue(sec.nextInSectionGroup, std::nullopt, &sec);
}
+
+ printWhyLive(ctx.symtab->find("foo"));
}
// Move the sections for some symbols to the main partition, specifically ifuncs
>From 33d45dabec5e48a9812435ac5da6ed72f9bba0e3 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 14:11:31 -0800
Subject: [PATCH 10/30] Add simple why-live printing fn
---
lld/ELF/MarkLive.cpp | 27 ++++++++++++++++++++++++---
1 file changed, 24 insertions(+), 3 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index a0f7deec45166b..a5549e93f6a7b7 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -56,7 +56,7 @@ template <class ELFT> class MarkLive {
private:
void enqueue(InputSectionBase *sec, std::optional<uint64_t> offset,
std::optional<LiveObject> parent);
- void printWhyLive(const Symbol *s) const;
+ void printWhyLive(Symbol *s) const;
void markSymbol(Symbol *sym);
void mark();
@@ -231,8 +231,29 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
queue.push_back(s);
}
-template <class ELFT>
-void MarkLive<ELFT>::printWhyLive(const Symbol *s) const {
+template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
+ std::string out = toString(*s) + " from " + toString(s->file);
+ int indent = 2;
+ LiveObject cur = s;
+ while (true) {
+ auto it = whyLive.find(cur);
+ if (it == whyLive.end())
+ if (auto *d = dyn_cast<Defined>(s))
+ it = whyLive.find(LiveObject{d->section});
+ if (it == whyLive.end())
+ break;
+ cur = it->second;
+ out += "\n" + std::string(indent, ' ');
+ if (std::holds_alternative<Symbol *>(cur)) {
+ auto *s = std::get<Symbol *>(cur);
+ out += toString(*s) + " from " + toString(s->file);
+ } else {
+ auto *s = std::get<InputSectionBase *>(cur);
+ // TODO: Fancy formatting
+ out += toString(s);
+ }
+ }
+ message(out);
}
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
>From b0829a0243c34c42aa8ed26a636bec6e8007ba94 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 14:13:23 -0800
Subject: [PATCH 11/30] Add missing cast
---
lld/ELF/MarkLive.cpp | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index a5549e93f6a7b7..a7510a361ed458 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -239,7 +239,8 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
auto it = whyLive.find(cur);
if (it == whyLive.end())
if (auto *d = dyn_cast<Defined>(s))
- it = whyLive.find(LiveObject{d->section});
+ if (auto *s = dyn_cast<InputSectionBase>(d->section))
+ it = whyLive.find(LiveObject{s});
if (it == whyLive.end())
break;
cur = it->second;
>From c9599fa60ef58d1daebebb2c14c791d192af2965 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 15:14:39 -0800
Subject: [PATCH 12/30] Explicitly mark roots as roots to break cycles
---
lld/ELF/MarkLive.cpp | 35 ++++++++++++++++++-----------------
1 file changed, 18 insertions(+), 17 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index a7510a361ed458..5a57765dcfca02 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -54,8 +54,9 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, std::optional<uint64_t> offset,
- std::optional<LiveObject> parent);
+ void enqueue(InputSectionBase *sec,
+ std::optional<uint64_t> offset = std::nullopt,
+ std::optional<LiveObject> parent = std::nullopt);
void printWhyLive(Symbol *s) const;
void markSymbol(Symbol *sym);
void mark();
@@ -77,7 +78,7 @@ template <class ELFT> class MarkLive {
// identifiers, so we just store a SmallVector instead of a multimap.
DenseMap<StringRef, SmallVector<InputSectionBase *, 0>> cNamedSections;
- DenseMap<LiveObject, LiveObject> whyLive;
+ DenseMap<LiveObject, std::optional<LiveObject>> whyLive;
};
} // namespace
@@ -143,7 +144,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
}
for (InputSectionBase *sec : cNamedSections.lookup(sym.getName()))
- enqueue(sec, std::nullopt, parent);
+ enqueue(sec);
}
// The .eh_frame section is an unfortunate special case.
@@ -217,13 +218,11 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
return;
sec->partition = sec->partition ? 1 : partition;
- if (parent) {
- whyLive.try_emplace(sec, *parent);
- if (offset) {
- Defined *sym = sec->getEnclosingSymbol(*offset);
- if (sym)
- whyLive.try_emplace(sym, *parent);
- }
+ whyLive.try_emplace(sec, parent);
+ if (offset) {
+ Defined *sym = sec->getEnclosingSymbol(*offset);
+ if (sym)
+ whyLive.try_emplace(sym, parent);
}
// Add input section to the queue.
@@ -241,9 +240,11 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
if (auto *d = dyn_cast<Defined>(s))
if (auto *s = dyn_cast<InputSectionBase>(d->section))
it = whyLive.find(LiveObject{s});
- if (it == whyLive.end())
+ assert(it != whyLive.end() &&
+ "all live objects should have a tracked reason for being live");
+ if (!it->second)
break;
- cur = it->second;
+ cur = *it->second;
out += "\n" + std::string(indent, ' ');
if (std::holds_alternative<Symbol *>(cur)) {
auto *s = std::get<Symbol *>(cur);
@@ -260,7 +261,7 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
if (auto *d = dyn_cast_or_null<Defined>(sym))
if (auto *isec = dyn_cast_or_null<InputSectionBase>(d->section))
- enqueue(isec, d->value, std::nullopt);
+ enqueue(isec, d->value);
}
// This is the main function of the garbage collector.
@@ -307,7 +308,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
}
for (InputSectionBase *sec : ctx.inputSections) {
if (sec->flags & SHF_GNU_RETAIN) {
- enqueue(sec, std::nullopt, std::nullopt);
+ enqueue(sec);
continue;
}
if (sec->flags & SHF_LINK_ORDER)
@@ -346,7 +347,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
// Preserve special sections and those which are specified in linker
// script KEEP command.
if (isReserved(sec) || ctx.script->shouldKeep(sec)) {
- enqueue(sec, std::nullopt, std::nullopt);
+ enqueue(sec);
} else if ((!ctx.arg.zStartStopGC || sec->name.starts_with("__libc_")) &&
isValidCIdentifier(sec->name)) {
// As a workaround for glibc libc.a before 2.34
@@ -406,7 +407,7 @@ template <class ELFT> void MarkLive<ELFT>::moveToMain() {
continue;
if (ctx.symtab->find(("__start_" + sec->name).str()) ||
ctx.symtab->find(("__stop_" + sec->name).str()))
- enqueue(sec, std::nullopt, std::nullopt);
+ enqueue(sec);
}
mark();
>From b127efb83b3e25204efd7b65d7cb9fad08896d42 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 15:26:49 -0800
Subject: [PATCH 13/30] Find bar actually, not foo
---
lld/ELF/MarkLive.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 5a57765dcfca02..10c572060da108 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -382,7 +382,7 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
enqueue(sec.nextInSectionGroup, std::nullopt, &sec);
}
- printWhyLive(ctx.symtab->find("foo"));
+ printWhyLive(ctx.symtab->find("bar"));
}
// Move the sections for some symbols to the main partition, specifically ifuncs
>From 584dfe8b98afc29a45a3cc0b65b97db8d9fa6473 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 15:38:53 -0800
Subject: [PATCH 14/30] Report section membership as a reason for being alive
---
lld/ELF/MarkLive.cpp | 43 +++++++++++++++++++++++++------------------
1 file changed, 25 insertions(+), 18 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 10c572060da108..8287ce3f51af8f 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -231,29 +231,36 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
}
template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
- std::string out = toString(*s) + " from " + toString(s->file);
- int indent = 2;
- LiveObject cur = s;
- while (true) {
- auto it = whyLive.find(cur);
- if (it == whyLive.end())
- if (auto *d = dyn_cast<Defined>(s))
- if (auto *s = dyn_cast<InputSectionBase>(d->section))
- it = whyLive.find(LiveObject{s});
- assert(it != whyLive.end() &&
- "all live objects should have a tracked reason for being live");
- if (!it->second)
- break;
- cur = *it->second;
- out += "\n" + std::string(indent, ' ');
- if (std::holds_alternative<Symbol *>(cur)) {
- auto *s = std::get<Symbol *>(cur);
+ std::string out;
+ int indent = 0;
+ for (std::optional<LiveObject> cur = s; cur; indent += 2) {
+ if (indent)
+ out += "\n" + std::string(indent, ' ');
+ if (std::holds_alternative<Symbol *>(*cur)) {
+ auto *s = std::get<Symbol *>(*cur);
out += toString(*s) + " from " + toString(s->file);
} else {
- auto *s = std::get<InputSectionBase *>(cur);
+ auto *s = std::get<InputSectionBase *>(*cur);
// TODO: Fancy formatting
out += toString(s);
}
+
+ auto it = whyLive.find(*cur);
+ if (it != whyLive.end()) {
+ // If there is a specific reason this object is live, report it.
+ if (!it->second)
+ break;
+ cur = *it->second;
+ } else {
+ // This object is live merely by being a member of its parent section, so
+ // report the parent.
+ InputSectionBase *parent = nullptr;
+ if (auto *d = dyn_cast<Defined>(s))
+ parent = dyn_cast<InputSectionBase>(d->section);
+ assert(parent &&
+ "all live objects should have a tracked reason for being live");
+ cur = LiveObject{parent};
+ }
}
message(out);
}
>From 3bcd86789bf57b59a8e19c40af4a8d1eac3063ab Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 4 Dec 2024 16:03:17 -0800
Subject: [PATCH 15/30] If a specific symbol is referenced, it's the reason its
section is alive
---
lld/ELF/MarkLive.cpp | 16 +++++++++++-----
1 file changed, 11 insertions(+), 5 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 8287ce3f51af8f..17447330941004 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -218,11 +218,17 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
return;
sec->partition = sec->partition ? 1 : partition;
- whyLive.try_emplace(sec, parent);
- if (offset) {
- Defined *sym = sec->getEnclosingSymbol(*offset);
- if (sym)
- whyLive.try_emplace(sym, parent);
+ Defined *sym = nullptr;
+ if (offset)
+ sym = sec->getEnclosingSymbol(*offset);
+ if (sym) {
+ // If a specific symbol is referenced, the parent makes it alive, and it
+ // (may) makes its section alive.
+ whyLive.try_emplace(sym, parent);
+ whyLive.try_emplace(sec, sym);
+ } else {
+ // Otherwise, the parent generically makes the section itself live.
+ whyLive.try_emplace(sec, parent);
}
// Add input section to the queue.
>From 94a480843697a9905e020e7fbc6d96b73b74ee6b Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <mysterymath at gmail.com>
Date: Thu, 5 Dec 2024 12:21:51 -0800
Subject: [PATCH 16/30] Determine canonical symbol for a reference
---
lld/ELF/MarkLive.cpp | 27 ++++++++++++++-------------
1 file changed, 14 insertions(+), 13 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 17447330941004..14f4694f69772a 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -54,8 +54,8 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec,
- std::optional<uint64_t> offset = std::nullopt,
+ void enqueue(InputSectionBase *sec, uint64_t offset = 0,
+ Symbol *sym = nullptr,
std::optional<LiveObject> parent = std::nullopt);
void printWhyLive(Symbol *s) const;
void markSymbol(Symbol *sym);
@@ -131,8 +131,13 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
// group/SHF_LINK_ORDER rules (b) if the associated text section should be
// discarded, marking the LSDA will unnecessarily retain the text section.
if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
- relSec->nextInSectionGroup)))
- enqueue(relSec, offset, parent);
+ relSec->nextInSectionGroup))) {
+ Symbol *canonicalSym = d;
+ if (offset >= d->value + d->size)
+ if (Symbol *s = relSec->getEnclosingSymbol(offset))
+ canonicalSym = s;
+ enqueue(relSec, offset, canonicalSym, parent);
+ }
return;
}
@@ -202,14 +207,13 @@ static bool isReserved(InputSectionBase *sec) {
}
template <class ELFT>
-void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
- std::optional<uint64_t> offset,
- std::optional<LiveObject> parent) {
+void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
+ Symbol *sym, std::optional<LiveObject> parent) {
// Usually, a whole section is marked as live or dead, but in mergeable
// (splittable) sections, each piece of data has independent liveness bit.
// So we explicitly tell it which offset is in use.
if (auto *ms = dyn_cast<MergeInputSection>(sec))
- ms->getSectionPiece(offset.value_or(0)).live = true;
+ ms->getSectionPiece(offset).live = true;
// Set Sec->Partition to the meet (i.e. the "minimum") of Partition and
// Sec->Partition in the following lattice: 1 < other < 0. If Sec->Partition
@@ -218,9 +222,6 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec,
return;
sec->partition = sec->partition ? 1 : partition;
- Defined *sym = nullptr;
- if (offset)
- sym = sec->getEnclosingSymbol(*offset);
if (sym) {
// If a specific symbol is referenced, the parent makes it alive, and it
// (may) makes its section alive.
@@ -388,11 +389,11 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
resolveReloc(sec, rel, false);
for (InputSectionBase *isec : sec.dependentSections)
- enqueue(isec, std::nullopt, &sec);
+ enqueue(isec, 0, nullptr, &sec);
// Mark the next group member.
if (sec.nextInSectionGroup)
- enqueue(sec.nextInSectionGroup, std::nullopt, &sec);
+ enqueue(sec.nextInSectionGroup, 0, nullptr, &sec);
}
printWhyLive(ctx.symtab->find("bar"));
>From 22d47b722e1ad57c592b238fe9bde40fb50699a8 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <mysterymath at gmail.com>
Date: Thu, 5 Dec 2024 12:33:45 -0800
Subject: [PATCH 17/30] Refer to sections rather than STT_SECTION symbols
---
lld/ELF/MarkLive.cpp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 14f4694f69772a..4ff4616a110319 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -136,6 +136,8 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
if (offset >= d->value + d->size)
if (Symbol *s = relSec->getEnclosingSymbol(offset))
canonicalSym = s;
+ if (canonicalSym->isSection())
+ canonicalSym = nullptr;
enqueue(relSec, offset, canonicalSym, parent);
}
return;
>From e956dd56dced83d90a6b8ac0ae93c5ce8f748e42 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <mysterymath at gmail.com>
Date: Thu, 5 Dec 2024 12:43:35 -0800
Subject: [PATCH 18/30] Encode parent for named sections
---
lld/ELF/MarkLive.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 4ff4616a110319..f3af48216111d9 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -151,7 +151,7 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
}
for (InputSectionBase *sec : cNamedSections.lookup(sym.getName()))
- enqueue(sec);
+ enqueue(sec, 0, nullptr, parent);
}
// The .eh_frame section is an unfortunate special case.
>From b7fecebe09c35fe03b5a0b38503c2d57a6d6bbe6 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <mysterymath at gmail.com>
Date: Thu, 5 Dec 2024 12:47:18 -0800
Subject: [PATCH 19/30] No defaults; was missing things
---
lld/ELF/MarkLive.cpp | 13 ++++++-------
1 file changed, 6 insertions(+), 7 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index f3af48216111d9..79834da67f5bce 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -54,9 +54,8 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, uint64_t offset = 0,
- Symbol *sym = nullptr,
- std::optional<LiveObject> parent = std::nullopt);
+ void enqueue(InputSectionBase *sec, uint64_t offset, Symbol *sym,
+ std::optional<LiveObject> parent);
void printWhyLive(Symbol *s) const;
void markSymbol(Symbol *sym);
void mark();
@@ -277,7 +276,7 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
if (auto *d = dyn_cast_or_null<Defined>(sym))
if (auto *isec = dyn_cast_or_null<InputSectionBase>(d->section))
- enqueue(isec, d->value);
+ enqueue(isec, d->value, sym, std::nullopt);
}
// This is the main function of the garbage collector.
@@ -324,7 +323,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
}
for (InputSectionBase *sec : ctx.inputSections) {
if (sec->flags & SHF_GNU_RETAIN) {
- enqueue(sec);
+ enqueue(sec, 0, nullptr, std::nullopt);
continue;
}
if (sec->flags & SHF_LINK_ORDER)
@@ -363,7 +362,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
// Preserve special sections and those which are specified in linker
// script KEEP command.
if (isReserved(sec) || ctx.script->shouldKeep(sec)) {
- enqueue(sec);
+ enqueue(sec, 0, nullptr, std::nullopt);
} else if ((!ctx.arg.zStartStopGC || sec->name.starts_with("__libc_")) &&
isValidCIdentifier(sec->name)) {
// As a workaround for glibc libc.a before 2.34
@@ -423,7 +422,7 @@ template <class ELFT> void MarkLive<ELFT>::moveToMain() {
continue;
if (ctx.symtab->find(("__start_" + sec->name).str()) ||
ctx.symtab->find(("__stop_" + sec->name).str()))
- enqueue(sec);
+ enqueue(sec, 0, nullptr, std::nullopt);
}
mark();
>From 715e0683ded64cdf58cf5b0d1979aabfe691c08b Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Fri, 6 Dec 2024 15:31:52 -0800
Subject: [PATCH 20/30] Add why-live option
---
lld/ELF/Options.td | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/lld/ELF/Options.td b/lld/ELF/Options.td
index c31875305952fb..93c7293c2ee151 100644
--- a/lld/ELF/Options.td
+++ b/lld/ELF/Options.td
@@ -559,6 +559,10 @@ defm wrap : Eq<"wrap", "Redirect symbol references to __wrap_symbol and "
"__real_symbol references to symbol">,
MetaVarName<"<symbol>">;
+defm why_live : EEq<"why-live", "Report a chain of references to <symbol-glob> that keeps it from "
+ "being garbage collected">,
+ MetaVarName<"<symbol-glob>">;
+
def z: JoinedOrSeparate<["-"], "z">, MetaVarName<"<option>">,
HelpText<"Linker option extensions">;
>From 6484056b14232501510ffc66aeac70ee2d98a06f Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Fri, 6 Dec 2024 15:34:06 -0800
Subject: [PATCH 21/30] toStr
---
lld/ELF/MarkLive.cpp | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 79834da67f5bce..025ff9b728ab62 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -246,11 +246,11 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
out += "\n" + std::string(indent, ' ');
if (std::holds_alternative<Symbol *>(*cur)) {
auto *s = std::get<Symbol *>(*cur);
- out += toString(*s) + " from " + toString(s->file);
+ out += toStr(ctx, *s) + " from " + toStr(ctx, s->file);
} else {
auto *s = std::get<InputSectionBase *>(*cur);
// TODO: Fancy formatting
- out += toString(s);
+ out += toStr(ctx, s);
}
auto it = whyLive.find(*cur);
>From b42bd9474f263180f652fcbb939b95078c539462 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Fri, 6 Dec 2024 15:36:35 -0800
Subject: [PATCH 22/30] Better wording
---
lld/ELF/Options.td | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/lld/ELF/Options.td b/lld/ELF/Options.td
index 93c7293c2ee151..2e7ee35e762f2d 100644
--- a/lld/ELF/Options.td
+++ b/lld/ELF/Options.td
@@ -559,9 +559,9 @@ defm wrap : Eq<"wrap", "Redirect symbol references to __wrap_symbol and "
"__real_symbol references to symbol">,
MetaVarName<"<symbol>">;
-defm why_live : EEq<"why-live", "Report a chain of references to <symbol-glob> that keeps it from "
- "being garbage collected">,
- MetaVarName<"<symbol-glob>">;
+defm why_live : EEq<"why-live", "Report a chain of references preventing garbage collection for "
+ "each symbol matching <glob>">,
+ MetaVarName<"<glob>">;
def z: JoinedOrSeparate<["-"], "z">, MetaVarName<"<option>">,
HelpText<"Linker option extensions">;
>From 7bca5d534c56e5b2c08a8886e5dcadc2b6d87193 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 9 Dec 2024 11:51:12 -0800
Subject: [PATCH 23/30] Parse whylive arg
---
lld/ELF/Config.h | 1 +
lld/ELF/Driver.cpp | 9 +++++++++
2 files changed, 10 insertions(+)
diff --git a/lld/ELF/Config.h b/lld/ELF/Config.h
index b2859486d58e93..12164f59993436 100644
--- a/lld/ELF/Config.h
+++ b/lld/ELF/Config.h
@@ -223,6 +223,7 @@ struct Config {
llvm::StringRef thinLTOCacheDir;
llvm::StringRef thinLTOIndexOnlyArg;
llvm::StringRef whyExtract;
+ llvm::SmallVector<llvm::GlobPattern, 0> whyLive;
llvm::StringRef cmseInputLib;
llvm::StringRef cmseOutputLib;
StringRef zBtiReport = "none";
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index 13e8f8ce6df207..db0b2ea8afcf0a 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -1472,6 +1472,15 @@ static void readConfigs(Ctx &ctx, opt::InputArgList &args) {
ctx.arg.warnSymbolOrdering =
args.hasFlag(OPT_warn_symbol_ordering, OPT_no_warn_symbol_ordering, true);
ctx.arg.whyExtract = args.getLastArgValue(OPT_why_extract);
+ for (opt::Arg *arg : args.filtered(OPT_why_live)) {
+ StringRef value(arg->getValue());
+ if (Expected<GlobPattern> pat = GlobPattern::create(arg->getValue())) {
+ ctx.arg.whyLive.emplace_back(std::move(*pat));
+ } else {
+ ErrAlways(ctx) << arg->getSpelling() << ": " << pat.takeError();
+ continue;
+ }
+ }
ctx.arg.zCombreloc = getZFlag(args, "combreloc", "nocombreloc", true);
ctx.arg.zCopyreloc = getZFlag(args, "copyreloc", "nocopyreloc", true);
ctx.arg.zForceBti = hasZOption(args, "force-bti");
>From e26bbf21ae29f2502d2b87b4aa616a8d8eaf93c0 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 9 Dec 2024 12:18:28 -0800
Subject: [PATCH 24/30] Connect whylive pattern matching
---
lld/ELF/MarkLive.cpp | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 025ff9b728ab62..952fa7b9fe2996 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -397,8 +397,13 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
enqueue(sec.nextInSectionGroup, 0, nullptr, &sec);
}
- printWhyLive(ctx.symtab->find("bar"));
-}
+ for (Symbol *sym : ctx.symtab->getSymbols()) {
+ if (llvm::any_of(ctx.arg.whyLive, [sym](const llvm::GlobPattern &pat) {
+ return pat.match(sym->getName());
+ }))
+ printWhyLive(sym);
+ }
+ }
// Move the sections for some symbols to the main partition, specifically ifuncs
// (because they can result in an IRELATIVE being added to the main partition's
>From 932cb8cefd0772fc3f703ffb895d28ba5471beec Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 9 Dec 2024 14:30:19 -0800
Subject: [PATCH 25/30] Don't trigger assertion for dead symbols
---
lld/ELF/MarkLive.cpp | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 952fa7b9fe2996..27131d04378093 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -241,6 +241,8 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
std::string out;
int indent = 0;
+ if (!whyLive.contains(s))
+ return;
for (std::optional<LiveObject> cur = s; cur; indent += 2) {
if (indent)
out += "\n" + std::string(indent, ' ');
@@ -403,7 +405,7 @@ template <class ELFT> void MarkLive<ELFT>::mark() {
}))
printWhyLive(sym);
}
- }
+}
// Move the sections for some symbols to the main partition, specifically ifuncs
// (because they can result in an IRELATIVE being added to the main partition's
>From cf794154fc7db9b9b61ab36d508dddbc736f0797 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Dec 2024 14:14:09 -0800
Subject: [PATCH 26/30] Optional args
---
lld/ELF/MarkLive.cpp | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 27131d04378093..1b05b5845a29e6 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -54,8 +54,9 @@ template <class ELFT> class MarkLive {
void moveToMain();
private:
- void enqueue(InputSectionBase *sec, uint64_t offset, Symbol *sym,
- std::optional<LiveObject> parent);
+ void enqueue(InputSectionBase *sec, uint64_t offset = 0,
+ Symbol *sym = nullptr,
+ std::optional<LiveObject> parent = std::nullopt);
void printWhyLive(Symbol *s) const;
void markSymbol(Symbol *sym);
void mark();
@@ -278,7 +279,7 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
if (auto *d = dyn_cast_or_null<Defined>(sym))
if (auto *isec = dyn_cast_or_null<InputSectionBase>(d->section))
- enqueue(isec, d->value, sym, std::nullopt);
+ enqueue(isec, d->value, sym);
}
// This is the main function of the garbage collector.
@@ -364,7 +365,7 @@ template <class ELFT> void MarkLive<ELFT>::run() {
// Preserve special sections and those which are specified in linker
// script KEEP command.
if (isReserved(sec) || ctx.script->shouldKeep(sec)) {
- enqueue(sec, 0, nullptr, std::nullopt);
+ enqueue(sec);
} else if ((!ctx.arg.zStartStopGC || sec->name.starts_with("__libc_")) &&
isValidCIdentifier(sec->name)) {
// As a workaround for glibc libc.a before 2.34
@@ -429,7 +430,7 @@ template <class ELFT> void MarkLive<ELFT>::moveToMain() {
continue;
if (ctx.symtab->find(("__start_" + sec->name).str()) ||
ctx.symtab->find(("__stop_" + sec->name).str()))
- enqueue(sec, 0, nullptr, std::nullopt);
+ enqueue(sec);
}
mark();
>From 6c0232b5ab1bd82c219fed2d1f9b5763efa0b79d Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Dec 2024 14:25:53 -0800
Subject: [PATCH 27/30] Simplify and correct canonical symbol detection
---
lld/ELF/MarkLive.cpp | 11 ++++++-----
1 file changed, 6 insertions(+), 5 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 1b05b5845a29e6..7c98ea00acd64f 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -132,12 +132,13 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
// discarded, marking the LSDA will unnecessarily retain the text section.
if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
relSec->nextInSectionGroup))) {
- Symbol *canonicalSym = d;
- if (offset >= d->value + d->size)
- if (Symbol *s = relSec->getEnclosingSymbol(offset))
+ Symbol *canonicalSym = nullptr;
+ if (!d->isSection()) {
+ if (offset < d->value + d->size)
+ canonicalSym = d;
+ else if (Symbol *s = relSec->getEnclosingSymbol(offset))
canonicalSym = s;
- if (canonicalSym->isSection())
- canonicalSym = nullptr;
+ }
enqueue(relSec, offset, canonicalSym, parent);
}
return;
>From bf7cb603e3687946ffe58db23a39643735722b36 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Dec 2024 14:46:18 -0800
Subject: [PATCH 28/30] Trust the symbol reference; it's all we've got
---
lld/ELF/MarkLive.cpp | 9 +++------
1 file changed, 3 insertions(+), 6 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 7c98ea00acd64f..ac1359a8f861d9 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -132,13 +132,10 @@ void MarkLive<ELFT>::resolveReloc(InputSectionBase &sec, RelTy &rel,
// discarded, marking the LSDA will unnecessarily retain the text section.
if (!(fromFDE && ((relSec->flags & (SHF_EXECINSTR | SHF_LINK_ORDER)) ||
relSec->nextInSectionGroup))) {
- Symbol *canonicalSym = nullptr;
- if (!d->isSection()) {
- if (offset < d->value + d->size)
- canonicalSym = d;
- else if (Symbol *s = relSec->getEnclosingSymbol(offset))
+ Symbol *canonicalSym = d;
+ if (d->isSection())
+ if (Symbol *s = relSec->getEnclosingSymbol(offset))
canonicalSym = s;
- }
enqueue(relSec, offset, canonicalSym, parent);
}
return;
>From 224769456be2d7e01ef45d122e58762a1d5e59b2 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 29 Jan 2025 15:46:25 -0800
Subject: [PATCH 29/30] Initial why-live test
---
lld/test/ELF/why-live.s | 17 +++++++++++++++++
1 file changed, 17 insertions(+)
create mode 100644 lld/test/ELF/why-live.s
diff --git a/lld/test/ELF/why-live.s b/lld/test/ELF/why-live.s
new file mode 100644
index 00000000000000..34a6c38ec88e49
--- /dev/null
+++ b/lld/test/ELF/why-live.s
@@ -0,0 +1,17 @@
+# REQUIRES: x86
+
+# RUN: llvm-mc -n -filetype=obj -triple=x86_64 %s -o %t.o
+# RUN: ld.lld %t.o -o /dev/null --gc-sections --why-live=a | FileCheck %s
+
+# CHECK: blah
+
+.globl _start
+.section ._start,"ax", at progbits
+_start:
+jmp a
+
+.globl a
+.section .a,"ax", at progbits
+a:
+jmp a
+
>From 2d50c974b6939847bddcdffb6c0917e7d4887e55 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 29 Jan 2025 16:16:35 -0800
Subject: [PATCH 30/30] Improve syntax for printing
---
lld/ELF/MarkLive.cpp | 24 +++++++++++++++---------
1 file changed, 15 insertions(+), 9 deletions(-)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index ac1359a8f861d9..214281c99e04b4 100644
--- a/lld/ELF/MarkLive.cpp
+++ b/lld/ELF/MarkLive.cpp
@@ -238,20 +238,23 @@ void MarkLive<ELFT>::enqueue(InputSectionBase *sec, uint64_t offset,
}
template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
- std::string out;
- int indent = 0;
if (!whyLive.contains(s))
return;
- for (std::optional<LiveObject> cur = s; cur; indent += 2) {
- if (indent)
- out += "\n" + std::string(indent, ' ');
+ auto diag = Msg(ctx);
+ bool first = true;
+ for (std::optional<LiveObject> cur = s; cur;) {
if (std::holds_alternative<Symbol *>(*cur)) {
auto *s = std::get<Symbol *>(*cur);
- out += toStr(ctx, *s) + " from " + toStr(ctx, s->file);
+ // Match the syntax for sections below.
+ diag << toStr(ctx, s->file) << ":(" << toStr(ctx, *s) << ')';
} else {
auto *s = std::get<InputSectionBase *>(*cur);
- // TODO: Fancy formatting
- out += toStr(ctx, s);
+ diag << toStr(ctx, s);
+ }
+
+ if (first) {
+ diag << " live because:";
+ first = false;
}
auto it = whyLive.find(*cur);
@@ -270,8 +273,11 @@ template <class ELFT> void MarkLive<ELFT>::printWhyLive(Symbol *s) const {
"all live objects should have a tracked reason for being live");
cur = LiveObject{parent};
}
+
+ if (cur)
+ diag << "\n>>> referenced by "
+ << (std::holds_alternative<Symbol *>(*cur) ? "symbol " : "section ");
}
- message(out);
}
template <class ELFT> void MarkLive<ELFT>::markSymbol(Symbol *sym) {
More information about the llvm-commits
mailing list