[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
Thu Dec 19 14:26:18 PST 2024
https://github.com/mysterymath updated https://github.com/llvm/llvm-project/pull/119279
>From a15fdeff8fe8aebfab0b7c8135e0a575ab33ed7e 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 01/27] 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 08d9ecedf2cc07..72e5ef14b73bca 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 a916b1761fcaa6e39928fb02c229075ab429092b 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 02/27] 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 72e5ef14b73bca..e49d105546b86d 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 4d7981a002e8e3331858c0f96a6066f3ec920e73 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 03/27] 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 e49d105546b86d..cb1ab59a0f1bd0 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 5ad1d0b5013253f7700dde78dd2c94648d06e06d 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 04/27] 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 cb1ab59a0f1bd0..3686b877c26b36 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 973dbba4f490a1e158b26ff5da8f7bc680867ced 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 05/27] 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 3686b877c26b36..5164c3c6df5a6f 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 9d3fd211464461ac061f3661d632097f0e6d341e 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 06/27] 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 5164c3c6df5a6f..d93f9ae187d337 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 1bb7a314e36f8ceeb0d1b0b50be7e905a0a45959 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 07/27] 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 d93f9ae187d337..3c654921a0dfdc 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 75a1c47a65f5f08ce40e9dc5d8d26710941c202e 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 08/27] 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 3c654921a0dfdc..6cba7a36988e4e 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 8a489a2b191990b1d67b1abd991de9f4f93b3b6f 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 09/27] Hax
---
lld/ELF/MarkLive.cpp | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/lld/ELF/MarkLive.cpp b/lld/ELF/MarkLive.cpp
index 6cba7a36988e4e..5ebe94bcf9de45 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 b9e3defbf36bc04c3d45e03a994f1355a1995770 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 10/27] 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 5ebe94bcf9de45..0092bc7ed45345 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 f76105c763511ef46d048b8392eb16ec4c472654 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 11/27] 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 0092bc7ed45345..bdefe6d0ab21a8 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 306ed975384a9f64296fb59f5838b04545e12eda 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 12/27] 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 bdefe6d0ab21a8..05fb6ce92f3e14 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 9c136667d034f19f20a163c112e6a6beed3ab6b6 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 13/27] 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 05fb6ce92f3e14..64878c158eeb8e 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 49ecd5728a9ee3bb97afbc2b39d10f551081f4eb 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 14/27] 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 64878c158eeb8e..db740551406d23 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 1ddc660a2ebc03d98adaa15ed2ee4429ef7cd94e 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 15/27] 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 db740551406d23..7669f99f9cd349 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 acf5e58b5442c0369d1303ab6eaef92f158ed73f 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 16/27] 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 7669f99f9cd349..4432eb592abc73 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 97d98224857e83eb43f7d93d492171589d278401 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 17/27] 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 4432eb592abc73..420f2f5f16cb13 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 21d4a90027e1c33db56eb5c86e41fdd301cdc125 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 18/27] 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 420f2f5f16cb13..a9e49cd32815c0 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 31980cf86f1e2b5bc4cd90ebe84760820f52b16a 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 19/27] 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 a9e49cd32815c0..06b822e14d9f35 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 5ad90dbccb8db44fbb8f78343cf5824a9d24875b 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 20/27] 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 ebe77204264210..9ad09fa3f89c40 100644
--- a/lld/ELF/Options.td
+++ b/lld/ELF/Options.td
@@ -556,6 +556,10 @@ 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 6c55f640684c093a564d66249acc774a4d2299c8 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 21/27] 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 06b822e14d9f35..796a50366f36f8 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 90a8f52f171f414207c892c202be9509ceb5c207 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 22/27] 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 9ad09fa3f89c40..4d0531005adb92 100644
--- a/lld/ELF/Options.td
+++ b/lld/ELF/Options.td
@@ -556,9 +556,9 @@ 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 50ca32647a88a817eed2b310aa1d168570890f0f 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 23/27] 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 a2836733c2715e..e3a0d90ceb5682 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 bc4b967ccbbbb4..b93178cebebd00 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -1469,6 +1469,15 @@ 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 94a10b29e8403cf7b448b426f3a7ff9c1f2bc6e5 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 24/27] 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 796a50366f36f8..8bdf454bf32b28 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 99ec544fa24dc479251fd9af9c685421ca30810c 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 25/27] 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 8bdf454bf32b28..2401ae056a6256 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 c03809d41962acf10a8ed2e7fba31af28909bc5b 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 26/27] 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 2401ae056a6256..d5b0b128fd1091 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 543001fedf5e5cdb1b58df971036ee9eee6a9ba6 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 27/27] 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 d5b0b128fd1091..d9f358d1ace00c 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;
More information about the llvm-commits
mailing list