[ELF] Implement -z dynamic-undefined-weak

The behavior of an undefined weak reference is implementation defined.
For static -no-pie linking, dynamic relocations are generally avoided (except
IRELATIVE). -shared linking generally emits dynamic relocations.

Dynamic -no-pie linking and -pie allow flexibility. Changes adjust the
behavior for better consistency and simpler internal representation,
e.g. https://reviews.llvm.org/D63003 https://reviews.llvm.org/D105164
(generalized to undefined non-weak in
2fcaa00).

GNU ld introduced -z [no]dynamic-undefined-weak option to fine-tune the
behavior. (The option is not very effective with -no-pie, e.g. on
x86-64, `ld.bfd a.o s.so -z dynamic-undefined-weak` generates
R_X86_64_NONE relocations instead of GLOB_DAT/JUMP_SLOT)

This patch implements -z [no]dynamic-undefined-weak option.
The effects are summarized as follows:

* Static -no-pie: no-op
* Dynamic -no-pie: nodynamic-undefined-weak suppresses GLOB_DAT/JUMP_SLOT
* Static -pie: dynamic-undefined-weak generates ABS/GLOB_DAT/JUMP_SLOT.
  https://discourse.llvm.org/t/lld-weak-undefined-symbols-in-vdso-only/86749
* Dynamic -pie: nodynamic-undefined-weak suppresses ABS/GLOB_DAT/JUMP_SLOT

The -pie behavior likely stays stable while -no-pie (`!ctx.arg.isPic` in
`isStaticLinkTimeConstant`) behavior will likely change in the future.
The current default value of ctx.arg.zDynamicUndefined is selected to
prevent behavior changes.

Pull Request: #143831

Author: MaskRay

lld/ELF/Config.h

@@ -368,6 +368,7 @@ struct Config {
   bool writeAddends;
   bool zCombreloc;
   bool zCopyreloc;
+  bool zDynamicUndefined;
   bool zForceBti;
   bool zForceIbt;
   bool zGlobal;

lld/ELF/Driver.cpp

@@ -591,6 +591,7 @@ static void checkZOptions(Ctx &ctx, opt::InputArgList &args) {
   args::getZOptionValue(args, OPT_z, "max-page-size", 0);
   args::getZOptionValue(args, OPT_z, "common-page-size", 0);
   getZFlag(args, "rel", "rela", false);
+  getZFlag(args, "dynamic-undefined-weak", "nodynamic-undefined-weak", false);
   for (auto *arg : args.filtered(OPT_z))
     if (!arg->isClaimed())
       Warn(ctx) << "unknown -z value: " << StringRef(arg->getValue());
@@ -3058,6 +3059,13 @@ template <class ELFT> void LinkerDriver::link(opt::InputArgList &args) {
   ctx.hasDynsym = !ctx.sharedFiles.empty() || ctx.arg.isPic;
   ctx.arg.exportDynamic &= ctx.hasDynsym;
 
+  // Preemptibility of undefined symbols when ctx.hasDynsym is true. Default is
+  // true for dynamic linking.
+  ctx.arg.zDynamicUndefined =
+      getZFlag(args, "dynamic-undefined-weak", "nodynamic-undefined-weak",
+               ctx.sharedFiles.size() || ctx.arg.shared) &&
+      ctx.hasDynsym;
+
   // If an entry symbol is in a static archive, pull out that file now.
   if (Symbol *sym = ctx.symtab->find(ctx.arg.entry))
     handleUndefined(ctx, sym, "--entry");

lld/ELF/Symbols.cpp

@@ -333,10 +333,13 @@ bool elf::computeIsPreemptible(Ctx &ctx, const Symbol &sym) {
   if (sym.visibility() != STV_DEFAULT)
     return false;
 
-  // At this point copy relocations have not been created yet, so any
-  // symbol that is not defined locally is preemptible.
+  // At this point copy relocations have not been created yet.
+  // Shared symbols are preemptible. Undefined symbols are preemptible
+  // when zDynamicUndefined (default in dynamic linking). Weakness is not
+  // checked, though undefined non-weak would typically trigger relocation
+  // errors unless options like -z undefs are used.
   if (!sym.isDefined())
-    return true;
+    return !sym.isUndefined() || ctx.arg.zDynamicUndefined;
 
   if (!ctx.arg.shared)
     return false;
@@ -360,7 +363,6 @@ void elf::parseVersionAndComputeIsPreemptible(Ctx &ctx) {
   // can contain versions in the form of <name>@<version>.
   // Let them parse and update their names to exclude version suffix.
   // In addition, compute isExported and isPreemptible.
-  bool maybePreemptible = ctx.sharedFiles.size() || ctx.arg.shared;
   for (Symbol *sym : ctx.symtab->getSymbols()) {
     if (sym->hasVersionSuffix)
       sym->parseSymbolVersion(ctx);
@@ -369,11 +371,11 @@ void elf::parseVersionAndComputeIsPreemptible(Ctx &ctx) {
       continue;
     }
     if (!sym->isDefined() && !sym->isCommon()) {
-      sym->isPreemptible = maybePreemptible && computeIsPreemptible(ctx, *sym);
+      sym->isPreemptible = computeIsPreemptible(ctx, *sym);
     } else if (ctx.arg.exportDynamic &&
                (sym->isUsedInRegularObj || !sym->ltoCanOmit)) {
       sym->isExported = true;
-      sym->isPreemptible = maybePreemptible && computeIsPreemptible(ctx, *sym);
+      sym->isPreemptible = computeIsPreemptible(ctx, *sym);
     }
   }
 }

lld/ELF/Writer.cpp

@@ -285,7 +285,6 @@ static void demoteDefined(Defined &sym, DenseMap<SectionBase *, size_t> &map) {
 static void demoteSymbolsAndComputeIsPreemptible(Ctx &ctx) {
   llvm::TimeTraceScope timeScope("Demote symbols");
   DenseMap<InputFile *, DenseMap<SectionBase *, size_t>> sectionIndexMap;
-  bool maybePreemptible = ctx.sharedFiles.size() || ctx.arg.shared;
   for (Symbol *sym : ctx.symtab->getSymbols()) {
     if (auto *d = dyn_cast<Defined>(sym)) {
       if (d->section && !d->section->isLive())
@@ -301,9 +300,8 @@ static void demoteSymbolsAndComputeIsPreemptible(Ctx &ctx) {
       }
     }
 
-    if (maybePreemptible)
-      sym->isPreemptible = (sym->isUndefined() || sym->isExported) &&
-                           computeIsPreemptible(ctx, *sym);
+    sym->isPreemptible = (sym->isUndefined() || sym->isExported) &&
+                         computeIsPreemptible(ctx, *sym);
   }
 }
 

lld/docs/ReleaseNotes.rst

@@ -25,6 +25,10 @@ Non-comprehensive list of changes in this release
 
 ELF Improvements
 ----------------
+* Added ``-z dynamic-undefined-weak`` to make undefined weak symbols dynamic
+  when the dynamic symbol table is present.
+  (`#143831 <https://github.com/llvm/llvm-project/pull/143831>`_)
+
 * For AArch64, added support for ``-zgcs-report-dynamic``, enabling checks for
   GNU GCS Attribute Flags in Dynamic Objects when GCS is enabled. Inherits value
   from ``-zgcs-report`` (capped at ``warning`` level) unless user-defined,

lld/docs/ld.lld.1

@@ -793,6 +793,14 @@ Specify how to report the missing GNU_PROPERTY_X86_FEATURE_1_IBT or GNU_PROPERTY
 .Cm none
 is the default, linker will not report the missing property otherwise will be reported as a warning or an error.
 .Pp
+.It Cm dynamic-undefined-weak
+Make undefined weak symbols dynamic when the dynamic symbol table is present, if they are referenced from
+relocatable object files and not forced local by symbol visibility or versioning. Do not make them dynamic when
+.Cm nodynamic-undefined-weak
+is specified.
+.Cm dynamic-undefined-weak
+is the default when building a shared object, or when an input shared object is present.
+.Pp
 .It Cm pauth-report Ns = Ns Ar [none|warning|error]
 Specify how to report the missing GNU_PROPERTY_AARCH64_FEATURE_PAUTH property.
 .Cm none

lld/test/ELF/driver.test

@@ -47,7 +47,8 @@
 # ERR9: error: cannot open output file utput=/no/such/file
 
 # RUN: ld.lld %t -z foo -o /dev/null 2>&1 | FileCheck -check-prefix=ERR10 %s --implicit-check-not=warning:
-# RUN: ld.lld %t -z foo -z rel -z rela -z max-page-size=1 -z common-page-size=1 -o /dev/null --version 2>&1 | \
+# RUN: ld.lld %t -z foo -z rel -z rela -z max-page-size=1 -z common-page-size=1 -z dynamic-undefined-weak \
+# RUN:   -z nodynamic-undefined-weak -o /dev/null --version 2>&1 | \
 # RUN:   FileCheck -check-prefix=ERR10 %s --implicit-check-not=warning:
 # ERR10: warning: unknown -z value: foo
 

lld/test/ELF/weak-undef-got-plt.s

@@ -6,11 +6,17 @@
 
 # RUN: ld.lld a.o -o a
 # RUN: llvm-readelf -r a | FileCheck %s --check-prefix=NORELOC
+# RUN: ld.lld a.o -o a -z dynamic-undefined-weak
+# RUN: llvm-readelf -r a | FileCheck %s --check-prefix=NORELOC
 # RUN: ld.lld a.o s.so -o as
 # RUN: llvm-objdump -dR as | FileCheck %s
+# RUN: ld.lld a.o s.so -o as -z nodynamic-undefined-weak
+# RUN: llvm-readelf -r a | FileCheck %s --check-prefix=NORELOC
 
 # RUN: ld.lld -pie a.o s.so -o as.pie
 # RUN: llvm-objdump -dR as.pie | FileCheck %s
+# RUN: ld.lld -pie a.o s.so -o as.pie -z nodynamic-undefined-weak
+# RUN: llvm-readelf -r as.pie | FileCheck --check-prefix=NORELOC %s
 
 # RUN: ld.lld -shared a.o -o a.so
 # RUN: llvm-objdump -dR a.so | FileCheck %s

lld/test/ELF/weak-undef-hidden.s

@@ -5,6 +5,10 @@
 // RUN: ld.lld %t.o -o %t -pie
 // RUN: llvm-readobj -r -S --section-data %t | FileCheck %s
 
+/// -z dynamic-undefined-weak does not affect hidden undefined symbols.
+// RUN: ld.lld %t.o -o %t.so -shared -z dynamic-undefined-weak
+// RUN: llvm-readobj -r -S --section-data %t.so | FileCheck %s
+
 /// This is usually guarded with a comparison. Don't report an error.
 call g
 

lld/test/ELF/weak-undef-rw.s

@@ -18,9 +18,22 @@
 ## gABI leaves the behavior of weak undefined references implementation defined.
 ## We choose to resolve them statically for static linking and produce dynamic relocations
 ## for dynamic linking (-shared or at least one input DSO).
-##
-## Note: Some ports of GNU ld support -z nodynamic-undefined-weak that we don't
-## implement.
+
+## -z dynamic-undefined-weak is ignored if .dynsym is absent (-no-pie without DSO)
+# RUN: ld.lld a.o -o a.d -z dynamic-undefined-weak 2>&1 | count 0
+# RUN: llvm-readelf -r --hex-dump=.data a.d | FileCheck %s --check-prefix=STATIC
+
+## Currently no effect for S+A relocations.
+# RUN: ld.lld a.o s.so -o as.d -z dynamic-undefined-weak
+# RUN: llvm-readelf -r --hex-dump=.data as.d | FileCheck %s --check-prefix=STATIC
+
+## -z dynamic-undefined-weak forces dynamic relocations if .dynsym is present.
+# RUN: ld.lld a.o -o a.pie.d -pie -z dynamic-undefined-weak
+# RUN: llvm-readelf -r a.pie.d | FileCheck %s --check-prefix=DYN
+
+## -z nodynamic-undefined-weak suppresses dynamic relocations.
+# RUN: ld.lld a.o -o a.so.n -shared -z dynamic-undefined-weak -z nodynamic-undefined-weak
+# RUN: llvm-readelf -r --hex-dump=.data a.so.n | FileCheck %s --check-prefix=STATIC
 
 # STATIC:      no relocations
 # STATIC:      Hex dump of section '.data':