[lld] 92f8a6e - [lld/mac] Fix --start-lib/--end-lib with split thinlto inputs
Nico Weber via llvm-commits
llvm-commits at lists.llvm.org
Mon Dec 5 07:51:33 PST 2022
Author: Nico Weber
Date: 2022-12-05T10:51:14-05:00
New Revision: 92f8a6eba7e87f0073a1aaf36250434152053b37
URL: https://github.com/llvm/llvm-project/commit/92f8a6eba7e87f0073a1aaf36250434152053b37
DIFF: https://github.com/llvm/llvm-project/commit/92f8a6eba7e87f0073a1aaf36250434152053b37.diff
LOG: [lld/mac] Fix --start-lib/--end-lib with split thinlto inputs
Fixes #59162. The test has a comment explaining what's going on.
See also Symbol::extract() in lld/ELF/Symbols.cpp.
The included test sadly also passes if I pass just bd448f01a62,
while doing that isn't enough to make my bigger repro case work
(if I port just that, something else asserts later on, but with
this fix here everything's fine in my bigger repro).
Differential Revision: https://reviews.llvm.org/D139199
Added:
lld/test/MachO/thinlto-split-unit-start-lib.ll
Modified:
lld/MachO/InputFiles.cpp
lld/test/MachO/Inputs/MacOSX.sdk/usr/lib/libc++abi.tbd
Removed:
################################################################################
diff --git a/lld/MachO/InputFiles.cpp b/lld/MachO/InputFiles.cpp
index b76303cc3fc73..eea700ea3aef8 100644
--- a/lld/MachO/InputFiles.cpp
+++ b/lld/MachO/InputFiles.cpp
@@ -2235,9 +2235,11 @@ void BitcodeFile::parseLazy() {
}
void macho::extract(InputFile &file, StringRef reason) {
- printArchiveMemberLoad(reason, &file);
- assert(file.lazy);
+ if (!file.lazy)
+ return;
file.lazy = false;
+
+ printArchiveMemberLoad(reason, &file);
if (auto *bitcode = dyn_cast<BitcodeFile>(&file)) {
bitcode->parse();
} else {
diff --git a/lld/test/MachO/Inputs/MacOSX.sdk/usr/lib/libc++abi.tbd b/lld/test/MachO/Inputs/MacOSX.sdk/usr/lib/libc++abi.tbd
index 2f2e79d01e897..51814385daa5b 100644
--- a/lld/test/MachO/Inputs/MacOSX.sdk/usr/lib/libc++abi.tbd
+++ b/lld/test/MachO/Inputs/MacOSX.sdk/usr/lib/libc++abi.tbd
@@ -6,5 +6,6 @@ install-name: '/usr/lib/libc++abi.dylib'
current-version: 1281
exports:
- archs: [ i386, x86_64, arm64 ]
- symbols: [ ___cxa_allocate_exception, ___cxa_begin_catch, ___cxa_end_catch, ___cxa_throw, ___gxx_personality_v0, __ZTIi ]
+ symbols: [ ___cxa_allocate_exception, ___cxa_begin_catch, ___cxa_end_catch, ___cxa_throw,
+ ___gxx_personality_v0, __ZTIi, __ZTVN10__cxxabiv117__class_type_infoE ]
...
diff --git a/lld/test/MachO/thinlto-split-unit-start-lib.ll b/lld/test/MachO/thinlto-split-unit-start-lib.ll
new file mode 100644
index 0000000000000..6c33a505c75cf
--- /dev/null
+++ b/lld/test/MachO/thinlto-split-unit-start-lib.ll
@@ -0,0 +1,96 @@
+; REQUIRES: x86
+; RUN: rm -rf %t; split-file %s %t
+
+;; Split-LTO bitcode files can have the same symbol (typeinfo) twice in the
+;; symbol table: First undefined in the main file, then defined in the index
+;; file.
+;; When used in --start-lib / --end-lib, ld64.lld creates lazy symbols
+;; for all non-undefined symbols in the bitcode file.
+;; In vtable.o below, the typeinfo __ZTI1S is present once as undefined and
+;; once as defined. The defined version is added as a added as a LazyObject
+;; symbol.
+;; When vtable.o gets loaded due to the __ZN1SC1Ev ref from vtable_use.o,
+;; the first __ZTI1S (undefined) used to cause vtable.o to be extracted
+;; a second time, which used to cause an assert.
+;; See PR59162 for details.
+
+; RUN: opt --thinlto-bc --thinlto-split-lto-unit -o %t/vtable.o %t/vtable.ll
+; RUN: opt --thinlto-bc --thinlto-split-lto-unit -o %t/vtable_use.o %t/vtable_use.ll
+
+; RUN: %lld -lc++ --start-lib %t/vtable.o --end-lib %t/vtable_use.o -o /dev/null
+
+;; Bitcode files created by:
+; % cat vtable.cc
+; struct S {
+; S();
+; virtual void f();
+; };
+; S::S() {}
+; void S::f() {}
+
+; % cat vtable_use.cc
+; struct S {
+; S();
+; virtual void f();
+; };
+; int main() { S s; }
+
+; % clang -c vtable_use.cc vtable.cc -emit-llvm -S -fno-exceptions -arch x86_64 -mmacos-version-min=11 -O1
+
+; ...and then manually ading `, !type !8, type !9` based on `clang -S -emit-llvm -flto=thin` output,
+; because splitAndWriteThinLTOBitcode() in ThinLTOBitcodeWriter.cpp only splits bitcode
+; if type annotations are present. While at it, also removed unneccessary metadata.
+; (NB: The first comment creates vtable.ll while the latter generates vtable.s! vtable.s
+; contains a few things opt complains about, so we can't use the output of that directly.)
+
+;--- vtable.ll
+; ModuleID = 'vtable.cc'
+source_filename = "vtable.cc"
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx11.0.0"
+
+ at _ZTV1S = unnamed_addr constant { [3 x ptr] } { [3 x ptr] [ptr null, ptr @_ZTI1S, ptr @_ZN1S1fEv] }, align 8
+ at _ZTVN10__cxxabiv117__class_type_infoE = external global ptr
+ at _ZTS1S = constant [3 x i8] c"1S\00", align 1
+ at _ZTI1S = constant { ptr, ptr } { ptr getelementptr inbounds (ptr, ptr @_ZTVN10__cxxabiv117__class_type_infoE, i64 2), ptr @_ZTS1S }, align 8, !type !0, !type !1
+
+; Function Attrs: mustprogress nofree norecurse nosync nounwind ssp willreturn memory(argmem: write) uwtable
+define void @_ZN1SC2Ev(ptr nocapture noundef nonnull writeonly align 8 dereferenceable(8) %this) unnamed_addr align 2 {
+entry:
+ store ptr getelementptr inbounds ({ [3 x ptr] }, ptr @_ZTV1S, i64 0, inrange i32 0, i64 2), ptr %this, align 8
+ ret void
+}
+
+; Function Attrs: mustprogress nofree norecurse nosync nounwind ssp willreturn memory(argmem: write) uwtable
+define void @_ZN1SC1Ev(ptr nocapture noundef nonnull writeonly align 8 dereferenceable(8) %this) unnamed_addr align 2 {
+entry:
+ store ptr getelementptr inbounds ({ [3 x ptr] }, ptr @_ZTV1S, i64 0, inrange i32 0, i64 2), ptr %this, align 8
+ ret void
+}
+
+; Function Attrs: mustprogress nofree norecurse nosync nounwind ssp willreturn memory(none) uwtable
+define void @_ZN1S1fEv(ptr nocapture nonnull align 8 %this) unnamed_addr align 2 {
+entry:
+ ret void
+}
+
+!0 = !{i64 16, !"_ZTS1S"}
+!1 = !{i64 16, !"_ZTSM1SFvvE.virtual"}
+
+;--- vtable_use.ll
+; ModuleID = 'vtable_use.cc'
+source_filename = "vtable_use.cc"
+target datalayout = "e-m:o-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx11.0.0"
+
+%struct.S = type { ptr }
+
+; Function Attrs: mustprogress noinline norecurse nounwind optnone ssp uwtable
+define noundef i32 @main() {
+entry:
+ %s = alloca %struct.S, align 8
+ call void @_ZN1SC1Ev(ptr noundef nonnull align 8 dereferenceable(8) %s)
+ ret i32 0
+}
+
+declare void @_ZN1SC1Ev(ptr noundef nonnull align 8 dereferenceable(8)) unnamed_addr
More information about the llvm-commits
mailing list