[ELF] Keep symbols specified by '-u' over LTO.

[ikudrin]

Summary

See D37332 for the discussion.

[ruiu]

I'm not sure if this is the right thing to do. It feels like I could argue either this is correct or this is incorrect.

I'd think this is probably okay, but please do not copy-n-paste the same comment.

[ikudrin]

In D38348#885488, @ruiu wrote:

I'm not sure if this is the right thing to do. It feels like I could argue either this is correct or this is incorrect.

Honestly, I feel the same. I can see the following pros and cons:

Pros:

• It fits POLA better because keeping symbols you explicitly asked for is just natural. I can remember that I was a bit puzzled once in the similar situation.
• It makes handling "-e" and "-u" symbols the same, see D37332#861693.
• It doesn't harm anything because without LTO those symbols would be preserved anyway.

Cons:

• As far as I know, nobody asked for it so far.
• "-u" is intended to only fetch asked symbols from archives. It says nothing about keeping symbols themselves.

Consequently, it's probably safe and makes things a bit better.

[ruiu]

LGTM. Please go ahead.

lld/trunk/ELF/SymbolTable.cpp

if (S->isWeak())
     replaceBody<LazyObject>(S, &Obj, Name, S->body()->Type);
   else if (InputFile *F = Obj.fetch())
     addFile<ELFT>(F);
 }
 
 // Process undefined (-u) flags by loading lazy symbols named by those flags.
 template <class ELFT> void SymbolTable::scanUndefinedFlags() {
   for (StringRef S : Config->Undefined)
-    if (auto *L = dyn_cast_or_null<Lazy>(find(S)))
+    if (SymbolBody *B = find(S)) {
+      // Mark the symbol not to be eliminated by LTO
+      // even if it is a bitcode symbol.
+      B->symbol()->IsUsedInRegularObj = true;
+      if (auto *L = dyn_cast_or_null<Lazy>(B))
         if (InputFile *File = L->fetch())
           addFile<ELFT>(File);
     }
+}
 
 // This function takes care of the case in which shared libraries depend on
 // the user program (not the other way, which is usual). Shared libraries
 // may have undefined symbols, expecting that the user program provides
 // the definitions for them. An example is BSD's __progname symbol.
 // We need to put such symbols to the main program's .dynsym so that
 // shared libraries can find them.
 // Except this, we ignore undefined symbols in DSOs.

lld/trunk/test/ELF/lto/keep-undefined.ll

+; REQUIRES: x86
+; This test checks that symbols which are specified in "-u" switches
+; are kept over LTO if we link an executable.
+; RUN: llvm-as %s -o %t.o
+; RUN: ld.lld -m elf_x86_64 %t.o -o %tout -u foo
+; RUN: llvm-nm %tout | FileCheck %s
+
+; CHECK: T foo
+
+target triple = "x86_64-unknown-linux-gnu"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @foo() {
+  ret void
+}
+
+define void @_start() {
+  call void @foo()
+  ret void
+}