[LLD][COFF] `start-lib` / `end-lib` incorrectly discards symbols referenced in the same library

[yuzhy8701]

When linking in a library with start-lib and end-lib, lld-link would discard a symbol unless it is already seen in the prior objects. It discards the symbol even if the symbol is referenced in a latter object from the same library (latter as the command line order).

A simple case to repro (Powershell on Windows):

1. Setup

   Set-Content -Path .\hello-time.h -Value @"
   #ifndef LIB_HELLO_TIME_H_
   #define LIB_HELLO_TIME_H_
   
   #include <string>
   
   std::string get_localtime();
   
   #endif
   "@
   
   Set-Content -Path .\hello-time.cc -Value @"
   #include "hello-time.h"
   #include <ctime>
   #include <string>
   
   std::string get_localtime() {
     std::time_t result = std::time(nullptr);
     return std::asctime(std::localtime(&result));
   }
   "@
   
   Set-Content -Path .\hello-greet.h -Value @"
   #ifndef MAIN_HELLO_GREET_H_
   #define MAIN_HELLO_GREET_H_
   
   #include <string>
   
   std::string get_greet(const std::string &thing);
   
   #endif
   "@
   
   Set-Content -Path .\hello-greet.cc -Value @"
   #include "hello-time.h"
   #include "hello-greet.h"
   #include <string>
   
   std::string get_greet(const std::string& who) {
     return "Hello " + who + "; " + get_localtime();
   }
   "@
   
   Set-Content -Path .\hello-world.cc -Value @"
   #include "hello-greet.h"
   #include <iostream>
   #include <string>
   
   int main(int argc, char** argv) {
     std::cout << get_greet("world") << std::endl;
     return 0;
   }
   "@

2. The following succeeds:

   If (Test-Path 'out') {
       Remove-Item -Path 'out' -Recurse -Force -Verbose
   }
   New-Item -Path 'out' -ItemType Directory
   
   clang-cl.exe /c 'hello-time.cc' /Fo'out/hello-time.obj'
   clang-cl.exe /c 'hello-greet.cc' /Fo'out/hello-greet.obj'
   clang-cl.exe /c 'hello-world.cc' /Fo'out/hello-world.obj'
   
   lld-link.exe /OUT:'out/hello-world.exe' `
       out/hello-world.obj `
       '/start-lib' `
       out/hello-greet.obj `
       out/hello-time.obj `
       '/end-lib'

3. The following fails with error: lld-link: error: relocation against symbol in discarded section: ?get_localtime@@YA?AV?$basic_string@DU?$char_traits@D@std@@V?$allocator@D@2@@std@@XZ

   $ErrorActionPreference = "Stop"
   
   If (Test-Path 'out') {
       Remove-Item -Path 'out' -Recurse -Force -Verbose
   }
   New-Item -Path 'out' -ItemType Directory
   
   clang-cl.exe /c 'hello-time.cc' /Fo'out/hello-time.obj'
   clang-cl.exe /c 'hello-greet.cc' /Fo'out/hello-greet.obj'
   clang-cl.exe /c 'hello-world.cc' /Fo'out/hello-world.obj'
   
   lld-link.exe /OUT:'out/hello-world.exe' `
       out/hello-world.obj `
       '/start-lib' `
       out/hello-time.obj `
       out/hello-greet.obj `
       '/end-lib'

Note that the only difference between success and fail cases is the object order between start-lib and end-lib. I'm able to reproduce with the latest llvm version (19.1.0).

[llvmbot]

@llvm/issue-subscribers-lld-coff

Author: None (yuzhy8701)

When linking in a library with `start-lib` and `end-lib`, `lld-link` would discard a symbol unless it is already seen in the prior objects. It discards the symbol even if the symbol is referenced in a latter object from the same library (latter as the command line order).

A simple case to repro (Powershell on Windows):

1. Setup

   Set-Content -Path .\hello-time.h -Value @"
   #ifndef LIB_HELLO_TIME_H_
   #define LIB_HELLO_TIME_H_
   
   #include <string>
   
   std::string get_localtime();
   
   #endif
   "@
   
   Set-Content -Path .\hello-time.cc -Value @"
   #include "hello-time.h"
   #include <ctime>
   #include <string>
   
   std::string get_localtime() {
     std::time_t result = std::time(nullptr);
     return std::asctime(std::localtime(&result));
   }
   "@
   
   Set-Content -Path .\hello-greet.h -Value @"
   #ifndef MAIN_HELLO_GREET_H_
   #define MAIN_HELLO_GREET_H_
   
   #include <string>
   
   std::string get_greet(const std::string &thing);
   
   #endif
   "@
   
   Set-Content -Path .\hello-greet.cc -Value @"
   #include "hello-time.h"
   #include "hello-greet.h"
   #include <string>
   
   std::string get_greet(const std::string& who) {
     return "Hello " + who + "; " + get_localtime();
   }
   "@
   
   Set-Content -Path .\hello-world.cc -Value @"
   #include "hello-greet.h"
   #include <iostream>
   #include <string>
   
   int main(int argc, char** argv) {
     std::cout << get_greet("world") << std::endl;
     return 0;
   }
   "@

2. The following succeeds:

   If (Test-Path 'out') {
       Remove-Item -Path 'out' -Recurse -Force -Verbose
   }
   New-Item -Path 'out' -ItemType Directory
   
   clang-cl.exe /c 'hello-time.cc' /Fo'out/hello-time.obj'
   clang-cl.exe /c 'hello-greet.cc' /Fo'out/hello-greet.obj'
   clang-cl.exe /c 'hello-world.cc' /Fo'out/hello-world.obj'
   
   lld-link.exe /OUT:'out/hello-world.exe' `
       out/hello-world.obj `
       '/start-lib' `
       out/hello-greet.obj `
       out/hello-time.obj `
       '/end-lib'

3. The following fails with error: lld-link: error: relocation against symbol in discarded section: ?get_localtime@@<!-- -->YA?AV?$basic_string@<!-- -->DU?$char_traits@<!-- -->D@<!-- -->std@@<!-- -->V?$allocator@<!-- -->D@<!-- -->2@@<!-- -->std@@<!-- -->XZ

   $ErrorActionPreference = "Stop"
   
   If (Test-Path 'out') {
       Remove-Item -Path 'out' -Recurse -Force -Verbose
   }
   New-Item -Path 'out' -ItemType Directory
   
   clang-cl.exe /c 'hello-time.cc' /Fo'out/hello-time.obj'
   clang-cl.exe /c 'hello-greet.cc' /Fo'out/hello-greet.obj'
   clang-cl.exe /c 'hello-world.cc' /Fo'out/hello-world.obj'
   
   lld-link.exe /OUT:'out/hello-world.exe' `
       out/hello-world.obj `
       '/start-lib' `
       out/hello-time.obj `
       out/hello-greet.obj `
       '/end-lib'

Note that the only difference between success and fail cases is the object order between start-lib and end-lib. I'm able to reproduce with the latest llvm version (19.1.0).

[aganea]

Should be fixed by #124294, if not, please reopen.