[gcc-16 regression] Linker scripts starting with `INPUT(AS_NEEDED(...))` do not work

[eseiler]

Hi,

gcc-trunk (g++-16 (GCC) 16.0.1 20260115 (experimental)) currently does not work with mold:

# test.cpp = `int main() {}`
g++-16 -fuse-ld=mold test.cpp     
mold: fatal: /usr/local/stow/gcc-16/lib64/libatomic_asneeded.so:3: INPUT ( AS_NEEDED( -latomic ) )
                                                                           ^ library not found: AS_NEEDED
collect2: error: ld returned 1 exit status

Background

GCC added some linker scripts (e.g., INPUT ( AS_NEEDED ( -latomic ) )):

• Commit
• Ticket

(Edit) The bug was also independently filed in GCC's bug tracker and now links to this issue.

Stacktrace

mold::Script<mold::X86_64>::resolve_path(mold::Script<mold::X86_64> * this, mold::Script<mold::X86_64> * this@entry, std::__1::string_view tok, bool check_target) (\mold\src\linker-script.cc:185)
mold::Script<mold::X86_64>::get_script_output_type(mold::Script<mold::X86_64> * this) (\mold\src\linker-script.cc:253)
mold::detect_machine_type<mold::X86_64>(mold::Context<mold::X86_64> & ctx) (\mold\src\main.cc:153)
mold::mold_main<mold::X86_64>(int argc, char ** argv) (\mold\src\main.cc:295)
libc.so.6![Unknown/Just-In-Time compiled code] (Unknown Source:0)
libc.so.6!__libc_start_main (Unknown Source:0)
_start (Unknown Source:0)

Misc

The error doesn't happen if:

• -m is set, or
• first entry in INPUT is a a path to a lib or uses -l

mold/src/main.cc

Lines 294 to 295 in 17956fd

	if (ctx.arg.emulation.empty())
	ctx.arg.emulation = detect_machine_type(ctx, file_args);

mold/src/main.cc

Lines 152 to 153 in 17956fd

	if (std::string_view target =
	Script(ctx, rctx, mf).get_script_output_type();

mold/src/linker-script.cc

Lines 251 to 253 in 17956fd

	if (tok.size() >= 3 && (tok[0] == "INPUT" || tok[0] == "GROUP") &&
	tok[1] == "(")
	if (MappedFile *mf = resolve_path(tok[2], false))

Trying to treat AS_NEEDED as a path then results in the error.
The code probably needs to handle groups, similar to what read_group does.

[ArsenArsen]

seems that linker scripts found via -l and specified directly are treated differently:

/tmp$ echo 'INPUT ( AS_NEEDED ( -latomic ) )' > libfoo.so
/tmp$ ld.mold -o a -L. -lfoo a.o
mold: fatal: ./libfoo.so:1: INPUT ( AS_NEEDED ( -latomic ) )
                                    ^ library not found: AS_NEEDED
/tmp 1 $ ld.mold -o a ./libfoo.so a.o
mold: fatal: library not found: atomic

... this would cause the error we're seeing with GCC

[jwakely]

The code probably needs to handle groups, similar to what read_group does.

N.B. GROUP( AS_NEEDED( -lx )) also doesn't work:

mold: fatal: /tmp/libscript.a:2: GROUP ( AS_NEEDED ( -latomic ) )
                                         ^ library not found: AS_NEEDED

As Arsen said, the problem seems to be that a linker script found via -l is handled differently. Maybe it isn't using the read_group code at all.

[jwakely]

Confirmed, mold::read_input_files<mold::X86_64> is never called for -lfoo arguments, so if it's a linker script, it won't use read_group.

Maybe the problem is here:

mold/src/linker-script.cc

Lines 251 to 254 in 17956fd

	if (tok.size() >= 3 && (tok[0] == "INPUT" || tok[0] == "GROUP") &&
	tok[1] == "(")
	if (MappedFile *mf = resolve_path(tok[2], false))
	return get_machine_type(ctx, rctx, mf);

Should this call read_group?

[jwakely]

This kluge solves the immediate problem for GCC:

--- a/src/linker-script.cc
+++ b/src/linker-script.cc
@@ -249,9 +249,12 @@ std::string_view Script<E>::get_script_output_type() {
   }
 
   if (tok.size() >= 3 && (tok[0] == "INPUT" || tok[0] == "GROUP") &&
-      tok[1] == "(")
+      tok[1] == "(") {
+    if (tok[2] == "AS_NEEDED" && tok[3] == "(")
+      tok = tok.subspan(2);
     if (MappedFile *mf = resolve_path(tok[2], false))
       return get_machine_type(ctx, rctx, mf);
+  }
   return "";
 }
 

This code will still fail for anything more complex than AS_NEEDED( -lx ) though.