Always emit symbols like `_GLOBAL_OFFSET_TABLE_`, `_PROCEDURE_LINKAGE_TABLE_` and others

[marxin]

I noticed these symbols are always emitted by the Mold linker and the presence of the symbols seems to me pretty hand when it comes to disassembly:

$ cat lib.c
void foo() {
  __builtin_printf("xxx\n");
}
$ gcc lib.c -fPIC -shared -o lib.so -fuse-ld=mold
$ readelf -sW lib.so
...
    67: 00000000000026f0     0 NOTYPE  LOCAL  DEFAULT   19 _DYNAMIC
    68: 00000000000028e8     0 NOTYPE  LOCAL  DEFAULT   21 _GLOBAL_OFFSET_TABLE_
    69: 00000000000015e0     0 NOTYPE  LOCAL  DEFAULT   16 _PROCEDURE_LINKAGE_TABLE_
...
    81: 0000000000000000     0 TLS     LOCAL  DEFAULT    1 _TLS_MODULE_BASE_

On the other hand, the Wild linker does not emit these symbols.

[davidlattimore]

For the most part, I've tried to make Wild match GNU ld as closely as possible. However, it seems reasonable to me to add extra symbols if they're useful for diagnostic purposes, especially given that another linker does it. So adding them sounds good to me

[marxin]

From debugging point of view, it's even more convenient for symbols like this:

   37: 00000000000015f0      0 FUNC    LOCAL  DEFAULT       14 __cxa_finalize$pltgot
   38: 000000000000c4e0      0 OBJECT  LOCAL  DEFAULT       20 __gmon_start__$got
   39: 000000000000c4e8      0 OBJECT  LOCAL  DEFAULT       20 _ITM_deregisterTMCloneTable$got
   40: 000000000000c4f0      0 OBJECT  LOCAL  DEFAULT       20 _ITM_registerTMCloneTable$got
   41: 000000000000c508      0 OBJECT  LOCAL  DEFAULT       20 __cxa_finalize$got
   42: 000000000000c4f8      0 OBJECT  LOCAL  DEFAULT       20 foo$tlsdesc

[davidlattimore]

Creating symbols for each GOT and PLT entry seems like a potentially useful feature. I'm unsure what effect that might have on performance and binary size and since GNU ld doesn't do it, I'm not sure if we want it on by default. I gather mold does it by default. I tried looking to see if it has flags to turn it on and off, but couldn't see any. In any case, I'm certainly fine with emitting such symbols if a flag enables it. If once we've done that, it turns out the effect on performance is negligible, then we could consider just turning it on by default.

[davidlattimore]

I'm going to take a look at this. It looks like GNU ld does emit symbols like _GLOBAL_OFFSET_TABLE_ in circumstances when we don't. We currently don't emit such symbols unless they're referenced. Probably I'll attempt to change it so that we also emit them if the section they're contained in is emitted. That's a little bit tricky since deciding whether to emit symbols affects section sizes, but we want to use the section sizes to decide whether to emit the symbol. But we do something similar with some of the headers, so I'm hopeful I'll be able to move the code there.

For the *$got, *plt etc symbols, I think that should be fairly easy to implement. Once I'm done with the built-in symbols, I'll probably write that up as a separate good-first-issue.

[marxin]

For completeness, just recently I was working on the RISC-V port, and it was pretty useful that the Mold linker provided extra symbols related to TLS, like:

   47: 0000000000002c10      0 OBJECT  LOCAL  DEFAULT       23 $tlsld
...
   49: 0000000000002d70      0 OBJECT  LOCAL  DEFAULT       23 local_tls7$tlsgd
   50: 0000000000002d80      0 OBJECT  LOCAL  DEFAULT       23 local_tls8$tlsgd
   51: 0000000000002d90      0 OBJECT  LOCAL  DEFAULT       23 local_tls9$tlsgd
...
   48: 0000000000002be8      0 OBJECT  LOCAL  DEFAULT       23 local_tls7$gottp
   49: 0000000000002bf0      0 OBJECT  LOCAL  DEFAULT       23 local_tls8$gottp