[lld] [lld][ELF] Add --why-live flag (inspired by Mach-O) (PR #127112)

[Daniel Thornburgh via llvm-commits]

mysterymath wrote:

> > I think we need Mach-O style `template <bool RecordWhyLive>` to prevent overhead with the regular --gc-sections code path.
> > https://reviews.llvm.org/D69607 is a previous attempt adding the option.
> 
> It seems prudent to collect some evidence that this might be a performance problem before moving to a less readable implementation. I'm planning to do a simple link of clang with `-ffunction-sections` and `-fdata-sections`; that should provide some cursory evidence. If that's not sufficient, I'm also totally okay with this patch being summarily reverted in case a regression occurs in something larger. Especially since it should be entirely mechanical to move this to use `template <bool RecordWhyLive>`.

I finally got back around to this, and I tried the above link of clang. Running a Student's t test on 5 trials yields a mean regression of 1.5%, and a very low p value for the means being the same (0.00024). That's definitely way too much, so I'll go through and add the templates. I'm surprised it was this big of a difference, but I'm still glad I measured.

https://github.com/llvm/llvm-project/pull/127112