"H.J. Lu" <hjl.tools@gmail.com> writes:

> We can use mmap to map in ELF section contents, instead of copying them
> into memory by hand.  We don't need to cache symbol nor relocation tables
> if they are mapped in.  Data to link the 3.5GB clang executable in LLVM
> 17 debug build on Linux/x86-64 with 32GB RAM is:
>
> 		stdio		mmap		improvement
> user		86.73		87.02		-0.3%
> system		9.55		9.21		3.6%
> total		100.40		97.66		0.7%
> maximum set(GB)	17.34		13.14		24%
> page faults	4047667		3042877		25%
>
> and data to link the 275M cc1plus executable in GCC 14 stage 1 build is:
>
> user		5.41		5.44		-0.5%
> system		0.80		0.76		5%
> total		6.25		6.26		-0.2%
> maximum set(MB)	1323		968		27%
> page faults	323451		236371		27%
>
> Data shows that these won't improve the single copy linker performance.
> But they improve the overall system performance when linker is used by
> reducing linker memory usage and page faults.  They allow more parallel
> linker jobs on LLVM debug build.
>
> Here is a quote from Noah Goldstein: "on a large project they are an
> extremely large speedup".

Thanks for working on this. I've suffered from OOMs when linking large
C++ projects with debug info a lot. I'm going to include it in my
regular test builds.