When PIC code uses a string, this creates a relocation that slows down program startup time. As long as the string is in the same compilation unit as the PIC code this relocation is unnecessary, as the offset of the compiled unit is stored in the program counter, and can be accessed by a stored subtraction/addition. This would mean more instructions at the call site, so the details I would image would require some careful coordination between llvm and lld.
I have a bug for this: https://bugs.llvm.org/show_bug.cgi?id=39049
and would greatly appreciate feedback, and pointers on how to do this.