Add better support for specific intrinsics

[phate]

We still have quite a few reasonably often used intrinsics in the benchmark code that currently do not match to a specific operation, and therefore are translate to a general call:

1. llvm.va_end (58 files, 354 calls)
2. llvm.va_start (58 files, 339 calls)
3. llvm.va_copy (4 files, 8 calls)
4. llvm.threadlocal.address (1 file, 54 calls) -- programs/gdb-15.2/bfd/bfd.c
5. llvm.stackrestore (3 files, 23 calls)
6. llvm.stacksave (3 files, 23 calls)
7. llvm.frameaddress (2 files, 2 calls)
8. llvm.eh.unwind.init (1 file, 1 call)
9. llvm.ssub.sat (1 file, 1 call)