[libclc] add --only-needed to llvm-link when INTERNALIZE flag is set

[wenju-he]

When -internalize flag is passed to llvm-link, we only need to link in needed symbols. This PR reduces size of linked bitcode, e.g. by removing following symbols:
Z12__clc_sw_fmaDv16_fS_S
Z12__clc_sw_fmaDv2_fS_S
Z12__clc_sw_fmaDv3_fS_S
Z12__clc_sw_fmaDv4_fS_S
Z12__clc_sw_fmaDv8_fS_S
_Z12__clc_sw_fmafff

[wenju-he]

@frasercrmck could you please review? thanks

[frasercrmck]

Interesting - I'm not seeing any change in bytecode modules before and after this change.

[wenju-he]

@frasercrmck could you please try this PR on https://github.com/intel/llvm repo?

My experiment:
clang --version
clang version 21.0.0git (https://github.com/llvm/llvm-project f5ee105)

before PR:
find . -name "builtins.link*.bc" -printf "%s\n" | paste -sd+ | bc
101593692

after PR:
find . -name "builtins.link*.bc" -printf "%s\n" | paste -sd+ | bc
101316928

This PR reduces bitcode file sizes by 0.27%

[frasercrmck]

You're right, my mistake: even on this repo.

I was using llvm-diff and only checking the error code. It reports e.g., function @_Z9__clc_minDv16_DhDh exists only in left module but still returns 0 so I missed it.

This only affects the builtin.link.* bytecode libraries and not the final ones, presumably due to optimizations.