[llvm-branch-commits] [lld] [lld][MachO] Change direct branch targets to ICF bodies (PR #239472)

Ellis Hoag via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Fri Nov 6 13:49:49 PST 2026


ellishg wrote:

> I'm not comfortable with direct calls bypassing the thunk by default. At the very least it should be behind a flag. These functions get a thunk because their address is significant, and that address is also where tools act on the function: a breakpoint, a DTrace pid entry probe, or an inline hook patching its entry. Today all of them see every call. With this change they only see calls through a pointer, so the same function behaves differently depending on how it's called.

I suppose breakpoints still work with `--icf=all` because both folded symbols point to the same address. With `--icf=safe_thunks`, the address significant symbol points to the thunk, and the other symbol points to the function body. So bypassing the thunk would break debuggers. I could guard this behind a flag, but I'm wondering if the benefit is worth it.

> Some of the cost described in the description comes from where the thunks are placed. They're appended after every other input section, so without an order file a call to a thunk-folded function first jumps to the end of __text, which could be a long jump in a large binary. I have a patch that instead puts each thunk where the function it replaces was in the input order. The caller's jump then reaches a thunk at the original function's distance.

We have logic in BP to order ICF thunks with their bodies, but of course that requires an orderfile.

> I also have another patch ready that keeps fewer functions address-significant. Safe ICF applies Apple ld's address-taken heuristic when __llvm_addrsig is missing or marks a function. A function then stays unique only if it's exported or something other than a branch references it. That should noticeably reduce the number of thunks, and with them the calls this PR is trying to shorten.

Yeah we do this too, but we only look at relocations in the `__text` section. This will miss function addresses stored in data sections, but apparently this isn't a problem for us. Do you do something similar? This reminds me of the discussion in https://github.com/llvm/llvm-project/pull/192636. 

https://github.com/llvm/llvm-project/pull/239472


More information about the llvm-branch-commits mailing list