LLD over-aligns strings and wastes space #54036
Assignees
Labels
Comments
|
@llvm/issue-subscribers-lld-macho |
|
#50135 explains why the hack was added in the first place. |
gbaraldi
pushed a commit
to JuliaLang/llvm-project
that referenced
this issue
Jan 6, 2023
Previously, we aligned every cstring to 16 bytes as a temporary hack to deal with llvm#50135. However, it was highly wasteful in terms of binary size. To recap, in contrast to ELF, which puts strings that need different alignments into different sections, `clang`'s Mach-O backend puts them all in one section. Strings that need to be aligned have the .p2align directive emitted before them, which simply translates into zero padding in the object file. In other words, we have to infer the alignment of the cstrings from their addresses. We differ slightly from ld64 in how we've chosen to align these cstrings. Both LLD and ld64 preserve the number of trailing zeros in each cstring's address in the input object files. When deduplicating identical cstrings, both linkers pick the cstring whose address has more trailing zeros, and preserve the alignment of that address in the final binary. However, ld64 goes a step further and also preserves the offset of the cstring from the last section-aligned address. I.e. if a cstring is at offset 18 in the input, with a section alignment of 16, then both LLD and ld64 will ensure the final address is 2-byte aligned (since `18 == 16 + 2`). But ld64 will also ensure that the final address is of the form 16 * k + 2 for some k (which implies 2-byte alignment). Note that ld64's heuristic means that a dedup'ed cstring's final address is dependent on the order of the input object files. E.g. if in addition to the cstring at offset 18 above, we have a duplicate one in another file with a `.cstring` section alignment of 2 and an offset of zero, then ld64 will pick the cstring from the object file earlier on the command line (since both have the same number of trailing zeros in their address). So the final cstring may either be at some address `16 * k + 2` or at some address `2 * k`. I've opted not to follow this behavior primarily for implementation simplicity, and secondarily to save a few more bytes. It's not clear to me that preserving the section alignment + offset is ever necessary, and there are many cases that are clearly redundant. In particular, if an x86_64 object file contains some strings that are accessed via SIMD instructions, then the .cstring section in the object file will be 16-byte-aligned (since SIMD requires its operand addresses to be 16-byte aligned). However, there will typically also be other cstrings in the same file that aren't used via SIMD and don't need this alignment. They will be emitted at some arbitrary address `A`, but ld64 will treat them as being 16-byte aligned with an offset of `16 % A`. I have verified that the two repros in llvm#50135 work well with the new alignment behavior. Fixes llvm#54036. Reviewed By: #lld-macho, oontvoo Differential Revision: https://reviews.llvm.org/D121342
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The over-alignment hack was added here: https://reviews.llvm.org/D104835
The overhead of over-alignment is minor on Chromium, but it seems significant on at least one of our internal builds (making
__cstringapprox. 50% larger).The text was updated successfully, but these errors were encountered: