pcc wrote: That's a fair point, I was also concerned about the number of options required here. We do need some way to communicate the preferred alignment separately from the minimum alignment. Without this, almost no sections would be eligible for jump table relaxation the way things work now.
An alternative approach would be to communicate both alignments in the object file. That way, the optimization would work by default with no flags passed to clang or lld but more work would be needed on the toolchain side. What I had in mind was that sh_addralign would contain the preferred alignment (which would also be controlled by -falign-functions), and we would introduce a new LLVM-specific ELF section type containing an array of (relocation pointing to section, minimum alignment) pairs. Another possibility is that we go back to the approach from the initial prototype of #147424 where the relocation from the jump table section gives the linker permission to reduce the alignment, but this is a bit of a hack and may cause problems e.g. if the code generator embeds constants in the code section that are read with movaps. Let me know what you think. https://github.com/llvm/llvm-project/pull/149448 _______________________________________________ llvm-branch-commits mailing list llvm-branch-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
