pcc wrote:
> This patch has a few problems
>
> * The check in `processAux` was moved to the wrong place (it will never get
> hit because this is after we deal with all symbols not defined locally). This
> patch moves it to approximately the right place but will need some logic
> changes to `isStaticLinkTimeConstant` to recognize ifuncs as not being static
> link time constants:
>
> ```
> diff --git a/lld/ELF/Relocations.cpp b/lld/ELF/Relocations.cpp
> index 1439fd859f51..969d71461cb1 100644
> --- a/lld/ELF/Relocations.cpp
> +++ b/lld/ELF/Relocations.cpp
> @@ -1127,6 +1127,8 @@ void RelocationScanner::processAux(RelExpr expr,
> RelType type, uint64_t offset,
> // handling of GOT-generating relocations.
> if (isStaticLinkTimeConstant(expr, type, sym, offset) ||
> (!ctx.arg.isPic && sym.isUndefWeak())) {
> + if (LLVM_UNLIKELY(isIfunc && !needsGot(expr) && !needsPlt(expr)))
> + sym.setFlags(HAS_DIRECT_RELOC);
> sec->addReloc({expr, type, offset, addend, &sym});
> return;
> }
> @@ -1194,9 +1196,6 @@ void RelocationScanner::processAux(RelExpr expr,
> RelType type, uint64_t offset,
> }
> }
>
> - if (LLVM_UNLIKELY(isIfunc && !needsGot(expr) && !needsPlt(expr)))
> - sym.setFlags(HAS_DIRECT_RELOC);
> -
> // When producing an executable, we can perform copy relocations (for
> // STT_OBJECT) and canonical PLT (for STT_FUNC) if sym is defined by a DSO.
> // Copy relocations/canonical PLT entries are unsupported for
> ```
>
> * Need to adjust how IRELATIVE relocations get emitted. Currently we emit
> relocations with the wrong type (assuming they will be demoted) and fix up
> the type in `RelocationBaseSection::partitionRels` if they turn out to not be
> demoted. But this isn't good enough for PDEs which need them to be static
> relocations. Here's what I think we should do instead:
>
> * Always emit ifunc relocations as IRELATIVE during `processAux`
> * In `handleNonPreemptibleIfunc`, when we demote an ifunc to an iplt entry,
> loop over all irelatives pointing to the symbol and rewrite them as either
> relative (for PIE/shared) or static relocations (for PDE). It's fine to loop
> over the irelatives every time because they are rare.
All of this is done in the latest change (except that I decided to handle the
demotion in a separate loop instead of `handleNonPreemptibleIfunc`, PTAL
https://github.com/llvm/llvm-project/pull/133531
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
