[ELF][PPC] Don't relax ifunc toc-indirect accesses to toc-relative

[MaskRay]

Summary

Fixes PR42759.

// If ifunc is taken address in -fPIC code, it may have a toc entry
.section .toc,"aw",@progbits
  .quad ifunc

// ifunc may be defined as STT_GNU_IFUNC in another object file
.type ifunc, %gnu_indirect_function

If ifunc is non-preemptable (e.g. when linking an executable), the toc
entry will be relocated by R_PPC64_IRELATIVE.

R_*_IRELATIVE represents the symbolic value of a
non-preemptable ifunc (not associated with a canonical PLT) in a writable location. It has an unknown value at
link time, so we cannot apply toc-indirect to toc-relative relaxation.

[luporl]

Nice, I've confirmed that it fixes PR42759 on FreeBSD.

[sfertile]

ELF/Arch/PPC64.cpp

175

minor nit: I think it would help readability if the new comment and check were added after the existing preemptible check.

176

It might be better to spell out why the address is not representable as a toc-relative value. IIUC: The symbol being referenced is a resolver function that the dynamic loader will run to determine what function should be used. Relaxing to a toc-relative reference will manifest the address of the resolver function as opposed to the address of the resolved function (which is not know until load time).

test/ELF/ppc64-toc-relax-ifunc.s

6–7

FileCheck run-step is missing.

[MaskRay]

Add a missed FileCheck

Improve a comment

[MaskRay]

ELF/Arch/PPC64.cpp

175

I think it is probably fine to have them on one line.. The first sentence makes it clear that preemptable and non-Defined symbols are excluded.

Only non-preemptable defined symbols can be relaxed.

The ifunc comment discusses the possible IRELATIVE case.

176

We have more detailed comment at Relocations.cpp:1231. I reworded the comment here but I think we probably don't have to delve too much into the details.

test/ELF/ppc64-toc-relax-ifunc.s

6–7

Thanks! Fixed.

[sfertile]

LGTM.

ELF/Arch/PPC64.cpp

175

Sure. My thoughts were that the Ifunc case is a special case of a defined symbol not being relaxable, splitting the code/comment out separately highlights that. I'm Ok with it the way it is as well.

176

Thanks, I agree: your reword covers it well enough for here.

ELF/Arch/PPC64.cpp

bool elf::tryRelaxPPC64TocIndirection(RelType type, const Relocation &rel,
   Defined *d;
   int64_t addend;
   auto *tocISB = cast<InputSectionBase>(defSym->section);
   std::tie(d, addend) =
       config->isLE ? getRelaTocSymAndAddend<ELF64LE>(tocISB, rel.addend)
                    : getRelaTocSymAndAddend<ELF64BE>(tocISB, rel.addend);
 
   // Only non-preemptable defined symbols can be relaxed.
-  if (!d || d->isPreemptible)
+  //
+  // The toc entry of a non-preemptable ifunc is relocated by R_PPC64_IRELATIVE,
+  // not representable as a toc-relative value. Exclude ifunc.
+  if (!d || d->isPreemptible || d->isGnuIFunc())
     return false;
 
   // Two instructions can materialize a 32-bit signed offset from the toc base.
   uint64_t tocRelative = d->getVA(addend) - getPPC64TocBase();
   if (!isInt<32>(tocRelative))
     return false;
 
   // Add PPC64TocOffset that will be subtracted by relocateOne().

test/ELF/ppc64-toc-relax-ifunc.s

+# REQUIRES: ppc
+
+# RUN: llvm-mc -filetype=obj -triple=powerpc64le %s -o %t.o
+# RUN: echo '.globl ifunc; .type ifunc, %gnu_indirect_function; ifunc:' | \
+# RUN:   llvm-mc -filetype=obj -triple=powerpc64le - -o %t1.o
+# RUN: ld.lld %t.o %t1.o -o %t
+
+## ifunc is a non-preemptable STT_GNU_IFUNC. Its toc entry will be
+## relocated by R_PPC64_IRELATIVE, not representable by a toc-relative value.
+## Check the toc-indirect access is not relaxed.
+
+# CHECK:      nop
+# CHECK-NEXT: ld 3, -32768(2)
+
+addis 3, 2, .toc@toc@ha
+ld 3, .toc@toc@l(3)
+
+.section .toc,"aw",@progbits
+  .quad ifunc