[PPC64] Consider localentry offset when computing branch distance

[MaskRay]

Summary

We don't take localentry offset into account, and thus may fail to
create long branch when the gap is just a few bytes smaller than 2^25.

relocation R_PPC64_REL24 out of range: 33554432 is not in [-33554432, 33554431]
relocation R_PPC64_REL24 out of range: 33554436 is not in [-33554432, 33554431]

Fix that.

[MaskRay]

This occurred extremely rarely (the distance must be exactly 33554432 or 33554436!) ... probably less than 0.01% or even 0.001%, but thanks to our numerous targets for exposing the bug 😊

[MaskRay]

.

[ruiu]

LGTM

Nice. Even if the probability of a failure for each relocation is very low, we have huge number of relocations. Apparently the latter overwhelms the former!

ELF/Arch/PPC64.cpp

bool PPC64::needsThunk(RelExpr Expr, RelType Type, const InputFile *File,
 
   // If a symbol is a weak undefined and we are compiling an executable
   // it doesn't need a range-extending thunk since it can't be called.
   if (S.isUndefWeak() && !Config->Shared)
     return false;
 
   // If the offset exceeds the range of the branch type then it will need
   // a range-extending thunk.
-  return !inBranchRange(Type, BranchAddr, S.getVA());
+  // See the comment in getRelocTargetVA() about R_PPC64_CALL.
+  return !inBranchRange(Type, BranchAddr,
+                        S.getVA() +
+                            getPPC64GlobalEntryToLocalEntryOffset(S.StOther));
 }
 
 bool PPC64::inBranchRange(RelType Type, uint64_t Src, uint64_t Dst) const {
   int64_t Offset = Dst - Src;
   if (Type == R_PPC64_REL14)
     return isInt<16>(Offset);
   if (Type == R_PPC64_REL24)
     return isInt<26>(Offset);

test/ELF/ppc64-long-branch-localentry-offset.s

+# REQUIRES: ppc
+
+# RUN: llvm-mc -filetype=obj -triple=powerpc64le-unknown-linux %s -o %t.o
+# RUN: ld.lld --no-toc-optimize %t.o -o %t
+# RUN: llvm-nm %t | FileCheck %s
+
+# CHECK-DAG: 0000000010010000 t __long_branch_callee
+# CHECK-DAG: 0000000010010010 T _start
+# CHECK-DAG: 0000000012010008 t callee
+
+# The bl instruction jumps to the local entry. The distance requires a long branch stub:
+# localentry(callee) - _start = 0x12010008+8 - 0x10010010 = 0x2000000
+
+# We used to compute globalentry(callee) - _start and caused a "R_PPC64_REL24
+# out of range" error because we didn't create the stub.
+
+.globl _start
+_start:
+  bl callee
+
+.space 0x1fffff4
+
+callee:
+.Lgep0:
+  addis 2, 12, .TOC.-.Lgep0@ha
+  addi 2, 2, .TOC.-.Lgep0@l
+.Llep0:
+  .localentry callee, .Llep0-.Lgep0
+  blr