[JITLink][PowerPC] Fix incorrect assertion of addend for R_PPC64_REL24

[lkail]

Summary

There is case that R_PPC64_REL24 with non-zero addend. The assertion is incorrectly triggered in such situation.

[arphaman]

@lkail the test you precommited (https://reviews.llvm.org/rG4468103bf68ce09820d0e31fa7c725004b5f2ca4) for this change is failing on Darwin -

https://green.lab.llvm.org/green/job/clang-stage1-cmake-RA-incremental/37181/

# CHECK: Assertion `Addend == 0 && "Addend is expected to be 0 for a function call"' failed
         ^
<stdin>:1:1: note: scanning from here
Assertion failed: (Addend == 0 && "Addend is expected to be 0 for a function call"), function addSingleRelocation, file /Users/buildslave/jenkins/workspace/clang-stage1-cmake-RA-incremental/llvm-project/llvm/lib/ExecutionEngine/JITLink/ELF_ppc64.cpp, line 349.
^

Input file: <stdin>
Check file: /Users/buildslave/jenkins/workspace/clang-stage1-cmake-RA-incremental/llvm-project/llvm/test/ExecutionEngine/JITLink/ppc64/ppc64-rel24-non-zero-addend.test

-dump-input=help explains the following input dump.

Input was:
<<<<<<
         1: Assertion failed: (Addend == 0 && "Addend is expected to be 0 for a function call"), function addSingleRelocation, file /Users/buildslave/jenkins/workspace/clang-stage1-cmake-RA-incremental/llvm-project/llvm/lib/ExecutionEngine/JITLink/ELF_ppc64.cpp, line 349. 
check:4     X~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ error: no match found

I will revert the precommitted test.

[lkail]

@arphaman Thanks for reverting this.

[lkail]

This comment was removed by lkail.

[lhames]

@lkail Were you able to determine what caused the issue on Darwin? I haven't had a chance to check it out yet.

[lkail]

In D158708#4623239, @lhames wrote:

@lkail Were you able to determine what caused the issue on Darwin? I haven't had a chance to check it out yet.

The assertion message on Darwin is

Assertion failed: (Addend == 0 && "Addend is expected to be 0 for a function call"), function addSingleRelocation,

While the culprit commit checks

Assertion `Addend == 0 && "Addend is expected to be 0 for a function call"' failed

[lkail]

Ping.

[lhames]

Apologies for the delay on this. LGTM!

llvm/include/llvm/ExecutionEngine/JITLink/ppc64.h

public:
 
   bool visitEdge(LinkGraph &G, Block *B, Edge &E) {
     bool isExternal = E.getTarget().isExternal();
     Edge::Kind K = E.getKind();
     if (K == ppc64::RequestCall) {
       if (isExternal) {
         E.setKind(ppc64::CallBranchDeltaRestoreTOC);
         this->StubKind = LongBranchSaveR2;
+        // FIXME: We assume the addend to the external target is zero. It's
+        // quite unusual that the addend of an external target to be non-zero as
+        // if we have known the layout of the external object.
         E.setTarget(this->getEntryForTarget(G, E.getTarget()));
-        // We previously set branching to local entry. Now reverse that
-        // operation.
+        // Addend to the stub is zero.
         E.setAddend(0);
       } else
         // TODO: There are cases a local function call need a call stub.
         // 1. Caller uses TOC, the callee doesn't, need a r2 save stub.
         // 2. Caller doesn't use TOC, the callee does, need a r12 setup stub.
         // 3. Branching target is out of range.
         E.setKind(ppc64::CallBranchDelta);
       return true;

llvm/lib/ExecutionEngine/JITLink/ELF_ppc64.cpp

Error addSingleRelocation(const typename ELFT::Rela &Rel,
     case ELF::R_PPC64_REL32:
       Kind = ppc64::Delta32;
       break;
     case ELF::R_PPC64_REL24_NOTOC:
       Kind = ppc64::RequestCallNoTOC;
       break;
     case ELF::R_PPC64_REL24:
       Kind = ppc64::RequestCall;
-      assert(Addend == 0 && "Addend is expected to be 0 for a function call");
-      // We assume branching to local entry, will reverse the addend if not.
-      Addend = ELF::decodePPC64LocalEntryOffset((*ObjSymbol)->st_other);
+      // Determining a target is external or not is deferred in PostPrunePass.
+      // We assume branching to local entry by default, since in PostPrunePass,
+      // we don't have any context to determine LocalEntryOffset. If it finally
+      // turns out to be an external call, we'll have a stub for the external
+      // target, the target of this edge will be the stub and its addend will be
+      // set 0.
+      Addend += ELF::decodePPC64LocalEntryOffset((*ObjSymbol)->st_other);
       break;
     case ELF::R_PPC64_REL64:
       Kind = ppc64::Delta64;
       break;
     case ELF::R_PPC64_PCREL34:
       Kind = ppc64::Delta34;
       break;
     case ELF::R_PPC64_GOT_TLSGD16_HA:

llvm/test/ExecutionEngine/JITLink/ppc64/ppc64-rel24-non-zero-addend.test

-# REQUIRES: asserts
 # RUN: yaml2obj %S/Inputs/rel24-non-zero-addend.yaml -o %t.o
-# RUN: not --crash llvm-jitlink -noexec %t.o 2>&1 | FileCheck %s
-# CHECK: Addend == 0 && "Addend is expected to be 0 for a function call"
-#
+# RUN: llvm-jitlink -noexec %t.o 2>&1
 # The object is generated from llvm/test/ExecutionEngine/MCJIT/test-global-ctors.ll,
 # containing an R_PPC64_REL24 whose addend is not zero.