[libomptarget][amdgpu] fix memory leak in find_metadata

[ThorBl]

Summary

Call to elf_memory had no corresponding elf_end calls. Added elf_end calls to deallocate the memory at every return point.

[ThorBl]

Received suggestion from Jonathan Chesterfield to use unique_ptrs.

[ronlieb]

@FabioLuporini might find this interesting, memory leak.

[ThorBl]

Previous fix was incomplete. Two exit points were overlooked (did not contain calls to elf_end).

[ThorBl]

Regarding Jon's suggestion to replace elf_end() by std::unique_ptr:

The libelf.h header file contains only a forward declaration of the Elf struct. The declaration itself is hidden in libelfP.h, which is not accessible. This causes an 'incomplete type' error in unique_ptr's dtor. Without a doubt, using unique_ptr would be the preferred solution. However, since it is not easily possible to make the declaration of the struct available, I would stick with the calls to elf_end().

[JonChesterfield]

Missing context in the diff (see https://llvm.org/docs/Phabricator.html) and missing reviewers.

Would be better to put an object on the stack the calls elf_end in the destructor than to wire it up to all the paths.

I think @jhuber6 has a patch in flight which will collide with this

[jhuber6]

Also changed in this https://reviews.llvm.org/D131401

[ThorBl]

https://reviews.llvm.org/D131401 already addresses the problem.

openmp/libomptarget/plugins/amdgpu/impl/system.cpp

 
   Elf *e = elf_memory(static_cast<char *>(binary), binSize);
   if (elf_kind(e) != ELF_K_ELF) {
+    elf_end(e);
     return failure;
   }
 
   size_t numpHdrs;
   if (elf_getphdrnum(e, &numpHdrs) != 0) {
+    elf_end(e);
     return failure;
   }
 

         address name = (address)&note[1];
 
         if (note->n_type == 7 || note->n_type == 8) {
+          elf_end(e);
           return failure;
         } else if (note->n_type == 10 /* NT_AMD_AMDGPU_HSA_METADATA */ &&
                    note->n_namesz == sizeof "AMD" &&
                    !memcmp(name, "AMD", note->n_namesz)) {
           // code object v2 uses yaml metadata, no longer supported
+          elf_end(e);
           return failure;
         } else if (note->n_type == 32 /* NT_AMDGPU_METADATA */ &&
                    note->n_namesz == sizeof "AMDGPU" &&

           unsigned char *metadata_start = (unsigned char *)ptr + offset;
           unsigned char *metadata_end =
               metadata_start + core::alignUp(note->n_descsz, 4);
+          elf_end(e);
           return {metadata_start, metadata_end};
         }
         ptr += sizeof(*note) + core::alignUp(note->n_namesz, sizeof(int)) +

       }
     }
   }
-
+  elf_end(e);
   return failure;
 }