Hi,

I noticed ld is failing to generate long branch stubs for aarch64 under certain conditions. After debugging the issue to try to understand what was happening, I observed that the function elfNN_aarch64_size_stubs() has a loop to compute the veneers, cache them into a special hash table and then redo layout (calling the layout_sections_again callback), doing this until no new stubs are needed. It then writes the stubs later by traversing the stub hash table with aarch64_build_one_stub() calls in elfNN_aarch64_build_stubs().

The problem is that layout_sections_again may change a symbol value (by changing its containing section's output_offset) but will not update the hash entry containing the stub destination, which was pre-computed with a now outdated value, thus generating a veneer with a bad target.

I could reproduce this on ld 2.31, but I wrote the patch for ld 2.36.

diff --git a/binutils-2.26.1/bfd/elfnn-aarch64.c b/binutils-2.26.1/bfd/elfnn-aarch64.c
--- a/binutils-2.26.1/bfd/elfnn-aarch64.c
+++ b/binutils-2.26.1/bfd/elfnn-aarch64.c
@@ -4163,6 +4163,9 @@ elfNN_aarch64_size_stubs (bfd *output_bfd,
 		    {
 		      /* The proper stub has already been created.  */
 		      free (stub_name);
+		      /* Always update this stub target since it may have
+			 changed after layout */
+		      stub_entry->target_value = sym_value + irela->r_addend;
 		      continue;
 		    }