[Bug gold/20529] New: Gold [powerpc] - relaxing loop never ends

[shenhan at google dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=20529

            Bug ID: 20529
           Summary: Gold [powerpc] - relaxing loop never ends
           Product: binutils
           Version: 2.28 (HEAD)
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: gold
          Assignee: ccoutant at gmail dot com
          Reporter: shenhan at google dot com
                CC: ian at airs dot com
  Target Milestone: ---

Here is one particular case that causes gold PPC linker trapped in a endless
relaxing loop. (I do not provide a reproducible case because the objects
involved are proprietary).

Here is the case - 

--------------
|            |
| 0x1b9446f0 |
|            | <- stub_table
|   ...      |
|------------|
|            |
|            |
|------------|
| 0x1c71e3d0 |-----
|------------|    |
|            |    |
|            |    |
|            |    |  jump delta = 0x2000000
|            |    | (>= max_branch_offset, which is 0x2000000)
|            |    |
|            |    |
|------------|    |
| 0x1e71e3d0 | <--|
|----------- |
|            |
|            |

So a long branch stub is added, which is 16-byte, and this pushes the section
forward, thus get to the following:

--------------
|            |
| 0x1b9446f0 |
|            | <- stub_table
| a new stub |
|   ...      |
|------------|
|            |
|            |
|------------|
| 0x1c71e3e0 |-----
|------------|    |
|            |    |
|            |    |
|            |    |  jump delta < 0x2000000
|            |    |    no stub needed
|            |    |
|            |    |
|------------|    |
| 0x1e71e3d0 | <--|  (the jump target is *NOT* pushed forward? why?)
|----------- |
|            |
|            |

Now for this particular branch_info, we do not add a branch stub. So stub table
shrinks by 16 bytes, and we get back to the first graph. Thus causes the
indefinite loop.

The solution is to only allow growth of stub tables after a few relax
iterations.

-- 
You are receiving this mail because:
You are on the CC list for the bug.