[Bug gas/25012] New: pushq/popq %gs/%fs in .code64 now unsupported

[christian.ehrhardt at canonical dot com]

https://sourceware.org/bugzilla/show_bug.cgi?id=25012

            Bug ID: 25012
           Summary: pushq/popq %gs/%fs in .code64 now unsupported
           Product: binutils
           Version: 2.33 (HEAD)
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: gas
          Assignee: unassigned at sourceware dot org
          Reporter: christian.ehrhardt at canonical dot com
  Target Milestone: ---

Hi,
I have found that with recent binutils ipxe was no more compiling.
It broke on a section doing essentially (simplified) doing:

.code64
pushq %gs
pushq %fs
popq %gs
popq %fs

The fail I got was like:
push.S:2: Error: unsupported instruction `push'
push.S:3: Error: unsupported instruction `push'
push.S:4: Error: unsupported instruction `pop'
push.S:5: Error: unsupported instruction `pop'

There is more about that in Ubuntu bug
https://bugs.launchpad.net/ubuntu/+source/binutils/+bug/1843394

But I'm reaching out to you here as I was wondering if it is a bug in gas
(unlikely, but it is a change in behavior for sure).

I found that even older 2.32 are ok.
2.32-7ubuntu4 - ok
2.32.51.20190905-0ubuntu1 - fails.

I was wondering and hoping you could help me on three things:
1. was the change to no more build that assembly intentional and could one
maybe link to a commit for that?
2. if the change was intentional is there a best practice how the code should
be modified to work again?
3. I was checking the behavior on push/pop %fs/gs and it confused me so I
wonder if the current state is correct (see attachment and below).

For #3 some details:
Summary D vs E:
- no suffix
  => works equally in both releases
  => same opcodes in all .code segments
- suffix "w"
  => works equally in both releases
  => opcodes in .code32/.code64 differ from .code16 (660f..)
  => .code16 matches the non-suffix opcodes (0f..)
- suffix "l"
  => failures in Disco, works in Eoan
  => .code16 opcodes match the non-.code16 of the "w" suffix (660f..)
  => .code32/.code64 are back to the base opcode (0f..)
  => If I remove the failing .code64 from disco then .code16/.code32 is the
same as in Eoan
- suffix "q"
  => different failures in Disco and Eoan
  => in Disco .code16/.code32 fails
  => in Disco .code64 generates the basic opcode (0f..)
  => in Eoan all three .code segments fail
I wonder if all of that is correct?

P.S. I can break this up into multiple bugs if you tell me which section you
want to discuss separately.
P.P.S. I have asked upstream IPXE for feedback for (a pure gut feeling) change,
but there was no response yet
(http://lists.ipxe.org/pipermail/ipxe-devel/2019-September/006763.html)

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