Re: [committed, PATCH] Remove Disp16|Disp32 from 64-bit direct branches

["H.J. Lu" <hjl.tools@gmail.com>]

On Tue, May 12, 2015 at 11:18 PM, Jan Beulich <JBeulich@suse.com> wrote:
>>>> On 12.05.15 at 18:08, <hjl.tools@gmail.com> wrote:
>> On Tue, May 12, 2015 at 9:03 AM, Michael Matz <matz@suse.de> wrote:
>>> Hi,
>>>
>>> On Tue, 12 May 2015, H.J. Lu wrote:
>>>
>>>> On Tue, May 12, 2015 at 8:47 AM, Michael Matz <matz@suse.de> wrote:
>>>> > Hi,
>>>> >
>>>> > On Tue, 12 May 2015, H.J. Lu wrote:
>>>> >
>>>> >> > Actually also that one is correctly printed I think (from a hello world
>>>> >> > main, where I added a jmprel16 +0):
>>>> >> >
>>>> >> > 000000000040055c <main>:
>>>> >> >   40055c:       55                      push   %rbp
>>>> >> >   40055d:       48 89 e5                mov    %rsp,%rbp
>>>> >> >   400560:       48 83 ec 30             sub    $0x30,%rsp
>>>> >> >   400564:       c6 45 d1 00             movb   $0x0,-0x2f(%rbp)
>>>> >> >   400568:       c6 45 d0 61             movb   $0x61,-0x30(%rbp)
>>>> >> >   40056c:       48 8d 45 d0             lea    -0x30(%rbp),%rax
>>>> >> >   400570:       48 89 c2                mov    %rax,%rdx
>>>> >> >   400573:       be 44 06 40 00          mov    $0x400644,%esi
>>>> >> >   400578:       66 e9 00 00             jmpw   57c <_init-0x3ffe8c>
>>>> >> >
>>>> >> > 000000000040057c <next>:
>>>> >> >   40057c:       bf 52 06 40 00          mov    $0x400652,%edi
>>>> >> >   ...
>>>> >> >
>>>> >> > It shows that rip is going to be truncated.
>>>> >> >
>>>> >>
>>>> >> This is the same issue as
>>>> >>
>>>> >> https://sourceware.org/bugzilla/show_bug.cgi?id=18386
>>>> >>
>>>> >> On Intel processors, 0x66 prefix before direct 32-bit unconditional
>>>> >> call/jmp is ignored.  Whatever we do is wrong on AMD or Intel
>>>> >> processors.
>>>> >
>>>> > Well, in that case I'd say the correct thing to do is to _not_ do any
>>>>
>>>> This is NO correct thing to do.
>>>
>>> Well, what do you suggest?  Your change is clearly wrong as well.
>>
>> I won't call it wrong since it implies there is a right.  Given that
>>
>> 0x66 jmp/call rel32
>>
>> works on Intel processors and crashes on AMD processors.
>
> What _works_ on Intel processors is secondary here. Fact is that
> the x86-64 design came from AMD, and hence Intel CPUs doing
> things differently than AMD's is - be honest - a flaw. The more

I don't think who came first is relevant here.  What relevant are

1. AMD and Intel specs are different.
2. There is no real usage for AMD spec.
3. There is a bug report against Intel spec.

> that by analogy with 32-bit mode, an operand size prefix on
> branches ought to truncate rIP. Plus (other than my own testing
> says) you seem to suggest that this isn't even consistent on Intel
> CPUs, as you specifically say "unconditional" above and you also
> only changed those.

Please open a bug report against Jcc and I will look into it.

>> I will keep my change in unlessl someone can show a real usage of
>>
>> 066 jmp/call rel16
>>
>> on AMD processors.
>
> That's the wrong position, you have to show that the change is
> useful - I certainly can't see why you'd need the operand size

https://sourceware.org/bugzilla/show_bug.cgi?id=18386

> prefix when (on Intel CPUs) it has no effect whatsoever.
> Together with it not being generally usable (due to the vendor
> differences), I view the change as pointless _and_ breaking
> compatibility (i.e. both by themselves a reason to revert).
>


-- 
H.J.