Re: [Valgrind-users] Unhandled instruction bytes (C++, x86 on amd64)

[Tom H. <to...@co...>]

On 16/03/18 17:16, John Reiser wrote:
> On 03/16/2018 1439 UTC, Zack Weinberg wrote:
>> On Fri, Mar 16, 2018 at 1323 UTC, John Reiser wrote:
>>>
>>> The code 0x67 0xE8 0x7D 0xFF is "addr32 callq ...".  Valgrind is 
>>> complaining
>>> about the 'addr32' opcode prefix.  The compiler for locale_init.cc did a
>>> not-so-good job.  The addr32 prefix is a total waste; remove it.
>>
>> This _could_ be an alignment hack.  I don't see any of these prefixes
>> in my copy of libstdc++ but the compiler could plausibly be trying to
>> make the instruction right after the call be on a tidy address
>> boundary.
> 
> The 0x67 was at 0x48cc63d which is (-3 mod 16), so the following 
> instruction
> is at (+3 mod 16) [1 for the 0x67, 1 for the 0xE8, 4 for the 32-bit 
> displacement].
> Alignment was not close to any auspicious boundary.

We figured out the cause on IRC and it's all down to this:

https://sourceware.org/ml/binutils/2016-05/msg00322.html

Basically code compiled with -fno-plt causes the linker to do
it when it needs to patch an indirect call without changing the
length of the instruction.

Tom

-- 
Tom Hughes (to...@co...)
http://compton.nu/