PATCH: Implement Intel Transactional Synchronization Extensions

[Jan Beulich]

>>> "H.J. Lu" <hjl.tools@gmail.com> 02/21/12 7:17 PM >>>
On Tue, Feb 21, 2012 at 12:56 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>> On 08.02.12 at 19:25, "H.J. Lu" <hongjiu.lu@intel.com> wrote:
>>>+      if (i.prefix[HLE_PREFIX] == XACQUIRE_PREFIX_OPCODE)
>>
>> With HLE_PREFIX == REP_PREFIX and XACQUIRE_PREFIX_OPCODE
>> == REPNE_PREFIX_OPCODE, how is this distinguished from the repne
>> case? Oh, I see, the *caller* checks i.have_hle - that's pretty bad to
>> be done this way imo.
>
>Any suggestions are welcome.

At least move the check into the function, so that an eventual second caller
doesn't have to ope code the check a second time.

>> Wouldn't it make sense to insert missing LOCK prefixes silently
>> (optionally verbosely) rather than complaining about their absence?
>> (I'd also question the value of displaying both prefixes by default
>> in the disassembly - this just clutters things rather than helping
>> with anything.)
>
>In some cases, it is very useful to know/control exactly how many
>prefixes are generated.

But there's no alternative here - the code is wrong without the prefixes.
And as written, for those who do care an optional warning could be
implemented.

>> Further, while the patch deals with CMPXCHG8B, for some reason
>> it leaves out CMPXCHG16B (perhaps because the instruction, oddly
>> enough, is considered an SSE3 one).
>>
>> Finally, albeit consistent with what is documented, I'm surprised that
>> MOV opcodes 0xA2 and 0xA3 aren't allowed with XRELEASE - is that
>> really the case?
>>
>
>Yes, they are implemented according to TSX spec.

Assuming you meant "are not", did you check that this isn't just an omission
in the spec?

Jan