porting reversible on arm/mips

Sean Chen sean.chen1234@gmail.com
Fri Dec 11 16:42:00 GMT 2009

On Fri, Dec 11, 2009 at 9:24 PM, paawan oza <paawan1982@yahoo.com> wrote:
> I am not sure how arm can be drastically slower than x86 ! considering arm 32 bit.
> but at the first point, if in some way if prec arch level stuff and abis related framework are in place,
> then optimization on the same may lead us to get faster recording like cache implementation and so on.
> But I am not sure of any specific reason why on arm it could be very slower, having the same conf as x86.
> Regards,
> Oza.
> ----- Original Message ----
> From: Jakob Engblom <jakob@virtutech.com>
> To: Sean Chen <sean.chen1234@gmail.com>; Michael Snyder <msnyder@vmware.com>
> Cc: paawan oza <paawan1982@yahoo.com>; Hui Zhu <teawater@gmail.com>; gdb@sourceware.org
> Sent: Fri, December 11, 2009 1:45:48 PM
> Subject: RE: porting reversible on arm/mips
>> I was interested in the porting on ARM. But later I found that the
>> performance impact on ARM might damage the usage of process record. In
>> my experiment, reversible debugging is about 20000x slower, which
>> might be endurable on the modern computer. However, ARM target is tens
>> of times (or even more if we consider the memory) slower than PC. So
>> recording instructions will be very slow, about thousands of
>> instructions per second.
> I just must pitch in and say that it depends on the simulator.
> An advantage to using a full simulator is that you simplify the system and no
> longer have to care about OS calls: the OS is just part of the context you save
> and reverse.  So the overhead actually goes down compared to native prec.  I
> think a reversible ARM simulator can be made to run within a factor of ten of
> native speed, easily.
> Best regards,
> /jakob
> _______________________________________________________
> Jakob Engblom, PhD, Technical Marketing Manager
> Virtutech                   Direct: +46 8 690 07 47
> Drottningholmsvägen 22      Mobile: +46 709 242 646
> 11243 Stockholm             Web:    www.virtutech.com
> Sweden
> ________________________________________________________

I think ARM does be at least tens of times slower than x86. Image a PC
and a phone, ARM architecture has to sacrifice the performance to gain
the advantage of power and size saving.

Here I assume that process record is more than 20000x slower on both
x86 and ARM. I agree with you and believe this can be improved a lot
in the future.

Best Regards,
Sean Chen

More information about the Gdb mailing list