On 08.09.2012, at 14:30, Blue Swirl <blauwir...@gmail.com> wrote:
> On Sat, Sep 8, 2012 at 12:13 PM, Alexander Graf <ag...@suse.de> wrote: >> >> >> On 08.09.2012, at 12:16, Blue Swirl <blauwir...@gmail.com> wrote: >> >>> On Sat, Sep 8, 2012 at 9:28 AM, Alexander Graf <ag...@suse.de> wrote: >>>> >>>> >>>> On 08.09.2012, at 10:06, Blue Swirl <blauwir...@gmail.com> wrote: >>>> >>>>> On Thu, Sep 6, 2012 at 8:44 AM, Avi Kivity <a...@redhat.com> wrote: >>>>>> On 09/05/2012 10:04 PM, Blue Swirl wrote: >>>>>>> >>>>>>> Reinventing a disassembler for ever growing x86 assembly is >>>>>>> no fun. >>>>>> >>>>>> We can try linking to a disassembler library. I use udis86 to >>>>>> disassemble instructions in kvm tracepoints >>>>>> (http://udis86.git.sourceforge.net/git/gitweb.cgi?p=udis86/udis86;a=shortlog), >>>>>> it's maintained but not heavily so. >>>>> >>>>> I think commonality with KVM would be preferred. The library looks >>>>> neat and based on changelog, more actively developed than BSD DDB. >>>>> >>>>>> >>>>>> Of course for non-x86 we'd need to continue using binutils; this is >>>>>> about copying code vs. libraries, not about licensing. >>>>> >>>>> For most architectures, pre-GPLv3 binutils is good enough since the >>>>> instruction set does not change anymore. Maybe only PPC and Sparc64 >>>>> still change besides x86. New CPUs types more recent than 2007 will >>>>> have problems. >>>> >>>> Alternatively we could try to run the disassembler in a different process, >>>> right? >>> >>> For qemu.log this would be doable and even improve performance since >>> only binary data would be transferred. >>> >>> But for monitor disassembly command x/i it may be too clumsy. >> >> Why would it be clumsy? We'd have to make sure we are communicating >> synchronously with the daemon, but apart from that it shouldn't be too >> different from the log, no? > > The log file should be written as binary which the disassembly tool > could read. The log file contains a lot more information than just the diassembly. You get cpu state dumps, tcg op dumps, and above all there are a very big amount of log writing bits throughout the code for debug purposes that write plain ascii. Do you think it's worth creating a 2-step process out of this? I was more thinking along the lines of a second process that qemu would spawn when log file is active / on monitor command which then would get binary opcodes voa a pipe and returns ascii disassembly that qemu cam use again. That second program could even be built as part of our build process, right? We would then be able to pull in gplv3 code from binutils into that program, but keep it out of the main project. Alex
