Re: step-by-step

[Andrew Cagney <cagney at redhat dot com>]

Yong,

I've been waiting for you to notice this ;-)  As Roland points out this 
is the PPC sequence:

    load word and reserve
    modify
    store word conditional on still being reserved

that is used to implement locking.    It isn't possible to single step 
this sequence since the step-instruction trap will clear the reservation 
resulting in the store-word-conditional always failing.

Something more complicated is needed - a step object implemented using 
frysk.proc's instruction-step and breakpoint primitives, switching 
between each as needed.  This more complicated object could also be 
expanded to implement operations such as step-until-branch-or-out-of-range.

Andrew

Yong Zheng wrote:
> On Fri, 2006-10-13 at 10:18 -0400, Andrew Cagney wrote:
>   
> > fstep is more than a toy.  Having been reduced to using an instruction 
> > tracing simulator to track down an SMP races, this forms the foundation 
> > of an extremely wicked tool!
> >
> >     
>
> Unfortunately, fstep doesnot work well for both 64-bit and 32-bit
> applications on our PPC64 box(with 4 processors). When we use fstep to
> step one applications, the application will run into never-ending loop
> like the following:
>
> 0xfe78cb8 (steps: 60000)
>         0xfe78cb8       lwarx   r9,0,r31
>         0xfe78cbc       cmpw    r9,r11
>         0xfe78cc0       bne-    0xfe78ccc
>         0xfe78cc4       stwcx.  r0,0,r31
>         0xfe78cc8       bne-    0xfe78cb8
> 0xfe78cb8 (steps: 70000)
>         0xfe78cb8       lwarx   r9,0,r31
>         0xfe78cbc       cmpw    r9,r11
>         0xfe78cc0       bne-    0xfe78ccc
>         0xfe78cc4       stwcx.  r0,0,r31
>         0xfe78cc8       bne-    0xfe78cb8
> 0xfe78cb8 (steps: 80000)
>         0xfe78cb8       lwarx   r9,0,r31
>         0xfe78cbc       cmpw    r9,r11
>         0xfe78cc0       bne-    0xfe78ccc
>         0xfe78cc4       stwcx.  r0,0,r31
>         0xfe78cc8       bne-    0xfe78cb8
>
> The instruction at 0xfe78cb8 is located in libc-2.4.so. The process of
> fstep may break down the lwarx/stwcx operations on PPC64.
>
>   
> > I created http://sourceware.org/bugzilla/show_bug.cgi?id=3364
> >     
> Fstep works well on our X86/X86_64 box(of course only with one
> processor). So I add our mail addresses in CC list. 
>
> Best regards
> Yong Zheng
>
>