Re: [PATCH] Linux Kernel Markers 0.5 for Linux 2.6.17 (with probe management)

[Jeremy Fitzhardinge <jeremy@goop.org>]

Ingo Molnar wrote:
> let me qualify that: parameters must be prepared there too - but no 
> actual function call inserted. (at most a NOP inserted). The register 
> filling doesnt even have to be function-calling-convention compliant - 
> that makes the symbolic probe almost zero-impact to register 
> allocation/scheduling, the only thing it should ensure is that the 
> parameters that are annotated to be available in register, stack or 
> memory _somewhere_. (i.e. not hidden or destroyed at that point by gcc) 
> Does a simple asm() that takes read-only parameters but only adds a NOP 
> achieve this result?

Do you mean using the asm to make sure gcc puts a reference to a 
variable into the DWARF info, or some other way of encoding the value 
locations?

Hm, another problem.  If the mark is in a loop, and gcc decides to 
unroll the loop, then you'll probably only get a mark in one iteration 
of the loop (or 1/Nth of the iterations).  Or worse, assembler errors.  
The only way I can see to deal with this is to not use symbols, but put 
records in a special section.  That way, if the asm() inserting them 
gets duplicated, you'll get duplicate records in the marker section.

I guess you'd get a similar problem with markers inserted in inlined 
functions.

(How does gdb deal with breakpoints in unrolled loops?)

    J