|
From: Jim C. <jim...@gm...> - 2008-01-02 20:19:18
|
hi valgrind-devs, I know your slideware suggested its bad to go looking for something to write, rather than looking for what info would help solve a problem & how valgrind could help.. but.. how about writing IR to disk-cache, and reloading it when code is unchanged. I suppose reuse might depend upon tool used, options (forex suppressions), etc. How much (or little) time is spent by the core doing the D&R ? How much is code IR vs tool IR, and is caching them separately have any value ? (sets upper bound of any benefit gained by caching IR) I suppose your points for D&R vs C&A already devalue this line of inquiry, but at this point its only a muse prompted by the no-libs question that J responded to earlier today. tia |
|
From: Nicholas N. <nj...@cs...> - 2008-01-02 20:57:04
|
On Wed, 2 Jan 2008, Jim Cromie wrote: > I know your slideware suggested its bad to go looking > for something to write, rather than looking for what info > would help solve a problem & how valgrind could help.. > > but.. > > how about writing IR to disk-cache, and reloading it > when code is unchanged. I suppose reuse might depend > upon tool used, options (forex suppressions), etc. > > How much (or little) time is spent by the core doing the D&R ? > How much is code IR vs tool IR, and is caching them separately > have any value ? > (sets upper bound of any benefit gained by caching IR) It varies greatly on the tool and the program. Eg. short-running programs vs. long-running programs. > I suppose your points for D&R vs C&A already devalue this > line of inquiry, but at this point its only a muse prompted by > the no-libs question that J responded to earlier today. The PLDI paper also points out that, roughly speaking, we think performance is an overrated virtue. Caching translations would make things more complicated, too. I believe the Pin people tried this, there's a paper about it somewhere that you could probably find. Nick |