Messages in this thread | | | Date | Mon, 28 Apr 2008 20:47:39 -0400 | From | "Frank Ch. Eigler" <> | Subject | Re: [patch 0/2] Immediate Values - jump patching update |
| |
Hi -
On Mon, Apr 28, 2008 at 04:06:14PM -0700, H. Peter Anvin wrote: > [...] > >>The call site is created with an asm() statement as opposed to a gcc > >>function call; it is up to the logging function to take the state and > >>mangle it into whatever format it wants to; the debugging information > >>(e.g. DWARF) should tell it all it needs to know [...] > > > >that would be rather impractical as we'd force DEBUG_INFO builds on > >anyone (it's HUGE) just to do some trivial tracing. Look at the ftrace > >plugin usage model - it wants to be widely available and easy to use. > > Otherwise you're forcing everyone to take the cost of additional cache > footprint, plus optimizer interference, just because they might want to > possibly do some trivial tracing.
The intent is for the tracing not to be trivial but useful.
> DEBUG_INFO is The Right Thing for this, as it carries all the > information you may want in a well-defined format. [...]
This would require either that DWARF processing be done far after kernel build, and thus the kernel cannot be self-sufficient / introspective without user-space assistance (like firmware); or that the DWARF data extraction (and systemtap-like $context-variable code generation) be part of the kernel build itself. It *might* be workable.
At least one complication though is that in the case of markers, tracing parameter evaluation is itself conditional (and placed out of the hot path due to -freorder-blocks). With your suggested kind of assembly ("g" constraints for all the expressions), those expressions would be evaluated unconditionally, just to make them live somewhere. That unconditional evaluation can easily entail memory reads and dependent arithmetic, which could swamp the savings of eliminating the marker-style conditional branch.
- FChE
| |