2010-03-16 Segher Boessenkool <segher@kernel.crashing.org>

[deliverable/binutils-gdb.git] / gprof / TODO

- add support for prof file format so that prof files can be displayed
  at the line-level (this is useful for the uprofile tool under DEC's
  OSF/1)
- take a hard look at --file-ordering (broken) and --function-ordering

+ documentation
+ optimize bfd_find_nearest_line_num() (or replace by different interface)
+ cleanup _bfd_ecoff_find_nearest_line_num() fixes & description
+ ensure "cc -pg" produces good files under OSF/1 v3.0
+ make sure gprof works together with OSF/1 v3.0's profiling libraries
+ implement symtab_parse(); modify sym_lookup() to consider addr_high
+ change gprof.c to collect lists, then invoke symtab_parse() for
  each list
+ Questions:
	o is -c (--static-call-graph) useful at all?  i can't see
	  how; if it were deleted, gprof would be completely machine
	  independent => yup, it is
	o are (long) option names appropriate?
	o -k (--exclude-arc) cannot be implemented with getopt();
	  is new syntax (-k from/to) acceptable?  If not, how to
	  fix it?
	o in the FSF output, the call-graph index now prints
	  the filename of static functions in parentheses; e.g.,
	  static function foo() that is defined in file bar.c
	  would be printed as:

			[4] foo (bar.c)

	  is this acceptable?  should it be done only optionally?
	o symbols with addresses that map back to a different
	  name are suppressed (happens with labels, for example);
	  is this acceptable?  should it be done only optionally?
+ generalize to allow arbitrary histograms (not just time histograms)
+ basic-block information currently replaces all symbols created from
  the core because of an ugly ordering conflict---for now, the current
  solution works, but something cleaner is desirable ==> cleaned up,
  but it's slower now
+ convert to very new file format (back to trivial format, that is :)
+ replace "dummy.h" for Alpha (if there is any use to it)
+ add support for execution time profiling at a basic-block level
+ fix filename-off-by-one bug for Alpha (see ~/tmp/d.[ch])---no longer
  relevant
+ "-pg -a" doesn't work as expected because mcleanup() will overwrite
  the file generated by __bb_exit_func() (or vice versa)
+ first basic-block of fac() seems to get credited to last basic-block
  of previous function => bug in basic_blocks.c
+ flat profile should provide automatic scaling for per-call times because
  otherwise they'll always be zero on a fast machine with tons of small
  functions
+ make "-a" imply to retain line number info (without actually generating
  the debugging information (unless -g is specified)---no, this is a
  bad idea, because it is not clear what level of debugging info should
  be requested (e.g., -g vs. -g3); leaving it up to the user seems best
+ add long options support (or at least use getopt instead of ad-hoc
  implementation)
+ split into files according to abstract objects that are manipulated
+ replace sccsid by rcsid & add "end of ..." to every .c file
+ use DBG() everywhere
+ fix spacing (" ," -> "," etc.)
+ use DEFUNs everywhere
+ make compile cleanly with -Wall
+ "gcc -pg -O2" doesn't work on tecc.c unless -fno-omit-frame-pointer is
  specified; find out why
+ make things portable (prototypes, const, etc.)
+ if NEW_GMON_OUT is not defined, have a flag that will allow to
  read new gmon.out style files.  The idea being that everyone
  will use the new format for basic-block style profiling but
  the old format for regular gpprofiling