Changes since release V3.8:

We have made two major additions: value profiling and upcall
handlers.

Dcpid now has builtin support for value profiling, which
collects data by periodically interpreting instructions and
saving, for each instruction, either its common result
values (dcpid -vprof) or the PCs of instructions that might
cause a replay trap on EV6 (dcpid -vreplay).

The two examples above, -vprof and -vreplay, are hardwired
into dcpid.  However, DCPI's value-profiling support (see
man dcpivprofiler) also allows you to specify what values to
capture, how to process those values before merging them
into the profile files, and how to format the values for
printing.  A value profiler is a dynamically loadable shared
library written by the user.  It encapsulates the code to
perform all user-specified processing.

Two value profilers are included in the DCPI distribution:
vp-classic.so collects the same information as the "classic
value profiling" hardwired into dcpid (-vprof), and
vp-addr.so collects the effective addresses of memory
operands in load and store instructions.

The second major addition is experimental support for
performing user-level upcalls that deliver profiling
interrupts directly to user-level handlers in profiled
applications (see man dcpiupcalls).  Kernel support for
upcalls is automatically included with the DCPI driver
kernel module.  A preliminary shared library named
libuvprof.so that performs limited user-level value
profiling is also included with the DCPI release.  Support
for upcalls is not complete and has not been extensively
tested.  It works on simple programs, but fails, e.g., for
programs that use exceptions in signal handlers.  See the
man page for many CAVEATS.

There were some minor changes: the default per-cpu hash
table size was doubled to 512KB to reduce overhead when
collecting profile-me data, and the dcpid man page
now includes 21264a (EV67) aggregate events.

In the release notes for V3.8, we neglected to report that
dcpid creates a "fake" PALcode image file (suitable for use
with dcpilist, dcpiprof, etc.) by reading memory.  The
PALcode image is stored in the profile directory.  The
disassembly of PALcode-specfic instructions is currently
very limited.


V3.9 is the last release of DCPI that SRC will be
publishing.  Future DCPI releases will be managed by product
groups in Spitbrook.  The SRC DCPI web pages will have prominent
link to the new external DCPI web site, which is
   http://www.unix.digital.com/dcpi
This site is expected to be up soon.  The mailing lists
   dcpi, dcpi-request, and dcpi-bugs
now forward to the product groups.

Thanks to all of our users who have helped make DCPI a success 
and everyone who has contributed (see the authors on the man
pages).

Happy performance tuning,
 -Carl, Mark, & Shun-Tak 
 


Changes since release V3.4:

- Supports ProfileMe counters on EV67 (Alpha 21264a) cpus!
  (See "man dcpiprofileme" and "man dcpid")

- Dcpiprof and dcpilist require "-help" to list descriptions of their
  many flags.

- Dcpid's command syntax for selecting events has changed (see man dcpid).
  The new syntax supports multiplexing of profile-me modes with
  aggregate (i.e., non-profileme) modes.

- A new tool, dcpitopcounts, lists instructions with the highest event
  counts (see man dcpitopcounts).

- Dcpid supports a new -nice option to adjust its scheduling priority,
  and a new -gc option to automatically delete old epochs.

- Support for Tru64 V4.0c and earlier has been dropped.  Use DCPI-V3.4
  for V4.0 through V4.0c.

Known problems:

- When collecting data using multiple "-slot"s, each slot must have at
  least one event that occurs at least every 2^30 cycles, e.g.,
  retires, cycles, issues, or profileme mode.  Dcpid does not
  check this requirement

Changes since release V2.12:

- Pcount driver is dynamically loadable.

- Dcpid loads pcount at startup and unloads it at exit.

- A new underlying profile file format is used to store sample data
  for all events in a single file per image.  Backwards compatibility
  has been retained so that all tools support profiles collected with
  previous versions of DCPI.

- Dcpid supports a new -nice option to adjust its scheduling priority,
  and a new -gc option to automatically delete old epochs.

- Fixed a rare assertion failure in instruction-level analysis tools.

- This version works with Digital Unix T5.0 646.2 (a.k.a. BL17)

- Support for Digital Unix 3.2 has been dropped.  Use DCPI-V2.11 and
  pcount-13.3 build 15 for machines running 3.2.

Changes since release V2.11:

- Added support for Alpha 21066 and other 21064-compatible cpus.

Changes since release V2.10:

- Adds support for compressed kernel modules.

- dcpiprof listings now contain information about which named
  procedure is near an anonymous procedure to help the user in finding
  anonymous procedures.

Changes since release V2.5:

- DCPI releases V2.6 through V2.9 each had incompatibilities
  with various versions of Digital Unix.  We believe that these
  problems have been corrected.
  
- New loader for 4.0c and higher.
  Dynamically linked images can be profiled on machines running
  DU V3.2 or V4.0*, but not V5.0*.  The INSTALL script tests the
  DCPI loader before installing it as the default system loader.

- Minor improvements to dcpid image identification.
  Image maps for common DU V4.0c and V4.0d binaries are compiled into dcpid.
  Fixed identification of large images started before dcpid.

- Major update to dcpi2pix to generate more accurate data.

- Added support for new 21264 instructions.