Prathamesh Kulkarni 7f114df909 Enable time profile function reordering with AutoFDO. ...

The patch enables time profile based reordering with AutoFDO with
-fauto-profile -fprofile-reorder-functions, by mapping timestamps obtained from perf
into node->tp_first_run.

The rationale for doing this is:
(1) GCC already implements time-profile function reordering with PGO, the patch enables
it with AutoFDO.
(2) While time profile ordering is primarly meant for optimizing startup time,
we've also observed good effects on code-locality for large internal workloads.
(3) Possibly useful for function reordering when accurate profile annotation is
hard with AutoFDO -- For eg, if branch samples are missing (due to absence of
LBR like structure).

On AutoFDO tools side, a corresponding patch extends gcov to emit 64-bit perf timestamp that
records first execution of function, which loosely corresponds to PGO's time_profile counter.
The timestamp is stored adjacent to head field in toplevel function info.

On GCC side, this patch makes the following changes:

(1) Changes to auto-profile pass:
The patch adds a new field timestamp to function_instance,
and populates it in read_function_instance.
It maintains a new timestamp_info_map from timestamp -> <name, tp_first_run>,
which maps timestamps sorted in ascending order to (1..N), so lowest ordered
timestamp is mapped to 1 and so on. The rationale for this is that
timestamps are 64-bit integers, and we don't need the full 64-bit range
for ordering by tp_first_run.

During annotation, the timestamp associated with function_instance is looked up
in timestamp_info_map, and corresponding mapped value is assigned
to node->tp_first_run.

Dhruv's sourcefile tracking patch already handles LTO privatized symbols.
The patch adds a workaround for mismatched/empty filenames, which should go away
when the issues with AutoFDO tools dwarf parsing are resolved.

(2) Param to disable profile driven opts.
The patch adds param auto-profile-reorder-only which only enables time-profile reordering with
AutoFDO:
(a) Useful as a debugging aid to isolate regression to either function reordering or profile driven opts.
(b) As a stopgap measure to avoid regressions with AutoFDO profile driven opts.
(c) Possibly useful for architectures which do not support branch sampling.

gcc/ChangeLog:
	* auto-profile.cc: (string_table::filenames): New method.
	(function_instance::timestamp_): New member.
	(function_instance::timestamp): New accessor for timestamp_ member.
	(function_instance::set_timestamp): New function.
	(function_instance::prop_timestamp): Likewise.
	(function_instance::prop_timestamp_1): Likewise.
	(function_instance::function_instance): Initialize timestamp_ to 0.
	(function_instance::read_function_instance): Adjust prototype by
	replacing head_count with toplevel param with default value true, and
	stream in head_count and timestamp values from gcov file.
	(autofdo::timestamp_info_map): New std::map.
	(autofdo_source_profile::get_function_instance_by_decl): New argument
	filename with default value NULL.
	(autofdo_source_profile::read): Populate timestamp_info_map and
	propagate timestamp to inlined instances from toplevel function.
	(afdo_annotate_cfg): Assign node->tp_first_run based on
	timestamp_info_map and bail out of annotation if
	param_auto_profile_reorder_only is enabled.
	* params.opt: New param auto-profile-reorder-only.

Signed-off-by: Prathamesh Kulkarni <prathameshk@nvidia.com>

2026-01-14 09:41:03 +00:00

66 KiB

Raw Permalink Blame History

View Raw