diff mbox

[1/5] Add gcc-auto-profile script

Message ID 1463848586-19259-2-git-send-email-andi@firstfloor.org
State New
Headers show

Commit Message

Andi Kleen May 21, 2016, 4:36 p.m. UTC
From: Andi Kleen <ak@linux.intel.com>

Using autofdo is currently something difficult. It requires using the
model specific branches taken event, which differs on different CPUs.
The example shown in the manual requires a special patched version of
perf that is non standard, and also will likely not work everywhere.

This patch adds a new gcc-auto-profile script that figures out the
correct event and runs perf.

This is needed to actually make use of autofdo in a generic way
in the build system and in the test suite.

Since maintaining the script would be somewhat tedious (needs changes
every time a new CPU comes out) I auto generated it from the online
Intel event database. The script to do that is in contrib and can be
rerun.

Right now there is no test if perf works in configure. This
would vary depending on the build and target system, and since
it currently doesn't work in virtualization and needs uptodate
kernel it may often fail in common distribution build setups.

So far the script is not installed.

gcc/:
2016-05-21  Andi Kleen  <ak@linux.intel.com>

	* doc/invoke.texi: Document gcc-auto-profile
	* config/i386/gcc-auto-profile: New file.

contrib/:

2016-05-21  Andi Kleen  <ak@linux.intel.com>

	* gen_autofdo_event.py: New file to regenerate
	gcc-auto-profile.
---
 contrib/gen_autofdo_event.py     | 155 +++++++++++++++++++++++++++++++++++++++
 gcc/config/i386/gcc-auto-profile |  70 ++++++++++++++++++
 gcc/doc/invoke.texi              |  31 ++++++--
 3 files changed, 251 insertions(+), 5 deletions(-)
 create mode 100755 contrib/gen_autofdo_event.py
 create mode 100755 gcc/config/i386/gcc-auto-profile

Comments

Bernhard Reutner-Fischer May 21, 2016, 5:58 p.m. UTC | #1
On May 21, 2016 6:36:22 PM GMT+02:00, Andi Kleen <andi@firstfloor.org> wrote:
>From: Andi Kleen <ak@linux.intel.com>

>+if [ "$1" = "--kernel" ] ; then
>+  FLAGS=k
>+  shift
>+fi
>+if [ "$1" == "--all" ] ; then

== is legacy, s/==/=/

>+  FLAGS=uk
>+  shift
>+fi
>+
>+if ! grep -q Intel /proc/cpuinfo ] ; then
>+  echo >&2 "Only Intel CPUs supported"
>+  exit 1
>+fi
>+
>+if grep -q hypervisor /proc/cpuinfo ; then
>+  echo >&2 "Warning: branch profiling may not be functional in VMs"
>+fi

grep && echo would do but OK.

>+
>+case `egrep -q "^cpu family\s*: 6" /proc/cpuinfo &&
>+  egrep "^model\s*:" /proc/cpuinfo | head -1` in'''

head and tail both require -n nowadays (in fact since susv2, IIRC), so please head -n1

thanks,
Jan Hubicka May 30, 2016, 12:39 a.m. UTC | #2
Andi,
thanks a lot for working on the auto-fdo bootstrap. It is badly needed to
have some coverage for this feature.  I don't think I can approve the
build machinery changes.

> From: Andi Kleen <ak@linux.intel.com>
> 
> Using autofdo is currently something difficult. It requires using the
> model specific branches taken event, which differs on different CPUs.
> The example shown in the manual requires a special patched version of
> perf that is non standard, and also will likely not work everywhere.
> 
> This patch adds a new gcc-auto-profile script that figures out the
> correct event and runs perf.
> 
> This is needed to actually make use of autofdo in a generic way
> in the build system and in the test suite.
> 
> Since maintaining the script would be somewhat tedious (needs changes
> every time a new CPU comes out) I auto generated it from the online
> Intel event database. The script to do that is in contrib and can be
> rerun.

I guess we need to figure out how to ship this to users.  At the moment
the script will tell you to rebuild when it meets new CPU, but it reffers
to gcc sources which is not the best place.

Also the script should insteall when it is documented in invoke.texi

What happens when you ahve wrong perf?

Honza
Andi Kleen May 30, 2016, 3:11 a.m. UTC | #3
On Mon, May 30, 2016 at 02:39:06AM +0200, Jan Hubicka wrote:
> > 
> > Since maintaining the script would be somewhat tedious (needs changes
> > every time a new CPU comes out) I auto generated it from the online
> > Intel event database. The script to do that is in contrib and can be
> > rerun.
> 
> I guess we need to figure out how to ship this to users.  At the moment
> the script will tell you to rebuild when it meets new CPU, but it reffers
> to gcc sources which is not the best place.

I don't have a better solution.

> 
> Also the script should insteall when it is documented in invoke.texi

I wasn't sure if it should be installed or not, but opted for not. 
For now I can remove the documentation.

> 
> What happens when you ahve wrong perf?

You mean with the manual example? It will just error out,
because it requires Google's special patched version.
Also the event name is not always the same, so it's only working
on some CPUs.

With my script perf should work, but you need LBR support in the kernel for
using -b. LBRs are a model specific feature, and this needs:

- The kernel is new enough and knows about your CPU model
(dmesg | grep Perf.*LBR outputs something)

- LBRs are typically not virtualized[1], so it usually does not work
in VMs

When the LBR support is not there the profiling run will not
work at all.

-Andi

[1] There are some patches for KVM, but they are not mainline so far.
diff mbox

Patch

diff --git a/contrib/gen_autofdo_event.py b/contrib/gen_autofdo_event.py
new file mode 100755
index 0000000..907430d
--- /dev/null
+++ b/contrib/gen_autofdo_event.py
@@ -0,0 +1,155 @@ 
+#!/usr/bin/python
+# Generate Intel taken branches Linux perf event script for autofdo profiling.
+
+# Copyright (C) 2016 Free Software Foundation, Inc.
+#
+# GCC is free software; you can redistribute it and/or modify it under
+# the terms of the GNU General Public License as published by the Free
+# Software Foundation; either version 3, or (at your option) any later
+# version.
+#
+# GCC is distributed in the hope that it will be useful, but WITHOUT ANY
+# WARRANTY; without even the implied warranty of MERCHANTABILITY or
+# FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+# for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with GCC; see the file COPYING3.  If not see
+# <http://www.gnu.org/licenses/>.  */
+
+# Run it with perf record -b -e EVENT program ...
+# The Linux Kernel needs to support the PMU of the current CPU, and
+# It will likely not work in VMs.
+# Add --all to print for all cpus, otherwise for current cpu.
+# Add --script to generate shell script to run correct event.
+#
+# Requires internet (https) access. This may require setting up a proxy
+# with export https_proxy=...
+#
+import urllib2
+import sys
+import json
+import argparse
+import collections
+
+baseurl = "https://download.01.org/perfmon"
+
+target_events = (u'BR_INST_RETIRED.NEAR_TAKEN',
+                 u'BR_INST_EXEC.TAKEN',
+                 u'BR_INST_RETIRED.TAKEN_JCC',
+                 u'BR_INST_TYPE_RETIRED.COND_TAKEN')
+
+ap = argparse.ArgumentParser()
+ap.add_argument('--all', '-a', help='Print for all CPUs', action='store_true')
+ap.add_argument('--script', help='Generate shell script', action='store_true')
+args = ap.parse_args()
+
+eventmap = collections.defaultdict(list)
+
+def get_cpu_str():
+    with open('/proc/cpuinfo', 'r') as c:
+        vendor, fam, model = None, None, None
+        for j in c:
+            n = j.split()
+            if n[0] == 'vendor_id':
+                vendor = n[2]
+            elif n[0] == 'model' and n[1] == ':':
+                model = int(n[2])
+            elif n[0] == 'cpu' and n[1] == 'family':
+                fam = int(n[3])
+            if vendor and fam and model:
+                return "%s-%d-%X" % (vendor, fam, model), model
+    return None, None
+
+def find_event(eventurl, model):
+    print >>sys.stderr, "Downloading", eventurl
+    u = urllib2.urlopen(eventurl)
+    events = json.loads(u.read())
+    u.close()
+
+    found = 0
+    for j in events:
+        if j[u'EventName'] in target_events:
+            event = "cpu/event=%s,umask=%s/" % (j[u'EventCode'], j[u'UMask'])
+            if u'PEBS' in j and j[u'PEBS'] > 0:
+                event += "p"
+            if args.script:
+                eventmap[event].append(model)
+            else:
+                print j[u'EventName'], "event for model", model, "is", event
+            found += 1
+    return found
+
+if not args.all:
+    cpu, model = get_cpu_str()
+    if not cpu:
+        sys.exit("Unknown CPU type")
+
+url = baseurl + "/mapfile.csv"
+print >>sys.stderr, "Downloading", url
+u = urllib2.urlopen(url)
+found = 0
+cpufound = 0
+for j in u:
+    n = j.rstrip().split(',')
+    if len(n) >= 4 and (args.all or n[0] == cpu) and n[3] == "core":
+        if args.all:
+            vendor, fam, model = n[0].split("-")
+            model = int(model, 16)
+        cpufound += 1
+        found += find_event(baseurl + n[2], model)
+u.close()
+
+if args.script:
+    print '''#!/bin/sh
+# Profile workload for gcc profile feedback (autofdo) using Linux perf.
+# Auto generated. To regenerate for new CPUs run
+# contrib/gen_autofdo_event.py --shell --all in gcc source
+
+# usages:
+# gcc-auto-profile program             (profile program and children)
+# gcc-auto-profile -a sleep X          (profile all for X secs, may need root)
+# gcc-auto-profile -p PID sleep X      (profile PID)
+# gcc-auto-profile --kernel -a sleep X (profile kernel)
+# gcc-auto-profile --all -a sleep X    (profile kernel and user space)
+
+# Identify branches taken event for CPU.
+#
+
+FLAGS=u
+
+if [ "$1" = "--kernel" ] ; then
+  FLAGS=k
+  shift
+fi
+if [ "$1" == "--all" ] ; then
+  FLAGS=uk
+  shift
+fi
+
+if ! grep -q Intel /proc/cpuinfo ] ; then
+  echo >&2 "Only Intel CPUs supported"
+  exit 1
+fi
+
+if grep -q hypervisor /proc/cpuinfo ; then
+  echo >&2 "Warning: branch profiling may not be functional in VMs"
+fi
+
+case `egrep -q "^cpu family\s*: 6" /proc/cpuinfo &&
+  egrep "^model\s*:" /proc/cpuinfo | head -1` in'''
+    for event, mod in eventmap.iteritems():
+        for m in mod[:-1]:
+            print "model*:\ %s|\\" % m
+        print 'model*:\ %s) E="%s$FLAGS" ;;' % (mod[-1], event)
+    print '''*)
+echo >&2 "Unknown CPU. Run contrib/gen_autofdo_event.py --all --script to update script."
+	exit 1 ;;'''
+    print "esac"
+    print 'exec perf record -e $E -b "$@"'
+
+if cpufound == 0 and not args.all:
+    sys.exit('CPU %s not found' % cpu)
+
+if found == 0:
+    sys.exit('Branch event not found')
diff --git a/gcc/config/i386/gcc-auto-profile b/gcc/config/i386/gcc-auto-profile
new file mode 100755
index 0000000..c6712b2
--- /dev/null
+++ b/gcc/config/i386/gcc-auto-profile
@@ -0,0 +1,70 @@ 
+#!/bin/sh
+# profile workload for gcc profile feedback (autofdo) using Linux perf
+# auto generated. to regenerate for new CPUs run
+# contrib/gen_autofdo_event.py --shell --all in gcc source
+
+# usages:
+# gcc-auto-profile program             (profile program and children)
+# gcc-auto-profile -a sleep X          (profile all for X secs, may need root)
+# gcc-auto-profile -p PID sleep X      (profile PID)
+# gcc-auto-profile --kernel -a sleep X (profile kernel)
+# gcc-auto-profile --all -a sleep X    (profile kernel and user space)
+
+# identify branches taken event for CPU
+#
+
+FLAGS=u
+
+if [ "$1" = "--kernel" ] ; then
+  FLAGS=k
+  shift
+fi
+if [ "$1" == "--all" ] ; then
+  FLAGS=uk
+  shift
+fi
+
+if ! grep -q Intel /proc/cpuinfo ] ; then
+  echo >&2 "Only Intel CPUs supported"
+  exit 1
+fi
+
+if grep -q hypervisor /proc/cpuinfo ; then
+  echo >&2 "Warning: branch profiling may not be functional in VMs"
+fi
+
+case `egrep -q "^cpu family\s*: 6" /proc/cpuinfo &&
+  egrep "^model\s*:" /proc/cpuinfo | head -1` in
+model*:\ 55|\
+model*:\ 77|\
+model*:\ 76) E="cpu/event=0xC4,umask=0xFE/p$FLAGS" ;;
+model*:\ 42|\
+model*:\ 45|\
+model*:\ 58|\
+model*:\ 62|\
+model*:\ 60|\
+model*:\ 69|\
+model*:\ 70|\
+model*:\ 63|\
+model*:\ 61|\
+model*:\ 71|\
+model*:\ 86|\
+model*:\ 78|\
+model*:\ 94) E="cpu/event=0xC4,umask=0x20/p$FLAGS" ;;
+model*:\ 46|\
+model*:\ 30|\
+model*:\ 31|\
+model*:\ 26|\
+model*:\ 47|\
+model*:\ 37|\
+model*:\ 44) E="cpu/event=0x88,umask=0x40/p$FLAGS" ;;
+model*:\ 28|\
+model*:\ 38|\
+model*:\ 39|\
+model*:\ 54|\
+model*:\ 53) E="cpu/event=0x88,umask=0x41/p$FLAGS" ;;
+*)
+echo >&2 "Unknown CPU. Run contrib/gen_autofdo_event.py --all --script to update script."
+	exit 1 ;;
+esac
+exec perf record -e $E -b "$@"
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index 926e1e6..5488877 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -8276,13 +8276,34 @@  which are generally profitable only with profile feedback available:
 If omitted, it defaults to @file{fbdata.afdo} in the current directory.
 
 Producing an AutoFDO profile data file requires running your program
-with the @command{perf} utility on a supported GNU/Linux target system.
+with the @command{gcc-auto-profile} utility on a supported GNU/Linux target system. @command{gcc-auto-profile} calls the @command{perf} utility.
+It also requires Last-Branch-Record support, which typically requires
+a new enough kernel not running virtualized.
+@command{gcc-auto-profile} accepts the same arguments as @command{perf record}.
 For more information, see @uref{https://perf.wiki.kernel.org/}.
 
-E.g.
 @smallexample
-perf record -e br_inst_retired:near_taken -b -o perf.data \
-    -- your_program
+gcc-auto-profile your_program
+@end smallexample
+
+On larger programs the resulting perf.data file may be very large.
+In this case it can be better to reduce the sampling rate.
+Collect samples every million taken branches:
+
+@smallexample
+gcc-auto-profile -c 1000000 program
+@end smallexample
+
+Or only profile representative run intervals of the program:
+
+@smallexample
+gcc-auto-profile -p PID-OF-PROGRAM sleep 5
+@end smallexample
+
+Profile complete system for 10 seconds (may require root)
+
+@smallexample
+gcc-auto-profile -a sleep 10
 @end smallexample
 
 Then use the @command{create_gcov} tool to convert the raw profile data
@@ -8293,7 +8314,7 @@  See @uref{https://github.com/google/autofdo}.
 E.g.
 @smallexample
 create_gcov --binary=your_program.unstripped --profile=perf.data \
-    --gcov=profile.afdo
+    --gcov=profile.afdo -gcov_version 1
 @end smallexample
 @end table