Fix the bitrotted profile feedback build

Fix the bitrotted profile feedback build

[Andi Kleen]

The profile feedback build support had bitrotted. This patchkit fixes
it and improves it slightly by adding a new profile-fast target.

-Andi

[PATCH 1/5] Use BASIC_FLAGS for profile feedback

[Andi Kleen]

From: Andi Kleen <ak@linux.intel.com>

Use BASIC_CFLAGS instead of CFLAGS to set up the profile feedback
option in the Makefile.

This allows still overriding CFLAGS on the make command line
without disabling profile feedback.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 Makefile | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/Makefile b/Makefile
index 07ea105..a9770ac 100644
--- a/Makefile
+++ b/Makefile
@@ -1552,13 +1552,13 @@ endif
 PROFILE_DIR := $(CURDIR)
 
 ifeq ("$(PROFILE)","GEN")
-	CFLAGS += -fprofile-generate=$(PROFILE_DIR) -DNO_NORETURN=1
+	BASIC_CFLAGS += -fprofile-generate=$(PROFILE_DIR) -DNO_NORETURN=1
 	EXTLIBS += -lgcov
 	export CCACHE_DISABLE = t
 	V = 1
 else
 ifneq ("$(PROFILE)","")
-	CFLAGS += -fprofile-use=$(PROFILE_DIR) -fprofile-correction -DNO_NORETURN=1
+	BASIC_CFLAGS += -fprofile-use=$(PROFILE_DIR) -fprofile-correction -DNO_NORETURN=1
 	export CCACHE_DISABLE = t
 	V = 1
 endif
-- 
2.0.1

[PATCH 2/5] Don't define away __attribute__ on gcc

[Andi Kleen]

From: Andi Kleen <ak@linux.intel.com>

Profile feedback sets -DNO_NORETURN, which causes the compat
header file to go into a default #else block. That #else
block defines away __attribute__(). Doing so causes all
kinds of problems with the Linux and gcc system headers:
in particular it makes the xmmintrin.h headers error out,
breaking the build.

Don't define away __attribute__ when __GNUC__ is set.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 git-compat-util.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/git-compat-util.h b/git-compat-util.h
index 96f5554..01e8695 100644
--- a/git-compat-util.h
+++ b/git-compat-util.h
@@ -291,10 +291,12 @@ extern char *gitbasename(char *);
 #else
 #define NORETURN
 #define NORETURN_PTR
+#ifndef __GNUC__
 #ifndef __attribute__
 #define __attribute__(x)
 #endif
 #endif
+#endif
 
 /* The sentinel attribute is valid from gcc version 4.0 */
 #if defined(__GNUC__) && (__GNUC__ >= 4)
-- 
2.0.1

[PATCH 3/5] Run the perf test suite for profile feedback too

[Andi Kleen]

From: Andi Kleen <ak@linux.intel.com>

Open: If the perf test suite is representative enough it may
be reasonable to only run that and skip the much longer full
test suite. Thoughts?

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 Makefile | 1 +
 1 file changed, 1 insertion(+)

diff --git a/Makefile b/Makefile
index a9770ac..ba64be9 100644
--- a/Makefile
+++ b/Makefile
@@ -1647,6 +1647,7 @@ ifeq ($(filter all,$(MAKECMDGOALS)),all)
 all:: profile-clean
 	$(MAKE) PROFILE=GEN all
 	$(MAKE) PROFILE=GEN -j1 test
+	$(MAKE) PROFILE=GEN -j1 perf
 endif
 endif
 
-- 
2.0.1

[PATCH 4/5] Fix profile feedback with -jN and add profile-fast

[Andi Kleen]

From: Andi Kleen <ak@linux.intel.com>

Profile feedback always failed for me with -jN. The problem
was that there was no implicit ordering between the profile generate
stage and the profile use stage. So some objects in the later stage
would be linked with profile generate objects, and fail due
to the missing -lgcov.

This adds a new profile target that implicitely enforces the
correct ordering by using submakes. Plus a profile-install target
to also install. This is also nicer to type that PROFILE=...

Plus I always run the performance test suite now for the full
profile run.

In addition I also added a profile-fast / profile-fast-install
target the only runs the performance test suite instead of the
whole test suite. This significantly speeds up the profile build,
which was totally dominated by test suite run time. However
it may have less coverage of course.

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 INSTALL  | 14 ++++++++++++--
 Makefile | 21 +++++++++++++++++----
 2 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/INSTALL b/INSTALL
index ba01e74..6ec7a24 100644
--- a/INSTALL
+++ b/INSTALL
@@ -28,7 +28,7 @@ set up install paths (via config.mak.autogen), so you can write instead
 If you're willing to trade off (much) longer build time for a later
 faster git you can also do a profile feedback build with
 
-	$ make prefix=/usr PROFILE=BUILD all
+	$ make prefix=/usr profile
 	# make prefix=/usr PROFILE=BUILD install
 
 This will run the complete test suite as training workload and then
@@ -36,10 +36,20 @@ rebuild git with the generated profile feedback. This results in a git
 which is a few percent faster on CPU intensive workloads.  This
 may be a good tradeoff for distribution packagers.
 
+Alternatively you can run profile feedback only with the git benchmark
+suite. This runs significantly faster than the full test suite, but
+has less coverage:
+
+	$ make prefix=/usr profile-fast
+	# make prefix=/usr PROFILE=BUILD install
+
 Or if you just want to install a profile-optimized version of git into
 your home directory, you could run:
 
-	$ make PROFILE=BUILD install
+	$ make profile-install
+
+or
+	$ make profile-fast-install
 
 As a caveat: a profile-optimized build takes a *lot* longer since the
 git tree must be built twice, and in order for the profiling
diff --git a/Makefile b/Makefile
index ba64be9..a760402 100644
--- a/Makefile
+++ b/Makefile
@@ -1643,13 +1643,20 @@ SHELL = $(SHELL_PATH)
 all:: shell_compatibility_test
 
 ifeq "$(PROFILE)" "BUILD"
-ifeq ($(filter all,$(MAKECMDGOALS)),all)
-all:: profile-clean
+all:: profile
+endif
+
+profile:: profile-clean
 	$(MAKE) PROFILE=GEN all
 	$(MAKE) PROFILE=GEN -j1 test
 	$(MAKE) PROFILE=GEN -j1 perf
-endif
-endif
+	$(MAKE) PROFILE=USE all
+
+profile-fast: profile-clean
+	$(MAKE) PROFILE=GEN all
+	$(MAKE) PROFILE=GEN -j1 perf
+	$(MAKE) PROFILE=USE all
+
 
 all:: $(ALL_PROGRAMS) $(SCRIPT_LIB) $(BUILT_INS) $(OTHER_PROGRAMS) GIT-BUILD-OPTIONS
 ifneq (,$X)
@@ -2336,6 +2343,12 @@ mergetools_instdir_SQ = $(subst ','\'',$(mergetools_instdir))
 
 install_bindir_programs := $(patsubst %,%$X,$(BINDIR_PROGRAMS_NEED_X)) $(BINDIR_PROGRAMS_NO_X)
 
+profile-install: profile
+	$(MAKE) install
+
+profile-fast-install: profile-fast
+	$(MAKE) install
+
 install: all
 	$(INSTALL) -d -m 755 '$(DESTDIR_SQ)$(bindir_SQ)'
 	$(INSTALL) -d -m 755 '$(DESTDIR_SQ)$(gitexec_instdir_SQ)'
-- 
2.0.1

[PATCH 5/5] Add a little script to compare two make perf runs

[Andi Kleen]

From: Andi Kleen <ak@linux.intel.com>

Signed-off-by: Andi Kleen <ak@linux.intel.com>
---
 diff-res | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)
 create mode 100755 diff-res

diff --git a/diff-res b/diff-res
new file mode 100755
index 0000000..90d57be
--- /dev/null
+++ b/diff-res
@@ -0,0 +1,26 @@
+#!/usr/bin/python
+# compare two make perf output file
+# this should be the results only without any header
+import argparse
+import math, operator
+from collections import OrderedDict
+
+ap = argparse.ArgumentParser()
+ap.add_argument('file1', type=argparse.FileType('r'))
+ap.add_argument('file2', type=argparse.FileType('r'))
+args = ap.parse_args()
+
+cmp = (OrderedDict(), OrderedDict())
+for f, k in zip((args.file1, args.file2), cmp):
+    for j in f:
+        num = j[59:63]
+        name = j[:59]
+        k[name] = float(num)
+
+for j in cmp[0].keys():
+    print j, cmp[1][j] - cmp[0][j]
+
+def geomean(l):
+   return math.pow(reduce(operator.mul, filter(lambda x: x != 0.0, l)), 1.0 / len(l))
+
+print "geomean %.2f -> %.2f" % (geomean(cmp[0].values()), geomean(cmp[1].values()))
-- 
2.0.1

Re: [PATCH 5/5] Add a little script to compare two make perf runs

[Bert Wesarg]

Hi,

On Sat, Jul 5, 2014 at 1:43 AM, Andi Kleen <andi@firstfloor.org> wrote:
> From: Andi Kleen <ak@linux.intel.com>
>
> Signed-off-by: Andi Kleen <ak@linux.intel.com>
> ---
>  diff-res | 26 ++++++++++++++++++++++++++
>  1 file changed, 26 insertions(+)
>  create mode 100755 diff-res
>
> diff --git a/diff-res b/diff-res
> new file mode 100755
> index 0000000..90d57be
> --- /dev/null
> +++ b/diff-res
> @@ -0,0 +1,26 @@
> +#!/usr/bin/python
> +# compare two make perf output file
> +# this should be the results only without any header
> +import argparse
> +import math, operator
> +from collections import OrderedDict
> +
> +ap = argparse.ArgumentParser()
> +ap.add_argument('file1', type=argparse.FileType('r'))
> +ap.add_argument('file2', type=argparse.FileType('r'))
> +args = ap.parse_args()
> +
> +cmp = (OrderedDict(), OrderedDict())
> +for f, k in zip((args.file1, args.file2), cmp):
> +    for j in f:
> +        num = j[59:63]
> +        name = j[:59]
> +        k[name] = float(num)
> +
> +for j in cmp[0].keys():
> +    print j, cmp[1][j] - cmp[0][j]
> +
> +def geomean(l):
> +   return math.pow(reduce(operator.mul, filter(lambda x: x != 0.0, l)), 1.0 / len(l))
> +
> +print "geomean %.2f -> %.2f" % (geomean(cmp[0].values()), geomean(cmp[1].values()))

a justification why the geometric mean is used here would increase my
confident significantly.

It calculates wrong values anyway iff there are zeros in the sampling set.

Thanks.

Bert

Re: [PATCH 5/5] Add a little script to compare two make perf runs

[Andi Kleen]

> a justification why the geometric mean is used here would increase my
> confident significantly.

It's just a standard way to summarize sets of benchmarks. For example SPEC 
uses the same approach.

Anyways the script is not essential to the rest of the profile feedback
feature.  Just ignore it if it's controversal.

-Andi

Re: [PATCH 5/5] Add a little script to compare two make perf runs

[Bert Wesarg]

On Sun, Jul 6, 2014 at 6:15 PM, Andi Kleen <ak@linux.intel.com> wrote:
>> a justification why the geometric mean is used here would increase my
>> confident significantly.
>
> It's just a standard way to summarize sets of benchmarks. For example SPEC
> uses the same approach.

No, SPEC would have calculated the geometric mean of the ratios
cmp[1][j] / cmp[0][j]. And this should also only be used under the
assumption that there is a multiplicative correlation.

Bert

Re: [PATCH 3/5] Run the perf test suite for profile feedback too

[Junio C Hamano]

Andi Kleen <andi@firstfloor.org> writes:

> From: Andi Kleen <ak@linux.intel.com>
>
> Open: If the perf test suite is representative enough it may
> be reasonable to only run that and skip the much longer full
> test suite. Thoughts?

I do not think it right now is representative, nor it was meant to
become so.  The operations are those that people cared about and
tuned, and hopefully it would cover stuff actual end users care
about in the real life, though.

>
> Signed-off-by: Andi Kleen <ak@linux.intel.com>
> ---
>  Makefile | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/Makefile b/Makefile
> index a9770ac..ba64be9 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -1647,6 +1647,7 @@ ifeq ($(filter all,$(MAKECMDGOALS)),all)
>  all:: profile-clean
>  	$(MAKE) PROFILE=GEN all
>  	$(MAKE) PROFILE=GEN -j1 test
> +	$(MAKE) PROFILE=GEN -j1 perf
>  endif
>  endif

Re: [PATCH 3/5] Run the perf test suite for profile feedback too

[Andi Kleen]

On Mon, Jul 07, 2014 at 02:06:57PM -0700, Junio C Hamano wrote:
> Andi Kleen <andi@firstfloor.org> writes:
> 
> > From: Andi Kleen <ak@linux.intel.com>
> >
> > Open: If the perf test suite is representative enough it may
> > be reasonable to only run that and skip the much longer full
> > test suite. Thoughts?
> 
> I do not think it right now is representative, nor it was meant to
> become so.  The operations are those that people cared about and
> tuned, and hopefully it would cover stuff actual end users care
> about in the real life, though.

I ended up answering the question by creating two separate 
makefile targets in the next patch.

profile to run the full test suite and profile-fast to run
only the performance test. profile-fast as the name implies
is a lot faster to build, and fast enough that it's not 
annoying.

I'll remove the "Open"

-Andi