Fix the bitrotted profile feedback support.
Changes to v1: - Remove obsolete comment - Remove controversal diff script.
[PATCH 1/4] Use BASIC_FLAGS for profile feedback
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=$(PROFILE_DIR) -fprofile-correction -DNO_NORETURN=1 export CCACHE_DISABLE = t V = 1 endif
-- 2.0.1
[PATCH 2/4] Don't define away __attribute__ on gcc
From: Andi Kleen <ak@linux.intel.com>
Profile feedback sets -DNO_NORETURN, which causes the compat header file to go into a default #else block. That #else block defines away __attribute__(). Doing so causes all kinds of problems with the Linux and gcc system headers: in particular it makes the xmmintrin.h headers error out, breaking the build.
Don't define away __attribute__ when __GNUC__ is set.
Signed-off-by: Andi Kleen <ak@linux.intel.com> --- git-compat-util.h | 2 ++ 1 file changed, 2 insertions(+)
diff --git a/git-compat-util.h b/git-compat-util.h index 96f5554..01e8695 100644 --- a/git-compat-util.h +++ b/git-compat-util.h @@ -291,10 +291,12 @@ extern char *gitbasename(char *); #else #define NORETURN #define NORETURN_PTR +#ifndef __GNUC__ #ifndef __attribute__ #define __attribute__(x) #endif #endif +#endif /* The sentinel attribute is valid from gcc version 4.0 */ #if defined(__GNUC__) && (__GNUC__ >= 4)
-- 2.0.1
[PATCH 3/4] Run the perf test suite for profile feedback too
From: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Andi Kleen <ak@linux.intel.com> --- Makefile | 1 + 1 file changed, 1 insertion(+)
diff --git a/Makefile b/Makefile index a9770ac..ba64be9 100644 --- a/Makefile +++ b/Makefile @@ -1647,6 +1647,7 @@ ifeq ($(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 4/4] Fix profile feedback with -jN and add profile-fast
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