Add support for building a pgo beam_emu

5 files changed, 212 insertions, 34 deletions

diff --git a/erts/Makefile b/erts/Makefile
index 73d8560c1e..0393ccc759 100644
--- a/erts/Makefile
+++ b/erts/Makefile
@@ -49,11 +49,6 @@ debug opt lcnt clean:
 	done
 	(cd preloaded/src && $(MAKE) ../ebin/erts.app)
 
-# ----------------------------------------------------------------------
-# These are "convenience targets", provided as shortcuts for developers
-# - don't use them in scripts or assume they will always stay like this!
-#
-
 .PHONY: smp
 smp:
 	$(V_at)for type in $(TYPES); do \
@@ -112,6 +107,11 @@ local_setup:
 		$(ERL_TOP)/bin/start_clean.script \
 		$(ERL_TOP)/bin/no_dot_erlang.script
 
+# ----------------------------------------------------------------------
+# These are "convenience targets", provided as shortcuts for developers
+# - don't use them in scripts or assume they will always stay like this!
+#
+
 # Run the configure script
 .PHONY: configure
 configure:
diff --git a/erts/aclocal.m4 b/erts/aclocal.m4
index 80bf236188..887babc13f 100644
--- a/erts/aclocal.m4
+++ b/erts/aclocal.m4
@@ -2726,6 +2726,21 @@ AC_DEFUN([LM_TRY_ENABLE_CFLAG], [
     fi
 ])
 
+AC_DEFUN([LM_CHECK_ENABLE_CFLAG], [
+    AC_MSG_CHECKING([whether $CC accepts $1...])
+    saved_CFLAGS=$CFLAGS;
+    CFLAGS="$1 $CFLAGS";
+    AC_TRY_COMPILE([],[return 0;],can_enable_flag=true,can_enable_flag=false)
+    CFLAGS=$saved_CFLAGS;
+    if test "X$can_enable_flag" = "Xtrue"; then
+        AS_VAR_SET($2, true)
+        AC_MSG_RESULT([yes])
+    else
+        AS_VAR_SET($2, false)
+        AC_MSG_RESULT([no])
+    fi
+])
+
 dnl ERL_TRY_LINK_JAVA(CLASSES, FUNCTION-BODY
 dnl                   [ACTION_IF_FOUND [, ACTION-IF-NOT-FOUND]])
 dnl Freely inspired by AC_TRY_LINK. (Maybe better to create a 
diff --git a/erts/configure.in b/erts/configure.in
index b765a8ffe4..9948e71b2d 100644
--- a/erts/configure.in
+++ b/erts/configure.in
@@ -554,6 +554,94 @@ AC_SUBST(WFLAGS)
 AC_SUBST(WERRORFLAGS)
 AC_SUBST(CFLAG_RUNTIME_LIBRARY_PATH)
 
+## Check if we can do profile guided optimization of beam_emu
+LM_CHECK_ENABLE_CFLAG([-fprofile-generate -Werror],[PROFILE_GENERATE])
+LM_CHECK_ENABLE_CFLAG([-fprofile-use -Werror],[PROFILE_USE])
+
+## Check if this is clang
+LM_CHECK_ENABLE_CFLAG([-fprofile-instr-generate -Werror],[PROFILE_INSTR_GENERATE])
+if test "X$PROFILE_INSTR_GENERATE" = "Xtrue"; then
+   # It was clang, now we also have to check if we have llvm-profdata and that
+   # we can link programs with -fprofile-instr-use
+   saved_CFLAGS=$CFLAGS;
+   CFLAGS="-fprofile-instr-generate -Werror $saved_CFLAGS"
+   AC_RUN_IFELSE([AC_LANG_PROGRAM([],[])],
+     [AC_CHECK_PROGS([LLVM_PROFDATA], [llvm-profdata])
+      AC_CHECK_PROGS([XCRUN], [xcrun])
+      if test "X$XCRUN" != "X" -a "X$LLVM_PROFDATA" = "X"; then
+        AC_MSG_CHECKING([for $XCRUN llvm-profdata])
+        if $XCRUN llvm-profdata --help 2>& AS_MESSAGE_LOG_FD >& AS_MESSAGE_LOG_FD; then
+          LLVM_PROFDATA="$XCRUN llvm-profdata"
+          AC_MSG_RESULT([yes])
+        else
+          AC_MSG_RESULT([no])
+        fi
+      fi
+      AC_SUBST(LLVM_PROFDATA)
+      if test "X$LLVM_PROFDATA" != "X"; then
+        CFLAGS="-fprofile-instr-use=default.profdata -Werror $saved_CFLAGS";
+        $LLVM_PROFDATA merge -output=default.profdata *.profraw;
+        AC_MSG_CHECKING([whether gcc accepts -fprofile-instr-use=default.profdata -Werror])
+        AC_COMPILE_IFELSE([],
+          [AC_MSG_RESULT([yes])
+           PROFILE_INSTR_USE=true],
+          [AC_MSG_RESULT([no])
+           PROFILE_INSTR_USE=false])
+        rm -f default.profdata
+      fi],
+     [])
+   rm -f *.profraw
+   CFLAGS=$saved_CFLAGS;
+fi
+
+AC_ARG_ENABLE(pgo,
+AS_HELP_STRING([--enable-pgo],
+               [build erts using PGO (profile guided optimization)]),
+[ case "$enableval" in
+    no) enable_pgo=no ;;
+    *)	enable_pgo=yes ;;
+  esac
+],enable_pgo=default)
+
+USE_PGO=false
+AC_MSG_CHECKING([whether to do PGO of erts])
+if test $enable_pgo = no; then
+  AC_MSG_RESULT([no, disabled by user])
+elif test $CROSS_COMPILING = yes; then
+  if $enable_pgo = yes; then
+    AC_MSG_ERROR(cannot use PGO when cross-compiling)
+  else
+    AC_MSG_RESULT([no, cross compiling])
+  fi
+elif test "X$host" = "Xwin32"; then
+  AC_MSG_RESULT([no, not supported in windows])
+elif test "X$PROFILE_GENERATE" = "Xtrue" -a "X$PROFILE_USE" = "Xtrue"; then
+  USE_PGO=true
+  AC_MSG_RESULT([yes, using -fprofile-generate])
+  PROFILE_COMPILER=gcc
+# check if $CC accepts -fprofile-correction, if so we can use PGO on multi-threaded files.
+  LM_CHECK_ENABLE_CFLAG([-fprofile-use -fprofile-correction -Werror],[PROFILE_CORRECTION])
+  if test "X$PROFILE_CORRECTION" = "Xtrue"; then
+    PROFILE_CORRECTION="-fprofile-correction"
+  else
+    PROFILE_CORRECTION=""
+  fi
+  AC_SUBST(PROFILE_CORRECTION)
+elif test "X$PROFILE_INSTR_GENERATE" = "Xtrue" -a "X$PROFILE_INSTR_USE" = "Xtrue"; then
+  USE_PGO=true
+  AC_MSG_RESULT([yes, using -fprofile-instr-generate])
+  PROFILE_COMPILER=clang
+else
+  if $enable_pgo = yes; then
+    AC_MSG_ERROR(cannot use PGO with this compiler)
+  else
+    AC_MSG_RESULT([no])
+  fi
+fi
+
+AC_SUBST(USE_PGO)
+AC_SUBST(PROFILE_COMPILER)
+
 AC_CHECK_SIZEOF(void *) # Needed for ARCH and smp checks below
 if test "x$ac_cv_sizeof_void_p" = x8; then
   AC_SUBST(EXTERNAL_WORD_SIZE, 64)
diff --git a/erts/emulator/Makefile.in b/erts/emulator/Makefile.in
index bc7eb72221..2b203c9ee6 100644
--- a/erts/emulator/Makefile.in
+++ b/erts/emulator/Makefile.in
@@ -63,6 +63,28 @@ ARFLAGS=rc
 OMIT_OMIT_FP=no
 TYPE_LIBS=
 
+PROFILE_COMPILER=@PROFILE_COMPILER@
+PROFILE_MARKER=
+ifeq ($(PROFILE),generate)
+PROFILE_MARKER=_pg
+else
+ifeq ($(PROFILE),use)
+PROFILE_MARKER=_pu
+endif
+endif
+
+ifeq ($(PROFILE_COMPILER), gcc)
+PROFILE_CORRECTION=@PROFILE_CORRECTION@
+PROFILE_GENERATE=-fprofile-generate
+PROFILE_USE=-fprofile-use $(PROFILE_CORRECTION)
+PROFILE_USE_DEPS=$(OBJDIR)/%_pu.gcda
+endif
+ifeq ($(PROFILE_COMPILER), clang)
+PROFILE_GENERATE=-fprofile-instr-generate
+PROFILE_USE=-fprofile-instr-use=$(OBJDIR)/default.profdata
+PROFILE_USE_DEPS=$(OBJDIR)/default.profdata
+endif
+
 DIRTY_SCHEDULER_SUPPORT=@DIRTY_SCHEDULER_SUPPORT@
 DIRTY_SCHEDULER_TEST=@DIRTY_SCHEDULER_TEST@
 
@@ -418,9 +440,20 @@ ifeq ($(TARGET), win32)
 EMULATOR_EXECUTABLE = beam$(TF_MARKER).dll
 else
 EMULATOR_EXECUTABLE = beam$(TF_MARKER)
+PROFILE_EXECUTABLE = beam.prof$(TF_MARKER)
 endif
 CS_EXECUTABLE = erl_child_setup$(TYPEMARKER)
 
+ifeq ($(PROFILE), generate)
+EMULATOR_EXECUTABLE = $(PROFILE_EXECUTABLE)
+ifeq ($(PROFILE_COMPILER), gcc)
+PROFILE_LDFLAGS = -fprofile-generate
+endif
+ifeq ($(PROFILE_COMPILER), clang)
+PROFILE_LDFLAGS = -fprofile-instr-generate
+endif
+endif
+
 # ----------------------------------------------------------------------
 
 ifeq ($(ERLANG_OSTYPE), unix)
@@ -687,16 +720,33 @@ $(OBJDIR)/beams.$(RES_EXT): $(TARGET)/beams.rc
 
 endif
 
-ifneq ($(filter tile-%,$(TARGET)),)
-$(OBJDIR)/beam_emu.o: beam/beam_emu.c
-	$(V_CC) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) \
-         $(INCLUDES) -c $< -o $@
-else
 # Usually the same as the default rule, but certain platforms (e.g. win32) mix
 # different compilers
 $(OBJDIR)/beam_emu.o: beam/beam_emu.c
 	$(V_EMU_CC) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
-endif
+
+$(OBJDIR)/%_pg.o: beam/%.c
+	$(V_CC) $(PROFILE_GENERATE) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+$(OBJDIR)/%_pu.o: beam/%.c $(PROFILE_USE_DEPS)
+	$(V_CC) $(PROFILE_USE) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
+
+$(OBJDIR)/PROFILE: $(BINDIR)/$(PROFILE_EXECUTABLE)
+	$(V_at)echo " PROFILE ${PROFILE_EXECUTABLE}"
+	$(V_at)rm -f $(OBJDIR)/erl*.profraw
+	$(V_at)set -e; LLVM_PROFILE_FILE="$(OBJDIR)/erlc-%m.profraw" \
+          ERL_FLAGS="-emu_type prof${TYPEMARKER} +S 1" $(ERLC) -DPGO \
+	  -o $(OBJDIR) test/estone_SUITE.erl > $(OBJDIR)/PROFILE_LOG
+	$(V_at)set -e; LLVM_PROFILE_FILE="$(OBJDIR)/erl-%m.profraw" \
+	  ERL_FLAGS="-emu_type prof${TYPEMARKER} +S 1" $(ERL) -pa $(OBJDIR) \
+	  -noshell -s estone_SUITE pgo -s init stop >> $(OBJDIR)/PROFILE_LOG
+	$(V_at)touch $@
+
+$(OBJDIR)/%_pu.gcda: $(OBJDIR)/PROFILE
+	$(V_at)mv $(OBJDIR)/$*_pg.gcda $@
+	$(V_at)touch $@
+
+$(OBJDIR)/default.profdata: $(OBJDIR)/PROFILE
+	$(V_LLVM_PROFDATA) merge -output $@ $(OBJDIR)/*.profraw
 
 $(OBJDIR)/%.o: beam/%.c
 	$(V_CC) $(subst -O2, $(GEN_OPT_FLGS), $(CFLAGS)) $(INCLUDES) -c $< -o $@
@@ -758,15 +808,23 @@ $(ERL_TOP)/lib/%.beam:
 
 INIT_OBJS = $(OBJDIR)/erl_main.o	$(PRELOAD_OBJ)
 
+# -fprofile-correction is needed in order to use PGO on erl_process
+# as multiple threads execute in that file.
+ifeq ($(PROFILE_CORRECTION),)
+PROFILE_OBJS = $(OBJDIR)/beam_emu.o
+RUN_OBJS = $(OBJDIR)/erl_process.o
+else
+PROFILE_OBJS = $(OBJDIR)/beam_emu.o	$(OBJDIR)/erl_process.o
+endif
+
 EMU_OBJS = \
-	$(OBJDIR)/beam_emu.o		$(OBJDIR)/beam_opcodes.o \
+	$(OBJDIR)/beam_opcodes.o \
 	$(OBJDIR)/beam_load.o		$(OBJDIR)/beam_bif_load.o \
 	$(OBJDIR)/beam_debug.o		$(OBJDIR)/beam_bp.o \
-	$(OBJDIR)/beam_catches.o \
-	$(OBJDIR)/code_ix.o \
+	$(OBJDIR)/beam_catches.o 	$(OBJDIR)/code_ix.o \
 	$(OBJDIR)/beam_ranges.o
 
-RUN_OBJS = \
+RUN_OBJS += \
 	$(OBJDIR)/erl_alloc.o		$(OBJDIR)/erl_mtrace.o \
 	$(OBJDIR)/erl_alloc_util.o	$(OBJDIR)/erl_goodfit_alloc.o \
 	$(OBJDIR)/erl_bestfit_alloc.o	$(OBJDIR)/erl_afit_alloc.o \
@@ -782,7 +840,7 @@ RUN_OBJS = \
 	$(OBJDIR)/utils.o		$(OBJDIR)/bif.o \
 	$(OBJDIR)/io.o 			$(OBJDIR)/erl_printf_term.o\
 	$(OBJDIR)/erl_debug.o		$(OBJDIR)/erl_md5.o \
-	$(OBJDIR)/erl_message.o		$(OBJDIR)/erl_process.o \
+	$(OBJDIR)/erl_message.o		 \
 	$(OBJDIR)/erl_process_dict.o	$(OBJDIR)/erl_process_lock.o \
 	$(OBJDIR)/erl_port_task.o	$(OBJDIR)/erl_arith.o \
 	$(OBJDIR)/time.o		$(OBJDIR)/erl_time_sup.o  \
@@ -923,21 +981,23 @@ ifdef HIPE_ENABLED
 EXTRA_BASE_OBJS += $(HIPE_OBJS)
 endif
 
-BASE_OBJS = $(EMU_OBJS) $(RUN_OBJS) $(OS_OBJS) $(EXTRA_BASE_OBJS) $(LTTNG_OBJS)
+BASE_OBJS = $(EMU_OBJS) $(RUN_OBJS) $(OS_OBJS) $(EXTRA_BASE_OBJS) \
+	$(LTTNG_OBJS) $(DRV_OBJS) $(NIF_OBJS)
 
-before_DTrace_OBJS =	$(BASE_OBJS) $(DRV_OBJS) $(NIF_OBJS)
+PROF_OBJS = $(patsubst %.o,%$(PROFILE_MARKER).o,$(PROFILE_OBJS)) $(BASE_OBJS)
+
+OBJS = $(PROF_OBJS)
 
-DTRACE_OBJS =
 ifdef DTRACE_ENABLED_2STEP
-DTRACE_OBJS = $(OBJDIR)/erlang_dtrace.o
-$(OBJDIR)/erlang_dtrace.o: $(before_DTrace_OBJS) $(TARGET)/erlang_dtrace.h
+# The $(PROFILE_MARKER) is placed in the object file name in order to
+# make sure we re-compile with the new object files for the profiled emulator
+OBJS += $(OBJDIR)/erlang$(PROFILE_MARKER)_dtrace.o
+$(OBJDIR)/erlang$(PROFILE_MARKER)_dtrace.o: $(PROF_OBJS) $(TARGET)/erlang_dtrace.h
 	dtrace -G -C -Ibeam \
 	  -s beam/erlang_dtrace.d \
-	  -o $@ $(before_DTrace_OBJS)
+	  -o $@ $(PROF_OBJS)
 endif
 
-OBJS = $(before_DTrace_OBJS) $(DTRACE_OBJS)
-
 $(INIT_OBJS): $(TTF_DIR)/GENERATED
 $(OBJS): $(TTF_DIR)/GENERATED
 
@@ -1029,8 +1089,8 @@ $(BINDIR)/$(EMULATOR_EXECUTABLE): $(INIT_OBJS) $(OBJS) $(DEPLIBS)
 
 else
 $(BINDIR)/$(EMULATOR_EXECUTABLE): $(INIT_OBJS) $(OBJS) $(DEPLIBS)
-	$(ld_verbose)$(PURIFY) $(LD) -o $(BINDIR)/$(EMULATOR_EXECUTABLE) \
-	$(HIPEBEAMLDFLAGS) $(LDFLAGS) $(DEXPORT) $(INIT_OBJS) $(OBJS) \
+	$(ld_verbose)$(PURIFY) $(LD) -o $@ \
+	$(HIPEBEAMLDFLAGS) $(PROFILE_LDFLAGS) $(LDFLAGS) $(DEXPORT) $(INIT_OBJS) $(OBJS) \
 	$(STATIC_NIF_LIBS) $(STATIC_DRIVER_LIBS) $(LIBS)
 
 endif
diff --git a/erts/emulator/test/estone_SUITE.erl b/erts/emulator/test/estone_SUITE.erl
index 8b336b366d..c4899967ca 100644
--- a/erts/emulator/test/estone_SUITE.erl
+++ b/erts/emulator/test/estone_SUITE.erl
@@ -20,7 +20,7 @@
 -module(estone_SUITE).
 %% Test functions
 -export([all/0, suite/0, groups/0,
-	 estone/1, estone_bench/1]).
+	 estone/1, estone_bench/1, pgo/0]).
 
 %% Internal exports for EStone tests
 -export([lists/1,
@@ -44,9 +44,9 @@
 	 links/1,lproc/1,
 	 run_micro/3,p1/1,ppp/3,macro/2,micros/0]).
 
-
--include_lib("common_test/include/ct.hrl").
+-ifndef(PGO).
 -include_lib("common_test/include/ct_event.hrl").
+-endif.
 
 %% EStone defines
 -define(TOTAL, (3000 * 1000 * 100)).   %% 300 secs
@@ -85,13 +85,28 @@ estone(Config) when is_list(Config) ->
 estone_bench(Config) ->
     DataDir = proplists:get_value(data_dir,Config),
     L = ?MODULE:macro(?MODULE:micros(),DataDir),
-    [ct_event:notify(
-       #event{name = benchmark_data, 
-	      data = [{name,proplists:get_value(title,Mark)},
-		      {value,proplists:get_value(estones,Mark)}]})
-     || Mark <- L],
+    {Total, Stones} = sum_micros(L, 0, 0),
+    notify([[{title,"ESTONES"}, {estones, Stones}] | L]),
     L.
 
+-ifndef(PGO).
+notify(Marks) ->
+    [ct_event:notify(
+       #event{name = benchmark_data,
+	      data = [{name,proplists:get_value(title, Mark)},
+		      {value,proplists:get_value(estones, Mark)}]})
+     || Mark <- Marks].
+-else.
+notify(_) ->
+    ok.
+-endif.
+
+%% The benchmarks to run in order to guide PGO (profile guided optimisation)
+pgo() ->
+    %% We run all benchmarks except the port_io as we don't want to
+    %% have to build a custom port.
+    Micros = ?MODULE:micros() -- [micro(port_io)],
+    ?MODULE:macro(Micros,[]).
 
 %%
 %% Calculate CPU speed