Merge pull request #727

.gitignore

@@ -24,6 +24,8 @@ pyston_grwl_dbg
 pyston_nosync
 pyston_gcc
 pyston_release_gcc
+pyston_release_gcc_pgo
+pyston_release_gcc_pgo_instrumented
 pystontmp*/
 /*_unittest
 

CMakeLists.txt

@@ -29,8 +29,11 @@ option(ENABLE_OPROFILE "enable oprofile support" OFF)
 option(ENABLE_SELF_HOST "use pyston to test pyston" OFF)
 option(ENABLE_VALGRIND "pyston valgrind support" OFF)
 
+option(ENABLE_PGO "enable -fprofile-generate/-fprofile-use" OFF)
+option(ENABLE_LTO "enable -flto" OFF)
+
 # automatically use ccache if found
-if(ENABLE_CCACHE)
+if(ENABLE_CCACHE AND NOT ENABLE_PGO)
   find_program(CCACHE ccache)
   if(CCACHE)
     message(STATUS "found ccache ${CCACHE}")
@@ -46,8 +49,11 @@ if(ENABLE_GOLD)
     message(STATUS "found the gold linker ${GOLD_LINKER}")
     set(CMAKE_LINKER "${GOLD_LINKER}")
     set(CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS} -B${CMAKE_SOURCE_DIR}/tools/build_system")
+    if(NOT ENABLE_PGO)
+      # let pgo determine the ordering
       set(CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS} -Wl,--section-ordering-file,${CMAKE_CURRENT_SOURCE_DIR}/section_ordering.txt")
     endif()
+  endif()
 endif()
 
 
@@ -59,7 +65,7 @@ else()
 endif()
 
 # initial clang flags (set here so they're used when building llvm)
-set(CLANG_FLAGS "-Qunused-arguments -fcolor-diagnostics" CACHE STRING "Clang specific C and CXX flags")
+set(CLANG_FLAGS "-Qunused-arguments -fcolor-diagnostics -Wno-profile-instr-unprofiled" CACHE STRING "Clang specific C and CXX flags")
 if("${CMAKE_C_COMPILER_ID}" STREQUAL "Clang")
   set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${CLANG_FLAGS}")
 endif()
@@ -67,6 +73,31 @@ if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${CLANG_FLAGS}")
 endif()
 
+if(ENABLE_LTO)
+  set(LTO_FLAGS "-flto")
+endif()
+
+if(ENABLE_PGO)
+  if(NOT "${PROFILE_DIR}" STREQUAL "")
+    set(PROFILE_STATE "${PROFILE_STATE}=${PROFILE_DIR}")
+  endif()
+  set(PGO_FLAGS "-fprofile-${PROFILE_STATE} -fprofile-correction")
+endif()
+
+macro(ADD_PROFILE_FLAGS)
+  set(PROFILE_FLAGS "${PGO_FLAGS} ${LTO_FLAGS}")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${PROFILE_FLAGS}")
+endmacro()
+
+# used in src/runtime/inline/CMakeLists.txt to remove the flags we added above
+macro(REMOVE_PROFILE_FLAGS)
+  string(REPLACE "${PROFILE_FLAGS}" " " CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
+  string(REPLACE "${PROFILE_FLAGS}" " " CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}")
+  string(REPLACE "${PROFILE_FLAGS}" " " CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE}")
+  string(REPLACE "${PROFILE_FLAGS}" " " CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS}")
+  string(REPLACE "${PROFILE_FLAGS}" " " CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS}")
+endmacro()
+
 # llvm disable debug info unless ENABLE_LLVM_DEBUG is ON
 if(${CMAKE_BUILD_TYPE} STREQUAL "Debug" AND NOT ENABLE_LLVM_DEBUG)
   set(CMAKE_CXX_FLAGS_DEBUG "-O3" CACHE STRING "" FORCE)
@@ -200,6 +231,7 @@ endif()
 
 if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-long-long -Wno-aggregate-return -Wno-inline -Wno-redundant-decls -Wno-strict-overflow")
+  set(CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS} -Wl,--no-as-needed") # without this, we don't seem to get jemalloc on gcc builds
 endif()
 
 add_definitions(${LLVM_DEFINITIONS})
@@ -213,6 +245,8 @@ find_package(LibLZMA REQUIRED)
 link_directories(${CMAKE_BINARY_DIR}/libunwind/lib)
 link_directories(${LLVM_LIBRARY_DIRS})
 
+ADD_PROFILE_FLAGS()
+
 add_subdirectory(lib_pyston)
 add_subdirectory(from_cpython)
 add_subdirectory(src)

Makefile

@@ -65,6 +65,8 @@ CMAKE_DIR_DBG := $(BUILD_DIR)/Debug
 CMAKE_DIR_RELEASE := $(BUILD_DIR)/Release
 CMAKE_DIR_GCC := $(BUILD_DIR)/Debug-gcc
 CMAKE_DIR_RELEASE_GCC := $(BUILD_DIR)/Release-gcc
+CMAKE_DIR_RELEASE_GCC_PGO := $(BUILD_DIR)/Release-gcc-pgo
+CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED := $(BUILD_DIR)/Release-gcc-pgo-instrumented
 CMAKE_SETUP_DBG := $(CMAKE_DIR_DBG)/build.ninja
 CMAKE_SETUP_RELEASE := $(CMAKE_DIR_RELEASE)/build.ninja
 
@@ -893,10 +895,10 @@ CMAKE_SHAREDMODS := sharedmods ext_pyston
 .PHONY: pyston_dbg pyston_release
 pyston_dbg: $(CMAKE_SETUP_DBG)
 	$(NINJA) -C $(CMAKE_DIR_DBG) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
-	ln -sf $(CMAKE_DIR_DBG)/pyston pyston_dbg
+	ln -sf $(CMAKE_DIR_DBG)/pyston $@
 pyston_release: $(CMAKE_SETUP_RELEASE)
 	$(NINJA) -C $(CMAKE_DIR_RELEASE) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
-	ln -sf $(CMAKE_DIR_RELEASE)/pyston pyston_release
+	ln -sf $(CMAKE_DIR_RELEASE)/pyston $@
 
 CMAKE_SETUP_GCC := $(CMAKE_DIR_GCC)/build.ninja
 $(CMAKE_SETUP_GCC):
@@ -906,7 +908,7 @@ $(CMAKE_SETUP_GCC):
 .PHONY: pyston_gcc
 pyston_gcc: $(CMAKE_SETUP_GCC)
 	$(NINJA) -C $(CMAKE_DIR_GCC) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
-	ln -sf $(CMAKE_DIR_GCC)/pyston pyston_gcc
+	ln -sf $(CMAKE_DIR_GCC)/pyston $@
 
 CMAKE_SETUP_RELEASE_GCC := $(CMAKE_DIR_RELEASE_GCC)/build.ninja
 $(CMAKE_SETUP_RELEASE_GCC):
@@ -916,7 +918,35 @@ $(CMAKE_SETUP_RELEASE_GCC):
 .PHONY: pyston_release_gcc
 pyston_release_gcc: $(CMAKE_SETUP_RELEASE_GCC)
 	$(NINJA) -C $(CMAKE_DIR_RELEASE_GCC) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
-	ln -sf $(CMAKE_DIR_RELEASE_GCC)/pyston pyston_release_gcc
+	ln -sf $(CMAKE_DIR_RELEASE_GCC)/pyston $@
+
+
+# GCC PGO build
+CMAKE_SETUP_RELEASE_GCC_PGO := $(CMAKE_DIR_RELEASE_GCC_PGO)/build.ninja
+$(CMAKE_SETUP_RELEASE_GCC_PGO):
+	@$(MAKE) cmake_check
+	@mkdir -p $(CMAKE_DIR_RELEASE_GCC_PGO)
+	cd $(CMAKE_DIR_RELEASE_GCC_PGO); CC='$(GCC)' CXX='$(GPP)' cmake -GNinja $(SRC_DIR) -DCMAKE_BUILD_TYPE=Release $(CMAKE_VALGRIND) -DENABLE_PGO=ON -DPROFILE_STATE=use
+.PHONY: pyston_release_gcc_pgo
+pyston_release_gcc_pgo: $(CMAKE_SETUP_RELEASE_GCC_PGO) $(CMAKE_DIR_RELEASE_GCC_PGO)/.trained
+	$(NINJA) -C $(CMAKE_DIR_RELEASE_GCC_PGO) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
+	ln -sf $(CMAKE_DIR_RELEASE_GCC_PGO)/pyston $@
+
+CMAKE_SETUP_RELEASE_GCC_PGO_INSTRUMENTED := $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED)/build.ninja
+$(CMAKE_SETUP_RELEASE_GCC_PGO_INSTRUMENTED):
+	@$(MAKE) cmake_check
+	@mkdir -p $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED)
+	cd $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED); CC='$(GCC)' CXX='$(GPP)' cmake -GNinja $(SRC_DIR) -DCMAKE_BUILD_TYPE=Release $(CMAKE_VALGRIND) -DENABLE_PGO=ON -DPROFILE_STATE=generate -DPROFILE_DIR=$(CMAKE_DIR_RELEASE_GCC_PGO)
+
+.PHONY: pyston_release_gcc_pgo_instrumented
+pyston_release_gcc_pgo_instrumented: $(CMAKE_SETUP_RELEASE_GCC_PGO_INSTRUMENTED)
+	$(NINJA) -C $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED) -f $(CMAKE_SETUP_RELEASE_GCC_PGO_INSTRUMENTED) pyston copy_stdlib copy_libpyston $(CMAKE_SHAREDMODS) ext_cpython $(NINJAFLAGS)
+	ln -sf $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED)/pyston $@
+
+$(CMAKE_DIR_RELEASE_GCC_PGO)/.trained: pyston_release_gcc_pgo_instrumented
+	@echo "Training pgo"
+	mkdir -p $(CMAKE_DIR_RELEASE_GCC_PGO)
+	(cd $(CMAKE_DIR_RELEASE_GCC_PGO_INSTRUMENTED) && ./pyston $(SRC_DIR)/minibenchmarks/combined.py && ./pyston $(SRC_DIR)/minibenchmarks/combined.py ) && touch $(CMAKE_DIR_RELEASE_GCC_PGO)/.trained
 
 .PHONY: format check_format
 format: $(CMAKE_SETUP_RELEASE)

src/runtime/inline/CMakeLists.txt

@@ -36,6 +36,10 @@
 
 # adapted from https://github.com/shadow/shadow/blob/master/cmake/LLVMTools.cmake
 
+# don't do pgo here, as our gcc command line flags are incompatible with clang, and we
+# don't want to bake the instrumentation into the resulting bitcode anyway.
+REMOVE_PROFILE_FLAGS()
+
 set(BC_INLINE_SRCS boxing.cpp dict.cpp link_forcer.cpp list.cpp tuple.cpp xrange.cpp)
 
 include_directories(${CMAKE_SOURCE_DIR}/src)

src/runtime/objmodel.cpp

@@ -3968,7 +3968,7 @@ Box* runtimeCallInternal(Box* obj, CallRewriteArgs* rewrite_args, ArgPassSpec ar
     } else if (obj->cls == instancemethod_cls) {
         BoxedInstanceMethod* im = static_cast<BoxedInstanceMethod*>(obj);
 
-        RewriterVar* r_im_func;
+        RewriterVar* r_im_func = NULL;
         if (rewrite_args) {
             r_im_func = rewrite_args->obj->getAttr(offsetof(BoxedInstanceMethod, func), Location::any());
         }