Merge pull request #574 from kmod/linkorder

Add a section-ordering script

CMakeLists.txt

@@ -45,9 +45,11 @@ 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")
+    set(CMAKE_CXX_LINK_FLAGS "${CMAKE_CXX_LINK_FLAGS} -Wl,--section-ordering-file,${CMAKE_CURRENT_SOURCE_DIR}/section_ordering.txt")
   endif()
 endif()
 
+
 # pyston self host mode
 if(ENABLE_SELF_HOST)
   set(PYTHON_EXE "pyston")
@@ -215,7 +217,11 @@ add_subdirectory(test/test_extension)
 add_subdirectory(test/unittests)
 add_subdirectory(tools)
 
-add_executable(pyston $<TARGET_OBJECTS:PYSTON_MAIN_OBJECT> $<TARGET_OBJECTS:PYSTON_OBJECTS> $<TARGET_OBJECTS:FROM_CPYTHON>)
+# There are supposed to be better ways [1] to add link dependencies, but none of them worked for me.
+# [1] http://www.cmake.org/pipermail/cmake/2010-May/037206.html
+add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/linkdeps_dummy.c COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_BINARY_DIR}/linkdeps_dummy.c DEPENDS ${CMAKE_SOURCE_DIR}/section_ordering.txt)
+
+add_executable(pyston $<TARGET_OBJECTS:PYSTON_MAIN_OBJECT> $<TARGET_OBJECTS:PYSTON_OBJECTS> $<TARGET_OBJECTS:FROM_CPYTHON> linkdeps_dummy.c)
 # Wrap the stdlib in --whole-archive to force all the symbols to be included and eventually exported
 target_link_libraries(pyston -Wl,--whole-archive stdlib -Wl,--no-whole-archive pthread m readline sqlite3 gmp ssl crypto unwind pypa liblz4 double-conversion ${LLVM_LIBS} ${LIBLZMA_LIBRARIES} ${OPTIONAL_LIBRARIES})
 

section_ordering.txt

+.text.div_i64_i64
+.text.boxInt
+.text.boxBool
+.text._ZN6pyston2gc10SmallArena6_allocEmi
+.text._ZNSt17_Function_handlerIFvPN6pyston2gc10SmallArena16ThreadBlockCacheEEZNS2_12freeUnmarkedERSt6vectorIPNS0_3BoxESaIS8_EEE3$_0E9_M_invokeERKSt9_Any_dataS4_
+.text.memset
+.text.gc_alloc
+.text.gc_realloc
+.text._ZN6pyston9BoxedList6ensureEi
+.text.PyNumber_AsSsize_t
+.text._ZN6pyston15objectNewNoArgsEPNS_10BoxedClassE
+.text._PyIndex_Check
+.text._ZN6pyston9threading21allowGLReadPreemptionEv
+.text._ZN6pyston9getOpNameEi
+.text._ZN6pyston8callFuncEPNS_17BoxedFunctionBaseEPNS_15CallRewriteArgsENS_11ArgPassSpecEPNS_3BoxES6_S6_PS6_PKSt6vectorIPKSsSaISA_EE
+.text._ZN6pyston2gc9GCVisitor5visitEPv
+.text.intLtInt
+.text.intModInt
+.text.intAddInt
+.text.intMulInt

tools/annotate.py

@@ -106,6 +106,7 @@ This will typically look like:
 Benchmark that was run.  '--heap-map-target BENCHMARK' is
 equivalent to '--heap-map-args ./pyston_release -i BENCHMARK'.
     """.strip())
+    parser.add_argument("--perf-data", default="perf.data")
     args = parser.parse_args()
 
     func = args.func_name
@@ -116,7 +117,7 @@ equivalent to '--heap-map-args ./pyston_release -i BENCHMARK'.
 
     objdump = get_objdump(func)
 
-    p = subprocess.Popen(["perf", "annotate", "-v", func], stdout=subprocess.PIPE, stderr=open("/dev/null", "w"))
+    p = subprocess.Popen(["perf", "annotate", "-i", args.perf_data, "-v", func], stdout=subprocess.PIPE, stderr=open("/dev/null", "w"))
     annotate = p.communicate()[0]
     assert p.wait() == 0
 

tools/perf_diff.py

@@ -5,7 +5,7 @@ def tally(fn):
 
     for l in open(fn):
         samples = int(l.split()[3])
-        func = l.split()[-1]
+        func = l.rsplit(']', 1)[1].strip()
         counts[func] = counts.get(func, 0) + samples
 
     return counts
@@ -23,12 +23,11 @@ def main():
 
     diff_thresh = (total1 + total2) / 2 / 100
 
-    names.sort(key=lambda n: counts1.get(n, 0) + counts2.get(n, 0), reverse=True)
-    for n in names:
+    names.sort(key=lambda n: abs(counts1.get(n, 0) - counts2.get(n, 0)), reverse=True)
+    for n in names[:10]:
         c1 = counts1.get(n, 0)
         c2 = counts2.get(n, 0)
-        if abs(c1 - c2) >= diff_thresh:
-            print n, counts1.get(n, 0), counts2.get(n, 0)
+        print n, counts1.get(n, 0), counts2.get(n, 0)
 
 if __name__ == "__main__":
     main()