[build] An awk script to extract PyMethodDef constants from C source.

Still needs some work to be useful.

- Show the total number of symbols.
- build/compile.sh: an example of how to remove dead code with GCC.  Not
  enabled yet, but I tested it and it makes a difference.

benchmarks/native-code.R

@@ -24,6 +24,9 @@ options(stringsAsFactors = F,
 Basic = function(ctx) {
   Banner('BASIC METRICS')
 
+  ctx$symbols %>% arrange(desc(filesize)) %>% head(20) -> f1
+  ShowValue('Number of Symbols: %d', nrow(ctx$symbols))
+
   ctx$symbols %>% arrange(desc(filesize)) %>% head(20) -> f1
   ShowFrame('By Size On Disk:', f1)
 

benchmarks/native-code.sh

@@ -90,10 +90,67 @@ run-for-release() {
   report
 }
 
-py-method-defs() {
+# TODO:
+# Write to a separate file like _build/pydefs/intobject.include
+# #ifdef OVM_MAIN
+# #include "intobject.include"
+# #else
+# ...
+# #end
+#
+# Should those files be checked in an edited by hand?  Or join them somehow
+# with oil-symbols.txt?
+# I think this is hard because of METHODS.
+# Maybe you should have a config file that controls it.  It takes a .include
+# file and then whitelist/blacklist, and then generates a new one.
+# could put it in build/pydefs-config.txt
+#
+# And then reprint the PyMethoDef without docstrings?  It shouldn't be that
+# hard to parse.  You can almost do it with a regex, since commas don't appear
+# in the string.
+
+extract-defs() {
+  # NOTE: PyMemberDef is also interesting, but we don't need it for the build.
+  gawk '
+  /static PyMethodDef/ {
+    if (printing != 0) {
+      printf("%s:%d Expected not to be printing\n", FILENAME, FNR);
+      exit 1;
+    }
+    printing = 1;
+    printf("\n");
+    printf("/* Extracted from %s */\n", FILENAME);
+  }
+
+  {
+    if (printing) {
+      print
+    }
+  }
+
+  /^[:space:]*\}/ {
+    printing = 0;
+  }
+  ' "$@"
+}
+
+source build/common.sh  # $PY27
+
+# TODO: Use PREPROC_FLAGS from build/compile.sh.
+preprocess() {
+  # What about stuff in pyconfig.h?
+  gcc -I $PY27 -E -D OVM_MAIN -
+}
+
+extract-all-defs() {
+  echo '#include "pyconfig.h"'
   # 52 different instances.  Sometimes multiple ones per file.
-  find _tmp/oil-tar-test -name '*.[ch]' \
-    | xargs grep -n 'static PyMethodDef'
+  find _tmp/oil-tar-test -name '*.[ch]' | xargs -- $0 extract-defs 
+}
+
+py-method-defs() {
+  extract-all-defs #| preprocess
 }
 
+
 "$@"

build/compile.sh

@@ -149,7 +149,16 @@ readonly CC=${CC:-cc}  # cc should be on POSIX systems
 # isn't necessary.  Python's configure.ac sometimes does it by compiling a test
 # file; at other times it does it by grepping $CC --help.
 
-readonly BASE_CFLAGS='-fno-strict-aliasing -fwrapv -Wall -Wstrict-prototypes'
+BASE_CFLAGS='-fno-strict-aliasing -fwrapv -Wall -Wstrict-prototypes'
+
+# TODO: Enable this once you test that Clang also accepts the flags?  Or should
+# that be detected by a config file?
+# https://stackoverflow.com/questions/6687630/how-to-remove-unused-c-c-symbols-with-gcc-and-ld
+#BASE_CFLAGS="$BASE_CFLAGS -fdata-sections -ffunction-sections"
+readonly BASE_CFLAGS
+
+#BASE_LDFLAGS='-Wl,--gc-sections'
+BASE_LDFLAGS=''
 
 # The user should be able to customize CFLAGS, but it shouldn't disable what's
 # in BASE_CFLAGS.
@@ -211,6 +220,7 @@ build() {
     $c_module_src_list \
     Modules/ovm.c \
     -l m \
+    ${BASE_LDFLAGS} \
     $readline_flags \
     "$@" \
     || true