[mypyc] Intern string literals (#9960)

This makes us more compatible with CPython, which interns (some)
string literals.

This speeds up some microbenchmarks slightly, by around 0.5-2%.

This may give a bigger speed win when calling functions using
vectorcalls and keyword arguments, but we'll see about that once that's
supported. Vectorcall argument parsing code has a fast path for interned
keyword argument names.

mypyc/codegen/emitmodule.py

@@ -831,6 +831,8 @@ def generate_globals_init(self, emitter: Emitter) -> None:
                 emitter.emit_line(
                     '{} = CPyTagged_FromObject({});'.format(actual_symbol, symbol)
                 )
+            elif isinstance(literal, str):
+                emitter.emit_line('PyUnicode_InternInPlace(&{});'.format(symbol))
 
         emitter.emit_lines(
             'is_initialized = 1;',

mypyc/test-data/run-strings.test

@@ -23,7 +23,10 @@ def match(x: str, y: str) -> Tuple[bool, bool]:
 
 [file driver.py]
 from native import f, g, tostr, booltostr, concat, eq, match
+import sys
+
 assert f() == 'some string'
+assert f() is sys.intern('some string')
 assert g() == 'some\a \v \t \x7f " \n \0string 🐍'
 assert tostr(57) == '57'
 assert concat('foo', 'bar') == 'foobar'