> Does it make sense to backport ac317700ce7439e38a8b420218d9a5035bba92ed to 3.6?

No, async was not a keyword in Python 3.6 on purpose. Making it a keyword can break a lot of code.

I confirm that Python 3.6 still crashs with a very high number of "async " prefixes: try attached async_parser_crash.py.

Extract of the gdb traceback on a crash:

(...)
#665 0x0000000000454867 in tok_get (tok=0x7fffff8b98c0, 
    p_start=0x7fffff8b9cb8, p_end=0x7fffff8b9cb0) at Parser/tokenizer.c:1571
#666 0x0000000000454867 in tok_get (tok=0x7fffff8b9d40, 
    p_start=0x7fffff8ba138, p_end=0x7fffff8ba130) at Parser/tokenizer.c:1571
#667 0x0000000000454867 in tok_get (tok=0x7fffff8ba1c0, 
    p_start=0x7fffff8ba5b8, p_end=0x7fffff8ba5b0) at Parser/tokenizer.c:1571
#668 0x0000000000454867 in tok_get (tok=0x7fffff8ba640, 
    p_start=0x7fffff8baa38, p_end=0x7fffff8baa30) at Parser/tokenizer.c:1571
#669 0x0000000000454867 in tok_get (tok=0x7fffff8baac0, 
    p_start=0x7fffff8baeb8, p_end=0x7fffff8baeb0) at Parser/tokenizer.c:1571
#670 0x0000000000454867 in tok_get (tok=0x7fffff8baf40, 
    p_start=0x7fffff8bb338, p_end=0x7fffff8bb330) at Parser/tokenizer.c:1571
#671 0x0000000000454867 in tok_get (tok=0x7fffff8bb3c0, 
    p_start=0x7fffff8bb7b8, p_end=0x7fffff8bb7b0) at Parser/tokenizer.c:1571
(...)

It looks like a stack overflow.

The tokenizer may fail earlier on "async async ".