unparser icon indicating copy to clipboard operation
unparser copied to clipboard

Chocking on interpolation with interpolation

Open akimd opened this issue 4 years ago • 22 comments

Hi Markus!

Weird things happen when there are interpolations within interpolations.

$ unparser -ve '"#{"#{true ? "1
" : "2"}
"}"'
(string)
Original-Source:
"#{"#{true ? "1
" : "2"}
"}"
Generated-Source:
"#{<<-HEREDOC}"
Original-Node:
(dstr
  (begin
    (dstr
      (begin
        (if
          (true)
          (str "1\n")
          (str "2")))
      (str "\n"))))
Generated-Node:
#<Parser::SyntaxError: unexpected token tLSHFT>
/opt/local/lib/ruby3.0/gems/3.0.0/gems/parser-3.0.1.0/lib/parser/diagnostic/engine.rb:72:in `process'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/parser-3.0.1.0/lib/parser/base.rb:286:in `on_error'
(eval):3:in `_racc_do_parse_c'
(eval):3:in `do_parse'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/parser-3.0.1.0/lib/parser/base.rb:190:in `parse'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser.rb:116:in `block in parse_either'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/either.rb:34:in `wrap_error'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser.rb:115:in `parse_either'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/either.rb:115:in `bind'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/validation.rb:63:in `from_string'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:40:in `validation'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:143:in `public_send'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:143:in `process_target'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:133:in `block in exit_status'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:132:in `each'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:132:in `exit_status'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/lib/unparser/cli.rb:64:in `run'
/opt/local/lib/ruby3.0/gems/3.0.0/gems/unparser-0.6.0/bin/unparser:10:in `<top (required)>'
/opt/local/bin/unparser:23:in `load'
/opt/local/bin/unparser:23:in `<main>'
Error: (string)

I agree no human being would write such code. But our generator does :(

Cheers!

akimd avatar May 07 '21 05:05 akimd

2 issues play in here:

First:

There are dynamic strings that produce AST that can only be the result of parsing heredocs. And the signature for these ASTs of "must have been a heredoc and can only be produced from a heredoc" are almost random.

Unparser has some heuristics to trigger heredoc emits, and sometimes (which happened here): We end up with triggering heredocs where we should not have. But somtimes emitting more heredocs makes the emitter code easier, while still producing the same AST on parsing.

Second:

Heredocs need 'parallel' dispatch, as they have a very dislocal effect on the token stream. Its likely that one of the parrallel all sides is missing hence not getting you the body of the heredoc, only the header.

Thanks for reporting: Right now, I'm super busy commercially, assume this one takes a very long time to fix.

mbj avatar May 07 '21 12:05 mbj

Markus, In trying to find something to reproduce our failure, I may well have written something actually more complex than our original use case. And so far, we can't reproduce the original issue, so we might just as well have used some older version of unparser somewhere. In other words: we are not blocked by this issue, and AFAIC you are welcome to let it rust. Cheers!

akimd avatar May 07 '21 15:05 akimd

@akimd its my end goal that unparser can round trip ALL ast that parser emits. Which may mean I'll have to ask the parser team to simplfy the ast.

I'm close to ask them to give me a dedicated node for heredocs in the original source, this would make my part so much easier.

mbj avatar May 07 '21 15:05 mbj

@akimd I recently spend a bit more time to think about the problem. Overall I've to conclude that the current parser AST, on any form of dynamic strings is currently lossy.

Not lossy for execution semantics, but lossy in terms of round trip semantics. Meaning you cloud certainly implement a ruby on top of that AST that executes correctly, but you cannot implement an unparser that produces source that produces the same AST, as critical information is lost/noisy.

I've got a local AST builder that subclasses the offiical parser builder producing an AST that does not suffer this compression losses.

Now the question is: Whats your use case? Do you round trip hand written ruby, (That may include heredocs / dynamic strings)? Or do you generate an AST from other sources to than unparse with unparser?

mbj avatar Apr 22 '22 15:04 mbj

Hey Markus, Oh, you have a new face :)

Unparser is the final part of a parse|transform|unparse pipe. We don't care about exact syntactic round tripping, we just care about the (execution) semantics.

Cheers!

akimd avatar Apr 23 '22 15:04 akimd

Oh, you have a new face :)

ageless version ;)

Unparser is the final part of a parse|transform|unparse pipe. We don't care about exact syntactic round tripping, we just care about the (execution) semantic

You are using Unparser.parse, if not would you consider using that one instead of Parser::CurrencyRuby.parse ?

mbj avatar Apr 23 '22 22:04 mbj

Markus, I have been told that since we have something that works (and works around that particular issue), we won't change it again. So I guess we'll stick to Parser's parser.

akimd avatar Apr 24 '22 09:04 akimd

@akimd Unparser.parse "is" the parsers parser, but it uses the modernized AST format. Thats all I want to know.

mbj avatar Apr 24 '22 17:04 mbj

Markus, Yes, I know, that was very clear. Yet this stuff is in production, and the guy in charge of that does not want to change it now. Maybe in the future this will be revisited, but not now. Sorry!

akimd avatar Apr 29 '22 08:04 akimd

So are you using the modernized AST Format or not?

mbj avatar Apr 29 '22 10:04 mbj

Hi Markus, No, we are still using the "native" AST from parser itself.

akimd avatar May 05 '22 12:05 akimd

kk, so as a recomemndation: The parsers default AST format is lossy on semantics, and unparser only supports the modern one for that reason.

Also all other major users of parser opt into the modern format for that reason, and it can very well be that you have semantic issues as you process the default format.

mbj avatar May 05 '22 15:05 mbj

Also a fix for this issue will probably only ever land in the modern AST format.

mbj avatar May 05 '22 20:05 mbj

I perfectly understand all these points, but the decision is not mine.  I'm just a proxy here.

akimd avatar May 09 '22 12:05 akimd

@akimd I'd like to emphasize here that the "default parser AST" is often wrong when it comes to execution semantics! If the semantics of the executed ruby are important: Its very much recommended to opt into modern AST format.

mbj avatar Jan 06 '23 23:01 mbj

Hi Markus, Happy new year! Thanks for the heads up. I will again forward your recommendation, and will tell you if we moved to the modernize'd AST format. FWIW, thanks to your message, I realized that in the code I maintain, I was lagging on several emit_*, which I have now enabled. Cheers!

akimd avatar Jan 07 '23 09:01 akimd

@akimd Is this still affecting you? I can't promise anything, but given what I ran into was a very similar problem (see the mention above), and I found a solution that worked for that case, I might see if I can get a chance to look into this one too. I reduced the error case at the start of this ticket to this (the ternary doesn't appear to make a difference at all):

"#{"#{}\n"}"

Which still produces this (after applying my fix from the other ticket):

Generated-Source:
"#{<<-HEREDOC}"
Original-Node:
(dstr
  (begin
    (dstr
      (begin)
      (str "\n"))))
Generated-Node:
#<Parser::SyntaxError: unexpected token tLSHFT>

That puts the location of the missing heredoc output somewhere in a relatively small number of places.

vidarh avatar Oct 13 '23 17:10 vidarh

Hi @vidarh, Thanks for the link!

I'm not in charge of the piece of code where this first happened (and I don't have access to it). So by necessity workarounds must have been installed, and I don't know which.

IOW, we no longer depend on this.

Cheers!

akimd avatar Oct 15 '23 08:10 akimd

@akimd I found a fix for this issue (in fact the entire dstr saga): https://github.com/mbj/unparser/pull/366

That branch does not pass CI at the time of me posting this comment (still working through mutation coverage)but check this out (run from that branch):

bundle exec unparser --verbose -e '"#{"#{true ? "1
" : "2"}
"}"'
(string)
Original-Source:
"#{"#{true ? "1
" : "2"}
"}"
Generated-Source:
"#{"#{if true
  "1\n"
else
  "2"
end}\n"}"
Original-Node:
(dstr
  (begin
    (dstr
      (begin
        (if
          (true)
          (str "1\n")
          (str "2")))
      (str "\n"))))
Generated-Node:
(dstr
  (begin
    (dstr
      (begin
        (if
          (true)
          (str "1\n")
          (str "2")))
      (str "\n"))))
Success: (string)

mbj avatar May 28 '24 02:05 mbj

Hi Markus,

I think I face the same issue again, under a different disguise.

require 'parser/current'
require 'unparser'

ast = Unparser.parse(<<~'RUBY')
  def get_val(var: nil, **opts)
    foo format: <<~JS
        a;
        #{transform};
        b;
    JS
  end
RUBY
p ast
puts Unparser.unparse(ast)
s(:def, :get_val,
  s(:args,
    s(:kwoptarg, :var,
      s(:nil)),
    s(:kwrestarg, :opts)),
  s(:send, nil, :foo,
    s(:kwargs,
      s(:pair,
        s(:sym, :format),
        s(:dstr,
          s(:str, "a;\n"),
          s(:begin,
            s(:send, nil, :transform)),
          s(:str, ";\n"),
          s(:str, "b;\n"))))))
def get_val(var: nil, **opts)
  foo(format: <<-HEREDOC)
end

Cheers!

akimd avatar Jun 21 '24 04:06 akimd

@akimd very likely this is fixed by #366 but that PR is not yet ready for prime time. It uses an exhaustive search algorithm to select round tripping dstr segments. And this has factorial runtime, its guaranteed to be correct but for large dstr constructs its to slow.

So I've to improve the search algorithm, still this entire issue class may be possible to be eliminated soon.

mbj avatar Jun 21 '24 13:06 mbj

@akimd This is fixed in #366 but #366 is not release ready yet, its a big change on how unparser does dynamic string literals and needs more testing.

$ cat x.rb
    foo format: <<~JS
        a;
        #{transform};
        b;
    JS
mutant-dev@mbj ~/devel/unparser (fix/dstr?) $ bundle exec bin/unparser --verbose x.rb
#<Unparser::CLI::Target::Path path=#<Pathname:x.rb>>
x.rb
Original-Source:
    foo format: <<~JS
        a;
        #{transform};
        b;
    JS

Generated-Source:
foo(format: "a;\n#{transform};
b;\n")
Original-Node:
(send nil :foo
  (kwargs
    (pair
      (sym :format)
      (dstr
        (str "a;\n")
        (begin
          (send nil :transform))
        (str ";\n")
        (str "b;\n")))))
Generated-Node:
(send nil :foo
  (kwargs
    (pair
      (sym :format)
      (dstr
        (str "a;\n")
        (begin
          (send nil :transform))
        (str ";\n")
        (str "b;\n")))))
Success: x.rb

As always your bugreports are appreciated!

mbj avatar Jun 22 '24 16:06 mbj