bare string concatenation uses lots o' stack

[enebo]

When JRuby is loading the serialized blob via Ruby stdlib unicode_normalize/tables.rb it hits large chains of strings built using string syntax concat chains:

  accents = "" \
    "[\u0300-\u034E" \
    "\u0350-\u036F" \
    "\u0483-\u0487" \
    "\u0591-\u05BD" \
    "\u05BF" \
    "\u05C1\u05C2" \
    "\u05C4\u05C5" \
    "\u05C7" \
    "\u0610-\u061A" \
    "\u064B-\u065F" \
    "\u0670" \
    "\u06D6-\u06DC" \
    "\u06DF-\u06E4" \
    "\u06E7\u06E8" \
    "\u06EA-\u06ED" \
    "\u0711" \
    "\u0730-\u074A" \
    "\u07EB-\u07F3" \
    "\u07FD" \
    "\u0816-\u0819" \
    #... a lot more.

This leads to a large amount of a stack trace like:

	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)
	at org.prism.Loader.loadNode(Loader.java:561)

If I increase our stack (which thankfully is possible to do with the JVM) then I can get this file to load. In legacy parsers this is all built up into a single string instead of putting it into a cons chain in the AST. JRuby does not want to grow stack to load this data file so I am wondering if we should rethink putting this into the tree or not?

[kddnewton]

Yup! That's still on my list. This would be fixed by #1479. If it's alright, I'm going to close this one in favor of that one.

[enebo]

@kddnewton hah. I forgot I opened that or I would have added this as a comment to it. I guess the only value of opening this was to point out a real-world problem with the current implementation.

[duerst]

We can also change the format of unicode_normalize/tables.rb if that helps.

[kddnewton]

Thank you @duerst that's very kind. I think we should fix it on our end, however, as I'm sure it's not just that file that has this issue and we want to be as performant as possible regardless of the source.