zsh: expression after subscript flags is incorrectly parsed as arithmetic

[LangLangBart]

When a subscript flag is present, shfmt seems to parse the expression after it
as an arithmetic expression instead of a pattern.

Reproduction

# (i) returns the index of the first double-dash/end of options operator
./shfmt -ln zsh <<<'array=(ls -ld -- "**/bar"); echo ${array[(i)--]}'
# <standard input>:1:45: `--` must be followed by a literal

# (I) returns the index of the last element matching the pattern '+'
./shfmt -ln zsh <<<'array=(foo + bar + baz); echo ${array[(I)+]}'
# <standard input>:1:42: `+` must be followed by an expression

# (R) returns the value of the last element matching the pattern '/*'
./shfmt -ln zsh <<<'array=(/usr/bin rel/path /etc ./local); echo ${array[(R)/*]}'
# <standard input>:1:57: `/` must follow an expression

# `%`, `^`, `<`, `>`, `|`, `&`, `=`, `:`, `,` also fail
./shfmt -ln zsh <<<'array=(5 "&" 6); echo ${array[(i)&]}'
# <standard input>:1:34: `&` must follow an expression

Note

The outer subscript containing a reverse subscript can still be
arithmetic. For example:

outer cmd_words[(i)--]-1] is arithmetic
inner (i) flag's argument -- is a pattern

# Returns the elements before the double-dash/end of options operator
cmd_words=(ls -a -- file.txt); echo ${cmd_words[1,cmd_words[(i)--]-1]}
# ls -a

zsh docs

https://zsh.sourceforge.io/Doc/Release/Parameters.html#Subscript-Flags

EDIT1: Delete redundant header

[mvdan]

Not sure if you've seen #1233 too. Parsing parameter expression subscripts in Zsh may be rather difficult to get right in these edge cases.

[LangLangBart]

Not sure if you've seen #1233 too.

Thanks for the pointer!

The problem in #1233 is that the parser can't know whether an array is
associative or indexed, so it has to guess arithmetic for [exp]. That's hard.

---

For this issue we know the deteced flags from zshSubFlags(), but always parses
the expression after them via arithmExprAssign.

Not all flags take patterns though. Per the subscript flags docs¹:

• r, R, i, I, k, K (with optional e): expression is a
  pattern/string
• w, f, s, p: expression is still a numeric index

# (w) keeps arithmetic
scalar="one two three"; echo ${scalar[(w)1+1]}
# two

# (wi) switches to pattern — i takes priority
scalar="one + two three"; echo ${scalar[(wi)+]}
# 5

Since the flag text is already captured, the parser could check for the presence
of certain flags to decide whether to parse the expression as arithmetic or as a
string ?

Footnotes

1. https://zsh.sourceforge.io/Doc/Release/Parameters.html#Subscript-Flags ↩