Change how and and or operations are compiled to IR to support custom values (#14653)

# Description

Because `and` and `or` are short-circuiting operations in Nushell, they
must be compiled to a sequence that avoids evaluating the RHS if the LHS
is already sufficient to determine the output - i.e., `false` for `and`
and `true` for `or`. I initially implemented this with `branch-if`
instructions, simply returning the RHS if it needed to be evaluated, and
returning the short-circuited boolean value if it did not.

Example for `$a and $b`:

```
   0: load-variable          %0, var 999 "$a"
   1: branch-if              %0, 3
   2: jump                   5
   3: load-variable          %0, var 1000 "$b" # label(0), from(1:)
   4: jump                   6
   5: load-literal           %0, bool(false) # label(1), from(2:)
   6: span                   %0          # label(2), from(4:)
   7: return                 %0
```

Unfortunately, this broke polars, because using `and`/`or` on custom
values is perfectly valid and they're allowed to define that behavior
differently, and the polars plugin uses this for boolean masks. But
without using the `binary-op` instruction, that custom behavior is never
invoked. Additionally, `branch-if` requires a boolean, and custom values
are not booleans. This changes the IR to the following, using the
`match` instruction to check for the specific short-circuit value
instead, and still invoking `binary-op` otherwise:

```
   0: load-variable          %0, var 125 "$a"
   1: match                  (false), %0, 4
   2: load-variable          %1, var 124 "$b"
   3: binary-op              %0, Boolean(And), %1
   4: span                   %0          # label(0), from(1:)
   5: return                 %0
```

I've also renamed `Pattern::Value` to `Pattern::Expression` and added a
proper `Pattern::Value` variant that actually contains a `Value`
instead. I'm still hoping to remove `Pattern::Expression` eventually,
because it's kind of a hack - we don't actually evaluate the expression,
we just match it against a few cases specifically for pattern matching,
and it's one of the cases where AST leaks into IR and I want to remove
all of those cases, because AST should not leak into IR.

Fixes #14518

# User-Facing Changes

- `and` and `or` now support custom values again.
- the IR is actually a little bit cleaner, though it may be a bit
slower; `match` is more complex.

# Tests + Formatting

The existing tests pass, but I didn't add anything new. Unfortunately I
don't think there's anything built-in to trigger this, but maybe some
testcases could be added to polars to test it.
This commit is contained in:
Devyn Cairns
2024-12-25 04:12:53 -08:00
committed by GitHub
parent 4b1f4e63c3
commit 35d2750757
9 changed files with 85 additions and 56 deletions

View File

@ -6286,7 +6286,11 @@ pub fn discover_captures_in_pattern(pattern: &MatchPattern, seen: &mut Vec<VarId
}
}
Pattern::Rest(var_id) => seen.push(*var_id),
Pattern::Value(_) | Pattern::IgnoreValue | Pattern::IgnoreRest | Pattern::Garbage => {}
Pattern::Expression(_)
| Pattern::Value(_)
| Pattern::IgnoreValue
| Pattern::IgnoreRest
| Pattern::Garbage => {}
}
}