IS NULL doesn't type-check its input

[neilconway]

Describe the bug

The input expression to IS NULL isn't type-checked, unless for other constructs (IS TRUE, binary operators, etc.)

To Reproduce

CREATE TABLE tt();
SELECT * FROM tt WHERE ((STARTS_WITH()) IS NULL);

Yields:

thread 'main' (41425691) panicked at datafusion/functions/src/string/starts_with.rs:168:54:
index out of bounds: the len is 0 but the index is 1
stack backtrace:
   0: __rustc::rust_begin_unwind
             at /rustc/ded5c06cf21d2b93bffd5d884aa6e96934ee4234/library/std/src/panicking.rs:698:5
   1: core::panicking::panic_fmt
             at /rustc/ded5c06cf21d2b93bffd5d884aa6e96934ee4234/library/core/src/panicking.rs:80:14
   2: core::panicking::panic_bounds_check
             at /rustc/ded5c06cf21d2b93bffd5d884aa6e96934ee4234/library/core/src/panicking.rs:276:5
   3: <usize as core::slice::index::SliceIndex<[T]>>::index
             at /Users/neilconway/.rustup/toolchains/1.92.0-aarch64-apple-darwin/lib/rustlib/src/rust/library/core/src/slice/index.rs:267:10
   4: core::slice::index::<impl core::ops::index::Index<I> for [T]>::index
             at /Users/neilconway/.rustup/toolchains/1.92.0-aarch64-apple-darwin/lib/rustlib/src/rust/library/core/src/slice/index.rs:18:15
   5: <alloc::vec::Vec<T,A> as core::ops::index::Index<I>>::index
             at /Users/neilconway/.rustup/toolchains/1.92.0-aarch64-apple-darwin/lib/rustlib/src/rust/library/alloc/src/vec/mod.rs:3628:9
   6: <datafusion_functions::string::starts_with::StartsWithFunc as datafusion_expr::udf::ScalarUDFImpl>::simplify
             at ./datafusion/functions/src/string/starts_with.rs:168:54

Expected behavior

The expected behavior is that STARTS_WITH() should be rejected at planning time because of a type mismatch -- which is what happens for other variants like STARTS_WITH() IS TRUE.

Additional context

No response

[neilconway]

I took a brief look at fixing this, but it was a bit out of my depth. Fixing the starts_with() implementation itself would be wrong because this case should be rejected at planning-time, as it is for IS TRUE and other variants. For IS TRUE, the type mismatch is detected during type coercion; I suppose we could add some type coercion for IS NULL and IS NOT NULL (e.g., by adding those cases to TreeNodeRewriter in type_coercion.rs), but I'd appreciate some comments on whether that makes sense or whether a different approach would be preferred.