fix(sub): error instead of panic when subtraction yields NaN

changelog.d/2893.fix.md

@@ -0,0 +1 @@
+Fix a panic in float subtraction that produces NaN values.

lib/tests/tests/expressions/arithmetic/subtraction/nan_result.vrl

@@ -0,0 +1,5 @@
+# result: can't subtract type float from float
+
+large = to_float!("1.7976931348623157e+308")
+inf_float = large + large
+inf_float - inf_float

src/compiler/value/arithmetic.rs

@@ -1,14 +1,13 @@
 #![deny(clippy::arithmetic_side_effects)]
 
-use std::ops::{Add, Mul, Rem, Sub};
-
-use bytes::{BufMut, Bytes, BytesMut};
+use std::ops::{Add, Mul, Rem};
 
 use crate::compiler::{
     value::{Kind, VrlValueConvert},
     ExpressionError,
 };
 use crate::value::{ObjectMap, Value};
+use bytes::{BufMut, Bytes, BytesMut};
 
 use super::ValueError;
 
@@ -60,6 +59,15 @@ pub trait VrlValueArithmetic: Sized {
     fn eq_lossy(&self, rhs: &Self) -> bool;
 }
 
+fn safe_sub(lhv: f64, rhv: f64) -> Option<Value> {
+    let result = lhv - rhv;
+    if result.is_nan() {
+        None
+    } else {
+        Some(Value::from_f64_or_zero(result))
+    }
+}
+
 impl VrlValueArithmetic for Value {
     /// Similar to [`std::ops::Mul`], but fallible (e.g. `TryMul`).
     fn try_mul(self, rhs: Self) -> Result<Self, ValueError> {
@@ -155,9 +163,9 @@ impl VrlValueArithmetic for Value {
                 let rhv = rhs.try_into_i64().map_err(|_| err())?;
                 i64::wrapping_sub(lhv, rhv).into()
             }
-            Value::Float(lhv) => {
-                let rhv = rhs.try_into_f64().map_err(|_| err())?;
-                lhv.sub(rhv).into()
+            Value::Float(lhs) => {
+                let rhs = rhs.try_into_f64().map_err(|_| err())?;
+                safe_sub(*lhs, rhs).ok_or_else(err)?
             }
             _ => return Err(err()),
         };