fix: incorrect negative offset in multi-byte string slicing (#15140)

crates/polars-ops/src/chunked_array/strings/substring.rs

@@ -6,48 +6,42 @@ fn substring_ternary(
     opt_offset: Option<i64>,
     opt_length: Option<u64>,
 ) -> Option<&str> {
-    match (opt_str_val, opt_offset) {
-        (Some(str_val), Some(offset)) => {
-            // If `offset` is negative, it counts from the end of the string.
-            let offset = if offset >= 0 {
-                offset as usize
-            } else {
-                let offset = (0i64 - offset) as usize;
-                str_val
-                    .char_indices()
-                    .rev()
-                    .nth(offset)
-                    .map(|(idx, _)| idx + 1)
-                    .unwrap_or(0)
-            };
+    let str_val = opt_str_val?;
+    let offset = opt_offset?;
 
-            let mut iter_chars = str_val.char_indices();
-            if let Some((offset_idx, _)) = iter_chars.nth(offset) {
-                let len_end = str_val.len() - offset_idx;
-
-                // Slice to end of str if no length given.
-                let length = if let Some(length) = opt_length {
-                    length as usize
-                } else {
-                    len_end
-                };
+    // Fast-path: always empty string.
+    if opt_length == Some(0) || offset >= str_val.len() as i64 {
+        return Some("");
+    }
 
-                if length == 0 {
-                    return Some("");
-                }
+    let mut indices = str_val.char_indices().map(|(o, _)| o);
+    let mut length_reduction = 0;
+    let start_byte_offset = if offset >= 0 {
+        indices.nth(offset as usize).unwrap_or(str_val.len())
+    } else {
+        // If `offset` is negative, it counts from the end of the string.
+        let mut chars_skipped = 0;
+        let found = indices
+            .inspect(|_| chars_skipped += 1)
+            .nth_back((-offset - 1) as usize);
 
-                let end_idx = iter_chars
-                    .nth(length.saturating_sub(1))
-                    .map(|(idx, _)| idx)
-                    .unwrap_or(str_val.len());
+        // If we didn't find our char that means our offset was so negative it
+        // is before the start of our string. This means our length must be
+        // reduced, assuming it is finite.
+        if let Some(off) = found {
+            off
+        } else {
+            length_reduction = (-offset) as usize - chars_skipped;
+            0
+        }
+    };
 
-                Some(&str_val[offset_idx..end_idx])
-            } else {
-                Some("")
-            }
-        },
-        _ => None,
-    }
+    let str_val = &str_val[start_byte_offset..];
+    let mut indices = str_val.char_indices().map(|(o, _)| o);
+    let stop_byte_offset = opt_length
+        .and_then(|l| indices.nth((l as usize).saturating_sub(length_reduction)))
+        .unwrap_or(str_val.len());
+    Some(&str_val[..stop_byte_offset])
 }
 
 pub(super) fn substring(

py-polars/tests/unit/namespaces/string/test_string.py

@@ -46,6 +46,22 @@ def test_str_slice_expr() -> None:
         df.select(pl.col("a").str.slice(0, -1))
 
 
+def test_str_slice_multibyte() -> None:
+    ref = "你好世界"
+    s = pl.Series([ref])
+
+    # Pad the string to simplify (negative) offsets starting before/after the string.
+    npad = 20
+    padref = "_" * npad + ref + "_" * npad
+    for start in range(-5, 6):
+        for length in range(6):
+            offset = npad + start if start >= 0 else npad + start + len(ref)
+            correct = padref[offset : offset + length].strip("_")
+            result = s.str.slice(start, length)
+            expected = pl.Series([correct])
+            assert_series_equal(result, expected)
+
+
 def test_str_len_bytes() -> None:
     s = pl.Series(["Café", None, "345", "東京"])
     result = s.str.len_bytes()