Let compilers vectorize v_complement()

The two's complement of a digit vector is ~a + 1, and the +1 only carries
through the trailing zero digits: they stay zero, the first nonzero digit
is negated, and the digits above it are just inverted.  Written that way
no carry is passed between loop iterations, so the loop can be vectorized.

v_complement() runs for every bitwise operation with a negative multi-digit
operand, up to three times per operation.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Author: eendebakpt

Objects/longobject.c

@@ -5552,14 +5552,20 @@ _PyLong_Lshift(PyObject *a, int64_t shiftby)
 static void
 v_complement(digit *z, digit *a, Py_ssize_t m)
 {
-    Py_ssize_t i;
-    digit carry = 1;
-    for (i = 0; i < m; ++i) {
-        carry += a[i] ^ PyLong_MASK;
-        z[i] = carry & PyLong_MASK;
-        carry >>= PyLong_SHIFT;
+    Py_ssize_t i = 0;
+
+    /* -a is ~a + 1, and the +1 only carries through the trailing zero
+       digits: they stay zero, the first nonzero digit is negated, and the
+       digits above it are just inverted.  No carry is passed between
+       iterations of the last loop, so it can be vectorized. */
+    while (i < m && a[i] == 0) {
+        z[i++] = 0;
+    }
+    assert(i < m);
+    z[i] = (0U - a[i]) & PyLong_MASK;
+    for (++i; i < m; ++i) {
+        z[i] = a[i] ^ PyLong_MASK;
     }
-    assert(carry == 0);
 }
 
 /* Bitwise and/xor/or operations */