On Mon, May 7, 2012 at 12:08 PM, victor.stinner
http://hg.python.org/cpython/rev/ab500b297900 changeset: 76821:ab500b297900 user: Victor Stinner
date: Mon May 07 13:02:44 2012 +0200 summary: Issue #14716: Change integer overflow check in unicode_writer_prepare() to compute the limit at compile time instead of runtime. Patch writen by Serhiy Storchaka.
if (newlen > PyUnicode_GET_LENGTH(writer->buffer)) { - /* overallocate 25% to limit the number of resize */ - if (newlen <= (PY_SSIZE_T_MAX - newlen / 4)) + /* Overallocate 25% to limit the number of resize. + Check for integer overflow: + (newlen + newlen / 4) <= PY_SSIZE_T_MAX */ + if (newlen <= (PY_SSIZE_T_MAX - PY_SSIZE_T_MAX / 5)) newlen += newlen / 4;
Hmm. Very clever, but it's not obvious that that overflow check is mathematically sound. As it turns out, the maths works provided that PY_SSIZE_T_MAX isn't congruent to 4 modulo 5; since PY_SSIZE_T_MAX will almost always be one less than a power of 2 and powers of 2 are always congruent to 1, 2 or 4 modulo 5, we're safe. Is the gain from this kind of micro-optimization really worth the cost of replacing obviously correct code with code whose correctness needs several minutes of thought? Mark