- if (message) {
- if (PL_in_eval & EVAL_KEEPERR) {
- static const char prefix[] = "\t(in cleanup) ";
- SV * const err = ERRSV;
- const char *e = NULL;
- if (!SvPOK(err))
- sv_setpvs(err,"");
- else if (SvCUR(err) >= sizeof(prefix)+msglen-1) {
- STRLEN len;
- e = SvPV_const(err, len);
- e += len - msglen;
- if (*e != *message || strNE(e,message))
- e = NULL;
- }
- if (!e) {
- SvGROW(err, SvCUR(err)+sizeof(prefix)+msglen);
- sv_catpvn(err, prefix, sizeof(prefix)-1);
- sv_catpvn(err, message, msglen);
- if (ckWARN(WARN_MISC)) {
- const STRLEN start = SvCUR(err)-msglen-sizeof(prefix)+1;
- Perl_warner(aTHX_ packWARN(WARN_MISC), "%s",
- SvPVX_const(err)+start);
- }
- }
- }
- else {
- sv_setpvn(ERRSV, message, msglen);
- }
+ /*
+ * Historically, perl used to set ERRSV ($@) early in the die
+ * process and rely on it not getting clobbered during unwinding.
+ * That sucked, because it was liable to get clobbered, so the
+ * setting of ERRSV used to emit the exception from eval{} has
+ * been moved to much later, after unwinding (see just before
+ * JMPENV_JUMP below). However, some modules were relying on the
+ * early setting, by examining $@ during unwinding to use it as
+ * a flag indicating whether the current unwinding was caused by
+ * an exception. It was never a reliable flag for that purpose,
+ * being totally open to false positives even without actual
+ * clobberage, but was useful enough for production code to
+ * semantically rely on it.
+ *
+ * We'd like to have a proper introspective interface that
+ * explicitly describes the reason for whatever unwinding
+ * operations are currently in progress, so that those modules
+ * work reliably and $@ isn't further overloaded. But we don't
+ * have one yet. In its absence, as a stopgap measure, ERRSV is
+ * now *additionally* set here, before unwinding, to serve as the
+ * (unreliable) flag that it used to.
+ *
+ * This behaviour is temporary, and should be removed when a
+ * proper way to detect exceptional unwinding has been developed.
+ * As of 2010-12, the authors of modules relying on the hack
+ * are aware of the issue, because the modules failed on
+ * perls 5.13.{1..7} which had late setting of $@ without this
+ * early-setting hack.
+ */
+ if (!(in_eval & EVAL_KEEPERR)) {
+ SvTEMP_off(exceptsv);
+ sv_setsv(ERRSV, exceptsv);