<html><head><meta http-equiv="Content-Type" content="text/html charset=us-ascii"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><div class=""><br class=""></div><div class="">Hi,</div><div class=""><br class=""></div><div class="">I have noticed the following behaviour while running the SPEC benchmark suite (built with llvm, of course). One of the tests was failing (Xalanc) with a segfault (because of stack overflow) because the terminating condition for a recursion involved NaN.</div><div class=""><br class=""></div><div class="">I've checked a little further and I have noticed that, with fast-math, the square roots of negative numbers will sometimes be replaced with "return undef". But only sometimes, and only the square roots (for example, log(x) does not manifest this behaviour).</div><div class=""><br class=""></div><div class="">It turns out a rather common way of getting the value of NaN in code is to assign it the return value of sqrt(<negative number>). Because of this bug, any such assignments will basically result undefined values.</div><div class=""><br class=""></div><div class="">These two patches should provide a fix for this. The first is a test to illustrate the situation (the negative functions will be replaced with return undef in the IR), while the other is a fix for this bug.</div><div class=""><br class=""></div><div class="">Best regards,</div><div class="">Octav</div><div class=""><br class=""></div><div class=""></div></body></html>