From: Martin Taibr <taibr.martin@gmail.com>
Date: Fri, 27 Mar 2020 09:12:00 +0000 (+0100)
Subject: update information about NaN bug (fixed in DP by f96918f0280)
X-Git-Tag: xonotic-v0.8.5~1144^2~4
X-Git-Url: https://git.rm.cloudns.org/?a=commitdiff_plain;h=82d2e1efe3030237a3a589111cf3d430cd2ea6cf;p=xonotic%2Fxonotic-data.pk3dir.git

update information about NaN bug (fixed in DP by f96918f0280)
---

diff --git a/qcsrc/lib/warpzone/mathlib.qc b/qcsrc/lib/warpzone/mathlib.qc
index cf86d97f9..2b1d00c2c 100644
--- a/qcsrc/lib/warpzone/mathlib.qc
+++ b/qcsrc/lib/warpzone/mathlib.qc
@@ -24,11 +24,12 @@ bool isinf(float e)
 }
 bool isnan(float e)
 {
-	// the sane way to detect NaN is broken because of a compiler bug
-	// (works with constants but breaks when assigned to variables)
-	// use conversion to string instead
-	//float f = e;
-	//return (e != f);
+	// The sane way to detect NaN is this:
+	// float f = e;
+	// return (e != f);
+	// but darkplaces used to be compiled with -ffinite-math-only which broke it.
+	// DP is fixed now but until all clients update (so after 0.8.3) we have to use the following workaround
+	// or they'd have issues when connecting to newer servers.
 
 	// Negative NaN ("-nan") is much more common but plain "nan" can be created by negating *some* -nans so we need to check both.
 	// DP's QCVM and GMQCC's QCVM behave differently - one needs ftos(-(0.0 / 0.0)), the other ftos(-sqrt(-1)).
@@ -225,7 +226,7 @@ float copysign(float e, float f)
 {
 	return fabs(e) * ((f>0) ? 1 : -1);
 }
-/// Always use `isnan` function to compare because `float x = nan(); x == x;` gives true
+
 float nan(string tag)
 {
 	return sqrt(-1);