Bug 16105 – `is` fails for init value of struct with float

The following test case only prints "B: works", although the first comparison should succeed, too: import std.stdio; struct Foo { int x; float f; } void main() { Foo foo; if(foo is Foo.init) "A: doesn't work".writeln; foo = Foo(); if(foo is Foo.init) "B: works".writeln; } Found by ArturG: https://forum.dlang.org/post/[email protected]

Addendum: It works correctly with LDC, so it's probably a codegen bug.

This isn't a codegen bug per se, but is rather a design bug. It happens because float.init is a signalling NaN. We should change it to be a quiet NaN. Using signalling NaNs seemed like a good idea at the time but it turns out that the hardware support for it just isn't good enough. Even on x86, Intel and AMD CPUs behave differently. And on other CPUs it is even worse. The situation at present is, the intended behaviour is a bit broken on some platforms, and totally broken on others. It is impossible to fix it. It is entirely my fault. When Walter released the backend I did this signalling NaN experiment that I had always wanted to do. I think it was my first compiler patch ever. It was a mistake. A failed experiment. We need to roll it back.

*** This issue has been marked as a duplicate of issue 15316 ***