Mark host_dirty() and device_dirty() with no_discard.

[mcourteaux]

Lost quite some time, because I needed set_host_dirty() instead of host_dirty().

[steven-johnson]

I wish this was the default in C++ :-/

[mcourteaux]

Heh, it's not working... It says "attribute ignored". Why is this? Is the macro not properly defined to check compiler support? Or am I misunderstanding the compiler output?

[steven-johnson]

Heh, it's not working... It says "attribute ignored". Why is this? Is the macro not properly defined to check compiler support? Or am I misunderstanding the compiler output?

Dunno, but for a lark, try putting it before HALIDE_ALWAYS_INLINE?

[mcourteaux]

Ah, it's because you apparently need to put all the attributes in a single [[...]] list... Meh... Let's think...

EDIT: This is not even true... You can do multiple [[ ]] attributes after each other.

[steven-johnson]

Ah, it's because you apparently need to put all the attributes in a single [[...]] list... Meh... Let's think...

We require C++17 at this point, we can just write it out explicitly without a macro

[mcourteaux]

Nevermind, you were right the order of the attributes mattered, as we mix __attribute__() with [[...]]. There is no always_inline attribute in for [[ ]] AFAICT, so this will be it for now.

[mcourteaux]

Is the cmake issue fixed by this merge of main? If so, the buildbots need to be retriggered. However, some of the builds did run and complete successfully. So, I think this PR is okay, as it's only a C++-level change, and not really anything to do with the logic or workings of Halide.