New code produces same result without loop at all, so
it cannot fall in infinite loop, and it is faster in
use cases requiring more than one loop in previous code.
The Unvanquished vega map is known to trigger the bug:
https://github.com/UnvanquishedAssets/map-vega_src.dpkdir
I reproduced it multiple time on various hardware (8 core FX-9590,
12 core/24 thread Ryzen 9 3900X) with commit af40508 and using
final compilation profile edited to use -fastbounce instead
of -fast option.
The symptom is simple, q3map2 stucks there:
--- Radiosity (bounce 1 of 8) ---
--- RadCreateDiffuseLights ---
0...1...2...3..
Or somewhere else in that progression bar given your hardware
and the amount of core your CPU has.
When stuck, all the CPU cores are running 100% but the thread
never returns (a strace can reveals it, a gdb backtrace too).
Thanks to @slipher for the precious advices and improving my first
attempt to fix it.
For more information on the issue, I asked:
> which negative value never can become positive
> when incremented infinitely?
slipher said:
> for a double, any value less than -2^53 would have this property
> don't know for float off the top of my head
But then, it means that's theorically verified this loop was able
to run forever in some case.
I don't know what this code is doing anyway, but at least we can
keep the behaviour without requiring to understand it.
-bsp -deep - include detail brushes into BSP tree generation (but at lowest possible priority), but still ignore them for vis
-vis -mergeportals - only merge vis portals on the same plane, but don't merge clusters (much faster vis, but only slightly worse - will evaluate later why it's worse vis at all)
git-svn-id: svn://svn.icculus.org/netradiant/trunk@244 61c419a2-8eb2-4b30-bcec-8cead039b335
- better BSP tree splitting (experimental, option -altsplit)
- also compare shaders when sorting surfaces (should give slightly more fps)
- misc_model spawnflag 32: set vertex alpha from vertex color (for terrain blending)
git-svn-id: svn://svn.icculus.org/netradiant/trunk@240 61c419a2-8eb2-4b30-bcec-8cead039b335
The logic isn't entirely legitimate, since fixedWinding_t is a
fixed-size type that is *sometimes* treated as a variable-sized type,
but it would require deeper refactoring to make this strictly
conforming. As it is, we just keep the offset computation as a
reasonable way to calculate the allocation size.
Fixes#583.
The arrays were always meant to be variably sized, and objects are only ever allocated dynamically. Object size computations are simplified with this change.
Flexible arrays were introduced in C99, so this change means that we will require a C99-conforming compiler henceforth.
