Code tries to force better optimization with -O3, -funroll-loops and similar, while disabling strict-aliasing, which may, or may not, give some nice speedup.
Created attachment 1899[details]
fixes for strict-aliasing
Patch fixes all cases of passing float argument as int, and vice-versa (uses union instead of punn+dereference). Also replaces -fno-strict-aliasing with -Wstrict-aliasing in Makefile.
I haven't noticed any problems with strict-aliasing after applying those changes. Tested only on Linux/PowerPC, other platforms may need some more work.
Would you mind cleaning up the mess by intoducing a global typedef for the conversion union? I'm guilty of creating some of those in various places already too :)
Hmm, you misunderstood me. I meant having one definition that union in a header like q_shared.h and getting rid of those local versions that have various different names. Just grep for union
Created attachment 1915[details]
fixes for strict-aliasing, v4
union type is now called "floatint_t" and it is defined globally in code/qcommon/q_shared.h; inside !Q3_VM block, because it may be problematic for lcc.
Patch also resolves Bug 3838.
Created attachment 1899 [details] fixes for strict-aliasing Patch fixes all cases of passing float argument as int, and vice-versa (uses union instead of punn+dereference). Also replaces -fno-strict-aliasing with -Wstrict-aliasing in Makefile. I haven't noticed any problems with strict-aliasing after applying those changes. Tested only on Linux/PowerPC, other platforms may need some more work.Created attachment 1901 [details] fixes for strict-aliasing, v2 typedef union fi_conv moved from qcommon.h to msg.cCreated attachment 1904 [details] fixes for strict-aliasing, v3 one more fix, found on x86_64Created attachment 1905 [details] fixes for strict-aliasing, v3b oops, correct one