Modulo and unsigned divisions by constant

[RamonUnch]

FastDoom could benefit from the optimisation of the modulo % operation. Internally WC generates an idiv operation. The optimization is similar to the divisions by a constant because it is almost the same thing: ie:

Modulo 10 for signed ints, according to GCC (fastcall convention)
int Mod10(int x)
    mov    eax, 1717986919
    imul   ecx
    mov    eax, edx
    mov    edx, ecx
    sar    edx, 31
    sar    eax, 2
    sub    eax, edx
    lea    edx, [eax+eax*4]
    mov    eax, ecx
    add    edx, edx
    sub    eax, edx
    ret
; Unsigned version:
unsigned ModU10(unsigned eax)
    mov    eax, ecx
    mov    edx, -858993459
    mul    edx
    mov    eax, edx
    shr    eax, 3
    lea    edx, [eax+eax*4]
    mov    eax, ecx
    add    edx, edx
    sub    eax, edx
    ret

There are very few places where the x % constant is used and it would almost not affect much performances.

Also there are some missing optimizations: signed and unsigned divisions should be separated and when you are dividing a value you know to be positive (ie: monster damage), then an unsigned variant of the division should be used: example for unsigned Div10:

; Again fastcall convention:
unsigned Div10u(unsigned x)
    mov eax, ecx
    mov edx, -858993459
    mul edx
    mov eax, edx
    shr eax, 3
    ret

unsigned versions save a few instructions and mul is faster than imul (presumably?) I simply use GCC with -O2 -march=i386 and -mtune=generic Those should not be hard to inline with OpenWatcom. EDIT: I did not check properly all instances of DivXX() clls, maybe you never need unsigned versions...

[maxxoccupancy]

For modulo by constants, masking the upper bits accomplishes that for powers of two: 2, 4, 8.... 256, etc. It may be worth it to build a separate lookup table for other values. Multiplication and division on these old, 1970s microprocessors was so slow that some CISCs didn't even bother with the microcode for them, let alone dedicated hardware.

[RamonUnch]

for modulo by a power of two WC already makes the optimization IIRC.