Some years ago, there was an attempt to optimize this function by using logarithmic division instead. Unfortunately, there were some glaring issues with it that prevented it from working correctly, notably the fact that it failed to take into account the fact that the function that was used as the basis for that optimization only supported 8-bit input values, whereas Sonic 1/2/3K supported 16-bit input values. Recently, I decided to take my own crack at it, and came up with this: Code (Text): ; ------------------------------------------------------------------------- ; 2-argument arctangent (angle between (0,0) and (x,y)) ; Based on http://codebase64.org/doku.php?id=base:8bit_atan2_8-bit_angle ; ------------------------------------------------------------------------- ; PARAMETERS: ; d1.w - X value ; d2.w - Y value ; RETURNS: ; d0.b - 2-argument arctangent value (angle between (0,0) and (x,y)) ; ------------------------------------------------------------------------- CalcAngle: moveq #0,d0 ; Default to bottom right quadrant tst.w d1 ; Is the X value negative? beq.s CalcAngle_XZero ; If the X value is zero, branch bpl.s CalcAngle_CheckY ; If not, branch not.w d1 ; If so, get the absolute value moveq #4,d0 ; Shift to left quadrant CalcAngle_CheckY: tst.w d2 ; Is the Y value negative? beq.s CalcAngle_YZero ; If the Y value is zero, branch bpl.s CalcAngle_CheckOctet ; If not, branch not.w d2 ; If so, get the absolute value addq.b #2,d0 ; Shift to top quadrant CalcAngle_CheckOctet: cmp.w d2,d1 ; Are we horizontally closer to the center? bcc.s CalcAngle_Divide ; If not, branch exg.l d1,d2 ; If so, divide Y from X instead addq.b #1,d0 ; Use octant that's horizontally closer to the center CalcAngle_Divide: move.w d1,-(sp) ; Shrink X and Y down into bytes moveq #0,d3 move.b (sp)+,d3 move.b WordShiftTable(pc,d3.w),d3 lsr.w d3,d1 lsr.w d3,d2 lea Log2Table(pc),a2 ; Perform logarithmic division move.b (a2,d2.w),d2 sub.b (a2,d1.w),d2 bne.s CalcAngle_GetAtan2Val move.w #$FF,d2 ; Edge case where X and Y values are too close for the division to handle CalcAngle_GetAtan2Val: lea Atan2Table(pc),a2 ; Get atan2 value move.b (a2,d2.w),d2 move.b OctantAdjust(pc,d0.w),d0 eor.b d2,d0 rts ; ------------------------------------------------------------------------- CalcAngle_YZero: tst.b d0 ; Was the X value negated? beq.s CalcAngle_End ; If not, branch (d0 is already 0, so no need to set it again on branch) moveq #$FFFFFF80,d0 ; 180 degrees CalcAngle_End: rts CalcAngle_XZero: tst.w d2 ; Is the Y value negative? bmi.s CalcAngle_XZeroYNeg ; If so, branch moveq #$40,d0 ; 90 degrees rts CalcAngle_XZeroYNeg: moveq #$FFFFFFC0,d0 ; 270 degrees rts ; ------------------------------------------------------------------------- OctantAdjust: dc.b %00000000 ; +X, +Y, |X|>|Y| dc.b %00111111 ; +X, +Y, |X|<|Y| dc.b %11111111 ; +X, -Y, |X|>|Y| dc.b %11000000 ; +X, -Y, |X|<|Y| dc.b %01111111 ; -X, +Y, |X|>|Y| dc.b %01000000 ; -X, +Y, |X|<|Y| dc.b %10000000 ; -X, -Y, |X|>|Y| dc.b %10111111 ; -X, -Y, |X|<|Y| WordShiftTable: dc.b $00, $01, $02, $02, $03, $03, $03, $03 dc.b $04, $04, $04, $04, $04, $04, $04, $04 dc.b $05, $05, $05, $05, $05, $05, $05, $05 dc.b $05, $05, $05, $05, $05, $05, $05, $05 dc.b $06, $06, $06, $06, $06, $06, $06, $06 dc.b $06, $06, $06, $06, $06, $06, $06, $06 dc.b $06, $06, $06, $06, $06, $06, $06, $06 dc.b $06, $06, $06, $06, $06, $06, $06, $06 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 dc.b $07, $07, $07, $07, $07, $07, $07, $07 Log2Table: dc.b $00, $00, $1F, $32, $3F, $49, $52, $59 dc.b $5F, $64, $69, $6E, $72, $75, $79, $7C dc.b $7F, $82, $84, $87, $89, $8C, $8E, $90 dc.b $92, $94, $95, $97, $99, $9A, $9C, $9E dc.b $9F, $A0, $A2, $A3, $A4, $A6, $A7, $A8 dc.b $A9, $AA, $AC, $AD, $AE, $AF, $B0, $B1 dc.b $B2, $B3, $B4, $B5, $B5, $B6, $B7, $B8 dc.b $B9, $BA, $BA, $BB, $BC, $BD, $BE, $BE dc.b $BF, $C0, $C0, $C1, $C2, $C2, $C3, $C4 dc.b $C4, $C5, $C6, $C6, $C7, $C8, $C8, $C9 dc.b $C9, $CA, $CA, $CB, $CC, $CC, $CD, $CD dc.b $CE, $CE, $CF, $CF, $D0, $D0, $D1, $D1 dc.b $D2, $D2, $D3, $D3, $D4, $D4, $D5, $D5 dc.b $D5, $D6, $D6, $D7, $D7, $D8, $D8, $D8 dc.b $D9, $D9, $DA, $DA, $DA, $DB, $DB, $DC dc.b $DC, $DC, $DD, $DD, $DE, $DE, $DE, $DF dc.b $DF, $DF, $E0, $E0, $E0, $E1, $E1, $E1 dc.b $E2, $E2, $E2, $E3, $E3, $E3, $E4, $E4 dc.b $E4, $E5, $E5, $E5, $E6, $E6, $E6, $E7 dc.b $E7, $E7, $E8, $E8, $E8, $E8, $E9, $E9 dc.b $E9, $EA, $EA, $EA, $EA, $EB, $EB, $EB dc.b $EC, $EC, $EC, $EC, $ED, $ED, $ED, $ED dc.b $EE, $EE, $EE, $EE, $EF, $EF, $EF, $F0 dc.b $F0, $F0, $F0, $F1, $F1, $F1, $F1, $F1 dc.b $F2, $F2, $F2, $F2, $F3, $F3, $F3, $F3 dc.b $F4, $F4, $F4, $F4, $F5, $F5, $F5, $F5 dc.b $F5, $F6, $F6, $F6, $F6, $F7, $F7, $F7 dc.b $F7, $F7, $F8, $F8, $F8, $F8, $F8, $F9 dc.b $F9, $F9, $F9, $F9, $FA, $FA, $FA, $FA dc.b $FA, $FB, $FB, $FB, $FB, $FB, $FC, $FC dc.b $FC, $FC, $FC, $FD, $FD, $FD, $FD, $FD dc.b $FE, $FE, $FE, $FE, $FE, $FE, $FF, $FF Atan2Table: dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $00, $00 dc.b $00, $00, $00, $00, $00, $00, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $01, $01, $01, $01, $01, $01, $01 dc.b $01, $02, $02, $02, $02, $02, $02, $02 dc.b $02, $02, $02, $02, $02, $02, $02, $02 dc.b $02, $02, $02, $02, $02, $02, $02, $02 dc.b $03, $03, $03, $03, $03, $03, $03, $03 dc.b $03, $03, $03, $03, $03, $03, $03, $03 dc.b $04, $04, $04, $04, $04, $04, $04, $04 dc.b $04, $04, $04, $05, $05, $05, $05, $05 dc.b $05, $05, $05, $05, $05, $06, $06, $06 dc.b $06, $06, $06, $06, $06, $07, $07, $07 dc.b $07, $07, $07, $08, $08, $08, $08, $08 dc.b $08, $09, $09, $09, $09, $09, $09, $0A dc.b $0A, $0A, $0A, $0B, $0B, $0B, $0B, $0B dc.b $0C, $0C, $0C, $0C, $0D, $0D, $0D, $0D dc.b $0E, $0E, $0E, $0F, $0F, $0F, $0F, $10 dc.b $10, $10, $11, $11, $11, $12, $12, $12 dc.b $13, $13, $13, $14, $14, $14, $15, $15 dc.b $16, $16, $16, $17, $17, $17, $18, $18 dc.b $19, $19, $1A, $1A, $1A, $1B, $1B, $1C dc.b $1C, $1C, $1D, $1D, $1E, $1E, $1F, $1F What I did to get around the input argument limitation was basically just to get the largest of the 2 values (which is needed anyways to do the atan2 calculation properly), and then determine how many bit shifts it would take to get it down into a byte, and then apply that number of bit shifts to both the X and Y (hence, the introduction of WordShiftTable). I'm also using my own generated log2 and atan2 tables. It also takes into account some edge cases regarding the edges of octants. Yes, it's also not save d3 or a2 on the stack. In Sonic 1, it's not really necessary, I found. I dunno about Sonic 2 or 3K, but that shouldn't be hard to get around anyways. If there's any further optimizations that can be made, or if there's any issues that I somehow missed, I'd like to hear it. Currently, the worst case scenario for this function is like 246 cycles, and the best case scenario (excluding if both X and Y are 0) in the original is like around the 272 cycles mark.
Good stuff. I've been wanting new ways to cut down on CPU usage. Doesn't this leave the stack pointer offset by 1 byte?
Reading/writing a byte on the stack actually advances the stack pointer by 2, not 1, in an attempt to keep it aligned at an even address (although, it doesn't really matter which address it's on, it will ALWAYS advance 2 bytes). This can be abused to do a slightly faster multiplication or division by $100 with a word value than with bit-shifting (20 cycles vs. 22 cycles). You can save an additional 4 cycles by not doing a register clear, but, you gotta make sure that it's safe to do that.
Another small update: did a quick optimization of the edge case where Y = 0, and it picks between 0 and 180 degrees (thanks @Spicy Bread SSR for pointing this out).
