1 files changed, 107 insertions, 143 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-22.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-22.asm
index 08421de..f362ccd 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-22.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-22.asm
@@ -3,170 +3,134 @@
 @in(r0.x)	in0
 @in(r0.y)	in1
 @in(r0.z)	in2
-@in(r0.w)	in4
-@in(r1.x)	in5
-@out(r3.x)	out0
-@out(r3.y)	out1
-@out(r3.z)	out2
-@out(r3.w)	out3
-@out(r2.x)	out4
-@out(r2.y)	out5
-@out(r2.z)	out6
-@out(r2.w)	out7
-(sy)(ss)floor.f r1.y, c11.z
+@in(r1.x)	in4
+@in(r1.y)	in5
+@out(r0.x)	out0
+@out(r0.y)	out1
+@out(r0.z)	out2
+@out(r0.w)	out3
+@out(r1.x)	out4
+@out(r1.y)	out5
+@out(r1.z)	out6
+@out(r1.w)	out7
+@const(c15.x)	0x3ecccccd, 0x00000000, 0x3eaaa64c, 0x3e800000
+@const(c16.x)	0x3f000000, 0x40000000, 0xbf800000, 0x40400000
+@const(c17.x)	0x399d4952, 0x3e4ccccd, 0x3f99999a, 0x3fd9999a
+@const(c18.x)	0x3c88ce70, 0x00000000, 0x00000000, 0x00000000
+@const(c19.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)floor.f r0.w, c11.z
 floor.f r1.z, c11.x
 absneg.f r1.w, (abs)c14.x
 absneg.f r2.x, (abs)c14.y
-add.f r1.y, c11.z, (neg)r1.y
+add.f r0.w, c11.z, (neg)r0.w
 add.f r1.z, c11.x, (neg)r1.z
 mul.f r2.y, r0.x, r0.z
 add.f r1.w, r1.w, r2.x
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.z, r1.z
-mul.f r2.x, r0.y, c17.x
-mov.f32f32 r1.w, r1.w
-max.f r1.y, r1.y, c15.y
+max.f r0.w, r0.w, c15.y
 max.f r1.z, r1.z, c15.y
-mul.f r2.x, r2.y, r2.x
-mul.f r2.y, c13.x, r1.w
-min.f r1.y, r1.y, c19.y
+mul.f r2.x, r0.y, c17.x
+mul.f r2.z, c13.x, r1.w
+min.f r0.w, r0.w, c19.y
 min.f r1.z, r1.z, c19.y
-mul.f r2.z, c8.z, r0.x
-mov.f32f32 r2.x, r2.x
-max.f r1.y, r1.y, c15.x
+mul.f r2.w, c8.z, r0.x
+mov.f32f32 r1.w, r1.w
+max.f r0.w, r0.w, c15.x
 max.f r1.z, r1.z, c15.x
-mad.f32 r2.z, c9.z, r0.y, r2.z
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.z, r1.z
-mad.f32 r2.z, c10.z, r0.z, r2.z
-mul.f r2.w, c8.x, r0.x
-mul.f r1.y, c13.x, r1.y
-mad.f32 r2.w, c9.x, r0.y, r2.w
-add.f r2.z, r2.z, c11.z
-mad.f32 r2.w, c10.x, r0.z, r2.w
-mov.f32f32 r1.y, r1.y
-mad.f32 r2.y, c15.w, r2.y, r2.z
-mov.f32f32 r2.x, r2.x
-add.f r2.w, r2.w, c11.x
-mad.f32 r1.y, c15.z, r1.y, c11.x
-mad.f32 r1.z, c13.x, r1.z, r2.w
-mov.f32f32 r2.y, r2.y
-max.f r2.x, r2.x, c15.y
-mov.f32f32 r1.y, r1.y
+mul.f r3.x, c8.x, r0.x
+mad.f32 r2.w, c9.z, r0.y, r2.w
+mul.f r0.w, c13.x, r0.w
+mad.f32 r3.x, c9.x, r0.y, r3.x
+mad.f32 r2.w, c10.z, r0.z, r2.w
+mad.f32 r3.x, c10.x, r0.z, r3.x
+mad.f32 r0.w, c15.z, r0.w, c11.x
+max.f r1.w, r1.w, c17.z
+mul.f r2.x, r2.y, r2.x
+add.f r2.y, r3.x, c11.x
+add.f r0.w, r0.w, c16.x
+mad.f32 r1.z, c13.x, r1.z, r2.y
+add.f r2.w, r2.w, c11.z
+min.f r1.w, r1.w, c17.w
+floor.f r3.x, r0.w
 add.f r1.z, r1.z, c16.x
-mov.f32f32 r2.y, r2.y
-min.f r2.x, r2.x, c19.y
-mov.f32f32 r1.y, r1.y
+mad.f32 r2.z, c15.w, r2.z, r2.w
+mov.f32f32 r3.y, r1.w
+add.f r0.w, r0.w, (neg)r3.x
 floor.f r3.x, r1.z
-add.f r2.y, r2.y, c16.x
-min.f r2.x, r2.x, c17.y
-add.f r1.y, r1.y, c16.x
+add.f r2.z, r2.z, c16.x
+max.f r2.x, r2.x, c15.y
+mad.f32 r0.w, c16.y, r0.w, c16.z
 add.f r1.z, r1.z, (neg)r3.x
-floor.f r3.x, r2.y
-mov.f32f32 r2.x, r2.x
-floor.f r3.y, r1.y
+floor.f r3.x, r2.z
+min.f r2.x, r2.x, c19.y
+absneg.f r0.w, (abs)r0.w
 mad.f32 r1.z, c16.y, r1.z, c16.z
-add.f r2.y, r2.y, (neg)r3.x
-mul.f r3.x, r0.y, c18.x
-add.f r1.y, r1.y, (neg)r3.y
+add.f r2.z, r2.z, (neg)r3.x
+min.f r2.x, r2.x, c17.y
+mul.f r3.x, c16.y, r0.w
 absneg.f r1.z, (abs)r1.z
-mad.f32 r2.y, c16.y, r2.y, c16.z
-mov.f32f32 r3.x, r3.x
-mad.f32 r1.y, c16.y, r1.y, c16.z
-mul.f r3.y, c16.y, r1.z
-absneg.f r2.y, (abs)r2.y
+mad.f32 r2.z, c16.y, r2.z, c16.z
+mul.f r0.w, r0.w, r0.w
+add.f r3.x, c16.w, (neg)r3.x
+mul.f r3.z, c16.y, r1.z
+absneg.f r2.z, (abs)r2.z
 mul.f r1.z, r1.z, r1.z
-absneg.f r1.y, (abs)r1.y
-add.f r3.y, c16.w, (neg)r3.y
-mul.f r3.z, c16.y, r2.y
-mul.f r2.y, r2.y, r2.y
-mul.f r3.w, c16.y, r1.y
-mul.f r1.z, r1.z, r3.y
-add.f r3.y, c16.w, (neg)r3.z
-mul.f r1.y, r1.y, r1.y
-add.f r3.z, c16.w, (neg)r3.w
-mov.f32f32 r1.z, r1.z
-mul.f r2.y, r2.y, r3.y
-mov.f32f32 r3.x, r3.x
-mul.f r1.y, r1.y, r3.z
-mul.f r1.z, r1.z, r2.x
-mov.f32f32 r2.y, r2.y
+mul.f r0.w, r0.w, r3.x
+mul.f r3.x, r0.y, c18.x
+add.f r3.z, c16.w, (neg)r3.z
+mul.f r3.w, c16.y, r2.z
+mul.f r2.z, r2.z, r2.z
 max.f r3.x, r3.x, c15.y
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.z, r1.z
-mul.f r2.x, r2.y, r2.x
-min.f r2.y, r3.x, c19.y
-max.f r1.w, r1.w, c17.z
-mul.f r3.x, c8.y, r0.x
+mul.f r1.z, r1.z, r3.z
+mov.f32f32 r3.z, r2.x
+add.f r3.w, c16.w, (neg)r3.w
+min.f r3.x, r3.x, c19.y
+mul.f r4.x, c8.y, r0.x
 mul.f r0.x, c8.w, r0.x
-min.f r2.y, r2.y, c15.w
-mov.f32f32 r1.w, r1.w
-mad.f32 r3.x, c9.y, r0.y, r3.x
+mad.f32 r4.x, c9.y, r0.y, r4.x
+min.f r3.x, r3.x, c15.w
+mul.f r1.z, r1.z, r3.z
+mul.f r2.z, r2.z, r3.w
+mad.f32 r3.z, c10.y, r0.z, r4.x
+mul.f r0.w, r0.w, r3.x
+mov.f32f32 r3.x, r1.z
+mul.f r2.x, r2.z, r2.x
+add.f r2.z, r3.z, c11.y
+mov.f32f32 r3.z, r0.w
+mad.f32 r0.w, c14.y, r0.w, r1.z
+mad.f32 r1.z, r2.x, r1.w, r2.w
+mad.f32 r1.w, r2.x, r3.y, r2.y
+mad.f32 r2.y, c14.x, r3.z, r3.x
+mad.f32 r2.x, r2.x, r3.y, r2.z
 mad.f32 r0.x, c9.w, r0.y, r0.x
-mov.f32f32 r0.y, r2.y
-min.f r1.w, r1.w, c17.w
-mad.f32 r2.y, c10.y, r0.z, r3.x
-mad.f32 r0.x, c10.w, r0.z, r0.x
-mul.f r0.y, r1.y, r0.y
-mov.f32f32 r0.z, r1.w
-add.f r1.y, r2.y, c11.y
-add.f r0.x, r0.x, c11.w
-mov.f32f32 r0.y, r0.y
-mad.f32 r1.w, r2.x, r0.z, r2.w
-mad.f32 r1.y, r2.x, r0.z, r1.y
-mad.f32 r0.z, r2.x, r0.z, r2.z
-mad.f32 r2.x, c14.x, r0.y, r1.z
-mad.f32 r0.y, c14.y, r0.y, r1.z
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r2.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.y, r1.x
-mov.f32f32 r2.x, r0.w
-add.f r0.w, r1.w, r1.z
-add.f r0.y, r0.z, r0.y
-(rpt1)nop
-mov.f32f32 r0.z, r0.w
-mov.f32f32 r0.y, r0.y
-(rpt1)nop
-mul.f r0.w, c0.w, r0.z
-mul.f r1.x, c0.z, r0.z
-mad.f32 r0.w, c1.w, r1.y, r0.w
-mad.f32 r1.x, c1.z, r1.y, r1.x
-mul.f r1.z, c0.y, r0.z
-mul.f r0.z, c0.x, r0.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.w, c2.w, r0.y, r0.w
-mad.f32 r1.x, c2.z, r0.y, r1.x
-mad.f32 r1.w, c3.w, r0.x, r0.w
-mad.f32 r3.x, c3.z, r0.x, r1.x
-mad.f32 r1.z, c1.y, r1.y, r1.z
-mad.f32 r0.z, c1.x, r1.y, r0.z
-mov.f32f32 r1.y, r1.w
-mov.f32f32 r1.w, r3.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r3.w, r1.y
-mov.f32f32 r3.z, r1.w
-mad.f32 r1.y, c2.y, r0.y, r1.z
-mad.f32 r0.y, c2.x, r0.y, r0.z
-mad.f32 r0.z, c3.y, r0.x, r1.y
-mad.f32 r0.x, c3.x, r0.x, r0.y
-mov.f32f32 r2.w, r0.w
-mov.f32f32 r2.z, r1.x
-mov.f32f32 r0.y, r0.z
-mov.f32f32 r0.x, r0.x
+nop
+add.f r0.y, r1.w, r2.y
+add.f r2.y, r1.z, r0.w
 (rpt1)nop
-mov.f32f32 r3.y, r0.y
-mov.f32f32 r3.x, r0.x
+mov.f32f32 r0.w, r0.y
+mul.f r0.y, c0.x, r0.y
+mov.f32f32 r2.z, r2.y
+mad.f32 r0.x, c10.w, r0.z, r0.x
+mul.f r0.z, c0.w, r0.w
+mul.f r1.z, c0.z, r0.w
+mad.f32 r0.z, c1.w, r2.x, r0.z
+mad.f32 r1.z, c1.z, r2.x, r1.z
+mad.f32 r1.w, c2.w, r2.z, r0.z
+add.f r2.w, r0.x, c11.w
+mad.f32 r1.z, c2.z, r2.z, r1.z
+mul.f r0.x, c0.y, r0.w
+mad.f32 r0.y, c1.x, r2.x, r0.y
+mad.f32 r0.w, c3.w, r2.w, r1.w
+mad.f32 r0.z, c3.z, r2.w, r1.z
+mad.f32 r0.x, c1.y, r2.x, r0.x
+mad.f32 r0.y, c2.x, r2.y, r0.y
+mad.f32 r2.x, c2.y, r2.z, r0.x
+mad.f32 r0.x, c3.x, r2.w, r0.y
+mad.f32 r0.y, c3.y, r2.w, r2.x
 end
 nop
 nop
 nop
-; VERT: outputs: r3.x (0:0) r2.x (5:9)
-; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r0.w (0:0,cm=3,il=12,b=0)
-; VERT: 152 instructions, 0 half, 4 full
-; pos: r3.x
+; VERT: outputs: r0.x (0:0) r1.x (5:9)
+; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r1.x (0:0,cm=3,il=12,b=0)
+; VERT: 110 instructions, 0 half, 5 full