1 files changed, 119 insertions, 166 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-59.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-59.asm
index 9f1027e..e7bcae3 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-59.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-59.asm
@@ -6,186 +6,139 @@
 @in(r0.w)	in4
 @in(r1.x)	in5
 @in(r1.y)	in6
-@in(r1.z)	in8
-@in(r1.w)	in9
-@out(r3.x)	out0
-@out(r3.y)	out1
-@out(r3.z)	out2
-@out(r3.w)	out3
-@out(r7.y)	out4
-@out(r7.z)	out5
-@out(r7.w)	out6
-@out(r8.x)	out7
+@in(r4.z)	in8
+@in(r4.w)	in9
+@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
 @out(r2.x)	out8
 @out(r2.y)	out9
 @out(r2.z)	out10
 @out(r2.w)	out11
-@out(r8.y)	out12
-@out(r8.z)	out13
-@out(r8.w)	out14
-@out(r9.x)	out15
+@out(r3.x)	out12
+@out(r3.y)	out13
+@out(r3.z)	out14
+@out(r3.w)	out15
 @out(r4.x)	out16
 @out(r4.y)	out17
 @out(r4.z)	out18
 @out(r4.w)	out19
-@out(r6.y)	out20
-@out(r6.z)	out21
-@out(r6.w)	out22
-@out(r7.x)	out23
-(sy)(ss)mul.f r2.x, c12.x, r0.x
-mul.f r2.y, c12.x, r0.w
-mad.f32 r2.x, c13.x, r0.y, r2.x
-mad.f32 r2.y, c13.x, r1.x, r2.y
-mad.f32 r2.x, c14.x, r0.z, r2.x
-mul.f r2.z, c12.z, r0.x
-mul.f r2.w, c12.z, r0.w
-mul.f r0.w, c12.y, r0.w
-add.f r2.x, r2.x, c15.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r2.z, c13.z, r0.y, r2.z
-mad.f32 r2.w, c13.z, r1.x, r2.w
-add.f r3.x, c4.x, (neg)r2.x
-mad.f32 r2.y, c14.x, r1.y, r2.y
-mul.f r3.y, c8.w, r2.x
-mul.f r3.z, c8.z, r2.x
-mul.f r3.w, r3.x, r3.x
-mul.f r4.x, c12.y, r0.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r4.x, c13.y, r0.y, r4.x
-absneg.f r4.y, (neg)c5.x
-mad.f32 r4.x, c14.y, r0.z, r4.x
-mul.f r4.z, c8.y, r2.x
-mul.f r4.w, c8.x, r2.x
-mul.f r5.x, r2.y, r4.y
-add.f r4.x, r4.x, c15.y
-mad.f32 r0.w, c13.y, r1.x, r0.w
-mul.f r1.x, c0.w, r2.x
-mul.f r5.y, c0.z, r2.x
-add.f r5.z, c4.y, (neg)r4.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r3.y, c9.w, r4.x, r3.y
-mad.f32 r3.z, c9.z, r4.x, r3.z
-mad.f32 r3.w, r5.z, r5.z, r3.w
-mad.f32 r0.w, c14.y, r1.y, r0.w
-mad.f32 r2.z, c14.z, r0.z, r2.z
-mad.f32 r4.z, c9.y, r4.x, r4.z
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r0.w, r0.w
-add.f r2.z, r2.z, c15.z
-absneg.f r5.w, (neg)c5.y
-mad.f32 r4.w, c9.x, r4.x, r4.w
-mad.f32 r1.x, c1.w, r4.x, r1.x
-add.f r6.x, c4.z, (neg)r2.z
-mad.f32 r5.x, r0.w, r5.w, r5.x
-mad.f32 r3.y, c10.w, r2.z, r3.y
-mad.f32 r3.z, c10.z, r2.z, r3.z
-mad.f32 r3.w, r6.x, r6.x, r3.w
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r2.w, r2.w
+@out(r5.x)	out20
+@out(r5.y)	out21
+@out(r5.z)	out22
+@out(r5.w)	out23
+@const(c17.x)	0x00000000, 0x00000000, 0x00000000, 0x00000000
+@const(c18.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)mul.f r1.z, c12.x, r0.x
+mul.f r1.w, c12.x, r0.w
+mad.f32 r1.z, c13.x, r0.y, r1.z
+mad.f32 r1.w, c13.x, r1.x, r1.w
+mad.f32 r1.z, c14.x, r0.z, r1.z
+mad.f32 r1.w, c14.x, r1.y, r1.w
+absneg.f r2.x, (neg)c5.x
+mul.f r2.y, c12.y, r0.x
+add.f r2.z, r1.z, c15.x
+mad.f32 r1.z, c13.y, r0.y, r2.y
+mul.f r2.y, c12.z, r0.x
+mul.f r2.w, r1.w, r2.x
+add.f r3.x, c4.x, (neg)r2.z
+mul.f r3.y, c12.y, r0.w
+mul.f r3.z, c8.y, r2.z
+mul.f r3.w, c8.x, r2.z
+mul.f r4.x, r3.x, r3.x
+mad.f32 r1.z, c14.y, r0.z, r1.z
+mad.f32 r3.y, c13.y, r1.x, r3.y
+mul.f r4.y, c8.w, r2.z
+mul.f r5.x, c8.z, r2.z
+add.f r5.w, r1.z, c15.y
+mad.f32 r6.x, c14.y, r1.y, r3.y
+absneg.f r1.z, (neg)c5.y
+mul.f r3.y, c0.w, r2.z
+add.f r5.y, c4.y, (neg)r5.w
+mad.f32 r3.z, c9.y, r5.w, r3.z
+mad.f32 r3.w, c9.x, r5.w, r3.w
+mad.f32 r2.w, r6.x, r1.z, r2.w
+mad.f32 r4.x, r5.y, r5.y, r4.x
+mad.f32 r2.y, c13.z, r0.y, r2.y
+mul.f r0.w, c12.z, r0.w
+mad.f32 r2.y, c14.z, r0.z, r2.y
+mad.f32 r0.w, c13.z, r1.x, r0.w
+mad.f32 r1.x, c9.w, r5.w, r4.y
+mad.f32 r4.y, c9.z, r5.w, r5.x
+add.f r2.y, r2.y, c15.z
+mad.f32 r6.y, c14.z, r1.y, r0.w
+absneg.f r0.w, (neg)c5.z
+mad.f32 r1.y, c1.w, r5.w, r3.y
+add.f r3.y, c4.z, (neg)r2.y
+mad.f32 r3.z, c10.y, r2.y, r3.z
+mad.f32 r3.w, c10.x, r2.y, r3.w
+mad.f32 r2.w, r6.y, r0.w, r2.w
+mad.f32 r4.x, r3.y, r3.y, r4.x
 mul.f r0.x, c12.w, r0.x
-mad.f32 r4.z, c10.y, r2.z, r4.z
-mad.f32 r4.w, c10.x, r2.z, r4.w
-mad.f32 r1.y, c14.z, r1.y, r2.w
-rsq r2.w, r3.w
-(ss)mov.f32f32 r2.w, r2.w
-(ss)absneg.f r3.w, (neg)c5.z
+mad.f32 r5.x, c10.w, r2.y, r1.x
+mad.f32 r4.y, c10.z, r2.y, r4.y
 mad.f32 r0.x, c13.w, r0.y, r0.x
-mov.f32f32 r0.y, r1.y
-mad.f32 r1.y, r3.x, r2.w, r4.y
-mad.f32 r3.x, r5.z, r2.w, r5.w
-mad.f32 r2.w, r6.x, r2.w, r3.w
-mad.f32 r3.w, r0.y, r3.w, r5.x
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.w, r3.w
-mul.f r4.y, r1.y, r1.y
+mad.f32 r0.y, c2.w, r2.y, r1.y
+mul.f r6.z, c0.z, r2.z
+rsq r1.x, r4.x
+(ss)mov.f32f32 r1.y, r1.x
+mad.f32 r2.x, r3.x, r1.x, r2.x
+max.f r1.x, c17.x, r2.w
 mad.f32 r0.x, c14.w, r0.z, r0.x
-mad.f32 r0.z, r3.x, r3.x, r4.y
-max.f r3.w, c17.x, r3.w
-mad.f32 r1.x, c2.w, r2.z, r1.x
-mad.f32 r4.y, c1.z, r4.x, r5.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r3.w, r3.w
-mad.f32 r0.z, r2.w, r2.w, r0.z
+mad.f32 r0.z, r5.y, r1.y, r1.z
+mov.f32f32 r1.z, r2.x
+mad.f32 r0.w, r3.y, r1.y, r0.w
+nop
+mov.f32f32 r2.w, r0.z
+mul.f r1.y, r1.z, r1.z
+(ss)mov.f32f32 r4.x, r0.w
+mov.f32f32 r3.x, r1.x
+mad.f32 r0.z, r0.z, r2.w, r1.y
 add.f r0.x, r0.x, c15.w
-mad.f32 r4.y, c2.z, r2.z, r4.y
-mul.f r5.x, c0.y, r2.x
-mul.f r5.y, c0.x, r2.x
-mad.f32 r5.z, c7.x, r2.z, c7.y
-mad.f32 r2.x, c7.x, r2.x, c7.y
+mad.f32 r0.z, r0.w, r4.x, r0.z
+mul.f r1.z, r3.x, c6.z
+mul.f r1.y, r3.x, c6.y
+mad.f32 r0.w, c11.y, r0.x, r3.z
+mad.f32 r3.x, c11.x, r0.x, r3.w
+mul.f r1.x, r1.x, c6.x
+mad.f32 r3.w, c11.w, r0.x, r5.x
 rsq r0.z, r0.z
-(ss)mov.f32f32 r0.z, r0.z
-mul.f r5.w, r3.w, c6.z
-mul.f r6.x, r3.w, c6.y
-mul.f r3.w, r3.w, c6.x
-mul.f r2.w, r2.w, r0.z
-mul.f r3.x, r3.x, r0.z
-mul.f r0.z, r1.y, r0.z
-mov.f32f32 r1.y, r5.w
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r0.z, r0.z
-nop
-mov.f32f32 r6.w, r2.w
-mov.f32f32 r6.z, r3.x
-mov.f32f32 r6.y, r0.z
-mov.f32f32 r7.w, r1.y
-mov.f32f32 r0.z, r6.x
-mov.f32f32 r1.y, r3.w
-(rpt1)nop
-mov.f32f32 r7.z, r0.z
-mov.f32f32 r7.y, r1.y
-mad.f32 r0.z, c11.w, r0.x, r3.y
-mad.f32 r1.y, c11.z, r0.x, r3.z
-mad.f32 r2.w, c11.y, r0.x, r4.z
-mad.f32 r3.x, c11.x, r0.x, r4.w
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.y, r1.y
-mul.f r2.w, r2.w, c16.y
+(ss)mov.f32f32 r3.z, r0.z
+mul.f r5.x, r2.x, r0.z
+mul.f r3.y, r0.w, c16.y
 mul.f r3.x, r3.x, c16.x
-mov.f32f32 r9.x, r0.z
-mov.f32f32 r8.w, r1.y
-mov.f32f32 r8.z, r2.w
-mov.f32f32 r8.y, r3.x
-mad.f32 r0.z, c3.w, r0.x, r1.x
-mad.f32 r1.x, c3.z, r0.x, r4.y
-mad.f32 r1.y, c1.y, r4.x, r5.x
-mad.f32 r2.w, c1.x, r4.x, r5.y
-mov.f32f32 r3.w, r0.z
-mov.f32f32 r3.z, r1.x
-mad.f32 r0.z, c2.y, r2.z, r1.y
-mad.f32 r1.x, c2.x, r2.z, r2.w
-mad.f32 r0.z, c3.y, r0.x, r0.z
-mad.f32 r0.x, c3.x, r0.x, r1.x
-mov.f32f32 r1.x, r5.z
-mov.f32f32 r1.y, r2.x
-mov.f32f32 r3.y, r0.z
-mov.f32f32 r3.x, r0.x
-mov.f32f32 r4.y, r1.x
-mov.f32f32 r4.x, r1.y
-mov.f32f32 r0.x, r0.y
-mov.f32f32 r0.y, r0.w
-mov.f32f32 r0.z, r2.y
-nop
-mov.f32f32 r2.z, r0.x
-mov.f32f32 r2.y, r0.y
-mov.f32f32 r2.x, r0.z
-mov.f32f32 r0.x, (0.000000)
-mov.f32f32 r0.y, r1.w
-mov.f32f32 r0.z, r1.z
-mov.f32f32 r0.w, (0.000000)
-mov.f32f32 r7.x, r0.x
-mov.f32f32 r4.w, r0.y
-mov.f32f32 r4.z, r0.z
-mov.f32f32 r2.w, r0.w
-mov.f32f32 r0.x, (0.000000)
-(rpt2)nop
-mov.f32f32 r8.x, r0.x
+mul.f r5.z, r4.x, r3.z
+mul.f r5.y, r2.w, r3.z
+mad.f32 r3.z, c11.z, r0.x, r4.y
+mad.f32 r0.w, c3.w, r0.x, r0.y
+mad.f32 r0.y, c1.z, r5.w, r6.z
+(ss)mul.f r0.z, c0.y, r2.z
+mad.f32 r0.y, c2.z, r2.y, r0.y
+mad.f32 r2.x, c1.y, r5.w, r0.z
+mad.f32 r0.z, c3.z, r0.x, r0.y
+mad.f32 r0.y, c2.y, r2.y, r2.x
+mul.f r2.x, c0.x, r2.z
+mad.f32 r0.y, c3.y, r0.x, r0.y
+mad.f32 r2.x, c1.x, r5.w, r2.x
+mad.f32 r4.y, c7.x, r2.y, c7.y
+mad.f32 r2.x, c2.x, r2.y, r2.x
+mad.f32 r4.x, c7.x, r2.z, c7.y
+mad.f32 r0.x, c3.x, r0.x, r2.x
+mov.f32f32 r2.z, r6.y
+mov.f32f32 r2.y, r6.x
+mov.f32f32 r2.x, r1.w
+mov.f32f32 r5.w, (0.000000)
+mov.f32f32 r2.w, (0.000000)
+mov.f32f32 r1.w, (0.000000)
 end
 nop
-; VERT: outputs: r3.x (0:0) r7.y (5:9) r2.x (5:10) r8.y (5:11) r4.x (5:12) r6.y (5:13)
-; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r0.w (0:0,cm=7,il=12,b=0) r1.z (0:0,cm=3,il=16,b=0)
-; VERT: 154 instructions, 0 half, 10 full
-; pos: r3.x
+nop
+; VERT: outputs: r0.x (0:0) r1.x (5:9) r2.x (5:10) r3.x (5:11) r4.x (5:12) r5.x (5:13)
+; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r0.w (0:0,cm=7,il=12,b=0) r4.z (0:0,cm=3,il=16,b=0)
+; VERT: 102 instructions, 0 half, 7 full