1 files changed, 143 insertions, 210 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-41.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-41.asm
index d73d764..5a712c6 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-41.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-41.asm
@@ -1,218 +1,151 @@
 ; options:
 ; VERT: new compiler
-@in(r0.x)	in0
-@in(r0.y)	in1
-@in(r0.z)	in2
-@in(r0.w)	in4
-@in(r1.x)	in5
-@in(r1.y)	in6
-@in(r1.z)	in8
-@in(r1.w)	in9
-@in(r2.x)	in10
-@in(r2.y)	in12
-@in(r2.z)	in13
-@out(r4.w)	out0
-@out(r5.x)	out1
-@out(r5.y)	out2
-@out(r5.z)	out3
-@out(r3.y)	out4
-@out(r3.z)	out5
-@out(r3.w)	out6
-@out(r4.x)	out7
-@out(r8.x)	out8
-@out(r8.y)	out9
-@out(r8.z)	out10
-@out(r8.w)	out11
-@out(r9.x)	out12
-@out(r9.y)	out13
-@out(r9.z)	out14
-@out(r9.w)	out15
-@out(r6.x)	out16
-@out(r6.y)	out17
-@out(r6.z)	out18
-@out(r6.w)	out19
-@out(r7.x)	out20
-@out(r7.y)	out21
-@out(r7.z)	out22
-@out(r7.w)	out23
-@out(r11.x)	out24
-@out(r11.y)	out25
-@out(r11.z)	out26
-@out(r11.w)	out27
-@out(r10.x)	out28
-@out(r10.y)	out29
-@out(r10.z)	out30
-@out(r10.w)	out31
-(sy)(ss)add.f r2.w, c4.x, (neg)r0.x
-mul.f r3.x, r0.w, r0.w
-absneg.f r3.y, (neg)c7.y
-mul.f r3.z, r0.x, c7.x
-mul.f r3.w, r2.w, r2.w
-add.f r4.x, c4.y, (neg)r0.y
-add.f r3.x, c14.x, (neg)r3.x
-mul.f r4.y, r0.x, (neg)r3.y
-mad.f32 r3.y, r0.z, (neg)r3.y, r3.z
-mad.f32 r3.z, r4.x, r4.x, r3.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.w, r4.y
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r3.z, r3.z
-add.f r4.y, c4.z, (neg)r0.z
-mul.f r4.z, r3.x, r3.x
-mul.f r4.w, r1.x, r0.w
-mad.f32 r3.w, (neg)c7.x, r0.z, r3.w
-mad.f32 r3.z, r4.y, r4.y, r3.z
-mov.f32f32 r3.y, r3.y
-mul.f r5.x, c9.w, r0.x
-mul.f r5.y, c9.z, r0.x
-mul.f r5.z, c9.y, r0.x
-mul.f r5.w, c9.x, r0.x
-mov.f32f32 r3.w, r3.w
-rsq r3.z, r3.z
-(ss)mov.f32f32 r3.z, r3.z
-add.f r4.w, c14.y, (neg)r4.w
-mov.f32f32 r6.x, r3.y
-mad.f32 r3.y, c10.w, r0.y, r5.x
-mad.f32 r2.w, r2.w, r3.z, (neg)c5.x
-mov.f32f32 r4.w, r4.w
-mad.f32 r4.x, r4.x, r3.z, (neg)c5.y
-mad.f32 r3.z, r4.y, r3.z, (neg)c5.z
-mov.f32f32 r2.w, r2.w
-mad.f32 r4.y, r4.w, r4.w, r4.z
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r3.z, r3.z
-mul.f r4.z, r2.w, r2.w
-mov.f32f32 r4.y, r4.y
-mad.f32 r4.z, r4.x, r4.x, r4.z
-mul.f r5.x, r1.y, r0.w
-mov.f32f32 r3.w, r3.w
-mad.f32 r3.y, c11.w, r0.z, r3.y
-mov.f32f32 r4.z, r4.z
-add.f r5.x, c14.y, (neg)r5.x
-mad.f32 r4.z, r3.z, r3.z, r4.z
-mov.f32f32 r3.w, r3.w
-(rpt4)nop
-rsq r4.z, r4.z
-(ss)mov.f32f32 r4.z, r4.z
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r6.y, r3.w
-add.f r3.y, r3.y, c12.w
-mul.f r3.z, r3.z, r4.z
-mul.f r3.w, r4.x, r4.z
-mul.f r2.w, r2.w, r4.z
-mad.f32 r4.x, r5.x, r5.x, r4.y
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r2.w, r2.w
-nop
-mov.f32f32 r7.w, r3.z
-mov.f32f32 r7.z, r3.w
-mov.f32f32 r7.y, r2.w
-rsq r2.w, r4.x
-(ss)mov.f32f32 r2.w, r2.w
-(ss)mov.f32f32 r4.x, r3.y
-mad.f32 r4.y, c10.z, r0.y, r5.y
-mad.f32 r4.z, c10.y, r0.y, r5.z
-mul.f r3.x, r3.x, r2.w
-mul.f r5.x, r5.x, r2.w
-mul.f r2.w, r4.w, r2.w
-mad.f32 r4.y, c11.z, r0.z, r4.y
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r4.w, r5.x
-mov.f32f32 r2.w, r2.w
-add.f r4.y, r4.y, c12.z
-mul.f r5.x, r1.x, r3.x
-mul.f r5.y, r0.w, r4.w
-mad.f32 r5.x, r0.w, r2.w, (neg)r5.x
-mad.f32 r5.y, r1.y, r3.x, (neg)r5.y
-mul.f r5.z, r1.y, r2.w
-mov.f32f32 r8.x, r4.w
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r5.y, r5.y
-mad.f32 r4.w, r1.x, r4.w, (neg)r5.z
-mov.f32f32 r8.z, r8.x
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r5.y, r5.y
-mov.f32f32 r4.w, r4.w
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r6.w, r5.y
-mov.f32f32 r4.w, r4.w
-nop
-mov.f32f32 r7.x, r5.x
-mov.f32f32 r8.y, r2.w
-mov.f32f32 r6.z, r4.w
-mov.f32f32 r2.w, r3.x
-mov.f32f32 r3.x, r4.y
-mad.f32 r4.y, c11.y, r0.z, r4.z
-mad.f32 r4.z, c10.x, r0.y, r5.w
-mov.f32f32 r8.x, r2.w
-mov.f32f32 r3.w, r3.x
-add.f r2.w, r4.y, c12.y
-mad.f32 r3.x, c11.x, r0.z, r4.z
-mul.f r4.y, c0.w, r0.x
-mul.f r4.z, c0.z, r0.x
-mul.f r2.w, r2.w, c13.y
-add.f r3.x, r3.x, c12.x
-mad.f32 r4.y, c1.w, r0.y, r4.y
-mad.f32 r4.z, c1.z, r0.y, r4.z
-mov.f32f32 r3.z, r2.w
-mul.f r2.w, r3.x, c13.x
-mad.f32 r3.x, c2.w, r0.z, r4.y
-mad.f32 r4.y, c2.z, r0.z, r4.z
-mul.f r4.z, c0.y, r0.x
-mov.f32f32 r3.y, r2.w
-add.f r2.w, r3.x, c3.w
-add.f r3.x, r4.y, c3.z
-mad.f32 r4.y, c1.y, r0.y, r4.z
-mul.f r4.z, c0.x, r0.x
-mov.f32f32 r5.z, r2.w
-mov.f32f32 r5.y, r3.x
-mad.f32 r2.w, c2.y, r0.z, r4.y
-mad.f32 r0.y, c1.x, r0.y, r4.z
-mad.f32 r3.x, c8.x, r0.z, c8.y
-mad.f32 r0.x, c8.x, r0.x, c8.y
-add.f r2.w, r2.w, c3.y
-mad.f32 r0.y, c2.x, r0.z, r0.y
-mov.f32f32 r0.z, r3.x
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r5.x, r2.w
-add.f r0.y, r0.y, c3.x
-mov.f32f32 r9.w, r0.z
-mov.f32f32 r9.z, r0.x
-mov.f32f32 r0.x, (0.000000)
-mov.f32f32 r4.w, r0.y
-mov.f32f32 r0.y, (0.000000)
-mov.f32f32 r0.z, r1.y
-mov.f32f32 r10.w, r0.x
-mov.f32f32 r0.x, r1.x
-mov.f32f32 r10.z, r0.y
-mov.f32f32 r10.y, r0.z
-mov.f32f32 r0.y, r0.w
-mov.f32f32 r10.x, r0.x
-mul.f r0.x, r2.x, c6.z
-mul.f r0.z, r1.w, c6.y
-mov.f32f32 r11.w, r0.y
-mul.f r0.y, r1.z, c6.x
-mov.f32f32 r11.z, r0.x
-mov.f32f32 r11.y, r0.z
+@in(r8.x)	in0
+@in(r8.y)	in1
+@in(r8.z)	in2
+@in(r6.w)	in4
+@in(r7.x)	in5
+@in(r7.y)	in6
+@in(r8.w)	in8
+@in(r9.x)	in9
+@in(r9.y)	in10
+@in(r3.x)	in12
+@in(r3.y)	in13
+@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(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(r5.x)	out20
+@out(r5.y)	out21
+@out(r5.z)	out22
+@out(r5.w)	out23
+@out(r6.x)	out24
+@out(r6.y)	out25
+@out(r6.z)	out26
+@out(r6.w)	out27
+@out(r7.x)	out28
+@out(r7.y)	out29
+@out(r7.z)	out30
+@out(r7.w)	out31
+@const(c14.x)	0x3f800000, 0x00000000, 0xbf800000, 0x00000000
+@const(c15.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)add.f r0.x, c4.x, (neg)r8.x
+mul.f r0.y, r6.w, r6.w
+mul.f r0.z, c9.y, r8.x
+mul.f r0.w, c9.x, r8.x
+mul.f r1.x, r0.x, r0.x
+add.f r1.z, c4.y, (neg)r8.y
+add.f r0.y, c14.x, (neg)r0.y
+mad.f32 r0.z, c10.y, r8.y, r0.z
+mad.f32 r0.w, c10.x, r8.y, r0.w
+mad.f32 r1.x, r1.z, r1.z, r1.x
+add.f r1.w, c4.z, (neg)r8.z
+mov.f32f32 r1.y, r0.y
+mad.f32 r0.z, c11.y, r8.z, r0.z
+mad.f32 r0.w, c11.x, r8.z, r0.w
+mad.f32 r1.x, r1.w, r1.w, r1.x
+mul.f r2.x, r1.y, r1.y
+mul.f r1.y, r7.x, r6.w
+add.f r0.z, r0.z, c12.y
+add.f r0.w, r0.w, c12.x
+mul.f r2.y, r7.y, r6.w
+mul.f r2.z, c9.w, r8.x
+rsq r1.x, r1.x
+(ss)mov.f32f32 r2.w, r1.x
+add.f r3.z, c14.y, (neg)r1.y
+mad.f32 r0.x, r0.x, r1.x, (neg)c5.x
+mul.f r1.y, r0.z, c13.y
+mad.f32 r0.z, r1.z, r2.w, (neg)c5.y
+mov.f32f32 r1.z, r3.z
+(ss)mov.f32f32 r1.x, r0.x
+mad.f32 r1.w, r1.w, r2.w, (neg)c5.z
+mov.f32f32 r2.w, r0.z
+mad.f32 r2.x, r3.z, r1.z, r2.x
+mul.f r1.x, r1.x, r1.x
+add.f r2.y, c14.y, (neg)r2.y
+mad.f32 r0.z, r0.z, r2.w, r1.x
+mov.f32f32 r3.z, r1.w
+mul.f r1.x, r0.w, c13.x
+mov.f32f32 r0.w, r2.y
+mad.f32 r2.z, c10.w, r8.y, r2.z
+mad.f32 r0.z, r1.w, r3.z, r0.z
+mad.f32 r1.w, c11.w, r8.z, r2.z
+mul.f r2.z, c9.z, r8.x
+mul.f r3.w, c0.w, r8.x
+mul.f r4.x, c0.z, r8.x
+mul.f r4.y, c0.y, r8.x
+mul.f r6.x, c0.x, r8.x
+rsq r0.z, r0.z
+(ss)mov.f32f32 r4.z, r0.z
+mul.f r5.y, r0.x, r0.z
+mad.f32 r0.x, r2.y, r0.w, r2.x
+add.f r1.w, r1.w, c12.w
+mul.f r5.w, r3.z, r4.z
+mul.f r5.z, r2.w, r4.z
+(ss)mad.f32 r0.z, c10.z, r8.y, r2.z
+mad.f32 r2.y, c1.w, r8.y, r3.w
+mad.f32 r2.w, c1.z, r8.y, r4.x
+rsq r0.x, r0.x
+(ss)mov.f32f32 r3.z, r0.x
+mul.f r2.x, r0.y, r0.x
+(ss)mad.f32 r0.x, c11.z, r8.z, r0.z
+mad.f32 r0.y, c2.w, r8.z, r2.y
+mul.f r2.z, r0.w, r3.z
+mul.f r2.y, r1.z, r3.z
+mov.f32f32 r0.z, r2.x
+add.f r1.z, r0.x, c12.z
 mov.f32f32 r0.x, r2.z
-mov.f32f32 r11.x, r0.y
-mov.f32f32 r0.y, r2.y
-mov.f32f32 r0.z, c14.z
-mov.f32f32 r9.y, r0.x
+mov.f32f32 r3.z, r2.y
+mul.f r3.w, r7.x, r0.z
+add.f r0.w, r0.y, c3.w
+mul.f r0.y, r6.w, r0.x
+mul.f r4.x, r7.y, r3.z
+mad.f32 r4.w, r7.y, r0.z, (neg)r0.y
+mad.f32 r4.z, r7.x, r0.x, (neg)r4.x
+mad.f32 r5.x, r6.w, r3.z, (neg)r3.w
+mad.f32 r0.x, c2.z, r8.z, r2.w
+mad.f32 r0.y, c1.y, r8.y, r4.y
+mad.f32 r2.w, c1.x, r8.y, r6.x
+absneg.f r3.z, (neg)c7.y
+add.f r0.z, r0.x, c3.z
+mad.f32 r0.x, c2.y, r8.z, r0.y
+mad.f32 r2.w, c2.x, r8.z, r2.w
+mul.f r3.w, r8.x, (neg)r3.z
 nop
-mov.f32f32 r9.x, r0.y
-mov.f32f32 r8.w, r0.z
+add.f r0.y, r0.x, c3.y
+add.f r0.x, r2.w, c3.x
+mad.f32 r4.y, (neg)c7.x, r8.z, r3.w
+mul.f r2.w, r8.x, c7.x
+mov.f32f32 r7.w, (0.000000)
+mad.f32 r4.x, r8.z, (neg)r3.z, r2.w
+mov.f32f32 r7.z, (0.000000)
+mul.f r6.z, r9.y, c6.z
+mul.f r6.y, r9.x, c6.y
+mul.f r6.x, r8.w, c6.x
+mad.f32 r3.w, c8.x, r8.z, c8.y
+mad.f32 r3.z, c8.x, r8.x, c8.y
+mov.f32f32 r2.w, c14.z
 end
 nop
 nop
-; VERT: outputs: r4.w (0:0) r3.y (5:9) r8.x (5:10) r9.x (5:11) r6.x (5:12) r7.x (5:13) r11.x (5:14) r10.x (5:15)
-; 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=7,il=16,b=0) r2.y (0:0,cm=3,il=20,b=0)
-; VERT: 170 instructions, 0 half, 12 full
-; pos: r4.w
+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) r6.x (5:14) r7.x (5:15)
+; VERT: inputs: r8.x (0:0,cm=7,il=8,b=0) r6.w (0:0,cm=7,il=12,b=0) r8.w (0:0,cm=7,il=16,b=0) r3.x (0:0,cm=3,il=20,b=0)
+; VERT: 97 instructions, 0 half, 10 full