1 files changed, 167 insertions, 227 deletions

diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-67.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-67.asm
index b65a363..63ee7bc 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-67.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-67.asm
@@ -6,242 +6,182 @@
 @in(r0.w)	in4
 @in(r1.x)	in5
 @in(r1.y)	in6
-@in(r1.z)	in8
-@in(r1.w)	in9
-@out(r5.x)	out0
-@out(r5.y)	out1
-@out(r5.z)	out2
-@out(r5.w)	out3
-@out(r4.x)	out4
-@out(r4.y)	out5
-@out(r4.z)	out6
-@out(r4.w)	out7
-@out(r2.z)	out8
-@out(r2.w)	out9
-@out(r3.x)	out10
-@out(r3.y)	out11
-@out(r0.z)	out12
-@out(r0.w)	out13
-@out(r1.x)	out14
-@out(r1.y)	out15
-(sy)(ss)floor.f r2.x, c14.z
-floor.f r2.y, c14.x
-absneg.f r2.z, (abs)c17.x
-absneg.f r2.w, (abs)c17.y
-add.f r2.x, c14.z, (neg)r2.x
-add.f r2.y, c14.x, (neg)r2.y
-mul.f r3.x, c11.x, r0.w
-add.f r2.z, r2.z, r2.w
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r2.w, c12.x, r1.x, r3.x
-mov.f32f32 r2.z, r2.z
-max.f r2.x, r2.x, c18.y
-max.f r2.y, r2.y, c18.y
-mov.f32f32 r2.w, r2.w
-mul.f r3.x, c16.x, r2.z
-min.f r2.x, r2.x, c22.y
-min.f r2.y, r2.y, c22.y
-mul.f r3.y, c11.z, r0.x
-mad.f32 r2.w, c13.x, r1.y, r2.w
-max.f r2.x, r2.x, c18.x
-max.f r2.y, r2.y, c18.x
-mad.f32 r3.y, c12.z, r0.y, r3.y
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r3.y, c13.z, r0.z, r3.y
-mul.f r3.z, c11.x, r0.x
-mul.f r2.x, c16.x, r2.x
-mad.f32 r3.z, c12.x, r0.y, r3.z
-add.f r3.y, r3.y, c14.z
-mad.f32 r3.z, c13.x, r0.z, r3.z
-mov.f32f32 r2.x, r2.x
-mad.f32 r3.x, c18.w, r3.x, r3.y
-mul.f r2.w, r2.w, (neg)c4.x
-mul.f r3.w, c11.y, r0.w
-mad.f32 r2.x, c18.z, r2.x, c14.x
-add.f r3.z, r3.z, c14.x
-mov.f32f32 r3.x, r3.x
-mad.f32 r2.y, c16.x, r2.y, r3.z
-mov.f32f32 r2.x, r2.x
-mad.f32 r3.w, c12.y, r1.x, r3.w
-mov.f32f32 r3.x, r3.x
-add.f r2.y, r2.y, c19.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r3.w, r3.w
-add.f r3.x, r3.x, c19.x
-mad.f32 r3.w, c13.y, r1.y, r3.w
-add.f r2.x, r2.x, c19.x
+@in(r3.z)	in8
+@in(r3.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(r3.x)	out12
+@out(r3.y)	out13
+@out(r3.z)	out14
+@out(r3.w)	out15
+@const(c18.x)	0x3ecccccd, 0x00000000, 0x3eaaa64c, 0x3e800000
+@const(c19.x)	0x3f000000, 0x40000000, 0xbf800000, 0x40400000
+@const(c20.x)	0x399d4952, 0x3e4ccccd, 0x3f99999a, 0x3fd9999a
+@const(c21.x)	0x3c88ce70, 0x00000000, 0x00000000, 0x00000000
+@const(c22.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)floor.f r1.z, c14.z
+floor.f r1.w, c14.x
+absneg.f r2.x, (abs)c17.x
+absneg.f r2.y, (abs)c17.y
+add.f r1.z, c14.z, (neg)r1.z
+add.f r1.w, c14.x, (neg)r1.w
+mov.f32f32 r2.z, c18.y
+mul.f r2.w, c11.x, r0.w
+max.f r1.z, r1.z, c18.y
+max.f r1.w, r1.w, c18.y
+add.f r2.x, r2.x, r2.y
+add.f r2.y, r2.z, c19.x
+min.f r1.z, r1.z, c22.y
+min.f r1.w, r1.w, c22.y
+mul.f r2.z, c16.x, r2.x
+mul.f r3.x, c11.z, r0.x
+max.f r1.z, r1.z, c18.x
+max.f r1.w, r1.w, c18.x
+mul.f r3.y, c11.x, r0.x
+mad.f32 r3.x, c12.z, r0.y, r3.x
+mul.f r1.z, c16.x, r1.z
+mad.f32 r3.y, c12.x, r0.y, r3.y
+mad.f32 r3.x, c13.z, r0.z, r3.x
+mad.f32 r3.y, c13.x, r0.z, r3.y
+mad.f32 r1.z, c18.z, r1.z, c14.x
 floor.f r4.x, r2.y
-floor.f r4.y, r3.x
-mov.f32f32 r3.w, r3.w
-floor.f r4.z, r2.x
+mad.f32 r2.w, c12.x, r1.x, r2.w
+add.f r3.y, r3.y, c14.x
+add.f r1.z, r1.z, c19.x
+mad.f32 r1.w, c16.x, r1.w, r3.y
+add.f r3.x, r3.x, c14.z
 add.f r2.y, r2.y, (neg)r4.x
-add.f r3.x, r3.x, (neg)r4.y
-mad.f32 r2.w, (neg)c4.y, r3.w, r2.w
-add.f r2.x, r2.x, (neg)r4.z
+floor.f r4.x, r1.z
+add.f r1.w, r1.w, c19.x
+mad.f32 r2.z, c18.w, r2.z, r3.x
 mad.f32 r2.y, c19.y, r2.y, c19.z
-mad.f32 r3.x, c19.y, r3.x, c19.z
-mov.f32f32 r2.w, r2.w
-mad.f32 r2.x, c19.y, r2.x, c19.z
+add.f r1.z, r1.z, (neg)r4.x
+floor.f r4.x, r1.w
+add.f r2.z, r2.z, c19.x
 absneg.f r2.y, (abs)r2.y
-absneg.f r3.x, (abs)r3.x
-mul.f r0.w, c11.z, r0.w
-absneg.f r2.x, (abs)r2.x
-mul.f r3.w, c19.y, r2.y
-mul.f r4.x, c19.y, r3.x
-mul.f r2.y, r2.y, r2.y
-mul.f r4.y, c19.y, r2.x
-add.f r3.w, c19.w, (neg)r3.w
+mad.f32 r1.z, c19.y, r1.z, c19.z
+add.f r4.x, r1.w, (neg)r4.x
+floor.f r4.y, r2.z
+mul.f r1.w, r2.y, r2.y
+absneg.f r1.z, (abs)r1.z
+mad.f32 r2.y, c19.y, r4.x, c19.z
+add.f r2.z, r2.z, (neg)r4.y
+mad.f32 r2.w, c13.x, r1.y, r2.w
+mul.f r4.x, c19.y, r1.z
+absneg.f r2.y, (abs)r2.y
+mad.f32 r2.z, c19.y, r2.z, c19.z
+mul.f r1.z, r1.z, r1.z
 add.f r4.x, c19.w, (neg)r4.x
-mul.f r3.x, r3.x, r3.x
+mul.f r4.y, c19.y, r2.y
+absneg.f r2.z, (abs)r2.z
+mul.f r2.y, r2.y, r2.y
+mul.f r1.z, r1.z, r4.x
+mul.f r4.x, r0.y, c21.x
 add.f r4.y, c19.w, (neg)r4.y
-mul.f r2.x, r2.x, r2.x
-mul.f r2.y, r2.y, r3.w
-mul.f r3.x, r3.x, r4.x
-mad.f32 r0.w, c12.z, r1.x, r0.w
-mul.f r1.x, r2.x, r4.y
-mov.f32f32 r2.x, r2.y
-mov.f32f32 r2.y, r3.x
-mul.f r3.x, r0.x, r0.z
-mov.f32f32 r1.x, r1.x
-mul.f r3.w, r0.y, c21.x
-mul.f r4.x, r0.y, c20.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r4.y, c18.y
-mov.f32f32 r3.w, r3.w
-mul.f r3.x, r3.x, r4.x
-mad.f32 r0.w, c13.z, r1.y, r0.w
-add.f r1.y, r4.y, c19.x
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r0.w, r0.w
-floor.f r4.x, r1.y
-max.f r3.w, r3.w, c18.y
-mov.f32f32 r3.x, r3.x
-mad.f32 r0.w, (neg)c4.z, r0.w, r2.w
-add.f r1.y, r1.y, (neg)r4.x
-min.f r2.w, r3.w, c22.y
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r1.y, c19.y, r1.y, c19.z
-min.f r2.w, r2.w, c18.w
-max.f r3.x, r3.x, c18.y
-max.f r0.w, c18.y, r0.w
-absneg.f r1.y, (abs)r1.y
-mov.f32f32 r2.w, r2.w
-min.f r3.x, r3.x, c22.y
-mov.f32f32 r0.w, r0.w
-mul.f r1.y, r1.y, r1.y
-mul.f r1.x, r1.x, r2.w
-min.f r2.w, r3.x, c20.y
-mul.f r3.x, r0.w, c5.z
-mul.f r3.w, r0.w, c5.y
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.w, r3.w
-mul.f r0.w, r0.w, c5.x
-mul.f r2.x, r2.x, r2.w
-mul.f r2.y, r2.y, r2.w
-max.f r2.z, r2.z, c20.z
-mov.f32f32 r4.z, r3.x
+mul.f r4.z, c19.y, r2.z
+mul.f r2.z, r2.z, r2.z
+max.f r4.x, r4.x, c18.y
+mul.f r2.y, r2.y, r4.y
+mul.f r4.y, r0.x, r0.z
+add.f r4.z, c19.w, (neg)r4.z
+min.f r4.x, r4.x, c22.y
+mul.f r4.w, r0.y, c20.x
+mul.f r2.w, r2.w, (neg)c4.x
+mul.f r5.x, c11.y, r0.w
+min.f r4.x, r4.x, c18.w
+mul.f r4.y, r4.y, r4.w
+mul.f r2.z, r2.z, r4.z
+mad.f32 r4.z, c12.y, r1.x, r5.x
+mul.f r1.z, r1.z, r4.x
+max.f r4.x, r4.y, c18.y
+mad.f32 r4.y, c13.y, r1.y, r4.z
 mov.f32f32 r2.x, r2.x
-mov.f32f32 r4.y, r3.w
-mad.f32 r2.w, c17.x, r1.x, r2.x
-mad.f32 r1.x, c17.y, r1.x, r2.x
-mov.f32f32 r2.x, r2.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r2.z, r2.w
-mov.f32f32 r1.x, r1.x
-min.f r2.x, r2.x, c20.w
-mov.f32f32 r4.x, r0.w
-mov.f32f32 r4.w, r1.y
-mul.f r0.w, c11.y, r0.x
-mov.f32f32 r1.y, r2.x
-mad.f32 r0.w, c12.y, r0.y, r0.w
+mov.f32f32 r4.z, r1.z
+min.f r4.x, r4.x, c22.y
+mad.f32 r2.w, (neg)c4.y, r4.y, r2.w
+mul.f r0.w, c11.z, r0.w
+max.f r2.x, r2.x, c20.z
+min.f r4.x, r4.x, c20.y
+mad.f32 r0.w, c12.z, r1.x, r0.w
+mul.f r1.x, c11.y, r0.x
 mul.f r0.x, c11.w, r0.x
-mov.f32f32 r1.w, r1.w
-mad.f32 r2.x, r2.y, r1.y, r3.z
-mad.f32 r2.w, r2.y, r1.y, r3.y
-mad.f32 r0.w, c13.y, r0.z, r0.w
+mov.f32f32 r4.y, r4.x
+mul.f r2.z, r2.z, r4.x
+min.f r2.x, r2.x, c20.w
+mad.f32 r0.w, c13.z, r1.y, r0.w
+mul.f r1.y, r2.y, r4.y
+mad.f32 r1.x, c12.y, r0.y, r1.x
 mad.f32 r0.x, c12.w, r0.y, r0.x
-add.f r0.y, r2.x, r2.z
-add.f r1.x, r2.w, r1.x
-add.f r0.w, r0.w, c14.y
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.w, r2.y, r1.y, r0.w
+mad.f32 r0.y, c13.y, r0.z, r1.x
+mov.f32f32 r1.x, r1.y
+mad.f32 r1.y, c17.y, r1.z, r1.y
+mad.f32 r1.x, c17.x, r4.z, r1.x
+mov.f32f32 r1.z, r2.x
+mad.f32 r2.x, r2.z, r2.x, r3.x
+add.f r0.y, r0.y, c14.y
+mad.f32 r0.w, (neg)c4.z, r0.w, r2.w
+mad.f32 r2.y, r2.z, r1.z, r3.y
+add.f r1.y, r2.x, r1.y
+mad.f32 r0.y, r2.z, r1.z, r0.y
+max.f r4.x, c18.y, r0.w
+add.f r0.w, r2.y, r1.x
+mov.f32f32 r1.x, r1.y
 mad.f32 r0.x, c13.w, r0.z, r0.x
-mul.f r0.z, c7.w, r0.y
-mul.f r1.y, c7.z, r0.y
-mad.f32 r0.z, c8.w, r0.w, r0.z
-mad.f32 r1.y, c8.z, r0.w, r1.y
-mul.f r2.x, c7.y, r0.y
-mul.f r2.y, c7.x, r0.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.y, r1.y
-mad.f32 r0.z, c9.w, r1.x, r0.z
-add.f r0.x, r0.x, c14.w
-mad.f32 r1.y, c9.z, r1.x, r1.y
-mad.f32 r2.x, c8.y, r0.w, r2.x
-mad.f32 r2.y, c8.x, r0.w, r2.y
-mad.f32 r0.z, c10.w, r0.x, r0.z
-mad.f32 r1.y, c10.z, r0.x, r1.y
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.y, r1.y
-mad.f32 r2.x, c9.y, r1.x, r2.x
-mad.f32 r2.y, c9.x, r1.x, r2.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.y, r1.y
-mad.f32 r2.x, c10.y, r0.x, r2.x
-mad.f32 r2.y, c10.x, r0.x, r2.y
-mov.f32f32 r3.y, r0.z
-mov.f32f32 r3.x, r1.y
-mov.f32f32 r0.z, r2.x
-mov.f32f32 r1.y, r2.y
-mul.f r2.x, c0.w, r0.y
-mul.f r2.y, c0.z, r0.y
-mul.f r0.z, r0.z, c15.y
-mul.f r1.y, r1.y, c15.x
-(rpt1)nop
-mov.f32f32 r2.w, r0.z
-mov.f32f32 r2.z, r1.y
-mad.f32 r0.z, c1.w, r0.w, r2.x
-mad.f32 r1.y, c1.z, r0.w, r2.y
-mad.f32 r0.z, c2.w, r1.x, r0.z
-mad.f32 r1.y, c2.z, r1.x, r1.y
-mad.f32 r0.z, c3.w, r0.x, r0.z
-mad.f32 r1.y, c3.z, r0.x, r1.y
-mul.f r2.x, c0.y, r0.y
-mul.f r2.y, c0.x, r0.y
-mov.f32f32 r5.w, r0.z
-mov.f32f32 r5.z, r1.y
-mad.f32 r0.z, c1.y, r0.w, r2.x
-mad.f32 r0.w, c1.x, r0.w, r2.y
-mad.f32 r0.z, c2.y, r1.x, r0.z
-mad.f32 r0.w, c2.x, r1.x, r0.w
-mad.f32 r0.z, c3.y, r0.x, r0.z
-mad.f32 r0.x, c3.x, r0.x, r0.w
-(rpt1)nop
-mov.f32f32 r5.y, r0.z
-mov.f32f32 r5.x, r0.x
-mad.f32 r0.x, c6.x, r1.x, c6.y
-mad.f32 r0.y, c6.x, r0.y, c6.y
-(rpt1)nop
-mov.f32f32 r0.w, r0.x
-mov.f32f32 r0.z, r0.y
-mov.f32f32 r1.y, r1.w
-mov.f32f32 r0.x, r1.z
-(rpt2)nop
-mov.f32f32 r1.x, r0.x
+nop
+mov.f32f32 r1.z, r0.w
+mul.f r3.x, c0.x, r0.w
+mad.f32 r3.y, c6.x, r1.x, c6.y
+mov.f32f32 r4.y, r4.x
+mul.f r0.z, c7.y, r1.z
+mul.f r0.w, c7.x, r1.z
+mad.f32 r0.z, c8.y, r0.y, r0.z
+mad.f32 r0.w, c8.x, r0.y, r0.w
+mad.f32 r0.z, c9.y, r1.x, r0.z
+add.f r4.z, r0.x, c14.w
+mad.f32 r0.x, c9.x, r1.x, r0.w
+mul.f r0.w, c7.w, r1.z
+mul.f r2.x, c7.z, r1.z
+mad.f32 r0.z, c10.y, r4.z, r0.z
+mad.f32 r0.x, c10.x, r4.z, r0.x
+mad.f32 r0.w, c8.w, r0.y, r0.w
+mad.f32 r2.z, c8.z, r0.y, r2.x
+mul.f r2.y, r0.z, c15.y
+mul.f r2.x, r0.x, c15.x
+mad.f32 r0.x, c9.w, r1.x, r0.w
+mad.f32 r0.z, c9.z, r1.x, r2.z
+mad.f32 r2.w, c10.w, r4.z, r0.x
+mad.f32 r2.z, c10.z, r4.z, r0.z
+mul.f r0.x, c0.w, r1.z
+mul.f r0.z, c0.z, r1.z
+mad.f32 r0.x, c1.w, r0.y, r0.x
+mad.f32 r0.z, c1.z, r0.y, r0.z
+mad.f32 r0.x, c2.w, r1.x, r0.x
+mad.f32 r0.z, c2.z, r1.x, r0.z
+mad.f32 r0.w, c3.w, r4.z, r0.x
+mad.f32 r0.z, c3.z, r4.z, r0.z
+mul.f r0.x, c0.y, r1.z
+mad.f32 r3.x, c1.x, r0.y, r3.x
+mad.f32 r0.x, c1.y, r0.y, r0.x
+mad.f32 r0.y, c2.x, r1.y, r3.x
+mad.f32 r1.x, c2.y, r1.x, r0.x
+mad.f32 r0.x, c3.x, r4.z, r0.y
+mad.f32 r0.y, c3.y, r4.z, r1.x
+mad.f32 r3.x, c6.x, r1.z, c6.y
+mul.f r1.z, r4.y, c5.z
+mul.f r1.y, r4.y, c5.y
+mul.f r1.x, r4.x, c5.x
 end
-; VERT: outputs: r5.x (0:0) r4.x (5:9) r2.z (5:10) r0.z (5:11)
-; 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: 221 instructions, 0 half, 6 full
-; pos: r5.x
+; VERT: outputs: r0.x (0:0) r1.x (5:9) r2.x (5:10) r3.x (5:11)
+; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r0.w (0:0,cm=7,il=12,b=0) r3.z (0:0,cm=3,il=16,b=0)
+; VERT: 152 instructions, 0 half, 6 full