1 files changed, 197 insertions, 304 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-40.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-40.asm
index e895f8d..15beb9d 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-40.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-40.asm
@@ -4,334 +4,227 @@
 @in(r0.w)	in3
 @in(r1.x)	in4
 @in(r1.y)	in5
-@out(r3.z)	out0
-@out(r3.w)	out1
-@out(r4.x)	out2
-@out(r4.y)	out3
+@out(r2.x)	out0
+@out(r2.y)	out1
+@out(r2.z)	out2
+@out(r2.w)	out3
+@const(c14.x)	0x3f000000, 0x00000000, 0x3f800000, 0x40000000
+@const(c15.x)	0xbf800000, 0xba03126f, 0xbf000000, 0x3f800000
+@const(c16.x)	0x3fb8aa65, 0x3de38866, 0x3cf5c28f, 0x00000000
+@const(c17.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
 (sy)(ss)bary.f r0.x, 12, r1.x
-bary.f r0.y, 0, r1.x
+bary.f r1.z, 0, r1.x
 add.f r0.w, r0.w, c14.y
-bary.f r1.z, 1, r1.x
-mov.f32f32 r1.w, r0.x
-add.f r2.x, r0.y, c15.z
-bary.f r2.y, 8, r1.x
+bary.f r1.w, 1, r1.x
+mov.f32f32 r2.x, r0.x
+bary.f r0.y, 13, r1.x
 add.f r2.z, r1.z, c15.z
-mov.f32f32 r2.w, r1.w
-bary.f r1.w, 13, r1.x
-floor.f r3.x, r2.x
+add.f r2.w, r1.w, c15.z
+bary.f r3.x, 10, r1.x
+mov.f32f32 r2.y, r0.y
+floor.f r3.z, r2.z
 rcp r0.w, r0.w
 add.f r0.z, r0.z, c14.y
-floor.f r3.y, r2.z
-mov.f32f32 r3.z, r1.w
-add.f r2.x, r2.x, (neg)r3.x
+floor.f r3.w, r2.w
+bary.f r3.y, 11, r1.x
+add.f r2.z, r2.z, (neg)r3.z
 (ss)mul.f r0.z, r0.z, r0.w
-(ss)add.f r0.w, r2.z, (neg)r3.y
-mov.f32f32 r3.x, r3.z
-mov.f32f32 r2.x, r2.x
+sam (f32)(xyz)r4.x, r2.x, s#3, t#3
+(sy)(ss)mad.f32 r0.w, c14.w, r4.x, c15.x
+absneg.f r2.x, (neg)c11.x
+mov.f32f32 r2.y, r2.z
+add.f r2.w, r2.w, (neg)r3.w
+mov.f32f32 r3.z, r0.w
+bary.f r3.w, 4, r1.x
+mul.f r4.x, c14.x, r2.y
+mul.f r2.x, r2.x, c11.x
+mov.f32f32 r4.w, r2.w
+mul.f r3.w, r3.w, r3.z
+mad.f32 r4.y, c14.w, r4.y, c15.x
+add.f r1.z, r1.z, (neg)r4.x
+mul.f r2.x, r2.x, r0.z
 mov.f32f32 r0.z, r0.z
-absneg.f r2.z, (neg)c11.x
-mov.f32f32 r0.w, r0.w
-mul.f r3.y, c14.x, r2.x
-mov.f32f32 r2.y, r2.y
-sam (f32)(xyz)r3.z, r2.w, s#3, t#3
-(sy)(ss)mad.f32 r2.w, c14.w, r3.z, c15.x
-mul.f r2.z, r2.z, c11.x
-mov.f32f32 r3.x, r3.y
-mul.f r3.y, c14.x, r0.w
-mov.f32f32 r2.w, r2.w
-bary.f r3.z, 4, r1.x
-add.f r0.y, r0.y, (neg)r3.x
-mul.f r2.z, r2.z, r0.z
-mov.f32f32 r3.x, r3.y
-mul.f r3.y, r3.z, r2.w
-mad.f32 r3.z, c14.w, r3.w, c15.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.z, r2.z
-add.f r1.z, r1.z, (neg)r3.x
-mov.f32f32 r3.x, r3.z
-bary.f r3.z, 14, r1.x
-add.f r3.w, c15.z, r0.y
-mul.f r0.z, r2.z, r0.z
-add.f r0.y, c15.w, r0.y
-mov.f32f32 r2.z, r3.z
-bary.f r3.z, 7, r1.x
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.y, r0.y
-mul.f r2.z, r2.z, (neg)r3.z
-mul.f r3.w, r3.w, c6.z
+mov.f32f32 r4.x, r4.y
+bary.f r5.x, 14, r1.x
+bary.f r5.y, 7, r1.x
+mov.f32f32 r5.z, r1.z
+mul.f r0.z, r2.x, r0.z
+mul.f r2.x, c14.x, r4.w
+mul.f r5.x, r5.x, (neg)r5.y
+add.f r5.z, c15.z, r5.z
 mul.f r0.z, r0.z, c16.x
-mul.f r0.y, r0.y, c6.z
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r4.y, r3.w
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r4.z, r0.y
-mad.f32 r2.z, r2.z, r3.x, r3.y
-mov.f32f32 r4.w, r4.y
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r3.y, r4.z
-mov.f32f32 r2.z, r2.z
-mad.f32 r4.x, c14.w, r4.x, c15.x
-add.f r4.y, c15.z, r1.z
-exp2 r0.z, r0.z
-(ss)mov.f32f32 r0.z, r0.z
-mov.f32f32 r5.z, r3.y
-mov.f32f32 r3.y, r4.x
-bary.f r4.x, 23, r1.x
-mov.f32f32 r4.y, r4.y
-add.f r4.z, c17.y, (neg)r0.z
+add.f r1.w, r1.w, (neg)r2.x
+mad.f32 r2.x, r5.x, r4.x, r3.w
+mad.f32 r3.w, c14.w, r4.z, c15.x
+mul.f r5.z, r5.z, c6.z
+mov.f32f32 r4.z, r1.w
 add.f r1.z, c15.w, r1.z
-mov.f32f32 r4.x, r4.x
-mul.f r4.y, r4.y, c6.w
-mul.f r4.z, r4.z, c11.y
-mul.f r0.z, r0.z, c14.z
-mad.f32 r2.z, r4.x, r3.y, r2.z
-mov.f32f32 r4.x, r4.y
-mov.f32f32 r1.z, r1.z
-add.f r0.z, r0.z, r4.z
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r5.x, r4.x
-bary.f r4.x, 2, r1.x
-mov.f32f32 r0.z, r0.z
-mul.f r4.z, r2.z, r2.z
-bary.f r5.y, 5, r1.x
-add.f r4.x, r4.x, c15.y
-mov.f32f32 r0.z, r0.z
-mul.f r1.z, r1.z, c6.w
-mul.f r5.w, r5.y, r2.w
-bary.f r5.y, 15, r1.x
-mov.f32f32 r6.x, r4.x
+mov.f32f32 r5.x, r3.w
+bary.f r5.w, 23, r1.x
+mov.f32f32 r6.y, r5.z
+exp2 r0.z, r0.z
+(ss)mov.f32f32 r6.x, r0.z
+add.f r4.z, c15.z, r4.z
+mad.f32 r2.x, r5.w, r5.x, r2.x
+mul.f r7.x, r1.z, c6.z
+add.f r1.z, c15.w, r1.w
+add.f r1.w, c17.y, (neg)r6.x
+mov.f32f32 r7.w, r2.x
+mul.f r8.y, r4.z, c6.w
+mov.f32f32 r8.x, r7.x
+mul.f r7.y, r1.z, c6.w
+mul.f r1.z, r2.x, r7.w
+bary.f r2.x, 5, r1.x
+mov.f32f32 r6.z, r8.y
+mul.f r1.w, r1.w, c11.y
+(ss)mul.f r0.z, r0.z, c14.z
+mul.f r2.x, r2.x, r3.z
+bary.f r3.z, 15, r1.x
+bary.f r4.z, 2, r1.x
+add.f r0.z, r0.z, r1.w
+mov.f32f32 r5.w, r7.y
+mul.f r1.w, r3.z, (neg)r5.y
+add.f r7.z, r4.z, c15.y
 max.f r0.z, r0.z, c14.y
-mov.f32f32 r6.y, r1.z
-mov.f32f32 r6.z, r5.y
-mov.f32f32 r5.y, r6.x
+add.f r2.y, c14.w, (neg)r2.y
+mad.f32 r1.w, r1.w, r4.x, r2.x
+bary.f r2.x, 24, r1.x
+mov.f32f32 r6.w, r7.z
 min.f r0.z, r0.z, c14.z
-mov.f32f32 r6.x, r6.y
-mul.f r6.y, r6.z, (neg)r3.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r3.w, r3.w
-add.f r6.z, c17.y, (neg)r0.z
-mov.f32f32 r6.y, r6.y
-sam.s (f32)(x)r4.w, r4.w, s#5, t#5
-(sy)(ss)mov.f32f32 r4.w, r4.w
-add.f r5.x, c17.y, (neg)r0.z
-mul.f r5.y, r6.z, c10.z
-mad.f32 r5.w, r6.y, r3.x, r5.w
-mov.f32f32 r4.w, r4.w
-add.f r6.y, c14.w, (neg)r2.x
-mul.f r5.x, r5.x, c10.y
-mov.f32f32 r6.z, r5.w
-bary.f r5.w, 24, r1.x
-mov.f32f32 r6.y, r6.y
-add.f r6.w, c14.w, (neg)r0.w
-add.f r7.x, c17.y, (neg)r0.z
-mov.f32f32 r7.y, r5.w
-mov.f32f32 r5.w, r6.x
-mov.f32f32 r6.x, r4.x
-mov.f32f32 r7.z, r0.y
-mad.f32 r0.y, r7.y, r3.y, r6.z
-mov.f32f32 r6.z, r6.w
-mul.f r6.w, r7.x, c10.x
-mov.f32f32 r6.x, r6.x
-mov.f32f32 r0.y, r0.y
-mul.f r7.x, r6.y, r6.z
-mov.f32f32 r4.y, r4.y
-mov.f32f32 r8.y, r3.w
-mad.f32 r3.w, r0.y, r0.y, r4.z
-mul.f r4.z, r7.x, r4.w
-mov.f32f32 r7.w, r4.y
-mov.f32f32 r4.y, r4.x
-mov.f32f32 r3.w, r3.w
-bary.f r4.w, 6, r1.x
-sam.s (f32)(x)r5.z, r5.z, s#5, t#5
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r8.x, r4.y
-(sy)mov.f32f32 r4.y, r5.z
-mul.f r2.w, r4.w, r2.w
-bary.f r4.w, 16, r1.x
-mov.f32f32 r8.z, r1.z
-mov.f32f32 r1.z, r4.x
-(ss)mov.f32f32 r5.z, r2.y
-mov.f32f32 r2.y, r4.w
-sam.s (f32)(x)r4.x, r7.z, s#5, t#5
-(sy)mov.f32f32 r4.x, r4.x
-add.f r2.x, r2.x, c14.z
-mov.f32f32 r8.w, r1.z
-mul.f r1.z, r2.y, (neg)r3.z
-bary.f r2.y, 9, r1.x
-mul.f r3.z, r2.x, r6.z
-add.f r0.w, r0.w, c14.z
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.y
-bary.f r4.w, 10, r1.x
-mad.f32 r3.z, r3.z, r4.x, r4.z
-mad.f32 r1.z, r1.z, r3.x, r2.w
-sam.s (f32)(x)r2.w, r8.y, s#5, t#5
-mov.f32f32 r5.w, r2.y
-mul.f r2.y, r6.y, r0.w
-mov.f32f32 r3.x, r3.z
-mov.f32f32 r1.z, r1.z
-bary.f r3.z, 25, r1.x
-(sy)mov.f32f32 r2.w, r2.w
-mov.f32f32 r4.x, r4.w
-sam (f32)(w)r5.z, r5.z, s#1, t#1
-(sy)add.f r4.z, c14.z, (neg)r6.y
-mov.f32f32 r3.z, r3.z
-mad.f32 r2.y, r2.y, r2.w, r3.x
-mov.f32f32 r2.w, r4.x
-mul.f r0.w, r2.x, r0.w
-mad.f32 r1.z, r3.z, r3.y, r1.z
-mov.f32f32 r2.x, r2.y
-mov.f32f32 r2.y, r4.z
-bary.f r3.x, 11, r1.x
-mov.f32f32 r1.z, r1.z
-mad.f32 r0.w, r0.w, r4.y, r2.x
-mov.f32f32 r2.x, r2.y
-mov.f32f32 r2.y, r3.x
-mad.f32 r3.x, r1.z, r1.z, r3.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r2.x, r2.x
+mov.f32f32 r8.z, r7.z
+mad.f32 r1.w, r2.x, r5.x, r1.w
+mov.f32f32 r6.x, r7.z
+add.f r2.x, c17.y, (neg)r0.z
+add.f r3.z, c17.y, (neg)r0.z
+mov.f32f32 r4.x, r1.w
+sam.s (f32)(x)r8.w, r6.y, s#5, t#5
+add.f r4.z, c17.y, (neg)r0.z
+mov.f32f32 r5.x, r2.y
+add.f r4.w, c14.w, (neg)r4.w
+mad.f32 r1.z, r1.w, r4.x, r1.z
+bary.f r1.w, 6, r1.x
+mul.f r2.x, r2.x, c10.z
+mul.f r3.z, r3.z, c10.y
+mul.f r4.z, r4.z, c10.x
+mul.f r0.w, r1.w, r0.w
+bary.f r1.w, 16, r1.x
+(ss)mov.f32f32 r6.y, r4.w
+sam.s (f32)(x)r9.x, r8.x, s#5, t#5
+sam.s (f32)(x)r9.y, r5.z, s#5, t#5
+sam.s (f32)(x)r6.z, r7.x, s#5, t#5
+add.f r2.z, r2.z, c14.z
+add.f r2.w, r2.w, c14.z
+mul.f r1.w, r1.w, (neg)r5.y
+mul.f r5.x, r5.x, r6.y
+mul.f r4.w, r2.z, r4.w
+mul.f r2.y, r2.y, r2.w
+mad.f32 r0.w, r1.w, r4.y, r0.w
+bary.f r1.w, 25, r1.x
+(sy)mul.f r4.y, r5.x, r8.w
+mul.f r2.z, r2.z, r2.w
+mad.f32 r2.w, r4.w, r9.x, r4.y
+mad.f32 r0.w, r1.w, r3.w, r0.w
+mad.f32 r1.w, r2.y, r9.y, r2.w
+(ss)nop
+sam (f32)(w)r4.w, r3.x, s#2, t#2
+(sy)cmps.f.lt r2.y, r5.z, c16.z
+mov.f32f32 r2.w, r0.x
+(ss)mov.f32f32 r3.y, r0.w
+mad.f32 r1.w, r2.z, r6.z, r1.w
+cov.u32f32 r2.y, r2.y
+mov.f32f32 r3.x, r0.y
+mad.f32 r1.z, r3.y, r3.y, r1.z
+mul.f r1.w, c16.y, r1.w
+cmps.f.ne r2.y, r2.y, c14.y
 (rpt3)nop
-rsq r3.x, r3.x
-(ss)mov.f32f32 r3.y, r3.x
-mul.f r0.w, c16.y, r0.w
-mov.f32f32 r4.y, r2.x
-(ss)mov.f32f32 r3.x, r2.y
-mul.f r2.x, r2.z, r3.y
-mov.f32f32 r0.w, r0.w
-mul.f r0.y, r0.y, r3.y
-mul.f r1.z, r1.z, r3.y
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r0.x
-sam (f32)(w)r5.z, r2.w, s#2, t#2
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.z, r1.z
-mul.f r2.z, r2.x, r2.x
-(ss)mul.f r2.w, (neg)c9.x, r2.x
-mad.f32 r2.z, r0.y, r0.y, r2.z
-mad.f32 r2.w, (neg)c9.y, r0.y, r2.w
-(sy)cmps.f.lt r3.x, r6.y, c16.z
-mov.f32f32 r3.y, r6.y
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r2.w, r2.w
-mad.f32 r2.z, r1.z, r1.z, r2.z
-mad.f32 r2.w, (neg)c9.z, r1.z, r2.w
-cov.u32f32 r3.x, r3.x
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r4.z, r2.y
-mov.f32f32 r2.y, r1.w
-mov.f32f32 r0.x, r0.x
+rsq r1.z, r1.z
+(ss)mov.f32f32 r2.z, r1.z
+mov.f32f32 r3.y, r1.w
+mul.f r0.w, r0.w, r1.z
+(ss)mov.f32f32 r1.z, c14.y
+mul.f r3.w, r7.w, r2.z
+mul.f r2.z, r4.x, r2.z
+mov.f32f32 r4.x, r0.w
+sel.b32 r1.z, r1.z, r2.y, r5.z
+mov.f32f32 r2.y, r3.w
+mul.f r3.w, (neg)c9.x, r3.w
+mov.f32f32 r4.y, r2.z
+mad.f32 r2.z, (neg)c9.y, r2.z, r3.w
+mul.f r3.w, r2.y, r2.y
+mad.f32 r0.w, (neg)c9.z, r0.w, r2.z
+mad.f32 r2.z, r4.y, r4.y, r3.w
+sam (f32)(xyz)r4.w, r2.w, s#4, t#4
+(sy)(ss)mul.f r3.x, c8.z, r5.y
+mad.f32 r2.z, r4.x, r4.x, r2.z
+max.f r0.w, r0.w, c14.y
+bary.f r2.w, 20, r1.x
+(rpt1)nop
+mov.f32f32 r3.w, r0.w
+bary.f r5.y, 21, r1.x
 rsq r2.z, r2.z
-(ss)mov.f32f32 r2.z, r2.z
-mov.f32f32 r2.w, r2.w
-cmps.f.ne r3.x, r3.x, c14.y
-nop
-mul.f r2.x, r2.x, r2.z
-max.f r2.w, r2.w, c14.y
-mul.f r0.y, r0.y, r2.z
-mul.f r1.z, r1.z, r2.z
-mov.f32f32 r2.x, r2.x
-bary.f r2.z, 17, r1.x
-mov.f32f32 r2.w, r2.w
-bary.f r4.w, 21, r1.x
-bary.f r5.z, 20, r1.x
-mul.f r2.x, r2.x, r2.z
-mov.f32f32 r0.y, r0.y
-bary.f r2.z, 18, r1.x
+(ss)mov.f32f32 r5.z, r2.z
 bary.f r5.w, 22, r1.x
-mad.f32 r6.x, c8.y, r2.w, (neg)r4.w
-mad.f32 r6.y, c8.x, r2.w, (neg)r5.z
-mad.f32 r0.y, r0.y, r2.z, r2.x
-mad.f32 r2.x, c8.z, r2.w, (neg)r5.w
-mov.f32f32 r2.z, r6.x
-mov.f32f32 r2.w, r6.y
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.z, r1.z
-bary.f (ei)r1.x, 19, r1.x
-mov.f32f32 r1.y, r2.x
-mad.f32 r2.x, c12.x, r2.z, r4.w
-mad.f32 r2.z, c12.x, r2.w, r5.z
-mad.f32 r0.y, r1.z, r1.x, r0.y
-mad.f32 r1.x, c12.x, r1.y, r5.w
-mov.f32f32 r1.y, r2.x
-mov.f32f32 r1.z, r2.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r2.x, c14.y
-mov.f32f32 r4.w, r2.y
-max.f r0.y, c14.y, r0.y
-mov.f32f32 r2.y, r0.x
-mov.f32f32 r0.x, r1.w
-sel.b32 r1.w, r2.x, r3.x, r3.y
-mov.f32f32 r0.y, r0.y
+mad.f32 r6.x, c8.y, r3.w, (neg)r5.y
+mad.f32 r3.w, c8.z, r3.w, (neg)r5.w
+mul.f r2.y, r2.y, r5.z
+bary.f r6.y, 17, r1.x
+mad.f32 r5.y, c12.x, r6.x, r5.y
+mul.f r4.y, r4.y, r5.z
+(ss)mul.f r2.z, r4.x, r2.z
+mul.f r2.y, r2.y, r6.y
+bary.f r4.x, 18, r1.x
+mad.f32 r3.w, c12.x, r3.w, r5.w
+mad.f32 r0.w, c8.x, r0.w, (neg)r2.w
+nop
+mad.f32 r2.y, r4.y, r4.x, r2.y
+bary.f r4.x, 19, r1.x
+mad.f32 r0.w, c12.x, r0.w, r2.w
+mul.f r4.y, c8.y, r5.x
+mul.f r4.w, c8.x, r4.w
+mad.f32 r2.y, r2.z, r4.x, r2.y
+bary.f r2.z, 8, r1.x
+bary.f (ei)r2.w, 9, r1.x
+sam (f32)(xyz)r5.z, r0.x, s#0, t#0
+nop
+(ss)max.f r0.x, c14.y, r2.y
 (rpt5)nop
-log2 r0.y, r0.y
-(ss)mul.f r0.y, c12.y, r0.y
-sam (f32)(xyz)r2.z, r4.z, s#4, t#4
-(sy)mul.f r2.x, c8.y, r2.w
-mul.f r2.w, c8.x, r2.z
-mul.f r3.x, c8.z, r3.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.z, r0.x
+log2 r0.x, r0.x
+(ss)mul.f r0.x, c12.y, r0.x
+sam (f32)(w)r6.y, r2.z, s#1, t#1
 (rpt4)nop
-exp2 r0.x, r0.y
+(sy)(ss)add.f r2.w, c14.z, (neg)r7.x
+exp2 r0.x, r0.x
 (ss)mul.f r0.y, r3.x, r0.x
-sam (f32)(xyz)r5.z, r2.y, s#0, t#0
-mul.f r2.x, r2.x, r0.x
-(sy)mad.f32 r0.y, r6.x, r1.x, r0.y
-mul.f r0.x, r2.w, r0.x
-mad.f32 r1.x, r5.w, r1.y, r2.x
-mad.f32 r0.x, r5.z, r1.z, r0.x
-mov.f32f32 r0.y, r0.y
-(rpt2)nop
-mul.f r0.y, r0.y, r0.w
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r0.x, r0.x
+mul.f r1.x, r4.y, r0.x
+mad.f32 r0.y, r6.x, r3.w, r0.y
+mad.f32 r1.x, r5.w, r5.y, r1.x
+(ss)mul.f r0.x, r4.w, r0.x
 nop
-mov.f32f32 r0.y, r0.y
-mul.f r1.x, r1.x, r0.w
+mul.f r0.y, r0.y, r3.y
+mul.f r1.x, r1.x, r3.y
 mad.f32 r0.y, c7.z, r6.x, r0.y
-mul.f r0.x, r0.x, r0.w
-(rpt1)nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r1.x
-mov.f32f32 r0.x, r0.x
-mad.f32 r0.w, c7.y, r5.w, r0.w
+mad.f32 r1.x, c7.y, r5.w, r1.x
+mad.f32 r0.x, r5.z, r0.w, r0.x
+nop
 mul.f r0.y, r0.z, r0.y
+mul.f r0.w, r0.z, r1.x
+mul.f r0.x, r0.x, r1.w
+nop
+add.f r0.y, r0.y, r2.x
+add.f r0.w, r0.w, r3.z
 mad.f32 r0.x, c7.x, r5.z, r0.x
-(rpt1)nop
-add.f r0.y, r0.y, r5.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r0.x, r0.x
 nop
-mul.f r0.y, r0.y, r1.w
-mul.f r0.w, r0.z, r0.w
+mul.f r2.z, r0.y, r1.z
+mul.f r2.y, r0.w, r1.z
 mul.f r0.x, r0.z, r0.x
-nop
-mov.f32f32 r0.y, r0.y
-add.f r0.z, r0.w, r5.x
-add.f r0.x, r0.x, r6.w
-nop
-mov.f32f32 r0.y, r0.y
-mul.f r0.z, r0.z, r1.w
-mul.f r0.x, r0.x, r1.w
-nop
-mov.f32f32 r4.x, r0.y
-mov.f32f32 r0.y, r0.z
-mov.f32f32 r0.x, r0.x
-(rpt1)nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.x, r0.x
-(rpt1)nop
-mov.f32f32 r3.w, r0.y
-mov.f32f32 r3.z, r0.x
+(rpt2)nop
+add.f r0.x, r0.x, r4.z
+(rpt2)nop
+mul.f r2.x, r0.x, r1.z
 end
 nop
 nop
-; FRAG: outputs: r3.z (1:0)
-; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r2.w (5:10,cm=f,il=12,b=1) r2.y (5:11,cm=f,il=16,b=1) r4.z (5:12,cm=f,il=20,b=1) r0.y (5:13,cm=f,il=24,b=1) r3.y (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
-; FRAG: 336 instructions, 0 half, 9 full
-; pos (bary): r1.x
-; color: r3.z
-; fragcoord: r0.x
+; FRAG: outputs: r2.x (1:0)
+; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r4.z (5:10,cm=f,il=12,b=1) r2.z (5:11,cm=f,il=16,b=1) r2.w (5:12,cm=f,il=20,b=1) r3.y (5:13,cm=f,il=24,b=1) r5.x (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
+; FRAG: 227 instructions, 0 half, 10 full