1 files changed, 98 insertions, 164 deletions
diff --git a/reference/0ad-alpine-valley/0ad-56.asm b/reference/0ad-alpine-valley/0ad-56.asm
index d4490e5..494f814 100644
--- a/reference/0ad-alpine-valley/0ad-56.asm
+++ b/reference/0ad-alpine-valley/0ad-56.asm
@@ -1,16 +1,16 @@
 ; options:
 ; VERT: new compiler
-@in(r2.w)	in0
-@in(r3.x)	in1
-@in(r3.y)	in2
-@in(r8.x)	in4
-@in(r8.y)	in5
-@in(r8.z)	in6
-@in(r6.y)	in8
-@in(r6.z)	in9
-@in(r6.w)	in10
-@in(r8.w)	in12
-@in(r9.x)	in13
+@in(r7.x)	in0
+@in(r7.y)	in1
+@in(r7.z)	in2
+@in(r6.y)	in4
+@in(r6.z)	in5
+@in(r6.w)	in6
+@in(r8.x)	in8
+@in(r8.y)	in9
+@in(r8.z)	in10
+@in(r3.x)	in12
+@in(r3.y)	in13
 @out(r0.x)	out0
 @out(r0.y)	out1
 @out(r0.z)	out2
@@ -43,167 +43,101 @@
 @out(r7.y)	out29
 @out(r7.z)	out30
 @out(r7.w)	out31
-(sy)(ss)add.f r0.x, c4.x, (neg)r2.w
-mul.f r0.y, r8.x, r8.x
-mul.f r0.z, c8.w, r2.w
-mul.f r0.w, c8.z, r2.w
+@const(c13.x)	0x3f800000, 0x00000000, 0xbf800000, 0x00000000
+@const(c14.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)add.f r0.x, c4.x, (neg)r7.x
+mul.f r0.y, r6.y, r6.y
+mul.f r0.z, c8.y, r7.x
+mul.f r0.w, c8.x, r7.x
 mul.f r1.x, r0.x, r0.x
-add.f r1.y, c4.y, (neg)r3.x
+add.f r1.z, c4.y, (neg)r7.y
 add.f r0.y, c13.x, (neg)r0.y
-mad.f32 r0.z, c9.w, r3.x, r0.z
-mad.f32 r0.w, c9.z, r3.x, r0.w
-mad.f32 r1.x, r1.y, r1.y, r1.x
-mov.f32f32 r0.y, r0.y
-mad.f32 r0.z, c10.w, r3.y, r0.z
-mad.f32 r0.w, c10.z, r3.y, r0.w
-mov.f32f32 r1.x, r1.x
-add.f r1.z, c4.z, (neg)r3.y
-mul.f r1.w, r0.y, r0.y
-mul.f r2.x, r8.y, r8.x
-add.f r0.z, r0.z, c11.w
+mad.f32 r0.z, c9.y, r7.y, r0.z
+mad.f32 r0.w, c9.x, r7.y, r0.w
 mad.f32 r1.x, r1.z, r1.z, r1.x
-add.f r0.w, r0.w, c11.z
-mul.f r2.y, c8.y, r2.w
-mul.f r2.z, c8.x, r2.w
-add.f r2.x, c13.y, (neg)r2.x
-mul.f r3.z, r6.w, c6.z
-mov.f32f32 r0.z, r0.z
+add.f r1.w, c4.z, (neg)r7.z
+mov.f32f32 r1.y, r0.y
+mad.f32 r0.z, c10.y, r7.z, r0.z
+mad.f32 r0.w, c10.x, r7.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, r6.z, r6.y
+add.f r0.z, r0.z, c11.y
+add.f r0.w, r0.w, c11.x
+mul.f r2.y, r6.w, r6.y
+mul.f r2.z, c8.w, r7.x
 rsq r1.x, r1.x
-(ss)mov.f32f32 r1.x, r1.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r2.y, c9.y, r3.x, r2.y
+(ss)mov.f32f32 r2.w, r1.x
+add.f r3.z, c13.y, (neg)r1.y
 mad.f32 r0.x, r0.x, r1.x, (neg)c5.x
-mad.f32 r1.w, r2.x, r2.x, r1.w
-mad.f32 r1.y, r1.y, r1.x, (neg)c5.y
-mad.f32 r1.x, r1.z, r1.x, (neg)c5.z
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r3.w, r1.w
-mul.f r1.z, r8.z, r8.x
-mov.f32f32 r1.y, r1.y
-mul.f r1.w, r0.x, r0.x
-mov.f32f32 r1.x, r1.x
-add.f r4.x, c13.y, (neg)r1.z
-mad.f32 r4.y, r1.y, r1.y, r1.w
-mov.f32f32 r1.w, r0.z
-mov.f32f32 r1.z, r0.w
-mad.f32 r0.z, c10.y, r3.y, r2.y
-mov.f32f32 r0.w, r4.y
-mov.f32f32 r2.y, r4.x
-mad.f32 r0.w, r1.x, r1.x, r0.w
-add.f r0.z, r0.z, c11.y
-mad.f32 r2.z, c9.x, r3.x, r2.z
-mov.f32f32 r3.z, r3.z
-mad.f32 r2.z, c10.x, r3.y, r2.z
-mul.f r5.z, c0.w, r2.w
-mul.f r5.w, c0.z, r2.w
-rsq r0.w, r0.w
-(ss)mov.f32f32 r0.w, r0.w
-mad.f32 r3.w, r2.y, r2.y, r3.w
-mul.f r0.z, r0.z, c12.y
-add.f r2.z, r2.z, c11.x
-mul.f r1.x, r1.x, r0.w
-mul.f r4.x, r1.y, r0.w
-mul.f r0.x, r0.x, r0.w
-mov.f32f32 r1.y, r0.z
-mov.f32f32 r0.z, r1.x
-mov.f32f32 r0.w, r4.x
-mov.f32f32 r0.x, r0.x
-rsq r1.x, r3.w
-(ss)mov.f32f32 r1.x, r1.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r0.x, r0.x
+mul.f r1.y, r0.z, c12.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, c13.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, c12.x
+mov.f32f32 r0.w, r2.y
+mad.f32 r2.z, c9.w, r7.y, r2.z
+mad.f32 r0.z, r1.w, r3.z, r0.z
+mad.f32 r1.w, c10.w, r7.z, r2.z
+mul.f r2.z, c8.z, r7.x
+mul.f r3.w, c0.w, r7.x
+mul.f r4.x, c0.z, r7.x
+mul.f r5.z, c0.y, r7.x
+mul.f r5.w, c0.x, r7.x
+rsq r0.z, r0.z
+(ss)mov.f32f32 r4.y, r0.z
+mul.f r4.w, r0.x, r0.z
+mad.f32 r0.x, r2.y, r0.w, r2.x
+add.f r1.w, r1.w, c11.w
+mul.f r5.y, r3.z, r4.y
+mul.f r5.x, r2.w, r4.y
+(ss)mad.f32 r0.z, c9.z, r7.y, r2.z
+mad.f32 r2.y, c1.w, r7.y, r3.w
+mad.f32 r2.w, c1.z, r7.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, c10.z, r7.z, r0.z
+mad.f32 r0.y, c2.w, r7.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, c11.z
+mov.f32f32 r0.x, r2.z
+mov.f32f32 r3.z, r2.y
+mul.f r3.w, r6.z, r0.z
+add.f r0.w, r0.y, c3.w
+mul.f r0.y, r6.y, r0.x
+mul.f r4.x, r6.w, r3.z
+mad.f32 r4.y, r6.w, r0.z, (neg)r0.y
+mad.f32 r4.x, r6.z, r0.x, (neg)r4.x
+mad.f32 r4.z, r6.y, r3.z, (neg)r3.w
+mad.f32 r0.x, c2.z, r7.z, r2.w
+mad.f32 r0.y, c1.y, r7.y, r5.z
+mad.f32 r2.w, c1.x, r7.y, r5.w
+mad.f32 r0.y, c2.y, r7.z, r0.y
+add.f r0.z, r0.x, c3.z
+mad.f32 r0.x, c2.x, r7.z, r2.w
 nop
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-(rpt1)nop
-mov.f32f32 r5.y, r0.z
-mov.f32f32 r5.x, r0.w
-mov.f32f32 r4.w, r0.x
-mul.f r0.x, r0.y, r1.x
-mul.f r0.y, r2.y, r1.x
-mul.f r0.z, r2.x, r1.x
-mul.f r0.w, r2.z, c12.x
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.x, r0.w
-mul.f r0.w, r8.y, r0.x
-mul.f r2.x, r8.x, r0.y
-mad.f32 r0.w, r8.x, r0.z, (neg)r0.w
-mad.f32 r2.x, r8.z, r0.x, (neg)r2.x
-mul.f r2.y, r8.z, r0.z
-mov.f32f32 r2.z, r0.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r2.x, r2.x
-mad.f32 r0.y, r8.y, r0.y, (neg)r2.y
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r4.z, r0.w
-mov.f32f32 r4.y, r2.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.x, r0.x
-nop
-mov.f32f32 r4.x, r0.y
-mov.f32f32 r2.y, r0.z
-mov.f32f32 r2.x, r0.x
-mov.f32f32 r0.x, r3.z
-mad.f32 r0.y, c1.w, r3.x, r5.z
-mad.f32 r0.z, c1.z, r3.x, r5.w
-mul.f r0.w, c0.y, r2.w
-mov.f32f32 r6.x, r0.x
-mad.f32 r0.x, c2.w, r3.y, r0.y
-mad.f32 r0.y, c2.z, r3.y, r0.z
-mad.f32 r0.z, c1.y, r3.x, r0.w
-mul.f r0.w, c0.x, r2.w
-add.f r0.x, r0.x, c3.w
-add.f r0.y, r0.y, c3.z
-mad.f32 r3.z, c2.y, r3.y, r0.z
-(ss)mad.f32 r3.w, c1.x, r3.x, r0.w
-mov.f32f32 r0.w, r0.x
-mov.f32f32 r0.z, r0.y
-add.f r0.x, r3.z, c3.y
-mad.f32 r3.z, c2.x, r3.y, r3.w
-mul.f r3.w, r6.z, c6.y
-mul.f r5.z, r6.y, c6.x
-mov.f32f32 r0.y, r0.x
-add.f r0.x, r3.z, c3.x
-mov.f32f32 r3.z, r3.w
-mov.f32f32 r3.w, r5.z
-mad.f32 r6.y, c7.x, r3.y, c7.y
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r5.w, r3.z
-mov.f32f32 r5.z, r3.w
-mov.f32f32 r3.z, r6.y
-mad.f32 r6.y, c7.x, r2.w, c7.y
-mov.f32f32 r6.z, (0.000000)
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r3.w, r3.z
-mov.f32f32 r3.z, r6.y
-mov.f32f32 r7.w, r6.z
-mov.f32f32 r7.z, r3.y
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.y, r8.z
-mov.f32f32 r7.y, r3.x
-mov.f32f32 r3.x, r8.y
-mov.f32f32 r7.x, r2.w
-mov.f32f32 r6.w, r3.y
-mov.f32f32 r2.w, r8.x
-mov.f32f32 r6.z, r3.x
-mov.f32f32 r3.x, r9.x
-mov.f32f32 r8.x, r8.w
-mov.f32f32 r6.y, r2.w
+add.f r0.y, r0.y, c3.y
+mov.f32f32 r7.w, (0.000000)
+add.f r0.x, r0.x, c3.x
+mul.f r6.x, r8.z, c6.z
+mul.f r5.w, r8.y, c6.y
+mul.f r5.z, r8.x, c6.x
+mad.f32 r3.w, c7.x, r7.z, c7.y
+mad.f32 r3.z, c7.x, r7.x, c7.y
 mov.f32f32 r2.w, c13.z
-mov.f32f32 r3.y, r3.x
-mov.f32f32 r3.x, r8.x
-nop
-mov.f32f32 r2.w, r2.w
 end
 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: r2.w (0:0,cm=7,il=8,b=0) r8.x (0:0,cm=7,il=12,b=0) r6.y (0:0,cm=7,il=16,b=0) r8.w (0:0,cm=3,il=20,b=0)
-; VERT: 160 instructions, 0 half, 10 full
+; VERT: inputs: r7.x (0:0,cm=7,il=8,b=0) r6.y (0:0,cm=7,il=12,b=0) r8.x (0:0,cm=7,il=16,b=0) r3.x (0:0,cm=3,il=20,b=0)
+; VERT: 91 instructions, 0 half, 9 full