1 files changed, 39 insertions, 78 deletions
diff --git a/reference/xon6.asm b/reference/xon6.asm
index 7603f89..d468c8e 100644
--- a/reference/xon6.asm
+++ b/reference/xon6.asm
@@ -1,24 +1,24 @@
 ; options:
 ; VERT: new compiler
-@in(r1.z)	in0
-@in(r1.w)	in1
-@in(r2.x)	in2
-@in(r2.y)	in3
+@in(r4.w)	in0
+@in(r5.x)	in1
+@in(r5.y)	in2
+@in(r5.z)	in3
 @in(r0.x)	in4
 @in(r0.y)	in5
 @in(r0.z)	in6
 @in(r0.w)	in7
-@in(r5.x)	in8
-@in(r5.y)	in9
-@in(r5.z)	in10
-@in(r2.z)	in12
-@in(r2.w)	in13
-@in(r3.x)	in14
-@in(r3.y)	in16
-@in(r3.z)	in17
-@in(r3.w)	in18
-@in(r5.w)	in20
-@in(r6.x)	in21
+@in(r2.x)	in8
+@in(r2.y)	in9
+@in(r2.z)	in10
+@in(r3.x)	in12
+@in(r3.y)	in13
+@in(r3.z)	in14
+@in(r4.x)	in16
+@in(r4.y)	in17
+@in(r4.z)	in18
+@in(r1.z)	in20
+@in(r1.w)	in21
 @out(r0.x)	out0
 @out(r0.y)	out1
 @out(r0.z)	out2
@@ -39,75 +39,36 @@
 @out(r4.y)	out17
 @out(r4.z)	out18
 @out(r4.w)	out19
+@const(c8.x)	0x00000000, 0x00000000, 0x00000000, 0x00000000
 (sy)(ss)mul.f r1.x, c4.y, r0.x
 mul.f r0.x, c4.x, r0.x
 mad.f32 r1.x, c5.y, r0.y, r1.x
 mad.f32 r0.x, c5.x, r0.y, r0.x
 mad.f32 r0.y, c6.y, r0.z, r1.x
 mad.f32 r0.x, c6.x, r0.z, r0.x
-mad.f32 r0.y, c7.y, r0.w, r0.y
-mad.f32 r0.x, c7.x, r0.w, r0.x
-mul.f r0.z, c0.w, r1.z
-mul.f r0.w, c0.z, r1.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.x, r0.x
-mad.f32 r0.z, c1.w, r1.w, r0.z
-mad.f32 r0.w, c1.z, r1.w, r0.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.x, r0.x
-mad.f32 r0.z, c2.w, r2.x, r0.z
-mad.f32 r0.w, c2.z, r2.x, r0.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.x, r0.x
-(rpt1)nop
-mov.f32f32 r1.y, r0.y
-mov.f32f32 r1.x, r0.x
-mad.f32 r0.x, c3.w, r2.y, r0.z
-mad.f32 r0.y, c3.z, r2.y, r0.w
-mul.f r4.x, c0.y, r1.z
-mul.f r1.z, c0.x, r1.z
-mov.f32f32 r0.w, r0.x
-mov.f32f32 r0.z, r0.y
-mad.f32 r0.x, c1.y, r1.w, r4.x
-mad.f32 r0.y, c1.x, r1.w, r1.z
-mad.f32 r0.x, c2.y, r2.x, r0.x
-mad.f32 r0.y, c2.x, r2.x, r0.y
-mad.f32 r0.x, c3.y, r2.y, r0.x
-mad.f32 r1.z, c3.x, r2.y, r0.y
-(rpt1)nop
-mov.f32f32 r0.y, r0.x
-mov.f32f32 r0.x, r1.z
-mov.f32f32 r1.z, c8.x
-mov.f32f32 r1.w, r3.w
-mov.f32f32 r2.x, r3.z
-mov.f32f32 r2.y, r3.y
-mov.f32f32 r4.w, r1.z
-mov.f32f32 r4.z, r1.w
-mov.f32f32 r4.y, r2.x
-mov.f32f32 r4.x, r2.y
-mov.f32f32 r1.z, c8.x
-mov.f32f32 r1.w, r3.x
-mov.f32f32 r2.x, r2.w
-mov.f32f32 r2.y, r2.z
-mov.f32f32 r3.w, r1.z
-mov.f32f32 r3.z, r1.w
-mov.f32f32 r3.y, r2.x
-mov.f32f32 r3.x, r2.y
-mov.f32f32 r1.z, c8.x
-mov.f32f32 r1.w, r5.z
-mov.f32f32 r2.x, r5.y
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r2.w, r1.z
-mov.f32f32 r2.z, r1.w
-mov.f32f32 r2.y, r2.x
-mov.f32f32 r2.x, r5.x
-mov.f32f32 r1.z, r6.x
-mov.f32f32 r5.x, r5.w
-(rpt1)nop
-mov.f32f32 r1.w, r1.z
-mov.f32f32 r1.z, r5.x
+mad.f32 r1.y, c7.y, r0.w, r0.y
+mad.f32 r1.x, c7.x, r0.w, r0.x
+mul.f r0.x, c0.w, r4.w
+mul.f r0.y, c0.z, r4.w
+mad.f32 r0.x, c1.w, r5.x, r0.x
+mad.f32 r0.y, c1.z, r5.x, r0.y
+mad.f32 r0.x, c2.w, r5.y, r0.x
+mad.f32 r0.y, c2.z, r5.y, r0.y
+mad.f32 r0.w, c3.w, r5.z, r0.x
+mad.f32 r0.z, c3.z, r5.z, r0.y
+mul.f r0.x, c0.y, r4.w
+mul.f r0.y, c0.x, r4.w
+mad.f32 r0.x, c1.y, r5.x, r0.x
+mad.f32 r0.y, c1.x, r5.x, r0.y
+mad.f32 r0.x, c2.y, r5.y, r0.x
+mad.f32 r2.w, c2.x, r5.y, r0.y
+mad.f32 r0.y, c3.y, r5.z, r0.x
+mad.f32 r0.x, c3.x, r5.z, r2.w
+mov.f32f32 r4.w, c8.x
+mov.f32f32 r3.w, c8.x
+mov.f32f32 r2.w, c8.x
 end
 ; VERT: outputs: r0.x (0:0) r1.x (5:20) r2.x (5:21) r3.x (5:22) r4.x (5:23)
-; VERT: inputs: r1.z (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=f,il=12,b=0) r5.x (0:0,cm=7,il=16,b=0) r2.z (0:0,cm=7,il=20,b=0) r3.y (0:0,cm=7,il=24,b=0) r5.w (0:0,cm=3,il=28,b=0)
-; VERT: 71 instructions, 0 half, 7 full
+; VERT: inputs: r4.w (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=f,il=12,b=0) r2.x (0:0,cm=7,il=16,b=0) r3.x (0:0,cm=7,il=20,b=0) r4.x (0:0,cm=7,il=24,b=0) r1.z (0:0,cm=3,il=28,b=0)
+; VERT: 28 instructions, 0 half, 6 full