From dfc864d131a9aa9d71df617eb5e31a200959ac1b Mon Sep 17 00:00:00 2001
From: vsariola <5684185+vsariola@users.noreply.github.com>
Date: Tue, 23 Mar 2021 15:07:24 +0200
Subject: [PATCH] optimize(templates): allow distort to flow into su_waveshaper
 (again)

---
 templates/amd64-386/effects.asm | 16 ----------
 templates/amd64-386/patch.asm   | 55 +++++++++++++++++++++++----------
 2 files changed, 38 insertions(+), 33 deletions(-)

diff --git a/templates/amd64-386/effects.asm b/templates/amd64-386/effects.asm
index bece80a..23007e6 100644
--- a/templates/amd64-386/effects.asm
+++ b/templates/amd64-386/effects.asm
@@ -1,19 +1,3 @@
-{{- if .HasOp "distort"}}
-;-------------------------------------------------------------------------------
-;   DISTORT opcode: apply distortion on the signal
-;-------------------------------------------------------------------------------
-;   Mono:   x   ->  x*a/(1-a+(2*a-1)*abs(x))            where x is clamped first
-;   Stereo: l r ->  l*a/(1-a+(2*a-1)*abs(l)) r*a/(1-a+(2*a-1)*abs(r))
-;-------------------------------------------------------------------------------
-{{.Func "su_op_distort" "Opcode"}}
-{{- if .Stereo "distort" -}}
-    {{.Call "su_effects_stereohelper"}}
-{{- end}}
-    fld     dword [{{.Input "distort" "drive"}}]
-    {{.TailCall "su_waveshaper"}}
-{{end}}
-
-
 {{- if .HasOp "hold"}}
 ;-------------------------------------------------------------------------------
 ;   HOLD opcode: sample and hold the signal, reducing sample rate
diff --git a/templates/amd64-386/patch.asm b/templates/amd64-386/patch.asm
index 9995a38..c34bf08 100644
--- a/templates/amd64-386/patch.asm
+++ b/templates/amd64-386/patch.asm
@@ -127,6 +127,44 @@ su_power:
 
 {{end}}
 
+{{- if .HasOp "distort"}}
+;-------------------------------------------------------------------------------
+;   DISTORT opcode: apply distortion on the signal
+;-------------------------------------------------------------------------------
+;   Mono:   x   ->  x*a/(1-a+(2*a-1)*abs(x))            where x is clamped first
+;   Stereo: l r ->  l*a/(1-a+(2*a-1)*abs(l)) r*a/(1-a+(2*a-1)*abs(r))
+;   This is placed here to be able to flow into waveshaper & also include
+;   wave shaper if needed by some other function; need to investigate the
+;   best way to do this
+;-------------------------------------------------------------------------------
+{{.Func "su_op_distort" "Opcode"}}
+{{- if .Stereo "distort" -}}
+    {{.Call "su_effects_stereohelper"}}
+{{- end}}
+    fld     dword [{{.Input "distort" "drive"}}]
+{{end}}
+
+{{- if or (.HasCall "su_waveshaper") (.HasOp "distort")}}
+{{- if .HasOp "distort"}}
+su_waveshaper:
+{{- else}}
+{{.Func "su_waveshaper"}}
+{{- end}}
+    fld     st0                             ; a a x
+    {{.Prepare (.Float 0.5)}}
+    fsub    dword [{{.Use (.Float 0.5)}}]                 ; a-.5 a x
+    fadd    st0                             ; 2*a-1 a x
+    fld     st2                             ; x 2*a-1 a x
+    fabs                                    ; abs(x) 2*a-1 a x
+    fmulp   st1                             ; (2*a-1)*abs(x) a x
+    fld1                                    ; 1 (2*a-1)*abs(x) a x
+    faddp   st1                             ; 1+(2*a-1)*abs(x) a x
+    fsub    st1                             ; 1-a+(2*a-1)*abs(x) a x
+    fdivp   st1, st0                        ; a/(1-a+(2*a-1)*abs(x)) x
+    fmulp   st1                             ; x*a/(1-a+(2*a-1)*abs(x))
+    ret
+{{end}}
+
 {{- if .HasCall "su_effects_stereohelper" }}
 ;-------------------------------------------------------------------------------
 ;   su_effects_stereohelper: moves the workspace to next, does the filtering for
@@ -145,23 +183,6 @@ su_effects_stereohelper_mono:
 
 {{end}}
 
-{{- if .HasCall "su_waveshaper" }}
-{{.Func "su_waveshaper"}}
-    fld     st0                             ; a a x
-    {{.Prepare (.Float 0.5)}}
-    fsub    dword [{{.Use (.Float 0.5)}}]                 ; a-.5 a x
-    fadd    st0                             ; 2*a-1 a x
-    fld     st2                             ; x 2*a-1 a x
-    fabs                                    ; abs(x) 2*a-1 a x
-    fmulp   st1                             ; (2*a-1)*abs(x) a x
-    fld1                                    ; 1 (2*a-1)*abs(x) a x
-    faddp   st1                             ; 1+(2*a-1)*abs(x) a x
-    fsub    st1                             ; 1-a+(2*a-1)*abs(x) a x
-    fdivp   st1, st0                        ; a/(1-a+(2*a-1)*abs(x)) x
-    fmulp   st1                             ; x*a/(1-a+(2*a-1)*abs(x))
-    ret
-{{end}}
-
 {{- if .HasCall "su_clip"}}
 {{.Func "su_clip"}}
     fld1                                    ; 1 x a