From 1b48ce3fe3410ef6862d34c7288a69ab201b532d Mon Sep 17 00:00:00 2001
From: UnknownShadow200 <unknownshadow200@gmail.com>
Date: Sat, 10 Aug 2024 10:08:06 +1000
Subject: [PATCH] Dreamcast: Slightly optimise clipping by avoiding colour
 interpolation unless necessary

---
 misc/dreamcast/VertexClip.S  | 10 ++++++++--
 misc/dreamcast/VertexClip2.S |  9 ++++++++-
 src/Platform_3DS.c           |  2 +-
 3 files changed, 17 insertions(+), 4 deletions(-)

diff --git a/misc/dreamcast/VertexClip.S b/misc/dreamcast/VertexClip.S
index cec3e069c..90df94d9e 100644
--- a/misc/dreamcast/VertexClip.S
+++ b/misc/dreamcast/VertexClip.S
@@ -123,10 +123,15 @@ _ClipLine:
 	add      #-4, OUT ! EX, OUT -= 4
 	
 	mov.l  @IN1,CL1   ! LS, ACOLOR = v1->bgra
+	mov.l  @IN2,CL2   ! LS, BCOLOR = v2->bgra
+! Bypass interpolation if unnecessary
+	cmp/eq  CL1,CL2   ! MT, T = ACOLOR == BCOLOR
+	bt.s    1f        ! BR, if (T) goto 1;
+	mov     CL1,CLO   ! MT, OUTCOLOR = ACOLOR (branch delay instruction)
+	
 	extu.b  CL1,TM1   ! EX, val = ACOLOR.b
 	lds     TM1,fpul  ! CO, FPUL = val
 	float   fpul,fr2  ! EX, fr2 = float(FPUL)
-	mov.l  @IN2,CL2   ! LS, BCOLOR = v2->bgra
 	extu.b  CL2,TM1   ! EX, val = BCOLOR.b
 	lds     TM1,fpul  ! CO, FPUL = val
 	float   fpul,fr3  ! EX, fr3 = float(FPUL)
@@ -181,8 +186,9 @@ _ClipLine:
 	shll16  TM2       ! EX, tmp <<= 16
 	shll8   TM2       ! EX, tmp <<= 8
 	or      TM2,CLO   ! EX, OUTCOLOR.a |= tmp
-	mov.l  CLO,@OUT   ! LS, OUT->color = OUTCOLOR
 
+1:
+	mov.l  CLO,@OUT   ! LS, OUT->color = OUTCOLOR
 	mov.l   @r15+,r5  ! LS, pop(r5)
 	rts               ! CO, return after executing instruction in delay slot
 	mov.l   @r15+,r4  ! LS, pop(r4)
diff --git a/misc/dreamcast/VertexClip2.S b/misc/dreamcast/VertexClip2.S
index ee04dc6c6..de372e33b 100644
--- a/misc/dreamcast/VertexClip2.S
+++ b/misc/dreamcast/VertexClip2.S
@@ -121,10 +121,15 @@ _ClipEdge:
 	add      #-4, OUT ! EX, OUT -= 4
 	
 	mov.l  @IN1,CL1   ! LS, ACOLOR = v1->bgra
+	mov.l  @IN2,CL2   ! LS, BCOLOR = v2->bgra
+! Bypass interpolation if unnecessary
+	cmp/eq  CL1,CL2   ! MT, T = ACOLOR == BCOLOR
+	bt.s    1f        ! BR, if (T) goto 1;
+	mov     CL1,CLO   ! MT, OUTCOLOR = ACOLOR (branch delay instruction)
+	
 	extu.b  CL1,TM1   ! EX, val = ACOLOR.b
 	lds     TM1,fpul  ! CO, FPUL = val
 	float   fpul,fr2  ! EX, fr2 = float(FPUL)
-	mov.l  @IN2,CL2   ! LS, BCOLOR = v2->bgra
 	extu.b  CL2,TM1   ! EX, val = BCOLOR.b
 	lds     TM1,fpul  ! CO, FPUL = val
 	float   fpul,fr3  ! EX, fr3 = float(FPUL)
@@ -179,5 +184,7 @@ _ClipEdge:
 	shll16  TM2       ! EX, tmp <<= 16
 	shll8   TM2       ! EX, tmp <<= 8
 	or      TM2,CLO   ! EX, OUTCOLOR.a |= tmp
+
+1:
 	rts               ! CO, return after executing instruction in delay slot
 	mov.l  CLO,@OUT   ! LS, OUT->color = OUTCOLOR
diff --git a/src/Platform_3DS.c b/src/Platform_3DS.c
index c2df0883f..5771b379e 100644
--- a/src/Platform_3DS.c
+++ b/src/Platform_3DS.c
@@ -34,9 +34,9 @@
 
 const cc_result ReturnCode_FileShareViolation = 1000000000; /* TODO: not used apparently */
 const cc_result ReturnCode_FileNotFound     = ENOENT;
+const cc_result ReturnCode_DirectoryExists  = EEXIST;
 const cc_result ReturnCode_SocketInProgess  = EINPROGRESS;
 const cc_result ReturnCode_SocketWouldBlock = EWOULDBLOCK;
-const cc_result ReturnCode_DirectoryExists  = EEXIST;
 
 const char* Platform_AppNameSuffix = " 3DS";
 cc_bool Platform_ReadonlyFilesystem;