From 69f0df9cbaed6620a190ac289958c5bf7c079d92 Mon Sep 17 00:00:00 2001
From: notaz <notasas@gmail.com>
Date: Thu, 16 Jun 2011 18:43:51 +0300
Subject: [PATCH] arm_utils: relax alignment requirements

the code needing alignment isn't reached often anyway.
---
 frontend/arm_utils.s       | 27 ++++++++++++++++++++++-----
 plugins/dfxvideo/draw_fb.c |  2 +-
 plugins/gpu_unai/gpu.cpp   |  2 +-
 3 files changed, 24 insertions(+), 7 deletions(-)

diff --git a/frontend/arm_utils.s b/frontend/arm_utils.s
index 4e561f93..d1319230 100644
--- a/frontend/arm_utils.s
+++ b/frontend/arm_utils.s
@@ -15,7 +15,8 @@
 bgr555_to_rgb565:
     mov         r3, #0x07c0
     vdup.16     q15, r3
-    sub         r2, r2, #64
+    subs        r2, r2, #64
+    blt         btr16_end64
 0:
     vldmia      r1!, {q0-q3}
     vshl.u16    q4, q0, #11
@@ -38,20 +39,36 @@ bgr555_to_rgb565:
     subs        r2, r2, #64
     bge         0b
 
+btr16_end64:
     adds        r2, r2, #64
     bxeq        lr
+    subs        r2, r2, #16
+    blt         btr16_end16
 
-    @ handle the remainder
+    @ handle the remainder (reasonably rare)
 0:
-    vld1.16     {q0}, [r1, :64]!
+    vld1.16     {q0}, [r1]!
     vshl.u16    q1, q0, #11
     vshl.u16    q2, q0, #1
     vsri.u16    q1, q0, #10
     vbit        q1, q2, q15
     subs        r2, r2, #16
-    vst1.16     {q1}, [r0, :64]!
-    bgt         0b
+    vst1.16     {q1}, [r0]!
+    bge         0b
 
+btr16_end16:
+    adds        r2, r2, #16
+    bxeq        lr
+    subs        r2, r2, #8
+    bxlt        lr
+
+    @ very rare
+    vld1.16     d0, [r1]!
+    vshl.u16    d1, d0, #11
+    vshl.u16    d2, d0, #1
+    vsri.u16    d1, d0, #10
+    vbit        d1, d2, d30
+    vst1.16     d1, [r0]!
     bx          lr
 
 
diff --git a/plugins/dfxvideo/draw_fb.c b/plugins/dfxvideo/draw_fb.c
index dbf0fc84..b5608130 100644
--- a/plugins/dfxvideo/draw_fb.c
+++ b/plugins/dfxvideo/draw_fb.c
@@ -21,7 +21,7 @@ unsigned long  lSetMask;
 
 static void blit(void)
 {
- int px = PSXDisplay.DisplayPosition.x & ~3; // XXX: align needed by bgr*_to_...
+ int px = PSXDisplay.DisplayPosition.x & ~1; // XXX: align needed by bgr*_to_...
  int py = PSXDisplay.DisplayPosition.y;
  int w = PreviousPSXDisplay.Range.x1;
  int h = PreviousPSXDisplay.DisplayMode.y;
diff --git a/plugins/gpu_unai/gpu.cpp b/plugins/gpu_unai/gpu.cpp
index 5152327a..0ef8f51f 100644
--- a/plugins/gpu_unai/gpu.cpp
+++ b/plugins/gpu_unai/gpu.cpp
@@ -879,7 +879,7 @@ static void blit(void)
 	u16 *srcs;
 	u8  *dest;
 
-	x0 = DisplayArea[0] & ~3; // alignment needed by blitter
+	x0 = DisplayArea[0] & ~1; // alignment needed by blitter
 	y0 = DisplayArea[1];
 	srcs = &((u16*)GPU_FrameBuffer)[FRAME_OFFSET(x0,y0)];
 
-- 
2.39.5