From 8b5f8f8bb57a38a958dea26a492a08a66640500d Mon Sep 17 00:00:00 2001
From: Dimitris Panokostas <midwan@gmail.com>
Date: Sat, 17 Dec 2016 11:42:10 +0100
Subject: [PATCH] Added missing arm_helper.s

Added missing file to allow Picasso96 on Pi 1/Zero again
---
 Makefile                    | 11 ++++-
 src/od-pandora/arm_helper.s | 97 +++++++++++++++++++++++++++++++++++++
 2 files changed, 107 insertions(+), 1 deletion(-)
 create mode 100644 src/od-pandora/arm_helper.s

diff --git a/Makefile b/Makefile
index 542a40c4..f7e71d1b 100644
--- a/Makefile
+++ b/Makefile
@@ -197,8 +197,13 @@ OBJS =	\
 OBJS += src/od-rasp/rasp_gfx.o
 OBJS += src/od-pandora/gui/sdltruetypefont.o
 OBJS += src/od-pandora/picasso96.o
-OBJS += src/od-pandora/neon_helper.o
 
+ifdef USE_ARMNEON
+	OBJS += src/od-pandora/neon_helper.o
+else
+	OBJS += src/od-pandora/arm_helper.o
+endif
+	
 OBJS += src/newcpu.o
 OBJS += src/newcpu_common.o
 OBJS += src/readcpu.o
@@ -215,6 +220,10 @@ OBJS += src/jit/compemu_support.o
 src/od-pandora/neon_helper.o: src/od-pandora/neon_helper.s
 	$(CXX) $(CPU_FLAGS) -Wall -o src/od-pandora/neon_helper.o -c src/od-pandora/neon_helper.s
 
+src/od-pandora/arm_helper.o: src/od-pandora/arm_helper.s
+	$(CXX) $(CPU_FLAGS) -Wall -o src/od-pandora/arm_helper.o -c src/od-pandora/arm_helper.s
+
+	
 src/trace.o: src/trace.c
 	$(CC) $(MORE_CFLAGS) -c src/trace.c -o src/trace.o
 
diff --git a/src/od-pandora/arm_helper.s b/src/od-pandora/arm_helper.s
new file mode 100644
index 00000000..9db26dd0
--- /dev/null
+++ b/src/od-pandora/arm_helper.s
@@ -0,0 +1,97 @@
+@ Some functions and tests to increase performance in drawing.cpp and custom.cpp
+
+.arm
+
+.global copy_screen_8bit
+.global copy_screen_16bit_swap
+.global copy_screen_32bit_to_16bit
+
+.text
+
+.align 8
+
+
+@----------------------------------------------------------------
+@ copy_screen_8bit
+@
+@ r0: uae_u8   *dst
+@ r1: uae_u8   *src
+@ r2: int      bytes always a multiple of 64: even number of lines, number of pixel per line is multiple of 32 (320, 640, 800, 1024, 1152, 1280)
+@ r3: uae_u32  *clut
+@
+@ void copy_screen_8bit(uae_u8 *dst, uae_u8 *src, int bytes, uae_u32 *clut);
+@
+@----------------------------------------------------------------
+copy_screen_8bit:
+  stmdb     sp!, {r4-r6, lr}
+copy_screen_8bit_loop:
+  pld       [r1, #192]
+  mov       lr, #64
+copy_screen_8bit_loop_2:
+  ldr       r4, [r1], #4
+  and       r5, r4, #255
+  ldr       r6, [r3, r5, lsl #2]
+  lsr r5, r4, #8
+  and r5, r5, #255
+  strh      r6, [r0], #2
+  ldr       r6, [r3, r5, lsl #2]
+  lsr r5, r4, #16
+  and r5, r5, #255
+  strh      r6, [r0], #2
+  ldr       r6, [r3, r5, lsl #2]
+  lsr r5, r4, #24
+  strh      r6, [r0], #2
+  ldr       r6, [r3, r5, lsl #2]
+  subs      lr, lr, #4
+  strh      r6, [r0], #2
+  bgt       copy_screen_8bit_loop_2
+  subs      r2, r2, #64
+  bgt       copy_screen_8bit_loop
+  ldmia     sp!, {r4-r6, pc}
+
+
+@----------------------------------------------------------------
+@ copy_screen_16bit_swap
+@
+@ r0: uae_u8   *dst
+@ r1: uae_u8   *src
+@ r2: int      bytes always a multiple of 128: even number of lines, 2 bytes per pixel, number of pixel per line is multiple of 32 (320, 640, 800, 1024, 1152, 1280)
+@
+@ void copy_screen_16bit_swap(uae_u8 *dst, uae_u8 *src, int bytes);
+@
+@----------------------------------------------------------------
+copy_screen_16bit_swap:
+ldr r3, [r1], #4
+rev16 r3, r3
+str r3, [r0], #4
+subs r2, r2, #4
+bne copy_screen_16bit_swap
+bx lr
+
+
+@----------------------------------------------------------------
+@ copy_screen_32bit_to_16bit
+@
+@ r0: uae_u8   *dst - Format (bits): rrrr rggg gggb bbbb
+@ r1: uae_u8   *src - Format (bytes) in memory rgba
+@ r2: int      bytes
+@
+@ void copy_screen_32bit_to_16bit(uae_u8 *dst, uae_u8 *src, int bytes);
+@
+@----------------------------------------------------------------
+copy_screen_32bit_to_16bit:
+stmdb sp!, {r4-r6, lr}
+copy_screen_32bit_to_16bit_loop:
+ldr r3, [r1], #4
+rev r3, r3
+lsr r4, r3, #27
+lsr r5, r3, #18
+and r5, r5, #63
+lsr r6, r3, #11
+and r6, r6, #31
+orr r6, r6, r5, lsl #5
+orr r6, r6, r4, lsl #11
+strh r6, [r0], #2
+subs r2, r2, #4
+bne copy_screen_32bit_to_16bit_loop
+ldmia sp!, {r4-r6, pc}
\ No newline at end of file