Add alignment to capture loops

hoglet67 · Nov 3, 2020 · fb786ae · fb786ae
1 parent 90541b2
commit fb786ae
Show file tree

Hide file tree

Showing 19 changed files with 62 additions and 69 deletions.
diff --git a/src/capture_line_atariXL_8bpp.S b/src/capture_line_atariXL_8bpp.S
@@ -309,7 +309,7 @@
         .ltorg
 
         // *** 8 bit atari ***
-
+        .align 6
         b       preload_capture_line_atari_8bpp
 capture_line_atari_8bpp:
         push    {lr}
@@ -353,7 +353,7 @@ preload_capture_line_atari_8bpp:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_atari_double_8bpp
 capture_line_atari_double_8bpp:
         push    {lr}
@@ -389,7 +389,7 @@ preload_capture_line_atari_double_8bpp:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_atari_sixbits_8bpp
 capture_line_atari_sixbits_8bpp:
         push    {lr}
@@ -451,7 +451,7 @@ preload_capture_line_atari_sixbits_8bpp:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_atari_sixbits_double_8bpp
 capture_line_atari_sixbits_double_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_4bpp_8bpp.S b/src/capture_line_default_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_default
 capture_line_default_4bpp:
         push    {lr}
@@ -56,7 +56,7 @@ preload_capture_line_default:
 
         // *** 8 bit ***
 
-
+        .align 6
         b       preload_capture_line_default_8bpp
 capture_line_default_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_double_4bpp_8bpp.S b/src/capture_line_default_double_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_default_double
 capture_line_default_double_4bpp:
         push    {lr}
@@ -52,7 +52,7 @@ preload_capture_line_default_double:
 
         // *** 8 bit ***
 
-
+        .align 6
         b       preload_capture_line_default_double_8bpp
 capture_line_default_double_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_onebit_4bpp_8bpp.S b/src/capture_line_default_onebit_4bpp_8bpp.S
@@ -159,7 +159,7 @@
 
 
 
-
+        .align 6
         b       preload_capture_line_default_onebit_4bpp
 capture_line_default_onebit_4bpp:
         push    {lr}
@@ -185,7 +185,7 @@ preload_capture_line_default_onebit_4bpp:
         b       capture_line_default_onebit_4bpp
 
         .ltorg
-
+        .align 6
         b       preload_capture_line_default_onebit_double_4bpp
 capture_line_default_onebit_double_4bpp:
         push    {lr}
@@ -212,7 +212,7 @@ preload_capture_line_default_onebit_double_4bpp:
         .ltorg
         // *** 8 bit ***
 
-
+        .align 6
         b       preload_capture_line_default_onebit_8bpp
 capture_line_default_onebit_8bpp:
         push    {lr}
@@ -244,7 +244,7 @@ preload_capture_line_default_onebit_8bpp:
         .ltorg
         // *** 8 bit ***
 
-
+        .align 6
         b       preload_capture_line_default_onebit_double_8bpp
 capture_line_default_onebit_double_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_sixbits_4bpp_8bpp.S b/src/capture_line_default_sixbits_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_default_sixbits
 capture_line_default_sixbits_4bpp:
         push    {lr}
@@ -65,7 +65,7 @@ preload_capture_line_default_sixbits:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_8bpp
 capture_line_default_sixbits_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_sixbits_8bpp_16bpp.S b/src/capture_line_default_sixbits_8bpp_16bpp.S
@@ -22,7 +22,7 @@
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
 
 // 4bpp not currently used but left in in case
-
+        .align 6
         b       preload_capture_line_default_sixbits
 capture_line_default_sixbits_4bpp:
         push    {lr}
@@ -67,7 +67,7 @@ preload_capture_line_default_sixbits:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_8bpp
 capture_line_default_sixbits_8bpp:
         push    {lr}
@@ -112,7 +112,7 @@ preload_capture_line_default_sixbits_8bpp:
         .ltorg
 
         // *** 16 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_16bpp
 capture_line_default_sixbits_16bpp:
         push    {lr}

diff --git a/src/capture_line_default_sixbits_double_4bpp_8bpp.S b/src/capture_line_default_sixbits_double_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_default_sixbits_double
 capture_line_default_sixbits_double_4bpp:
         push    {lr}
@@ -57,7 +57,7 @@ preload_capture_line_default_sixbits_double:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_double_8bpp
 capture_line_default_sixbits_double_8bpp:
         push    {lr}

diff --git a/src/capture_line_default_sixbits_double_8bpp_16bpp.S b/src/capture_line_default_sixbits_double_8bpp_16bpp.S
@@ -22,7 +22,7 @@
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
 
 // 4bpp not currently used but left in in case
-
+        .align 6
         b       preload_capture_line_default_sixbits_double
 capture_line_default_sixbits_double_4bpp:
         push    {lr}
@@ -58,7 +58,7 @@ preload_capture_line_default_sixbits_double:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_double_8bpp
 capture_line_default_sixbits_double_8bpp:
         push    {lr}
@@ -93,7 +93,7 @@ preload_capture_line_default_sixbits_double_8bpp:
         .ltorg
 
         // *** 16 bit ***
-
+        .align 6
         b       preload_capture_line_default_sixbits_double_16bpp
 capture_line_default_sixbits_double_16bpp:
         push    {lr}

diff --git a/src/capture_line_default_twelvebits_8bpp_16bpp.S b/src/capture_line_default_twelvebits_8bpp_16bpp.S
@@ -26,7 +26,7 @@
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_eightbits_8bpp
 capture_line_default_eightbits_8bpp:
         push    {lr}
@@ -88,7 +88,7 @@ preload_capture_line_default_eightbits_8bpp:
 
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_default_twelvebits_16bpp
 capture_line_default_twelvebits_16bpp:
@@ -127,7 +127,7 @@ preload_capture_line_default_twelvebits_16bpp:
 
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_default_ninebitslo_16bpp
 capture_line_default_ninebitslo_16bpp:
@@ -168,7 +168,7 @@ preload_capture_line_default_ninebitslo_16bpp:
 
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_default_ninebitshi_16bpp
 capture_line_default_ninebitshi_16bpp:

diff --git a/src/capture_line_default_twelvebits_double_8bpp_16bpp.S b/src/capture_line_default_twelvebits_double_8bpp_16bpp.S
@@ -26,7 +26,7 @@
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_default_eightbits_double_8bpp
 capture_line_default_eightbits_double_8bpp:
         push    {lr}
@@ -69,13 +69,9 @@ preload_capture_line_default_eightbits_double_8bpp:
         b       capture_line_default_eightbits_double_8bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_default_twelvebits_double_16bpp
-
-
-
-
 capture_line_default_twelvebits_double_16bpp:
         push    {lr}
         SETUP_VSYNC_DEBUG_16BPP_R11
@@ -106,6 +102,7 @@ preload_capture_line_default_twelvebits_double_16bpp:
 
 
         // *** 16 bit ***
+        .align 6
         b       preload_capture_line_default_ninebitslo_double_16bpp
 capture_line_default_ninebitslo_double_16bpp:
         push    {lr}
@@ -137,6 +134,7 @@ preload_capture_line_default_ninebitslo_double_16bpp:
 
 
         // *** 16 bit ***
+        .align 6
         b       preload_capture_line_default_ninebitshi_double_16bpp
 capture_line_default_ninebitshi_double_16bpp:
         push    {lr}

diff --git a/src/capture_line_fast_4bpp_8bpp.S b/src/capture_line_fast_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_fast
 capture_line_fast_4bpp:
         push    {lr}
@@ -56,7 +56,7 @@ preload_capture_line_fast:
 
         // *** 8 bit ***
 
-
+        .align 6
         b       preload_capture_line_fast_8bpp
 capture_line_fast_8bpp:
         push    {lr}

diff --git a/src/capture_line_fast_simple_16bpp.S b/src/capture_line_fast_simple_16bpp.S
@@ -57,7 +57,7 @@ loop_16bpp_simple\@:
 
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_trailing_pos_16bpp
 capture_line_fast_simple_trailing_pos_16bpp:
@@ -72,7 +72,7 @@ preload_capture_line_fast_simple_trailing_pos_16bpp:
         b       capture_line_fast_simple_trailing_pos_16bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_leading_pos_16bpp
 capture_line_fast_simple_leading_pos_16bpp:
@@ -87,7 +87,7 @@ preload_capture_line_fast_simple_leading_pos_16bpp:
         b       capture_line_fast_simple_leading_pos_16bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_trailing_neg_16bpp
 capture_line_fast_simple_trailing_neg_16bpp:
@@ -102,7 +102,7 @@ preload_capture_line_fast_simple_trailing_neg_16bpp:
         b       capture_line_fast_simple_trailing_neg_16bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_leading_neg_16bpp
 capture_line_fast_simple_leading_neg_16bpp:
@@ -117,7 +117,7 @@ preload_capture_line_fast_simple_leading_neg_16bpp:
         b       capture_line_fast_simple_leading_neg_16bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_trailing_both_16bpp
 capture_line_fast_simple_trailing_both_16bpp:
@@ -131,7 +131,7 @@ preload_capture_line_fast_simple_trailing_both_16bpp:
         b       capture_line_fast_simple_trailing_both_16bpp
 
         .ltorg
-
+        .align 6
         // *** 16 bit ***
         b       preload_capture_line_fast_simple_leading_both_16bpp
 capture_line_fast_simple_leading_both_16bpp:

diff --git a/src/capture_line_fast_sixbits_4bpp_8bpp.S b/src/capture_line_fast_sixbits_4bpp_8bpp.S
@@ -20,7 +20,7 @@
 //   r8 = frame buffer height (=param_fb_height)
 //
 // All registers are available as scratch registers (i.e. nothing needs to be preserved)
-
+        .align 6
         b       preload_capture_line_fast_sixbits
 capture_line_fast_sixbits_4bpp:
         push    {lr}
@@ -63,7 +63,7 @@ preload_capture_line_fast_sixbits:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_fast_sixbits_8bpp
 capture_line_fast_sixbits_8bpp:
         push    {lr}

diff --git a/src/capture_line_fast_sixbits_8bpp_16bpp.S b/src/capture_line_fast_sixbits_8bpp_16bpp.S
@@ -23,7 +23,7 @@
 
 
 // 4bpp not currently used but left in in case
-
+        .align 6
         b       preload_capture_line_fast_sixbits
 capture_line_fast_sixbits_4bpp:
         push    {lr}
@@ -67,7 +67,7 @@ preload_capture_line_fast_sixbits:
         .ltorg
 
         // *** 8 bit ***
-
+        .align 6
         b       preload_capture_line_fast_sixbits_8bpp
 capture_line_fast_sixbits_8bpp:
         push    {lr}
@@ -112,7 +112,7 @@ preload_capture_line_fast_sixbits_8bpp:
         .ltorg
 
         // *** 16 bit ***
-
+        .align 6
         b       preload_capture_line_fast_sixbits_16bpp
 capture_line_fast_sixbits_16bpp:
         push    {lr}