cscpace: fix more alignment issues

libretro/pcsx_rearmed#719

diff --git a/frontend/blit320.s b/frontend/blit320.s
index 201fdea..2d50dfd 100644 (file)
--- a/frontend/blit320.s
+++ b/frontend/blit320.s
@@ -25,6 +25,7 @@
 blit320_640:
     stmfd   sp!, {r4-r8,lr}
     mov     r12, #40
+    bic     r1, r1, #3
 0:
     ldmia   r1!, {r2-r8,lr}
     lhw_str r2, r3
@@ -40,6 +41,7 @@ blit320_640:
 blit320_512:
     stmfd   sp!, {r4-r8,lr}
     mov     r12, #32
+    bic     r1, r1, #3
 0:
     ldmia   r1!, {r2-r8,lr}
     lsl     r2, #16
@@ -73,6 +75,7 @@ blit320_512:
 blit320_368:
     stmfd   sp!, {r4-r8,lr}
     mov     r12, #23
+    bic     r1, r1, #3
 0:
     ldmia   r1!, {r2-r8,lr}
     unaligned_str r2, r3         @ 1,2

diff --git a/frontend/cspace.c b/frontend/cspace.c
index fb7fba4..06167a9 100644 (file)
--- a/frontend/cspace.c
+++ b/frontend/cspace.c
@@ -8,6 +8,7 @@
  * See the COPYING file in the top-level directory.
  */
 
+#include <stdint.h>
 #include "cspace.h"
 
 /*
@@ -30,7 +31,6 @@
         || (defined(__GNUC__) && __GNUC__ >= 5)) \
        && __BYTE_ORDER__ != __ORDER_BIG_ENDIAN__
 
-#include <stdint.h>
 #include <assert.h>
 
 #if defined(__ARM_NEON) || defined(__ARM_NEON__)
@@ -93,7 +93,8 @@ void bgr555_to_rgb565(void * __restrict__ dst_, const void *  __restrict__ src_,
 
 void bgr555_to_rgb565(void *dst_, const void *src_, int bytes)
 {
-    const unsigned int *src = src_;
+    // source can be misaligned, but it's very rare, so just force
+    const unsigned int *src = (const void *)((intptr_t)src_ & ~3);
     unsigned int *dst = dst_;
     unsigned int x, p, r, g, b;
 

diff --git a/frontend/cspace_arm.S b/frontend/cspace_arm.S
index 67778da..177b085 100644 (file)
--- a/frontend/cspace_arm.S
+++ b/frontend/cspace_arm.S
@@ -34,6 +34,12 @@ FUNCTION(bgr555_to_rgb565): @ void *dst, const void *src, int bytes
     orr      lr, lr, lsl #16
     blt      1f
 
+    @ src can be unaligned, but that's very rare, so just force it.
+    @ The manual says unaligned ldm should fault, and it does on
+    @ cortex-a78's 32bit mode, but curiously on cortex-a8 it just
+    @ works and loads the data correctly.
+    bic      r1, r1, #3
+
 0:
     ldmia    r1!, {r3-r10}
     subs     r2, #4*8

diff --git a/frontend/cspace_neon.S b/frontend/cspace_neon.S
index 4928b44..3a89fdb 100644 (file)
--- a/frontend/cspace_neon.S
+++ b/frontend/cspace_neon.S
@@ -40,7 +40,8 @@ FUNCTION(bgr555_to_rgb565): @ dst, src, bytes
     pld         [r1, #64*2]
     @ Pulls 15-bit BGR color values (which are actually 16 bits) into q0-q3.
     @ example:  q0 = 0111 1110 0101 0011
-    vldmia      r1!, {q0-q3}
+    vld1.16     {d0-d3}, [r1]!
+    vld1.16     {d4-d7}, [r1]!
     @ Shift BGR color 1 bit to the left, discarding MSB and preparing for vbit.
     @ MSB is used for transparency (not needed here, and can mess with green).
     @ example:  q0 = 1111 1100 1010 0110
@@ -113,7 +114,8 @@ FUNCTION(bgr555_to_rgb565_b): @ dst, src, bytes, int brightness2k // 0-0x0800
     vdup.16     q14, r3
 0:
     pld         [r1, #64*2]
-    vldmia      r1!, {q0-q3}
+    vld1.16     {d0-d3}, [r1]!
+    vld1.16     {d4-d7}, [r1]!
     vand.u16    q8,  q0, q14
     vand.u16    q9,  q1, q14
     vand.u16    q10, q2, q14
@@ -270,4 +272,4 @@ FUNCTION(rgb888_to_rgb565): @ dst, src, bytes
     bx          lr
 
 
-@ vim:filetype=armasm
+@ vim:filetype=armasm:expandtab

diff --git a/frontend/plugin_lib.c b/frontend/plugin_lib.c
index 171296d..d5cec76 100644 (file)
--- a/frontend/plugin_lib.c
+++ b/frontend/plugin_lib.c
@@ -402,8 +402,6 @@ static void pl_vout_flip(const void *vram, int stride, int bgr24, int w, int h)
 #endif
        else
        {
-               src = (void *)((uintptr_t)src & ~3); // align for the blitter
-
                for (; h1-- > 0; dest += dstride * 2, src += stride)
                {
                        bgr555_to_rgb565(dest, src, w * 2);