lsws/ppc/yuv2rgb_altivec: Replace vec_lvsl/vec_perm with vec_xl
authorChip Kerchner <ckerchne@linux.vnet.ibm.com>
Tue, 13 Aug 2019 00:21:24 +0000 (02:21 +0200)
committerCarl Eugen Hoyos <ceffmpeg@gmail.com>
Tue, 13 Aug 2019 00:21:24 +0000 (02:21 +0200)
gcc 6.x and 7.x generate wrong code for little endian machines
for the vec_lvsl/vec_perm instruction combos in some cases.
The bug was fixed in version 8.x
If these instructions are replaced with vec_xl, the problem goes
away for all versions of the compilers.

Fixes ticket #7124.

libswscale/ppc/yuv2rgb_altivec.c

index c1e2852..5365452 100644 (file)
@@ -305,9 +305,6 @@ static int altivec_ ## name(SwsContext *c, const unsigned char **in,          \
     vector signed short R1, G1, B1;                                           \
     vector unsigned char R, G, B;                                             \
                                                                               \
-    const vector unsigned char *y1ivP, *y2ivP, *uivP, *vivP;                  \
-    vector unsigned char align_perm;                                          \
-                                                                              \
     vector signed short lCY       = c->CY;                                    \
     vector signed short lOY       = c->OY;                                    \
     vector signed short lCRV      = c->CRV;                                   \
@@ -338,26 +335,13 @@ static int altivec_ ## name(SwsContext *c, const unsigned char **in,          \
         vec_dstst(oute, (0x02000002 | (((w * 3 + 32) / 32) << 16)), 1);       \
                                                                               \
         for (j = 0; j < w / 16; j++) {                                        \
-            y1ivP = (const vector unsigned char *) y1i;                       \
-            y2ivP = (const vector unsigned char *) y2i;                       \
-            uivP  = (const vector unsigned char *) ui;                        \
-            vivP  = (const vector unsigned char *) vi;                        \
-                                                                              \
-            align_perm = vec_lvsl(0, y1i);                                    \
-            y0 = (vector unsigned char)                                       \
-                     vec_perm(y1ivP[0], y1ivP[1], align_perm);                \
+            y0 = vec_xl(0, y1i);                                              \
                                                                               \
-            align_perm = vec_lvsl(0, y2i);                                    \
-            y1 = (vector unsigned char)                                       \
-                     vec_perm(y2ivP[0], y2ivP[1], align_perm);                \
+            y1 = vec_xl(0, y2i);                                              \
                                                                               \
-            align_perm = vec_lvsl(0, ui);                                     \
-            u = (vector signed char)                                          \
-                    vec_perm(uivP[0], uivP[1], align_perm);                   \
+            u = (vector signed char) vec_xl(0, ui);                           \
                                                                               \
-            align_perm = vec_lvsl(0, vi);                                     \
-            v = (vector signed char)                                          \
-                    vec_perm(vivP[0], vivP[1], align_perm);                   \
+            v = (vector signed char) vec_xl(0, vi);                           \
                                                                               \
             u = (vector signed char)                                          \
                     vec_sub(u,                                                \