neon code for bgra to rgba conversion for copyFTBitmap

src/opts/SkFontHost_FreeType_common_opts_neon.cpp

Index: src/opts/SkFontHost_FreeType_common_opts_neon.cpp
diff --git a/src/opts/SkFontHost_FreeType_common_opts_neon.cpp b/src/opts/SkFontHost_FreeType_common_opts_neon.cpp
new file mode 100644
index 0000000000000000000000000000000000000000..f881ca50dfa4e9e049c4206189c18e8eb71296d9
--- /dev/null
+++ b/src/opts/SkFontHost_FreeType_common_opts_neon.cpp
@@ -0,0 +1,72 @@
+#include <arm_neon.h>
+
+#include "SkColorPriv.h"
+
+void SkPackARGB32_neon(const uint32_t* src, SkPMColor* dst, int height, int width, 
+                       size_t dstRowBytes, size_t srcRowBytes)
+{
+    asm volatile (
+
+        // if (height == 0) exit
+        "cmp    %[height], #0                                           \n\t"
+        "beq    End                                                     \n\t"
+
+        "RowLoop:"
+
+            // reset r1 = width for each new line
+            "mov    r1, %[width]                                        \n\t"
+            "mov    r2, %[src]                                          \n\t"
+            "mov    r3, %[dst]                                          \n\t"
+
+            "ColumnBatchLoop:"
+
+                // while (r1 >= 8)
+                "cmp    r1, #8                                          \n\t"
+                "blt    PixelLoopStart                                  \n\t"
+
+                // load src, 8 pixels at a time, each pixel taking 32bit value
+                // d0: B0B1B2B3...B7 d1: G0G1G2G3...G7 d2: R0R1R2R3...G7 d3: A0A1A2A3...A7
+                "vld4.8 {d0, d1, d2, d3}, [r2]!                         \n\t"
+                //swap d0 and d2 so we have RGBA format if SkPMColor is RGBA, otherwise keep BGRA
+#ifdef SK_PMCOLOR_IS_RGBA
+                "vswp d0, d2                                            \n\t"
+#endif
+                "vst4.8 {d0, d1, d2, d3}, [r3]!                         \n\t"
+                //r1 -= 8 because we are processing 8 pixels at a time
+                "sub    r1, r1, #8                                      \n\t"
+                "b      ColumnBatchLoop                                 \n\t"
+
+                "PixelLoopStart:                                        \n\t"
+                    // while (r1 > 0)
+                    "cmp    r1, #0                                      \n\t"
+                    "beq    PixelLoopEnd                                \n\t"
+
+                "PixelLoop:"
+                    //load 1 32bits pixel at a time and swap again if needed
+                    "vld4.8  {d0[0], d1[0], d2[0], d3[0]}, [r2]!        \n\t"
+#ifdef SK_PMCOLOR_IS_RGBA
+                    "vswp d0, d2                                        \n\t"
+#endif
+                    "vst4.8  {d0[0], d1[0], d2[0], d3[0]}, [r3]!        \n\t"
+                    // pixel counter decrement
+                    "subs    r1, r1, #1                                 \n\t"
+                    "bne     PixelLoop                                  \n\t"
+
+                "PixelLoopEnd:                                          \n\t"
+
+            "add    %[src], %[src], %[srcRowBytes]                      \n\t"
+            "add    %[dst], %[dst], %[dstRowBytes]                      \n\t"
+            // while (height) process next line
+            "subs   %[height], %[height], #1                            \n\t"
+            "bne    RowLoop                                             \n\t"
+
+        "End:                                                           \n\t"
+
+            : [height] "+r" (height), [dst] "+r" (dst), [src] "+r" (src)
+            : [width] "r" (width), [srcRowBytes] "r" (srcRowBytes), [dstRowBytes] "r" (dstRowBytes)
+            : "cc", "memory",
+              "r1", "r2", "r3",
+              "d0", "d1", "d2", "d3");
+}
+
+