avcodec/get_bits: Avoid reading multiple times in get_bits_long

Due to non-byte-alignment a read of 32 bits guarantees only
25 usable bits; therefore get_bits_long() (which is used to
potentially read more than 25 bits) performs two reads in case
it needs to read more than 25 bits and combines the result.

Yet this is not necessary: One can just read 64 bits at a time
to get 32 usable bits (57 would be possible). This commit does so.

This reduced the size of .text by 30144B for GCC 11.4 and 5648B
for Clang 14 (both with -O3).

(get_bits_long() is a building block of show_bits_long()
and get_ue_golomb_long(), so this patch affects these, too.)

Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>

libavcodec/get_bits.h

@@ -187,21 +187,28 @@ static inline unsigned int show_bits(GetBitContext *s, int n);
 
 #define CLOSE_READER(name, gb) (gb)->index = name ## _index
 
+#define UPDATE_CACHE_BE_EXT(name, gb, bits, dst_bits) name ## _cache = \
+    AV_RB ## bits((gb)->buffer + (name ## _index >> 3)) << (name ## _index & 7) >> (bits - dst_bits)
+
+#define UPDATE_CACHE_LE_EXT(name, gb, bits, dst_bits) name ## _cache = \
+    (uint ## dst_bits ## _t)(AV_RL ## bits((gb)->buffer + (name ## _index >> 3)) >> (name ## _index & 7))
+
+/* Using these two macros ensures that 32 bits are available. */
+# define UPDATE_CACHE_LE_32(name, gb) UPDATE_CACHE_LE_EXT(name, (gb), 64, 32)
+
+# define UPDATE_CACHE_BE_32(name, gb) UPDATE_CACHE_BE_EXT(name, (gb), 64, 32)
+
 # ifdef LONG_BITSTREAM_READER
 
-# define UPDATE_CACHE_LE(name, gb) name ## _cache = \
-      AV_RL64((gb)->buffer + (name ## _index >> 3)) >> (name ## _index & 7)
+# define UPDATE_CACHE_LE(name, gb) UPDATE_CACHE_LE_32(name, (gb))
 
-# define UPDATE_CACHE_BE(name, gb) name ## _cache = \
-      AV_RB64((gb)->buffer + (name ## _index >> 3)) >> (32 - (name ## _index & 7))
+# define UPDATE_CACHE_BE(name, gb) UPDATE_CACHE_BE_32(name, (gb))
 
 #else
 
-# define UPDATE_CACHE_LE(name, gb) name ## _cache = \
-      AV_RL32((gb)->buffer + (name ## _index >> 3)) >> (name ## _index & 7)
+# define UPDATE_CACHE_LE(name, gb) UPDATE_CACHE_LE_EXT(name, (gb), 32, 32)
 
-# define UPDATE_CACHE_BE(name, gb) name ## _cache = \
-      AV_RB32((gb)->buffer + (name ## _index >> 3)) << (name ## _index & 7)
+# define UPDATE_CACHE_BE(name, gb) UPDATE_CACHE_BE_EXT(name, (gb), 32, 32)
 
 #endif
 
@@ -209,12 +216,14 @@ static inline unsigned int show_bits(GetBitContext *s, int n);
 #ifdef BITSTREAM_READER_LE
 
 # define UPDATE_CACHE(name, gb) UPDATE_CACHE_LE(name, gb)
+# define UPDATE_CACHE_32(name, gb) UPDATE_CACHE_LE_32(name, (gb))
 
 # define SKIP_CACHE(name, gb, num) name ## _cache >>= (num)
 
 #else
 
 # define UPDATE_CACHE(name, gb) UPDATE_CACHE_BE(name, gb)
+# define UPDATE_CACHE_32(name, gb) UPDATE_CACHE_BE_32(name, (gb))
 
 # define SKIP_CACHE(name, gb, num) name ## _cache <<= (num)
 
@@ -414,15 +423,26 @@ static inline unsigned int get_bits_long(GetBitContext *s, int n)
     av_assert2(n>=0 && n<=32);
     if (!n) {
         return 0;
-    } else if (n <= MIN_CACHE_BITS) {
+    } else if ((!HAVE_FAST_64BIT || av_builtin_constant_p(n <= MIN_CACHE_BITS))
+               && n <= MIN_CACHE_BITS) {
         return get_bits(s, n);
     } else {
+#if HAVE_FAST_64BIT
+        unsigned tmp;
+        OPEN_READER(re, s);
+        UPDATE_CACHE_32(re, s);
+        tmp = SHOW_UBITS(re, s, n);
+        LAST_SKIP_BITS(re, s, n);
+        CLOSE_READER(re, s);
+        return tmp;
+#else
 #ifdef BITSTREAM_READER_LE
         unsigned ret = get_bits(s, 16);
         return ret | (get_bits(s, n - 16) << 16);
 #else
         unsigned ret = get_bits(s, 16) << (n - 16);
         return ret | get_bits(s, n - 16);
+#endif
 #endif
     }
 }