MIPS: silk: optimize silk_SMLAWB for MIPS32+

MIPS32 has 32x32=>64bit multiplication, although shifting 64-bit result
isn't trivial so its worth to shift right 64-bit value to 32 it means
just drop LSB register of the result.

Since third argument of silk_SMLAWB is 16-bit wide we can shift it
left by 16 before multiplication to apply technique above to the result.

Signed-off-by: Siarhei Volkau <[email protected]>
Signed-off-by: Jean-Marc Valin <[email protected]>

Author: SiarheiVolkau

silk/mips/macros_mipsr1.h

@@ -82,6 +82,15 @@ static inline int silk_SMULWB(int a, int b)
     return ac >> 32;
 }
 
+/* a32 + (b32 * (opus_int32)((opus_int16)(c32))) >> 16 output have to be 32bit int */
+#undef silk_SMLAWB
+static inline int silk_SMLAWB(int a, int b, int c)
+{
+    long long ac = (long long)b * (int)(c << 16);
+
+    return a + (ac >> 32);
+}
+
 #endif
 
 #if defined (__mips_isa_rev) /* MIPS32r1+ */