Optimize for size, while unrolling the hottest loops

[kazuho]

The attempt of this PR is to reduce the code size while manually unrolling the loops, as sugessted by @herumi.

While manual unrolling does not provide performance improvement on my testbed (Core i5 9400, Ubuntu 19.10), we see comparable performance with less than half the code size.

	object size	1440 bytes	16384 bytes
original (-O3)	5696	0.293	2.804
original (-O2 -fno-unroll-loops)	4480	0.332	3.245
this PR (-O3)	5808	0.294	2.813
this PR (-O2 -fno-unroll-loops)	2480	0.297	2.893

Object sizes are in bytes, other cells are in seconds (i.e. smaller the better). In addition to the command options listed above, -DNDEBUG -march-native were used in common.