Optimization using bytestring

[msakai]

Attempt for optimization using bytestring (#24).

• read file as lazy ByteString and feed it to Alex directly without going through String (Note that Alex internally uses UTF-8),
• dump formula into bytestring Builder instead of ShowS.
• use ByteString for internal representation for some types (V, AtomicWord, Role, DistinctObjectTerm)

Compatibility is kept to some extent using PatternSynonyms extension.

As I show below, maximum residency is reduced significantly, but I'm still not sure if it is worth its implementation complexity.

Experiments

I compiled TestImportExportImportFile using GHC 8.6.5 + Stackage LTS 14.20 using stack build --executable-profiling --flag logic-TPTP:BuildTestPrograms, and ran echo TPTP-v7.3.0/Problems/HWV/HWV122-1.p | TestImportExportImportFile False +RTS -p -s on master (4ef0ec9) and on this bytestring branch (6cb34a6) three times for each and take average of profile information.

Result

Time/allocation profile (-p):

	before	after
total time	11.14 secs	10.66 secs
total alloc (excludes profiling overheads)	6,864,859,344 bytes	7,661,934,288 bytes

Summary GC statistics (-s):

	before	after
allocated in the heap	11,053,555,346 bytes	12,196,062,090 bytes
copied during GC	2,644,563,434 bytes	2,403,806,042 bytes
maximum residency	459,088,146 bytes	266,768,733 bytes
maximum slop	2,519,392 bytes	2,328,909 bytes
total memory in use	437 MB	254 MB

Discussion

Total allocation increased a little (+11.6 %). I guess this is caused by increased conversion between types.

On the other hands, maximum residency is reduced significantly (-41.9 %).

Execution time is reduced a little (-4.3 %), but this may be by chance.