Optimize component part construction

[ancapdev]

Component part construction performs some redundant conversions on constants.

The conversions are more costly than they look, as the decimal to integer path range checks against type ranges on the integer types, which in turn performs conversions of these to the decimal type. For Dec128/UInt128 this is particularly bad as the integer to decimal conversion falls back to conversion via strings.

Fixed by precomputing integer constants.

Simple example benchmark in the form of

@btime T($(Ref(123))[], $(Ref(-8))[])

T	before	after
Dec32	10.701 ns (0 allocations: 0 bytes)	4.200 ns (0 allocations: 0 bytes)
Dec64	11.912 ns (0 allocations: 0 bytes)	3.810 ns (0 allocations: 0 bytes)
Dec128	394.069 ns (6 allocations: 288 bytes)	6.510 ns (0 allocations: 0 bytes)

[stevengj]

Would be good to add tests of the form:

@test DecFP._int_maxintfloat(Dec32) === UInt32(maxintfloat(Dec32))
@test DecFP._int_maxintfloat(Dec64) === UInt64(maxintfloat(Dec64))
@test DecFP._int_maxintfloat(Dec128) === UInt128(maxintfloat(Dec128))

[ancapdev]

Would be good to add tests of the form:

@test DecFP._int_maxintfloat(Dec32) === UInt32(maxintfloat(Dec32))
@test DecFP._int_maxintfloat(Dec64) === UInt64(maxintfloat(Dec64))
@test DecFP._int_maxintfloat(Dec128) === UInt128(maxintfloat(Dec128))

Makes sense, done

[codecov]

Codecov Report

Merging #155 (4e27f04) into master (e3c133d) will increase coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #155      +/-   ##
==========================================
+ Coverage   92.32%   92.38%   +0.05%     
==========================================
  Files           1        1              
  Lines         417      420       +3     
==========================================
+ Hits          385      388       +3     
  Misses         32       32              

Impacted Files	Coverage Δ
src/DecFP.jl	92.38% <100.00%> (+0.05%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e3c133d...4e27f04. Read the comment docs.

[stevengj]

LGTM. The 32-bit failures are not a regression (see #91).