Avoid use of Uint128 in sampling from a range #10606
Conversation
Introduces the tryparse method:
- tryparse{T<:Integer}(::Type{T<:Integer},s::AbstractString)
- tryparse(::Type{Float..},s::AbstractString)
- a few variants of the above
And:
- tryparse(Float.., ...) call the corresponding C functions jl_try_strtof, jl_try_substrtof, jl_try_strtod and jl_try_substrtod.
- The parseint, parsefloat, float64_isvalid and float32_isvalid methods wrap the corresponding tryparse methods.
- The jl_strtod, jl_strtof, ... functions are wrappers over the jl_try_str... functions.
This should fix #10498 as well.
Ref: discussions at #9316, #3631, #5704
|
+1 |
|
I love the stackoverflow link in the comment. We should have more of those. |
Internet hive mind, solve my problem for me! Appveyor failure was an intermittent unrelated fluke (#9572), OSX Travis timed out. LGTM. |
|
PS: Master PR |
|
Woot. That's a lovely speedup. |
| maxmultiplemix(k::UInt64) = (div((k >> 32 != 0)*0x0000000000000000FFFFFFFF00000000 + 0x0000000100000000, k + (k == 0))*k - 1) % UInt64 | ||
| # that is 0xFFFF...FFFF if k = typemax(T) - typemin(T) with intentional underflow | ||
| # see http://stackoverflow.com/questions/29182036/integer-arithmetic-add-1-to-uint-max-and-divide-by-n-without-overflow | ||
| maxmultiple{T<:Unsigned}(k::T) = (div(typemax(T) - k + one(k), k + (k == 0))*k + k - one(k))::T |
There was a problem hiding this comment.
is there another (slight) improvement to use zero(T) here? (and below)
There was a problem hiding this comment.
I guess k == 0 is a julia idiom, the result is of type Bool anyway adding a Boolean does not change the type of k.
There was a problem hiding this comment.
Are the one(k) even needed here?
There was a problem hiding this comment.
In most situations it's exactly equivalent, or better, because numeric types typically have a specialized methods for comparison with Int. (Eg BigInt)
one(k) is required to not widen the result for UInt8, UInt16 (and usually UInt32).
Avoid use of Uint128 in sampling from a range
This is desired because of 2c327d3#commitcomment-10326954
Note that this also allows to use the full entropy of Uint128 random integers to sample a Uint128 range, so I included this immediately. This means for certain cases, rand(uint128(0:i)) will be more efficient (but produce different random numbers.)
cc @rfourquet