Unsafe literal constants with powers (e.g. 10ˆ10 in 32 bit systems)

[mschauer]

I think this example might fail on 32 bit systems.

julia/base/math.jl

Line 533 in df1ee8c

julia> a = 10^10;

[mschauer]

This is also an instance, right?

julia/base/libc.jl

Line 387 in df1ee8c

rand(::Type{Float64}) = rand(UInt32) / 2^32

[mschauer]

For reference, I looked for examples for example in

grep -r --include="*.jl" [^\.][0-9]\^[0-9] . | less

[mschauer]

julia/test/rounding.jl

Line 136 in df1ee8c

for x = 2^53-10:2^53+10

[mschauer]

julia/test/hashing.jl

Line 75 in df1ee8c

Any[-2^53-1, 0.0, 2^53+1], [-2^53-1, 0, 2^53+1], (-2^53-1):2^53+1:(2^53+1),

[mschauer]

... and I did this to convince you to reopen #5573 ;-) This problem permeates the entire ecosystem, https://pkg.julialang.org/docs/search?q=2%5C%5E%5B3-9%5D%5B0-9%5D%7C%5B%5E%5C.%5D%5B0-9%5D%7B2,%7D%5C%5E%5B0-9%5D%7B2,%7D

[StefanKarpinski]

As a somewhat broader point, maybe we should run doctests on both a 64- and 32-bit platform?

[StefanKarpinski]

Also, what would you like the behavior here to be?

[mschauer]

As a somewhat broader point, maybe we should run doctests on both a 64- and 32-bit platform?

A somewhat minimal requirement is to make passing 32-bit doctests mandatory for the release candidates?

[ararslan]

@staticfloat, could we just add an extra CI build for 32-bit doctests?

[mschauer]

Also, what would you like the behavior here to be?

I think I would like ^ with literal arguments use a checked version of power_by_squaring (which would make sense if there is no performance penalty due to constant propagation), but I cannot say if this realistic.