f32/f64 MIN_EXP is either wrong or badly documented.

[eddyb]

They are both -MAX_EXP + 3 while core::num::dec2flt::RawFloat uses -MAX_EXP + 1.
Also, APFloat (being ported to Rust in #43554) uses the same -MAX_EXP + 1 value.
That is the value of the exponent for the smallest normal, AFAIK.

cc @rust-lang/libs @est31 @nagisa @lifthrasiir @rkruppe

[hanna-kruppe]

I chased the git blame chain and apparently -MAX_EXP + 3 matches, and historically stems from C's (or at least GCC's) {FLT,DBL}_MIN_EXP 🤔

[kennytm]

MIN_EXP is "One greater than the minimum possible normal power of 2 exponent". For f64, the smallest normal number is 2^-1022, so MIN_EXP = -1022 + 1 = -1021 as expected.

I don't know why such a strange definition is picked, but this follows how ***_MIN_EXP is defined in C and C++.

C11 (n1570) §5.2.4.2.2/11

minimum negative integer such that FLT_RADIX raised to one less than that power is a normalized floating-point number, e_min

FLT_MIN_EXP
DBL_MIN_EXP
LDBL_MIN_EXP

C++14 (n3936) §18.3.2.4/27:

static constexpr int  min_exponent;

Minimum negative integer such that radix raised to the power of one less than that integer is a normalized floating point number.

[eddyb]

@kennytm That doesn't add up. APFloat has 2^MIN_EXP as the smallest normal number.
So 2^-1023 for f64, which would make std::f64::MIN_EXP -1022. What am I missing?

[hanna-kruppe]

Uhh, the smallest normal f64 is indeed 2^-1022, not 2^-1023. The latter is a subnormal with a significand of 0b1000....

[eddyb]

Oh, figured it out. APFloat MIN_EXP is indeed -1022. What I had in mind was that it was defined as -MAX_EXP + 1 which is true but MAX_EXP also is off-by-one from libstd.
rawfp uses the same values as APFloat, so this is only some C nonsense. My bad.