[llvm-branch-commits] [llvm] [InstCombine] Fold exp2(uitofp iN x) to ldexp when N equals the width of int (PR #227380)

[Eric Ross via llvm-branch-commits]

ZERICO2005 wrote:

> It always holds today (Emax + 1 <= 16384), but it keeps the fold safe if a wider FP type is ever added. :-)

For `float128`/`float80`, that value should be `16384 + 16384 + 112 = 32880` or somewhere close to there. This is because you have to consider what exponent is needed for `ldexp(FLT128_TRUE_MIN, exponent)` to guarantee overflow.


https://github.com/llvm/llvm-project/pull/227380