New features available on some targets, in particular support for IEEE-compatible 128-bit floating-point arithmetic on Power (and Intel), are starting to expose long-standing GDB limitations. In particular, GDB currently performs all expression evaluation on floating-point types using host "long double" arithmetic. This may get very different results compared with actual target execution if the target data type has greater range and/or precision than the host's long double type. In this talk, I'll discuss options how to fix this problem. In particular, I'll propose a solution to use floating-point emulation, similar to what is done by GCC for constant-folding, to precisely mirror target floating-point arithmetic.