[Bug c/110656] New: Floating point to char/short or bitfield conversions

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110656

            Bug ID: 110656
           Summary: Floating point to char/short or bitfield conversions
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org
  Target Milestone: ---

While implementing floating point to _BitInt convertions, I'm wondering if it
is correct that we implement say float -> signed char conversion effectively as
float -> int -> signed char.
signed char foo (float x) { return x; }
with foo (512.0f).
ISO C17 has in F.4 says
"or if the integral part of the floating value
exceeds the range of the integer type, then the “invalid” floating-point
exception is raised and the resulting value is unspecified."
but we even with -ftrapping-math trigger invalid say for 8589934592.0f (when it
doesn't fit into int), but the above just does
        cvttss2sil      %xmm0, %eax
        movsbl  %al, %eax
on x86_64.
In the scope of _BitInt, I wonder if I should care, i.e. whether I should call
say a library routine for floating point to _BitInt(N)/unsigned _BitInt(N)
conversions when
N is smaller than bit width of __int128 (if it exists, long long otherwise) and
is not equal to width of int, long or long long, or if I can implement those
like it is implemented for char/short and bitfields.
I guess for now I'll try to implement the library side so that it handles
correctly all valid Ns and we can decide on the compiler side later.