Only "Rounding towards 0" is supported and "Rounding towards Nearest", "Rounding +/- inf." are not supported. The result of "Rounding towards 0" can differ from IEEE 754 in the least significant bit.
How the rounding looks like in codes? - is it an one assembler opcode "round(x,y)"?
If yes - it is in binutils code, right? Can opcodes be rewritten to a 'C'-function at all? If yes - the mathematic of rounding can be remade to alternative with 3 cases, can't be?
However, i am rushing clever words to wind))) I can't imagine the math. examples for Nearest and +/- inf... or just translating the meaning of these words wrong.
We also can use a deprecation for rounding and test the system for bugs to obtain how often/whenever the system use rounding at all)