RISC-V round_away () handling of non canonical rounding modes

[Vineet Gupta]

On 1/16/25 11:21, Florian Weimer wrote:
> * Vineet Gupta:
>
>> On RISC-V we are running into an obscure issue where glibc FP print
>> call stack is hitting abort.
>>
>> __snprintf
>>    __vsnprintf_internal
>>      __printf_buffer
>>         __printf_fp_l_buffer
>>             __printf_fp_buffer_1
>>                  *round_away*  + get_rounding_mode
>>                       abort
>>
>> The reason is round_away () only handles the 4 canonical rounding
>> modes (Up, Down, Towards zero, Nearest- ties to even), while RISC-V
>> ISA supports an additional  *Nearest - ties to Max magnitude* [1]. The
>> last one being set implicitly by various FP instructions.
>>
>> With unrelated gcc changes, SPEC 2017 cam4 benchmark runs are hitting
>> the abort because printf_dp_buffer_1 ()  just reads the last set
>> rounding mode which happens to be the fifth one set by some FP insn.
> Is it really correct if the rounding mode changes spontaneously?  Or
> is this an -ffast-math thing?

Indeed it is -Ofast and thus -ffast-math

Thx,
-Vineet