On 8/14/25 15:11, Andrew Waterman wrote:
>> Yes it needed adjustment due to additional save/restore. Fixed that and will
>> push shortly.
> While you're at it, does it make sense to skip the flag save/restore
> if !flag_trapping_math? It's an admittedly minor point, since the
> non-Zfa case is more about correctness than it is about performance,
> but if it's easy to do, you might as well.
Sure, its easy enough. FWIW it seems to be one of the things checked already to
gate the pattern, but once inside we can check again for the flags specifically.
(define_expand "<round_pattern><ANYF:mode>2"
[(set (match_operand:ANYF 0 "register_operand" "=f")
(unspec:ANYF
[(match_operand:ANYF 1 "register_operand" " f")]
ROUND))]
"(TARGET_HARD_FLOAT
&& (TARGET_ZFA || flag_fp_int_builtin_inexact || !flag_trapping_math)
<------
&& (TARGET_ZFA || TARGET_64BIT || <ANYF:MODE>mode != DFmode)
&& (TARGET_ZFA || <ANYF:MODE>mode != HFmode))"
I'll submit a v2 so that CI pre-commit can give it a spin !
Thx,
-Vineet
