On 2/19/25 7:52 PM, Jin Ma wrote:
Are you suggesting that we should emit the rounding mode insn earlier or
incorporate the rounding mode into the pattern (in fact, there are already
operands[9]/reg FRM_REGNUM)? However, this doesn't seem to be effective
because the side effects of the rounding mode do not take effect in
float_extend, and CSE will always optimize away
pred_single_widen_subrvvmf2sf_scalar,
just like before :)
That was my original thought, but as others have pointed out,
-frounding-math is probably a better path forward.
jeff
