On 2/13/2026 8:48 AM, Robin Dapp wrote:
I might be completely off here, but a quick idea which came up is
whether a secondary reload could be utilized in order to get a scratch
register, rendering the initial mov into an ordinary mov with two
operands so that LRA would again detect a cycle.
Yes, Peter tried a secondary reload and it got rid of the ICE. Just the
resulting code was worse. If nothing else helps, we could still go that
route.
My sense is we probably should be defining secondary reloads for some
scenarios, but I'd be hesitant to introduce them right now. Improving
cycle detection seems like the lower risk solution.
Jeff
