https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119702
--- Comment #7 from Avinash Jayakar <avinashd at linux dot ibm.com> --- (In reply to Peter Bergner from comment #6) > If the vaddudm is the fastest sequence, then yes. Got it, I referred the power10 user manual to check the latency information, and vaddudm has better cpi, min latency and max latency, compared to combination of the 2 instructions. So I will work on making this change (using vaddudm in place of splat and shift). > I think you should fire off a git bisect to track down which commit changed > the behavior from GCC 14 and then you can decide how to fix it. Is this to check why vextsb2d was being generated? Because apart from that, the instruction sequence is the same right?
