https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66790
--- Comment #41 from Pierre-Marie de Rodat <derodat at adacore dot com> --- Thank you again for spotting this! Yes, these artificial defs break the consevativeness for the MIR analysis. I guess your proposal would work: considering as uninintialized registers than aren’t is conservative although it will prevent more optimizations that one could expect in REE. I don’t know if it’s a problem in practice but your example shows that we have a lot of registers in this situation. At this point, though, I wonder why artificial defs do not describe what actually happens for this function. Shouldn’t we instead fix DF to emit artificial defs only for actual input registers? I’ll investigate this. … and then I will followup on gcc-patches as you suggested. Once more, thank you all for your help in this! :-)