https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85961
--- Comment #4 from bki at hacon dot de --- (In reply to Alexander Monakov from comment #3) > You'd need to disable IPA-RA after forcing -O2 with the pragma, i.e.: > > #pragma GCC optimize "O2" > #pragma GCC optimize "no-ipa-ra" Yes, this changes the problematic instruction sequence to: 7e: 48 89 c5 mov %rax,%rbp 81: e8 7a ff ff ff callq 0 <_ZNKSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEE6lengthEv.isra.0> 86: 48 8d 54 05 00 lea 0x0(%rbp,%rax,1),%rdx Which is better, since rbp is not a scratch register. > [...]. Do we want to adjust it given that "pragma optimized" is documented as > "not suitable for production use"? I appreciate the explanation which makes sense to me. In the original context, I did introduce the pragma to reduce compilation times like this: #pragma GCC optimize "-fno-var-tracking", "-fno-var-tracking-assignments" So, with instrumentation but without the optimize pragma, scratch registers will always be considered changed across a function call? This would then both fix my issue and more general doubts on the feasibility of my ABI messing. I also agree that the pragma is clearly documented as not fit for production use and would remove it accordingly. I think I can manage to pass these options on the command line somehow.