https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120427
--- Comment #11 from Harald van Dijk <harald at gigawatt dot nl> ---
(In reply to H.J. Lu from comment #10)
> (In reply to Harald van Dijk from comment #9)
> > (In reply to H.J. Lu from comment #2)
> > > We shouldn't transform "mov $-1,reg" to "push $-1; pop reg".
> > > We should transform "mov $-1,reg" to "or $-1,reg" instead.
> >
> > Is that really an improvement? For 32-bit registers, sure, it's fine, but
> > for 64-bit registers this increases code size, does it not? orq $-1, %rax is
> > four bytes, when pushq $-1; popq %rax is only three.
>
> By default, GCC optimizes for performance. "mov $-1,reg" is write. Other
> encodings are write after read and maybe more. If you want the smallest code
> size, please use -Oz.
It was -Oz I was asking about. Look at the generated code for
long f() { return -1; }
at https://godbolt.org/z/z3shahbYK, comparing GCC 15.1 (without your patch) to
GCC trunk (with your patch). This function used to be 4 bytes at -Oz, it is now
5 bytes.
