https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120034
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
We get (-O2 -mbmi2):
bzhi_emulation1:
.LFB1:
.cfi_startproc
movq $-1, %rax
shlx %rsi, %rax, %rax
notl %eax
andl %edi, %eax
ret
that has single uses only. The issue is likely that 1ULL << index is
well-defined for value >= 32 and we likely fail to "saturate" this
case when it's demoted to SImode for the AND.
