Extended Description

To reproduce

clang++ -c hot.cc -o dbg.o -O3 -g
clang++ -c hot.cc -o rel.o -O3

Then diff the objdump outputs.

• c2: 48 8d 6c 24 28 lea 0x28(%rsp),%rbp
• c7: 48 89 eb mov %rbp,%rbx
• ca: 0f 11 45 00 movups %xmm0,0x0(%rbp)

• c2: 48 8d 5c 24 28 lea 0x28(%rsp),%rbx
• c7: 0f 11 03 movups %xmm0,(%rbx)
• ca: 8b 44 24 20 mov 0x20(%rsp),%eax

For those two instructions:

lea 0x28(%rsp), %rbp
mov %rbp, %rbx

can be applied with backward copy propagation to eliminate the rbp register.