Hello! Commit 37a5970c19ca7ad2b5de2f667748c840c199f878 changes the VM engine such that the jump table address is kept in a register, at least on x86_64.
Using GCC 4.6.0 on x86_64, before the change, the dispatch code looked like this: movzbl 0x0(%rbp),%edx ;; %edx <- *IP lea 0x0(%rip),%rax ;; %rax <- &jump_table add $0x1,%rbp ;; IP++ mov (%rax,%rdx,8),%rax ;; %rax <- jump_table[*IP] jmpq *%rax After, there’s 3 instructions left, instead of 4, with the jump table address saved in r12: movzbl 0x0(%rbp),%eax add $0x1,%rbp mov (%r12,%rax,8),%rax jmpq *%rax This leads a ~7% speedup on srfi-1.bm, for instance. Interestingly, defining IP_REG and SP_REG on x86_64 is not necessary since GCC already makes the right choice. Thanks, Ludo’.
