Re: Faster tagged integer arithmetic

[Ludovic Courtès]

Hi!

There’s now an x86_64-optimized version of fixnum addition and
subtraction, which provide a 17% speedup according to the
micro-benchmarks in ‘arithmetic.bm’.

I’ve tried something similar for multiplication, but that yielded only
5% (which is sort of expected, since the code is very close to what the
naive C code does), so I decided not to include it.  Here’s the code,
for posterity:

# define ASM_MUL(x, y)                                                  \
    {                                                                   \
      asm volatile goto ("mov %1, %%rcx; mov %0, %%rbx; "               \
                         "test %[tag], %%cl; je %l[slow_mul]; "         \
                         "test %[tag], %%bl; je %l[slow_mul]; "         \
                         "sar %[tag_size], %%rcx; "                     \
                         "sub %[tag], %%rbx; "                          \
                         "imul %%rbx, %%rcx; jo %l[slow_mul]; "         \
                         "add %[tag], %%rcx; "                          \
                         "mov %%rcx, (%[vsp])\n"                        \
                         : /* no outputs */                             \
                         : "r" (x), "r" (y),                            \
                           [vsp] "r" (sp), [tag] "i" (scm_tc2_int),     \
                           [tag_size] "i" (2)                           \
                         : "rcx", "rbx", "memory"                       \
                         : slow_mul);                                   \
      NEXT;                                                             \
    }                                                                   \
  slow_mul:                                                             \
    do { } while (0)

Comments welcome!

Thanks,
Ludo’.