Re: VIRT_ADDR_VARIES

[Paul Eggert <eggert@cs.ucla.edu>]

On 11/06/11 09:18, Andreas Schwab wrote:
> Is there any reason not to define VIRT_ADDR_VARIES?

I expect the main reason is performance.  For example,
on x86-64 with GCC 4.6.2 -O2, the function:

int pure (Lisp_Object obj) { return PURE_P (obj); }

has 8 instructions (not counting the 'ret') if VIRT_ADDR_VARIES
is defined, and 4 instructions if it is not defined.  The key difference
is that VIRT_ADDRESS_VARIES generates this:

        cmpq    $pure+1000000, %rdi
        jge     .L2
        xorl    %eax, %eax
        cmpq    $pure, %rdi
        setge   %al
  .L2:  rep

(where the "1000000" is a function of PURESIZE),
whereas omitting VIRT_ADDRESS_VARIES generates this:

        cmpq    $my_edata, %rdi
        setl    %al


> The only difference
> it makes is that the bounds of the pure array are checked accurately

True, but for Emacs it shouldn't matter whether PURE_P checks
accurately or loosely -- either way Emacs should operate correctly.