Re: VIRT_ADDR_VARIES

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

On 11/10/11 01:29, Andreas Schwab wrote:
>> +  ((uintptr_t) XPNTR (obj) - (uintptr_t) pure <= PURESIZE)
> 
> There is no need for obfuscation, the compiler should find that out by
> itself.

Perhaps it should, but it typically doesn't.  For example, given this:

	#include <inttypes.h>
	#include <stddef.h>
	#define TYPEMASK 7
	#define XPNTR(a) ((a) & ~TYPEMASK)
	#define PURESIZE 1000000

	extern long pure[];

	#define PURE_P1(obj) \
	  ((char *) pure <= (char *) XPNTR (obj) \
	   && (char *) XPNTR (obj) < (char *) pure + PURESIZE)

	#define PURE_P2(obj) \
	  ((uintptr_t) XPNTR (obj) - (uintptr_t) pure <= PURESIZE)

	int pure1 (intptr_t obj) { return PURE_P1 (obj); }
	int pure2 (intptr_t obj) { return PURE_P2 (obj); }

gcc -O2 (x86-64, GCC 4.6.2) generates this for the comparisons of PURE_P1:

		xorl	%eax, %eax
		cmpq	$pure, %rdi
		jb	.L2
		xorl	%eax, %eax
		cmpq	$pure+1000000, %rdi
		setb	%al
	.L2:	rep

and this for PURE_P2:

		xorl	%eax, %eax
		subq	$pure, %rdi
		cmpq	$1000000, %rdi
		setbe	%al

The PURE_P1 version has more instructions, and has a conditional jump.
The PURE_P2 version is nearly as fast as the old code without
VIRT_ADDR_VARIES.  I observed similar savings with the other
compilers I tried (Sun C 5.11 sparcv9 cc -xO4, clang 2.8 x86-64 -O2).
So we should go with PURE_P2.