Re: PCH support on Solaris

["Bonzini" <bonzini at gnu dot org>]

> Yes, that's why Geoff suggested to first make sure that we can safely pass
> MAP_FIXED by using OS-specific hooks.

The problem is not that MAP_FIXED is not portable, it is that it is dangerous.  It's
too easy to get a crash if MAP_FIXED is too zealous in getting the address you asked
for: in my experience I had segfaults and SIGILLs in libgcc, which were even dependant
on the system libraries, for instance on the file sizes and on what libc mmap-ed at
startup.  This weird behavior is not good for gcc where bugs are seeked to be as easily
reproducible as possible.

I was attempting to make it safer. Knowing more about how the code's work, I can detail
more about my suggestion.

- when loading the pre-compiled header, you first ask the OS for memory for the PCH,
preferrably at the base address that the PCH was translated to, with

#ifndef MAP_NORESERVE
#define MAP_NORESERVE 0
#endif

guess = mmap (mmi.preferred_base, mmi.size, PROT_NONE,
        MAP_PRIVATE|MAP_ANON|MAP_NORESERVE, -1, 0);

if (guess == (caddr_t) -1)
    {
        warning ("PCH not usable");
        return;
    }

_If getting that area is not possible, the OS will not unmap anything!_

- then if the OS likes the position, you are sure that mapping the file at that
location is not risky, so you can use MAP_FIXED like

addr = mmap (guess, mmi.size, PROT_READ | PROT_WRITE,
        MAP_PRIVATE|MAP_FIXED, fileno (f), mmi.offset);

if (addr != guess || addr != mmi.preferred_base)
   {
        munmap (guess, mmi.size);
        assert (addr == (caddr_t) -1);    /* else MAP_FIXED did not work! */
        warning ("PCH not usable");
        return;
    }

/* ok, PCH was mmap-ed */

With this approach you have no OS dependent heuristic, you just ask the OS in advance
whether what you want can be achieved with the current memory layout.

Paolo