On Sat, 2012-11-10 at 22:53 -0200, Henrique de Moraes Holschuh wrote: > On Sun, 11 Nov 2012, Ben Hutchings wrote: > > On Sat, 2012-11-10 at 20:14 -0200, Henrique de Moraes Holschuh wrote: > > > On Sat, 10 Nov 2012, Bastian Blank wrote: > > > > On Sat, Nov 10, 2012 at 02:15:14PM -0200, Henrique de Moraes Holschuh wrote: > > > > > Yes, I know :) Our amavisd-box at work has 16GiB RAM and 16 cores, > > > > > we need at least that much to be able to run 64 instances with the > > > > > scratch directories on tmpfs... > > > > > x32 would be most likely a _MAJOR_ win for that box. > > > > > > > > Is it _likely_ or is it? 16GiB can be used with the 32bit i386 also. > > > > > > I assume you mean a amd64-kernel with i386 userspace combination. > > > > > > Compared to x32, i*86 userspace on a 64-bit kernel has the added aggravation > > > of the full kernel syscall compat layer, which has been a source of problems > > > in the past, and all the wasted performance of our i486-optimized userland. > > [...] > > > > The compat layer for x32 is almost entirely the same as for i386 on > > x86_64; indeed it is almost entirely the same as Linux uses for *every* > > 64-bit architecture that supports a 32-bit userland. I think the only > > interesting difference is that x32 has 64-bit time_t. > > It has changed, then... I thought it used the 64-bit syscalls plus ~84 that > were x32-specific, while ia32 had to go through compat for all syscalls... This is correct, but for the syscalls that x32 can use directly the 'compat layer' for i386 consists of zero-extending some registers. On x32 the same zero-extension must be done, but in userland instead. Ben. -- Ben Hutchings Design a system any fool can use, and only a fool will want to use it.
Description: This is a digitally signed message part