Re: New regression on ARM Linux

[Richard Biener <rguenther@suse.de>]

On Tue, 31 Mar 2015, Alan Lawrence wrote:

> Jakub Jelinek wrote:
> > On Tue, Mar 31, 2015 at 11:47:37AM +0100, Alan Lawrence wrote:
> > > Richard Biener wrote:
> > > > But I find it odd that on ARM passing *((aligned_int *)p) as
> > > > vararg (only as varargs?) changes calling conventions independent
> > > > of the functions type signature.
> > > Does it? Do you have a testcase, and compilation flags, that'll make this
> > > show up in an RTL dump? I've tried numerous cases, including AFAICT yours,
> > > and I always get the value being passed in the expected ("unaligned")
> > > register?
> > 
> > If the integral type alignment right now matters, I'd try something like:
> > 
> > typedef int V __attribute__((aligned (8)));
> > V x;
> > 
> > int foo (int x, ...)
> > {
> >   int z;
> >   __builtin_va_list va;
> >   __builtin_va_start (va, x);
> >   switch (x)
> >     {
> >     case 1:
> >     case 3:
> >     case 6:
> >       z = __builtin_va_arg (va, int);
> >       break;
> >     default:
> >       z = __builtin_va_arg (va, V);
> >       break;
> >     }
> >   __builtin_va_end (va);
> >   return z;
> > }
> > 
> > int
> > bar (void)
> > {
> >   V v = 3;
> >   int w = 3;
> >   foo (1, (int) v);
> >   foo (2, (V) w);
> >   v = 3;
> >   w = (int) v;
> >   foo (3, w);
> >   foo (4, (V) w);
> >   v = (V) w;
> >   foo (5, v);
> >   foo (6, (int) v);
> >   foo (7, x);
> >   return 0;
> > }
> > 
> > (of course, most likely with passing a different value each time and
> > verification of the result).
> > As the compiler treats all those casts there as useless, I'd expect
> > that the types of the passed argument would be pretty much random.
> > And, note that even on x86_64, the __builtin_va_arg with V expands into
> >   # addr.1_3 = PHI <addr.1_27(9), _31(10)>
> >   z_35 = MEM[(V * {ref-all})addr.1_3];
> > using exactly the same address for int as well as V va_arg - if you increase
> > the overalignment arbitrarily, it will surely be a wrong IL because nobody
> > really guarantees anything about the overalignment.
> > 
> > So, I think the tree-sra.c patch is a good idea - try to keep using the main
> > type variants as the types in the IL where possible except for the MEM_REF
> > first argument (i.e. even the lhs of the load should IMHO not be
> > overaligned).
> > 
> > As Eric Botcazou said, GCC right now isn't really prepared for under or
> > overaligned scalars, only when they are in structs (or for middle-end in
> > *MEM_REFs).
> > 
> > 	Jakub
> > 
> 
> On ARM, I get the arguments being passed in r0 & r1 for every call in bar()
> above. It sounds as if this is because the casts are being removed as useless;
> so the only way for overalignment info to be present, is when SRA puts it
> there.
> 
> The only way I can get a register to be skipped, is by providing a prototype
> with alignment specified via a typedef:
> 
> typedef int aligned_int __attribute__((aligned((8))));
> int foo(int a, aligned_int b) {...} //compiles ok
> 
> whereas specifying alignment directly, is rejected:
> 
> nonvar.c:2:20: error: alignment may not be specified for 'b'
>  int foo(int a, int b __attribute__((aligned((8)))))
>                     ^
> Note this is using the GNU __attribute__((aligned)) extension. Trying to use
> C11 _Alignas results in a frontend error either way; IIUC the C11 spec deems
> that sort of thing illegal.
> 
> (1) If we wish to keep the AAPCS principle that varargs are passed just as
> named args, we should use TYPE_MAIN_VARIANT inside
> arm_needs_doubleword_alignment, which will then ignore overalignment on both
> varargs _and named args_. However this would be silently ABI-changing....?
> 
> (2) It seems to me that SRA is the only way for overalignment info to be
> present on a value, so the patch to tree-sra.c/create_access_replacement seems
> to make things more consistent?

I'm not so sure about (2), SCCVN records the type of a reference
and PRE uses it to create the LHS temporaries to insert them.
You'd need some tricky order of optimizations to expose that to
a call argument though (copy-propagating the inserted value to
a call argument).  LIM may have similar issues (when doing store-motion),
so may predictive commoning and loop distribution (and maybe others I
forgot).

Richard.

-- 
Richard Biener <rguenther@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)