[Bug target/59803] New: [4.8 Regression] s390x -march=z10 reload ICE

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59803

            Bug ID: 59803
           Summary: [4.8 Regression] s390x -march=z10 reload ICE
           Product: gcc
           Version: 4.8.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: jakub at gcc dot gnu.org

extern void baz (void) __attribute__ ((__noreturn__));
struct A { int g, h; };
extern struct A a;
struct B { unsigned char i, j, k, l, m; };
int c, d, e;
static int f;

void
foo (void)
{
  f = 1;
}

void
bar (struct B *x)
{
  x->i = e;
  x->k = c;
  x->l = d;
  x->j = a.h;
  x->m = f;
  if (x->i != e) baz ();
  if (x->k != c) baz ();
  if (x->j != a.h) baz ();
}

ICEs with -O2 -march=z10 on 4.8 branch (verified with x86_64-linux ->
s390x-linux cross at r206599).  Works in 4.6 and works on the trunk, though it
is unclear if it just isn't latent.

Before reload we have:
(insn 10 9 11 2 (set (reg:SI 58 [ d ])
        (mem/c:SI (symbol_ref:DI ("d")  <var_decl 0x7f11a63617b8 d>) [2 d+0 S4
A32])) rh1052372.ii:19 67 {*movsi_zarch}
     (expr_list:REG_EQUIV (mem/c:SI (symbol_ref:DI ("d")  <var_decl
0x7f11a63617b8 d>) [2 d+0 S4 A32])
        (nil)))
(insn 11 10 13 2 (set (mem:QI (plus:DI (reg/v/f:DI 57 [ x ])
                (const_int 3 [0x3])) [0 x_4(D)->l+0 S1 A8])
        (subreg:QI (reg:SI 58 [ d ]) 3)) rh1052372.ii:19 74 {*movqi}
     (expr_list:REG_DEAD (reg:SI 58 [ d ])
        (nil)))
and the ICE is about unrecognized instruction:
(insn 61 10 11 2 (set (reg:DI 12 %r12)
        (const:DI (plus:DI (symbol_ref:DI ("d") <var_decl 0x7fbc425267b8 d>)
                (const_int 3 [0x3])))) rh1052372.ii:19 -1
     (nil))
If I look at trunk dumps, it seems the difference there is already at expansion
time, while on the trunk *.expand has separate insns to set (symbol_ref:DI
("d")
to a pseudo and load (mem:SI (that pseudo)), 4.8 branch uses movsi_zarch insn
which is load from (mem:SI (symbol_ref:DI ("d"))).  Then, at combine time
trunk combines the store with the load (and not symbol_ref larl), while 4.8
fails to combine the store with the load because I suppose lowest bit in larl
loaded value can't be set?