Hi all,

This ICE has a similar cause to PR 64600.
The arm backend creates a const_int with a large value and doesn't properly sign-extend it.
If that rtx then happens to pass through the simplify-rtx machinery and ends up in wide-int land the asserts there catch it and ICE. In this case it was 0x7fffffff having 1 added to it to make
0x80000000 which should have been sign-extended to 0xffffffff80000000.

The code that caused this is:
class G {
public:
   void allocate(int p1) {
     if (p1 > max_size())
       operator new(sizeof(short));
   }
   unsigned max_size() { return -1 / sizeof(short); }
};


but it also needs to go through simplify-rtx to ICE, so the
testcase contains the context as well.

Similarly, the solution here is to use ARM_SIGN_EXTEND on the
SImode constants and trunc_int_for_mode for the DImode ones.

Bootstrapped and tested on arm-none-linux-gnueabihf.
Built SPEC2006 with it.
No code changes anywhere except in one instruction where a:
   cmp     r6, #2147483648

is transformed into a:
   cmp     r6, #-2147483648

Both assemble to the same thing, the disassembly is:
   cmp.w   r6, #2147483648 ; 0x80000000


Ok for trunk?

Thanks,
Kyrill

2015-04-09  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * config/arm/arm.c (arm_canonicalize_comparison): Use ARM_SIGN_EXTEND
     when creating +1 values for SImode and trunc_int_for_mode for similar
     DImode operations.

2015-04-09  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * g++.dg/torture/pr65694.C: New test.