www.digitalmars.com         C & C++   DMDScript  

D.gnu - Compiler bug? Addition/subtraction code

reply Era Scarecrow <rtcvb32 yahoo.com> writes:
Passes in DMD compiler, but breaks in GDC during assert tests.

B4S1L3 suggested it might be an old compiler bug. I need 
confirmation if this is an old (not yet caught up) or a new one 
(to which I'll have to write or have a work-around until the bug 
is fixed).

// http://dpaste.com/0M7M8NK
uint[] sub(uint[] lhs, const (uint)[] rhs) {
     long t;    //temporary, doubles as carry

     foreach(i, ref v; lhs) {
         t += v;
         t -= rhs[i];
         v = cast(uint) t;
         //reset carry
         t >>= uint.sizeof*8;
     }

     return lhs;
}

unittest {
     uint[3] lhs = [99, 201, 300],
             rhs = [-1, 0, 0];

     assert(sub(lhs, rhs) == [100, 200, 300]);
}
Jun 19 2017
parent reply Era Scarecrow <rtcvb32 yahoo.com> writes:
On Tuesday, 20 June 2017 at 01:58:22 UTC, Era Scarecrow wrote:

    long t;    //temporary, doubles as carry
    <snip>
         t >>= uint.sizeof*8;



  To note, looking in compiler explorer (GDC 5.2.0), the following 
output is present (for the above line):

         mov     rax, QWORD PTR [rbp-24]
         shr     rax, 32
         mov     QWORD PTR [rbp-24], rax

  'shr' is unsigned right shift, when it should be 'sar' signed 
right shift. That is my guess where the bug is.
Jun 19 2017
parent reply Era Scarecrow <rtcvb32 yahoo.com> writes:
On Tuesday, 20 June 2017 at 02:14:06 UTC, Era Scarecrow wrote:

  To note, looking in compiler explorer (GDC 5.2.0), the 
 following output is present (for the above line):


  Apparently on GDC 6.3 this doesn't fail. So likely an old 
compiler bug.

  Adding the following after the shift more or less confirms the 
unsigned shift is the error.

     t >>= uint.sizeof*8;
     version(GNU) {
         if (t)
             t |= 0xffffffff_00000000L;
     }
Jun 19 2017
parent reply Era Scarecrow <rtcvb32 yahoo.com> writes:
On Tuesday, 20 June 2017 at 03:31:24 UTC, Era Scarecrow wrote:

  Apparently on GDC 6.3 this doesn't fail. So likely an old 
 compiler bug.


  Got a message not long ago, apparently the assertion still 
happens on GDC 6.3. So...
Jun 20 2017
parent reply Iain Buclaw <ibuclaw gdcproject.org> writes:
On Tuesday, 20 June 2017 at 18:49:24 UTC, Era Scarecrow wrote:

 On Tuesday, 20 June 2017 at 03:31:24 UTC, Era Scarecrow wrote:

  Apparently on GDC 6.3 this doesn't fail. So likely an old 
 compiler bug.


  Got a message not long ago, apparently the assertion still 
 happens on GDC 6.3. So...


OK. Bugzilla? I'll have a look.

(By the way, you can inspect disassembly here: 
https://explore.dgnu.org/)

Regards,
Iain.
Jun 20 2017
parent Iain Buclaw <ibuclaw gdcproject.org> writes:
On Tuesday, 20 June 2017 at 19:06:22 UTC, Iain Buclaw wrote:

 On Tuesday, 20 June 2017 at 18:49:24 UTC, Era Scarecrow wrote:

 On Tuesday, 20 June 2017 at 03:31:24 UTC, Era Scarecrow wrote:

  Apparently on GDC 6.3 this doesn't fail. So likely an old 
 compiler bug.


  Got a message not long ago, apparently the assertion still 
 happens on GDC 6.3. So...


 OK. Bugzilla? I'll have a look.


OK, no need.  It's close enough to a prior bug (171).

https://github.com/D-Programming-GDC/GDC/pull/501

Regards,
Iain.
Jun 20 2017