#2329 long long mul / mod broken on gbz80

[Philipp Klaus Krause]

Tests longlong.c and gcc-torture-execute-20080529-1.c fail with the new version of long long multiplication. Since gbz80 probably has much fewer long long users than stm8, I don't want to revert to the old one. And the gbz80 issue needs to be fixed anyway.

Philipp

[Ben Shi]

The mod operation also fails (only) for gbz80 with the following tests (if added in longlong.c).

  y = 0x1122334455667700ull;
  x = 0x7ll;
  ASSERT (y % x == 0x1122334455667700ull % 0x7ll);
  x = 0x1122334455667700ll;
  ASSERT (x % 0x7ll == 0x1122334455667700ll % 0x7ll);

[Ben Shi]

other failed longlong test cases on gbz80 besides longlong.c

gcc-torture-execute-20000717-2.c
gcc-torture-execute-20001108-1.c
gcc-torture-execute-20011126-1.c
gcc-torture-execute-20041011-1.c
gcc-torture-execute-20050104-1.c
gcc-torture-execute-20060110-2.c
gcc-torture-execute-20080529-1.c
gcc-torture-execute-920829-1.c
gcc-torture-execute-921202-2.c
gcc-torture-execute-950511-1.c
gcc-torture-execute-950612-1.c
gcc-torture-execute-950704-1.c
gcc-torture-execute-961122-1.c
gcc-torture-execute-961213-1.c
gcc-torture-execute-961213-1.c

[Philipp Klaus Krause]

THis is an issue due to the use of ldhl overwriting the carry byte. It can also be reproduced with long:

long acc2;

addhi2 (short a)
{
  acc2 += (long) a << 16;
}

is affected by the same issue. I.e. regtest gcc-torture-execute-961122-1.c would fail for gbz80, even if it used long instead of longlong.

Philipp

[Philipp Klaus Krause]

I fixed the ldhl carry issues. As of revision #9563 everything except for multiplication works.

Philipp

[Philipp Klaus Krause]

The remaining issue turned out to be a bug in register tracking for shifts by 0. It is fixed in [r10170].

Philipp