[Bug middle-end/53823] [4.8 Regression] FAIL: gcc.c-torture/execute/930921-1.c execution at -O0 and -O1

[danglin at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53823

--- Comment #21 from John David Anglin <danglin at gcc dot gnu.org> 2012-08-01 18:44:04 UTC ---
The issue is with the handling of negative constants.

In this bit of code,

          max_cost = (set_src_cost (gen_rtx_MULT (mode, fake_reg, op1), speed)
                      - neg_cost(speed, mode));
          if (max_cost > 0
              && choose_mult_variant (mode, -coeff, &algorithm,
                                      &variant, max_cost))

max_cost is computed to be 24 and choose_mult_variant returns 0.  This
causes the following hunk to be executed:

      max_cost = set_src_cost (gen_rtx_MULT (mode, fake_reg, op1), speed);
      if (choose_mult_variant (mode, coeff, &algorithm, &variant, max_cost))
        return expand_mult_const (mode, op0, coeff, target,
                                  &algorithm, variant);

max_cost is now 32 and choose_mult_variant succeeds.  However,
expand_mult_const
generates incorrect code when coeff is negative.

I hacked the max_cost in the negative case to be 32, and this produces
the correct result and code similar to your cross-compiled output.

If I remember correctly, the xmpyu instruction was introduced in PA 1.1
and this results in different costs for PA 1.0 and 1.1.