[Bug tree-optimization/22116] [4.1 Regression] PRE of COMPLEX_EXPR causes ICE

[rth at gcc dot gnu dot org]

------- Additional Comments From rth at gcc dot gnu dot org  2005-06-19 03:12 -------
(In reply to comment #7)
> It actually does.

Ah.  I failed to add -vops to the dump and so mis-diagnosed the problem.

> The first problem here is that PRE doesn't mark anything for renaming
> because it's never inserted things that may change vuses/v_may_defs before.

It shouldn't have to change vuse/vdef now either because ...

> I fixed it to do this, and it still fails, because we create a PHI of a
> COMPLEX_TYPE, which it thinks is a virtual phi for a gimple register,
> because DECL_COMPLEX_GIMPLE_REG_P is not set on it.

... DECL_COMPLEX_GIMPLE_REG_P should be set on the new variable.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22116