[Bug middle-end/86554] Incorrect code generation with signed/unsigned comparison

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86554

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue seems to be that we do

Value numbering ret_10 stmt = ret_10 = (int) _9;
Setting value number of ret_10 to ret_11 (changed)

so we correctly figure both ret are the same.  We then continue

Value numbering ret_12 stmt = ret_12 = PHI <ret_10(3), ret_11(4)>
Setting value number of ret_12 to ret_11
...
Visiting control stmt ending BB 5: if (ret_12 <= 0)
Marking all edges out of BB 5 but (5 -> 6) as not executable

oops.  We seem to have simplified this as

 if (_8 + -2147483647 <= 0)

which given strict overflow rules is always true.

The issue here is that we did expression simplification of a value expression,
and this is a really really hard issue to fix.  I'm currently working on a
complete rewrite of FRE for GCC9 which doesn't exhibit this issue (by design).

For some reason PRE code-hoisting re-introduces the issue by hoisting
the signed operation.