[Bug tree-optimization/105189] [9/10/11/12 Regression] Wrong code with -O1

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105189

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
I think the bug is in make_range_step.
When we see
(unsigned) foo () >= 0U, first make_range_step determines +[0U, -] range
for (unsigned) foo (), that is correct (though equivalent to +[-, -] aka always
true).
But then we handle the NOP_EXPR in another make_range_step call, exp_type is
unsigned int, arg0_type is int.  There is code that handles signed exp_type and
unsigned arg0_type and we punt when arg0_type is not integral, or has higher
precision than exp_type (i.e. narrowing conversion) or if either of the bounds
(if specified) doesn't fit into arg0_type.
None of that is the case here, so we incorrectly change the range to
foo () +[0, -] (but e.g. (unsigned) foo () +[-, -] would be handled correctly
as foo () +[-, -]).
At least for the TYPE_PRECISION equal case, the unsigned exp_type signed
arg0_type case is IMHO valid as is only if high is non-NULL or both low and
high are NULL (if high is non-NULL and higher than signed maximum, we already
punt), because if high is NULL, the range includes the largest unsigned value,
which is -1 in signed and when low is non-NULL, it is necessarily [0,
largest_signed].