simplify_relational_operation and unsigned char

simplify_relational_operation and unsigned char

[Amir Gonnen]

(gcc-4.4.1)

Hi,
I found out some strange behavior of simplify_relational_operation.
Apparently it assumes QImode is signed even though this could not be
deduced from the EQ operator.
Here is an example:

rtx reg = gen_reg_rtx(QImode);
...
rtx test = simplify_relational_operation(EQ, VOIDmode, QImode, reg,
GEN_INT(129));

In this case simplify_relational_operation returns const_int 0 because
of comparisons optimization with upper and lower bounds performed on
simplify_const_relational_operation.
The problem with this behavior was the removal of a "case 129:" in a
switch during cse pass (proprietary target and backend).

Perhaps EQ and NE could be treated as unsigned, like the
GEU,LEU,GTU,LTU operators?

Amir

Re: simplify_relational_operation and unsigned char

[Richard Henderson]

On 10/29/2009 03:35 AM, Amir Gonnen wrote:
> rtx test = simplify_relational_operation(EQ, VOIDmode, QImode, reg,
> GEN_INT(129));

CONST_INT is always signed.

If you really wrote this as you quote here, that would be the breakage. 
  Try gen_int_mode (129, QImode) instead.



r~

Re: simplify_relational_operation and unsigned char

[Dave Korn]

Amir Gonnen wrote:

> Perhaps EQ and NE could be treated as unsigned, like the
> GEU,LEU,GTU,LTU operators?

  :) What exactly is "signed equality" and how does/would it differ from
"unsigned equality"?  I'm not sure the concept makes any sense!  (And that's
always a warning sign that your diagnosis has gone astray somewhere and led
you to a false understanding of the problem you're working on.)

    cheers,
      DaveK