[Bug c/12654] [3.3/3.4 regression] Incorrect comparison code generated for Alpha

[falk at debian dot org]

PLEASE REPLY TO gcc-bugzilla@gcc.gnu.org ONLY, *NOT* gcc-bugs@gcc.gnu.org.

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=12654


falk at debian dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|normal                      |critical
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |wrong-code
            Summary|Incorrect comparison code   |[3.3/3.4 regression]
                   |generated for Alpha         |Incorrect comparison code
                   |                            |generated for Alpha


------- Additional Comments From falk at debian dot org  2003-10-17 08:52 -------
This boils down to 

foo(MIN_LONG)

(the posted test case is technically invalid).

It is caused by gcc generating

lda     v0,-1023(a0)
cmplt   zero,v0,v0

instead of

lda     v0,1024
cmple   v0,a0,v0

Happens only in some contexts, e.g. for

int foo (long x) { if (x >= 1024) return 1; else return 0; }

but not for

int bar (long x) { return x >= 1024; }

gcc 2.95 got it right, 3.2 not. Probably target dependent, but I don't have 
anything else to test currently.

The bogus variant is actually slightly better WRT register pressure, so we
might want to retain it for int arguments.