[Bug rtl-optimization/26479] New: suboptimal register allocation for return register

[Bug rtl-optimization/26479] New: suboptimal register allocation for return register

[hp at gcc dot gnu dot org]

For this code, at -O2, R13 is needlessly used, where R10 would suffice.
(Leading to two extra move insns.)  Curiously, if-conversion exposes
the suboptimal allocation; with -fno-if-conversion you get the better
allocation, but looking at the assembled code, there's no reason to not
use r10.  Observed with 108225 and 111226.  Code is from bzip2recover.c.

extern int _IO_getc (void *) ;
extern int *__errno_location (void) __attribute__ ((__const__));
void readError ( void );
typedef
   struct {
      void* handle;
      int buffer;
      int buffLive;
      char mode;
   }
   BitStream;

int bsGetBit ( BitStream* bs )
{  
   if (bs->buffLive > 0) {
      bs->buffLive --;
      return ( ((bs->buffer) >> (bs->buffLive)) & 0x1 );
   } else {
      int retVal = _IO_getc (bs->handle);
      if ( retVal == (-1) ) {
         if ((*__errno_location ()) != 0) readError();
         return 2;
      }
      bs->buffLive = 7;
      bs->buffer = retVal;
      return ( ((bs->buffer) >> 7) & 0x1 );
   }
}


-- 
           Summary: suboptimal register allocation for return register
           Product: gcc
           Version: 4.2.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: rtl-optimization
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hp at gcc dot gnu dot org
GCC target triplet: cris-axis-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[pinskia at gcc dot gnu dot org]

------- Comment #1 from pinskia at gcc dot gnu dot org  2006-03-03 18:54 -------
Interesting it shows up on powerpc64 also with the work around of turning off
if conversion fixing the issue too.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
 GCC target triplet|cris-axis-linux-gnu         |cris-axis-linux-gnu,
                   |                            |powerpc64-*
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2006-03-03 18:54:10
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2006-03-03 18:56 -------
With -O2:
L4:
        addi r1,r1,128
        mr r3,r2
        ld r0,16(r1)
        ld r31,-8(r1)
        mtlr r0
        blr

With -O2 -fno-if-conversion:
L4:
        addi r1,r1,128
        ld r0,16(r1)
        ld r31,-8(r1)
        mtlr r0
        blr

I have not looked enough into the RTL to figure out why this is going wrong.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |ra


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2006-03-03 19:00 -------
Oh, if I turn off the first pass of scheduling it also works so that means this
is truely a ra issue.  if conversion is just exposing the problem by moving the
load of constant above the branch.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[pinskia at gcc dot gnu dot org]

------- Comment #4 from pinskia at gcc dot gnu dot org  2008-09-14 03:16 -------
I think this was fixed for 4.3.0:
.L3:
        lwz 0,20(1)
        lwz 31,12(1)
        addi 1,1,16
        mtlr 0
        blr


HP,
can you try this again on cris?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[hp at gcc dot gnu dot org]

------- Comment #5 from hp at gcc dot gnu dot org  2008-09-24 16:49 -------
(In reply to comment #4)
> I think this was fixed for 4.3.0:
> HP,
> can you try this again on cris?

At 140627, the problem is still there on the 4.3 branch for CRIS.  I'll attach
a diff of the generated assembly of -O2 compared to -O2 -fno-if-conversion.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[hp at gcc dot gnu dot org]

------- Comment #6 from hp at gcc dot gnu dot org  2008-09-24 16:51 -------
Created an attachment (id=16402)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16402&action=view)
Diff of asembly of -O2 vs -O2 -fno-if-conversion at r.140627

See previous comment.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479

[Bug rtl-optimization/26479] suboptimal register allocation for return register

[hp at gcc dot gnu dot org]

------- Comment #7 from hp at gcc dot gnu dot org  2008-09-24 16:57 -------
I guess I should turn the code in the description into a testcase in gcc.dg so
target maintainers can add their xfailed scan-assembler-not and we'd see
xpasses if/when this is magically fixed...
(After the IRA adjustments for CRIS!)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=26479