[Bug other/16800] New: PowerPC - Unnecessary rldicl

[Bug other/16800] New: PowerPC - Unnecessary rldicl

[gcc-bugzilla at gcc dot gnu dot org]

Description:
A non-optimal code sequence is illustrated.  The rldicl noted below is not needed (and doesn't occur for the "==" comparison).  Duplicate using gcc 3.5 and command line:

gcc -O3 -m64 -c test.c

Testcase:
long l1,l2;
float f1,f2,f3,f4;

void logic_func4() {
  l1 = (f1 == f2);
  l2 = (f3 > f4);
}

Assembly:
.logic_func4:
	ld 8,.LC5@toc(2)
	ld 6,.LC4@toc(2)
	lfs 13,0(8)
	lfs 11,0(6)
	ld 5,.LC1@toc(2)
	ld 4,.LC2@toc(2)
	lfs 12,0(5)
	fcmpu 7,11,13
	lfs 0,0(4)
	ld 11,.LC0@toc(2)
	ld 10,.LC3@toc(2)
	fcmpu 6,12,0
	mfcr 3
	rlwinm 3,3,30,1
	rldicl 0,3,0,32 <-- Unnecessary.
	mfcr 9
	rlwinm 9,9,27,1
	std 0,0(10)
	std 9,0(11)
	blr



-- 
           Summary: PowerPC - Unnecessary rldicl
           Product: gcc
           Version: 3.5.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P1
         Component: other
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: steinmtz at us dot ibm dot com
                CC: gcc-bugs at gcc dot gnu dot org,steinmtz at us dot ibm
                    dot com
 GCC build triplet: powerpc64-linux
  GCC host triplet: powerpc64-linux
GCC target triplet: powerpc64-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[bangerth at dealii dot org]

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|other                       |target


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-08-02 01:02 -------
Confirmed, I think this is a target problem:
(insn:TI 36 20 37 0 (set (reg:SI 0 0 [134])
        (gt:SI (reg:CCFP 75 7 [135])
            (const_int 0 [0x0]))) 422 {*rs6000.md:11384} (insn_list 34 (nil))
    (expr_list:REG_DEAD (reg:CCFP 75 7 [135])
        (nil)))

(insn:TI 37 36 23 0 (set (reg:DI 0 0 [131])
        (zero_extend:DI (reg:SI 0 0 [134]))) 12 {*rs6000.md:374} (insn_list 36 (nil))
    (expr_list:REG_EQUIV (mem/i:DI (reg/f:DI 10 10 [128]) [4 l2+0 S8 A64])
        (nil)))

(insn:TI 23 37 38 0 (set (reg:DI 9 9 [122])
        (eq:DI (reg:CCFP 74 6 [126])
            (const_int 0 [0x0]))) 423 {*rs6000.md:11416} (insn_list:REG_DEP_ANTI 32 
(insn_list:REG_DEP_OUTPUT 29 (insn_list 20 
(nil))))
    (expr_list:REG_DEAD (reg:CCFP 74 6 [126])
        (expr_list:REG_EQUIV (mem/i:DI (reg/f:DI 11 11 [119]) [4 l1+0 S8 A64])
            (nil))))

But I could be wrong.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2004-08-02 01:02:22
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-01 00:27 -------
Fixed on the mainline.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[steinmtz at us dot ibm dot com]

------- Additional Comments From steinmtz at us dot ibm dot com  2004-11-22 22:27 -------
I rechecked this one with a version of mainline built today (11/22/04) and the 
problem still appears.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |REOPENED
         Resolution|FIXED                       |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[amodra at bigpond dot net dot au]

------- Additional Comments From amodra at bigpond dot net dot au  2004-11-24 07:00 -------
I believe this problem is caused by rs6000_rtx_costs.  The two following insns
are successfully combined:

(insn 20 19 21 0 (set (reg:DI 126)
        (eq:DI (reg:CCFP 125)
            (const_int 0 [0x0]))) 410 {*rs6000.md:11408} (insn_list:REG_DEP_TRUE
19 (nil))
    (expr_list:REG_DEAD (reg:CCFP 125)
        (nil)))

(insn 21 20 22 0 (set (reg:SI 124)
        (subreg:SI (reg:DI 126) 4)) 280 {*movsi_internal1}
(insn_list:REG_DEP_TRUE 20 (nil))
    (expr_list:REG_DEAD (reg:DI 126)
        (nil)))

to

(insn 21 20 22 0 (set (reg:SI 124)
        (eq:SI (reg:CCFP 125)
            (const_int 0 [0x0]))) 409 {*rs6000.md:11376} (insn_list:REG_DEP_TRUE
20 (nil))
    (expr_list:REG_DEAD (reg:DI 126)
        (nil)))

Following that, the new insn 21 and

(insn 22 21 23 0 (set (reg:DI 121)
        (zero_extend:DI (reg:SI 124))) 24 {*rs6000.md:379}
(insn_list:REG_DEP_TRUE 21 (nil))
    (expr_list:REG_DEAD (reg:SI 124)
        (nil)))

is a combine candidate, with a good result pattern of

(set (reg:DI 121)
    (eq:DI (reg:CCFP 125)
        (const_int 0 [0x0])))

However, rs6000_rtx_costs decides this patten is more expensive than the
previous two instructions.

rejecting combination of insns 21 and 22
original costs 12 + 4 = 16
replacement cost 20

Code in question is as follows.  I don't understand why these particular
comparisons have been singled out, or why the mode == Pmode test.
This contributes 12

    case EQ:
    case GTU:
    case LTU:
      if (mode == Pmode)
	{
	  switch (outer_code)
	    {
	    case SET:
	      *total = COSTS_N_INSNS (3);
	      break;

The (const_int 0) contributes 8!  This is because outer_code is EQ, which
doesn't match anything in the rather long set of outer_code tests for CONST_INT.

An obvious fix is to handle EQ and other comparisons here.



-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dje at gcc dot gnu dot org
   Last reconfirmed|2004-08-02 01:02:22         |2004-11-24 07:00:16
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[amodra at bigpond dot net dot au]

------- Additional Comments From amodra at bigpond dot net dot au  2004-11-24 09:40 -------
Of course, fixing the odd rtx_cost for the (const_int 0) inside an EQ doesn't
help this particular bug.  More or less the same happens:

rejecting combination of insns 21 and 22
original costs 4 + 4 = 8
replacement cost 12

It's the 12 counted for EQ that's the problem.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[dje at watson dot ibm dot com]

------- Additional Comments From dje at watson dot ibm dot com  2004-11-24 15:21 -------
Subject: Re:  PowerPC - Unnecessary rldicl 

	The test for mode == Pmode is present because most of the fast scc
patterns depend on carry.  PowerPC carry depends on the mode of the
processor.  If the mode of the comparison is not the mode of the processor
(Pmode), the scc patterns cannot be applied and the modified cost should
not be applied.

	Similarly, only EQ, GTU, and LTU costs are handled because those
are the fast scc patterns active for PowerPC.  I did not include the costs
for the original POWER architecture.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-11-27 23:01 -------
Subject: Bug 16800

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	dje@gcc.gnu.org	2004-11-27 23:01:00

Modified files:
	gcc            : ChangeLog 
	gcc/config/rs6000: rs6000.c 

Log message:
	PR target/16800
	* config/rs6000/rs6000.c (rs6000_rtx_costs): Improve accuracy of
	EQ, GTU, and LTU costs.  Add costs for GT, LT, and UNORDERED.
	Distinguish between SImode and DImode CONST_INT.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6593&r2=2.6594
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/config/rs6000/rs6000.c.diff?cvsroot=gcc&r1=1.756&r2=1.757



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800

[Bug target/16800] PowerPC - Unnecessary rldicl

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-11-28 00:07 -------
Fixed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|REOPENED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=16800