[Bug c++/22568] New: Should use cmov in some stituations

[Bug c++/22568] New: Should use cmov in some stituations

[arndt at jjj dot de]

With (ulong a, ulong b) and code like
if ( a<b )  { ulong t=a; a=b; b=t; }
gcc should (if a and b are in registers already)
not emit a conditional jump but code like
MOV a, t;
CMP a, b;
CMOVcc b, a;
CMOVcc t, b;

Other such examples are easily found.  If I got it correctly, gcc
emits CMOVcc in just one situation, the conditional assignment.

Machine is AMD64, OS is SuSE Linux 9.3
gcc (GCC) 3.3.5 20050117 (prerelease) (SUSE Linux)

-- 
           Summary: Should use cmov in some stituations
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P2
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: arndt at jjj dot de
                CC: gcc-bugs at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug c++/22568] Should use cmov in some stituations

[matz at suse dot de]

------- Additional Comments From matz at suse dot de  2005-07-20 14:20 -------
This still happens with 4.1.  I also can't make it use two cmovs, by changing 
the source a bit, e.g. like: 
 
typedef unsigned long ulong; 
extern ulong use (ulong, ulong); 
ulong f(ulong a, ulong b) 
{ 
  ulong tmp = a; 
  if (a < b) { 
    a = b; 
    b = tmp; 
  } 
  return use (a, b); 
} 
 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |matz at suse dot de


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug rtl-optimization/22568] Should use cmov in some stituations

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-07-20 15:11 -------
IIRC ifcvt is not smart enough to do this.

Confirmed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
          Component|c++                         |rtl-optimization
     Ever Confirmed|                            |1
           Keywords|                            |missed-optimization
   Last reconfirmed|0000-00-00 00:00:00         |2005-07-20 15:11:41
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug rtl-optimization/22568] Should use cmov in some stituations

[rguenth at gcc dot gnu dot org]

------- Additional Comments From rguenth at gcc dot gnu dot org  2005-08-31 09:40 -------
With new tree-codes instead of using COND_EXPR we may use the tree-vectorizers
if-conversion and make expand preserve the conditional moves.

Also it shouldn't be too hard to hack rtl if-conversion to handle the case of
exactly two set's in the then/else block, too, and that may turn out to be
profitable always.  Is there any arch whose conditional move will kill
condition codes?  RTL is not my best friend (yet), but ifcvt doens't look too bad
either ;)

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug rtl-optimization/22568] Should use cmov in some stituations

[rguenth at gcc dot gnu dot org]

------- Additional Comments From rguenth at gcc dot gnu dot org  2005-09-01 12:38 -------
I have a patch that for

int a,b;
void foo(void)
{
  int x;
  if (a<b) x=a,a=b,b=x;
}

produces

bar:
        movl    a, %eax
        movl    b, %edx
        cmpl    %edx, %eax
        movl    %eax, %ecx
        cmovl   %edx, %ecx
        cmovge  b, %eax
        movl    %ecx, a
        movl    %eax, b
        ret

but it produces horrible code if the stuff is not available in pseudos
already.  (i.e. we miss load/store sinking)

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug rtl-optimization/22568] Should use cmov in some stituations

[bonzini at gcc dot gnu dot org]

------- Additional Comments From bonzini at gcc dot gnu dot org  2005-09-02 09:47 -------
Richard, can you write a case where it produces awful code?

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568

[Bug rtl-optimization/22568] Should use cmov in some stituations

[rguenth at gcc dot gnu dot org]

------- Additional Comments From rguenth at gcc dot gnu dot org  2005-09-02 09:54 -------
Blindly applying ifcvt to something like

int a,b;
void foo(int flag)
{
  int x;
  if (flag)
    x=a,a=b,b=x;
}

because we're presented with

  if (flag)
  {
    int reg_a = a;
    x = reg_a;
    int reg_b = b;
    a = reg_b;
    b = x;
  }

and we get cmovs for loading a,b into pseudos instead of
loading them unconditionally.  F.i.

int a,b;
void foobar(int flag)
{
  if (flag)
    a = b;
}

will become

foobar:
        movl    4(%esp), %eax   # flag, flag
        testl   %eax, %eax      # flag
        cmovne  b, %edx # b,, b
        movl    a, %eax # a, tmp61
        cmovne  %edx, %eax      # b,, tmp61
        movl    %eax, a # tmp61, a
        ret

notice how we special-cased the _store_ to use a temporary.  I'll
try applying the same for loads.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22568