[sdcc-devel] [ sdcc-Feature Requests-1162453 ] condition (i<0x1000) could be (i&0xf000)

[SourceForge.net <no...@so...>]

Feature Requests item #1162453, was opened at 2005-03-13 15:11
Message generated for change (Comment added) made by frief
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=350599&aid=1162453&group_id=599

Category: None
Group: None
Status: Open
Priority: 1
Submitted By: Frieder Ferlemann (frief)
Assigned to: Nobody/Anonymous (nobody)
Summary: condition (i<0x1000) could be (i&0xf000)

Initial Comment:
An unsigned compare to a literal which is a power of 2
and >=256 could be reduced to anding a bitmask.


unsigned char t;

void f(unsigned int i)
{
  if( i < 0x1000 )
     t++;

  /* functionally equivalent, but only upper byte(s)
need(s) to be accessed */
  if( !( i & 0xf000 ) )
     t++;
}

On the mcs51:

;lessthan.c:6: if( i < 0x1000 )
        clr     c
        mov     a,r2
        subb    a,#0x00
        mov     a,r3
        subb    a,#0x10
        jnc     00102$


;lessthan.c:10: if( !( i & 0xf000 ) )

        mov     a,r3
        anl     a,#0xF0
        jz      00111$


----------------------------------------------------------------------

>Comment By: Frieder Ferlemann (frief)
Date: 2005-03-29 20:54

Message:
Logged In: YES 
user_id=589052

attaching the file Hubert sent to me. 

----------------------------------------------------------------------

Comment By: Hubert Sack (hsack)
Date: 2005-03-29 20:46

Message:
Logged In: YES 
user_id=1160854

I corrected the source and pleased Frieder to do the 
attachment.


----------------------------------------------------------------------

Comment By: Hubert Sack (hsack)
Date: 2005-03-29 16:30

Message:
Logged In: YES 
user_id=1160854

I detected that I made a mistake. The jump is inverted! I'll 
post a correct version soon.
My error was found by running regression tests of two more 
solutions of RFE's.
The example above is also wrong, if 00111$ is the label 
following the conditional block!

----------------------------------------------------------------------

Comment By: Erik Petrich (epetrich)
Date: 2005-03-16 07:08

Message:
Logged In: YES 
user_id=635249

I fixed the bug of x = x being discarded when x is volatile.

> left  = ((volatile unsigned int)left);
> Shouldn't this provoke a dummy read?

I don't think so. The cast modifies the type of the value of
the expression following, but has no control over how the
value is initially derived. Since the object "left" is not
itself declared as volatile, the compiler is still free to
determine its value by either direct or optimized means.


----------------------------------------------------------------------

Comment By: Hubert Sack (hsack)
Date: 2005-03-15 19:02

Message:
Logged In: YES 
user_id=1160854

Hello Frieder,
I think, you're right. Volatile is checked in "genAssign" for 
this. But the modifier "volatile" has no effect. With the option "-
-i-code-in-asm" there is no "volatile" shown and therefore no 
code is generated.
But there is a much more strange error (i think it should be 
reported as bug!):
If the variable is declared as volatile, the elemination is 
already done some stage(s) earlier - there is no icode 
generated. This is a bug! Reading a volatile variable must be 
done and can not be simply optimized away...


----------------------------------------------------------------------

Comment By: Frieder Ferlemann (frief)
Date: 2005-03-14 21:07

Message:
Logged In: YES 
user_id=589052

Hello Hubert,
I added test cases for this in
sdcc/support/regression/tests/compare.c (patch is appended).
There's a problem lurking somewhere.

Off-topic: Another (unrelated) strangeness (with my attached
patch) is that there is no code generated for:
 left  = ((volatile unsigned int)left);
Shouldn't this provoke a dummy read?


----------------------------------------------------------------------

Comment By: Hubert Sack (hsack)
Date: 2005-03-13 21:08

Message:
Logged In: YES 
user_id=1160854

Here is the solution for the RFE:

Index: sdcc/src/SDCCopt.c
=============================================
======================
RCS file: /cvsroot/sdcc/sdcc/src/SDCCopt.c,v
retrieving revision 1.54
diff -c -r1.54 SDCCopt.c
*** sdcc/src/SDCCopt.c	9 Mar 2005 19:18:47 -0000
	1.54
--- sdcc/src/SDCCopt.c	13 Mar 2005 21:02:08 -0000
***************
*** 582,587 ****
--- 581,611 ----
  
+           if (ic->op == '<' && isOperandLiteral(IC_RIGHT(ic)) 
&& IS_UNSIGNED(operandType(IC_LEFT(ic))))
+             {
+               unsigned litVal = abs(operandLitValue(IC_RIGHT
(ic)));
+ 
+               // See if literal value is greather 255 and a power 
of 2.
+               if (litVal > 255)
+                 {
+                   while (litVal && !(litVal & 1))
+                     {
+                       litVal >>= 1;
+                     }
+                   if (litVal)
+                     {
+                       // discard lowest set bit.
+                       litVal >>= 1;
+                     }
+                 }
+               if (!litVal)
+                 {
+                   ic->op = BITWISEAND;
+                   IC_RIGHT(ic) = operandFromLit(0 - 
operandLitValue(IC_RIGHT(ic)));
+                   continue;
+                 }    
+             }

It's tested and o.k.


----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=350599&aid=1162453&group_id=599