Patch for jump optimization

Patch for jump optimization

[Weiwen Liu]

On alpha-dec-osf4.0, the current egcs in the cvs tree fails to build
libgcc2.a, because the jump optimization (modified_in_p) fails to
process the following insn.

(jump_insn 52 51 54 (set (pc)
        (if_then_else (eq (reg:DI 102)
                (const_int 0))
            (label_ref 56)
            (pc))) 178 {umindi3+2} (nil)
    (nil))

I use the following patch to get the bootstrap going.

Mon Feb  2 1998 11:00:20  Weiwen Liu <weiwen.liu@yale.edu>
	* rtlanal.c (modified_in_p): Do not check modified_in_p if
	XEXP (x, i) is NULL.

Index: rtlanal.c
===================================================================
RCS file: /egcs/carton/cvsfiles/egcs/gcc/rtlanal.c,v
retrieving revision 1.6
diff -u -p -r1.6 rtlanal.c
--- rtlanal.c	1998/01/28 19:21:48	1.6
+++ rtlanal.c	1998/02/03 03:46:30
@@ -594,7 +594,7 @@ modified_in_p (x, insn)
   fmt = GET_RTX_FORMAT (code);
   for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
     {
-      if (fmt[i] == 'e' && modified_in_p (XEXP (x, i), insn))
+      if (fmt[i] == 'e' && XEXP (x, i) && modified_in_p (XEXP (x, i), insn))
 	return 1;
 
       if (fmt[i] == 'E')

Re: Patch for jump optimization

[Jeffrey A Law]

  In message < Pine.OSF.3.96.980202225739.10800A-100000@ylws34 >you write:
  > On alpha-dec-osf4.0, the current egcs in the cvs tree fails to build
  > libgcc2.a, because the jump optimization (modified_in_p) fails to
  > process the following insn.
  > 
  > (jump_insn 52 51 54 (set (pc)
  >         (if_then_else (eq (reg:DI 102)
  >                 (const_int 0))
  >             (label_ref 56)
  >             (pc))) 178 {umindi3+2} (nil)
  >     (nil))
  > 
  > I use the following patch to get the bootstrap going.
  > 
  > Mon Feb  2 1998 11:00:20  Weiwen Liu <weiwen.liu@yale.edu>
  > 	* rtlanal.c (modified_in_p): Do not check modified_in_p if
  > 	XEXP (x, i) is NULL.
What field number was NULL (ie, what is the value of 'i' when it failed)?

The real bug might be elsewhere...

jeff

Re: Patch for jump optimization

[Richard Henderson]

Err, not PATTERN (temp) but 

	modifued_in_p (PATTERN (insn), temp)

certainly that fixes the immediate problem.


r~

Re: Patch for jump optimization

[Jeffrey A Law]

  In message < 19980203000517.65224@dot.cygnus.com >you write:
  > Err, not PATTERN (temp) but 
  > 
  > 	modifued_in_p (PATTERN (insn), temp)
  > 
  > certainly that fixes the immediate problem.
Agreed.  The first arg should be a pattern, the second an actual insn.

I'll go fix it :-)

jeff

Re: Patch for jump optimization

[Richard Henderson]

On Tue, Feb 03, 1998 at 12:06:56AM -0700, Jeffrey A Law wrote:
>   > (jump_insn 52 51 54 (set (pc)
>   >         (if_then_else (eq (reg:DI 102)
>   >                 (const_int 0))
>   >             (label_ref 56)
>   >             (pc))) 178 {umindi3+2} (nil)
>   >     (nil))
>
> What field number was NULL (ie, what is the value of 'i' when it failed)?

6 -- that (nil) up there.

> The real bug might be elsewhere...

I'm inclined to think those modified_in_p calls you added
shouldn't be passing in temp, but rather PATTERN(temp).


r~

Re: Patch for jump optimization

[Weiwen Liu]

On Tue, 3 Feb 1998, Jeffrey A Law wrote:

> 
>   In message < Pine.OSF.3.96.980202225739.10800A-100000@ylws34 >you write:
>   > On alpha-dec-osf4.0, the current egcs in the cvs tree fails to build
>   > libgcc2.a, because the jump optimization (modified_in_p) fails to
>   > process the following insn.
>   > 
>   > (jump_insn 52 51 54 (set (pc)
>   >         (if_then_else (eq (reg:DI 102)
>   >                 (const_int 0))
>   >             (label_ref 56)
>   >             (pc))) 178 {umindi3+2} (nil)
>   >     (nil))
>   > 
>   > I use the following patch to get the bootstrap going.
>   > 
>   > Mon Feb  2 1998 11:00:20  Weiwen Liu <weiwen.liu@yale.edu>
>   > 	* rtlanal.c (modified_in_p): Do not check modified_in_p if
>   > 	XEXP (x, i) is NULL.
> What field number was NULL (ie, what is the value of 'i' when it failed)?
> 
> The real bug might be elsewhere...
> 
> jeff
> 

When i is 6, XEXP (x, 6) = (nil)

Patch for jump optimization

[Michael P. Hayes]

Weiwen Liu writes:
 > On alpha-dec-osf4.0, the current egcs in the cvs tree fails to build
 > libgcc2.a, because the jump optimization (modified_in_p) fails to
 > process the following insn.
 > 
 > (jump_insn 52 51 54 (set (pc)
 >         (if_then_else (eq (reg:DI 102)
 >                 (const_int 0))
 >             (label_ref 56)
 >             (pc))) 178 {umindi3+2} (nil)
 >     (nil))

The problem is that Jeff got a little carried away installing an old
patch of mine that disabled some jump optimisations if the inserted
insn clobbered something required by the following insn.
I fixed the problem for GCC 2.8.0 by looking for safer places that
the insn could be inserted, and then only rejecting the optimisation
if it couldn't find a suitable place.

The calls to modified_in_p aren't required.


Index: jump.c
===================================================================
RCS file: /egcs/carton/cvsfiles/egcs/gcc/jump.c,v
retrieving revision 1.10
diff -p -u -r1.10 jump.c
--- jump.c	1998/02/03 00:09:57	1.10
+++ jump.c	1998/02/03 08:38:26
@@ -995,7 +995,6 @@ jump_optimize (f, cross_jump, noop_moves
 	      && GET_CODE (SET_SRC (temp1)) != REG
 	      && GET_CODE (SET_SRC (temp1)) != SUBREG
 	      && GET_CODE (SET_SRC (temp1)) != CONST_INT
-	      && ! modified_in_p (PATTERN (insn), temp)
 	      && ! side_effects_p (SET_SRC (temp1))
 	      && ! may_trap_p (SET_SRC (temp1))
 	      && rtx_cost (SET_SRC (temp1), SET) < 10)
@@ -1035,13 +1034,11 @@ jump_optimize (f, cross_jump, noop_moves
 	      && GET_MODE_CLASS (GET_MODE (temp2)) == MODE_INT
 	      && (! SMALL_REGISTER_CLASSES
 		  || REGNO (temp2) >= FIRST_PSEUDO_REGISTER)
-	      && ! modified_in_p (PATTERN (insn), temp)
 	      && ! side_effects_p (SET_SRC (temp1))
 	      && ! may_trap_p (SET_SRC (temp1))
 	      && rtx_cost (SET_SRC (temp1), SET) < 10
 	      && (temp4 = single_set (temp3)) != 0
 	      && rtx_equal_p (SET_DEST (temp4), temp2)
-	      && ! modified_in_p (PATTERN (insn), temp3)
 	      && ! side_effects_p (SET_SRC (temp4))
 	      && ! may_trap_p (SET_SRC (temp4))
 	      && rtx_cost (SET_SRC (temp4), SET) < 10)
@@ -1094,7 +1091,6 @@ jump_optimize (f, cross_jump, noop_moves
 	      && REGNO (temp5) >= FIRST_PSEUDO_REGISTER
 	      && REGNO_FIRST_UID (REGNO (temp5)) == INSN_UID (temp)
 	      && REGNO_LAST_UID (REGNO (temp5)) == INSN_UID (temp3)
-	      && ! modified_in_p (PATTERN (insn), temp)
 	      && ! side_effects_p (SET_SRC (temp1))
 	      && ! may_trap_p (SET_SRC (temp1))
 	      && rtx_cost (SET_SRC (temp1), SET) < 10
@@ -1104,7 +1100,6 @@ jump_optimize (f, cross_jump, noop_moves
 	      && (! SMALL_REGISTER_CLASSES
 		  || REGNO (temp2) >= FIRST_PSEUDO_REGISTER)
 	      && rtx_equal_p (SET_DEST (temp4), temp2)
-	      && ! modified_in_p (PATTERN (insn), temp3)
 	      && ! side_effects_p (SET_SRC (temp4))
 	      && ! may_trap_p (SET_SRC (temp4))
 	      && rtx_cost (SET_SRC (temp4), SET) < 10)

Re: Patch for jump optimization

[Jeffrey A Law]

  In message < 199802030846.VAA11549@ongaonga.chch.cri.nz >you write:
  > Weiwen Liu writes:
  >  > On alpha-dec-osf4.0, the current egcs in the cvs tree fails to build
  >  > libgcc2.a, because the jump optimization (modified_in_p) fails to
  >  > process the following insn.
  >  > 
  >  > (jump_insn 52 51 54 (set (pc)
  >  >         (if_then_else (eq (reg:DI 102)
  >  >                 (const_int 0))
  >  >             (label_ref 56)
  >  >             (pc))) 178 {umindi3+2} (nil)
  >  >     (nil))
  > 
  > The problem is that Jeff got a little carried away installing an old
  > patch of mine that disabled some jump optimisations if the inserted
  > insn clobbered something required by the following insn.
  > I fixed the problem for GCC 2.8.0 by looking for safer places that
  > the insn could be inserted, and then only rejecting the optimisation
  > if it couldn't find a suitable place.
  > 
I must have missed that -- I didn't remember seeing any related changes go
into gcc-2.8.  I must have missed them.

Just so we're both clear, the stuff to handle constants in memory should be
OK right?

  > The calls to modified_in_p aren't required.
I'll install your updated patch :-)

Thanks for clarifying things!

jeff

Re: Patch for jump optimization

[Michael P. Hayes]

Jeffrey A Law writes:
 >   > I fixed the problem for GCC 2.8.0 by looking for safer places that
 >   > the insn could be inserted, and then only rejecting the optimisation
 >   > if it couldn't find a suitable place.
 >   > 
 > I must have missed that -- I didn't remember seeing any related changes go
 > into gcc-2.8.  I must have missed them.
 > 
 > Just so we're both clear, the stuff to handle constants in memory should be
 > OK right?

Yes, the stuff to handle conditional floating point loads and
conditional loads of constants in memory should be OK.  We could
probably generalise things to any operand supported by the backend.

Michael.