Re: Reload patch v4

[Bernd Schmidt <crux at Pool dot Informatik dot RWTH-Aachen dot DE>]

>   > > I think we need to fix forget_old_reloads_1 to handle cases where we
>   > > store into a pseudo which happens to be allocated to a hard reg that
>   > > has also been used as a spill reg.  Imagine
>   > > 
>   > > 	Use reg 0 as an output reload for some value V
>   > > 
>   > > 	Psuedo X is made live and is allocated to reg 0
>   > > 	[ ... ]
>   > > 	Pseudo X dies
>   > > 
>   > > 	Use of V which needs an input reload.
>   > > 
>   > > In this case we might try and use reg0 to satisfy the input reload
>   > > even though it no longer holds the value we need.
>   > 
>   > I'm not convinced yet that anything is broken.  The situation you describe
>   > could already occur with the old reload code, since it has some simple code
>   > to avoid spilling within one basic block.
> We only avoid spilling a pseudo in a block if the block needed no spilling of
> any class that contains the hard register the pseudo was allocated to.
> 
> Thus if we use reg0 to hold an output reload, then no pseudos can be allocated
> to reg0 in that block and thus the above situation can not currently occur.
> 
> [ Remember, we can't inherit a reload across basic block boundaries, at least
>   we shouldn't do it right now :-) ]

Are you sure about this?  It should be perfectly legal to inherit across
a basic block boundary, unless there's a CODE_LABEL at the start of the second
basic block.  I can't find any code that forgets old reloads when a new basic
block starts, only a piece of code that does so at a label.

The situation might look like this:

	[first basic block]
 	Use reg 0 as an output reload for some value V

	conditional jump to some place	

	[second basic block, this one didn't spill reg 0]
 	Psuedo X is made live and is allocated to reg 0
 	[ ... ]
 	Pseudo X dies
 
	conditional jump to some place

	[third basic block]
 	Use of V which needs an input reload.

>   > In the above example, if pseudo X is allocated to reg 0, its rtl will have
>   > been modified to look like hard reg 0 by the time forget_old_reloads_1 runs > .
>   > And in that case, the following piece of code in find_old_reloads_1 should
>   > run.
> Normally I would agree with you, but if the pseudo's rtl had been changed,
> then why does forget_old_reload_1 check for regno >= FIRST_PSEUDO_REGISTER?
>
> I *know* pseudos can be passed to that function.  I've seen it happen.  When
> they do, they do not go through the code you quoted.

Maybe because of this (from eliminate_regs):

        /* If SET_DEST changed from a REG to a MEM and INSN is an insn,
           write a CLOBBER insn.  */
        if (GET_CODE (SET_DEST (x)) == REG && GET_CODE (new0) == MEM
            && insn != 0 && GET_CODE (insn) != EXPR_LIST
            && GET_CODE (insn) != INSN_LIST)
          emit_insn_after (gen_rtx_CLOBBER (VOIDmode, SET_DEST (x)), insn);

This appears to emit CLOBBERs for some pseudos that didn't get hard registers.
I think I remember that there was code in reload.c that did something similar,
but I can't find it right now.


Bernd