Bug in simplify_subreg?

Bug in simplify_subreg?

[Lars Brinkhoff]

In simplify-rtx.c:simplify_subreg, there is some code to simplify a
SUBREG of a hard register (starting around line 2557).  This code is
careful not to make the simplification if the register is the stack,
frame, or argument pointer.  However, if the (inner)mode is larger
than SImode, no check is made to ensure that the simplification isn't
made such that the result is one of the above pointers.

I have a situation where the second word of a DI register is allocated
to the frame pointer register (the function doesn't need a frame
pointer).  In this case, the ORIGINAL_REGNO of the frame_pointer_rtx
will be clobbered with the old pseudo REGNO (around line 2597), which
later will be referenced, causing a crash.

If I insert the same checks against final_regno as are made to
REGNO (op), everything seems to be fine.  Is this the right thing
to do?  If so, I'll test and submit a patch.

-- 
Lars Brinkhoff          http://lars.nocrew.org/     Linux, GCC, PDP-10,
Brinkhoff Consulting    http://www.brinkhoff.se/    HTTP programming

Re: Bug in simplify_subreg?

[Richard Henderson]

On Fri, Aug 02, 2002 at 01:02:26PM +0200, Lars Brinkhoff wrote:
> If I insert the same checks against final_regno as are made to
> REGNO (op), everything seems to be fine.  Is this the right thing
> to do?

No.  If the frame pointer isn't being used, then
gen_reg_RTX isn't supposed to return frame_pointer_rtx.


r~

Re: Bug in simplify_subreg?

[Lars Brinkhoff]

Richard Henderson <rth@redhat.com> writes:
> On Fri, Aug 02, 2002 at 01:02:26PM +0200, Lars Brinkhoff wrote:
> > If I insert the same checks against final_regno as are made to
> > REGNO (op), everything seems to be fine.  Is this the right thing
> > to do?
> No.  If the frame pointer isn't being used, then gen_reg_RTX isn't
> supposed to return frame_pointer_rtx.

But it does.  Since outermode==SImode==Pmode and reload_in_progress==0,
gen_reg_RTX returns frame_pointer_rtx to simplify_subreg.  And so
ORIGINAL_REGNO of frame_pointer_rtx is clobbered.

Would the right solution be to check frame_pointer_needed in gen_reg_RTX?

-- 
Lars Brinkhoff          http://lars.nocrew.org/     Linux, GCC, PDP-10,
Brinkhoff Consulting    http://www.brinkhoff.se/    HTTP programming

Re: Bug in simplify_subreg?

[Richard Earnshaw]

> Richard Henderson <rth@redhat.com> writes:
> > On Fri, Aug 02, 2002 at 01:02:26PM +0200, Lars Brinkhoff wrote:
> > > If I insert the same checks against final_regno as are made to
> > > REGNO (op), everything seems to be fine.  Is this the right thing
> > > to do?
> > No.  If the frame pointer isn't being used, then gen_reg_RTX isn't
> > supposed to return frame_pointer_rtx.
> 
> But it does.  Since outermode==SImode==Pmode and reload_in_progress==0,
> gen_reg_RTX returns frame_pointer_rtx to simplify_subreg.  And so
> ORIGINAL_REGNO of frame_pointer_rtx is clobbered.
> 
> Would the right solution be to check frame_pointer_needed in gen_reg_RTX?

We cannot know for certain whether a frame pointer is needed before reload 
starts.  Even during reload the decision might change -- on some machines 
the decision may depend on which registers are being used by the function 
-- and that can change as we push more reloads.

It's probable that on many machines we can accurately guess whether a 
frame pointer is needed or not before reload starts to run, but there's no 
way to be sure.

R.

Re: Bug in simplify_subreg?

[Lars Brinkhoff]

Richard Earnshaw <rearnsha@arm.com> writes:
> > Since outermode==SImode==Pmode and reload_in_progress==0,
> > gen_reg_RTX returns frame_pointer_rtx to simplify_subreg.  And so
> > ORIGINAL_REGNO of frame_pointer_rtx is clobbered.
> It's probable that on many machines we can accurately guess whether a 
> frame pointer is needed or not before reload starts to run, but there's no 
> way to be sure.

Ok, then how about doing
    ORIGINAL_REGNO (frame_pointer_rtx) = FRAME_POINTER_REGNUM;
after compilation of every function?

Any other suggestions?

-- 
Lars Brinkhoff          http://lars.nocrew.org/     Linux, GCC, PDP-10,
Brinkhoff Consulting    http://www.brinkhoff.se/    HTTP programming

Re: Bug in simplify_subreg?

[Richard Henderson]

On Mon, Aug 05, 2002 at 02:43:06PM +0200, Lars Brinkhoff wrote:
> Any other suggestions?

gen_reg_RTX should check reload_compled && frame_pointer_needed.


r~

Re: Bug in simplify_subreg?

[Richard Earnshaw]

> On Mon, Aug 05, 2002 at 02:43:06PM +0200, Lars Brinkhoff wrote:
> > Any other suggestions?
> 
> gen_reg_RTX should check reload_compled && frame_pointer_needed.
> 
> 
> r~

After reload we should certainly not be returning frame_pointer_rtx on the 
ARM, or any other platform that defines HARD_FRAME_POINTER_REGNUM.

R.