Re: RFA: recognize new instructions in rs6000 prologues

[Kevin Buettner <kevinb@redhat.com>]

On 26 Mar 2004 11:56:14 -0500
Jim Blandy <jimb@redhat.com> wrote:

> Kevin Buettner <kevinb@redhat.com> writes:
> > On 24 Mar 2004 10:10:04 -0500
> > Jim Blandy <jimb@redhat.com> wrote:
> > 
> > > 
> > > These prologues are generated by a not-yet-released compiler, but the
> > > test suite does catch the problem.
> > > 
> > > 2004-02-25  Jim Blandy  <jimb@redhat.com>
> > > 
> > > 	* rs6000-tdep.c (skip_prologue): Recognize moves from argument
> > > 	registers to temp register r0 and byte stores as prologue
> > > 	instructions.
> > > 
> > > *** gdb/rs6000-tdep.c.~2~	2004-02-25 15:14:13.000000000 -0500
> > > --- gdb/rs6000-tdep.c	2004-02-25 15:15:43.000000000 -0500
> > > ***************
> > > *** 772,777 ****
> > > --- 772,785 ----
> > >   
> > >   	  /* store parameters in stack */
> > >   	}
> > > +       /* Move parameters from argument registers to temporary register.  */
> > > +       else if ((op & 0xfc0007fe) == 0x7c000378 &&	/* mr(.)  Rx,Ry */
> > > +                (((op >> 21) & 31) >= 3) &&              /* R3 >= Ry >= R10 */
> > > +                (((op >> 21) & 31) <= 10) &&
> > > +                (((op >> 16) & 31) == 0)) /* Rx: scratch register r0 */
> > > +         {
> > > +           continue;
> > > +         }
> > 
> > Is this case really needed?  I would have thought that the catchall
> > case at the end would handle this situation.  I'm concerned that
> > adding this case may cause us to overshoot the prologue in some
> > circumstances.  (Of course, there's a danger of doing that anyway...)
> 
> Here's the code for the function in question:
> 
> 	.align 2
> 	.globl arg_passing_test2
> 	.type	arg_passing_test2, @function
> arg_passing_test2:
> .LFB107:
> 	.loc 1 62 0
> 	stwu 1,-64(1)
> .LCFI11:
> 	stw 31,60(1)
> .LCFI12:
> 	mr 31,1
> .LCFI13:
> 	mr 0,3
> 	evstdd 4,16(31)
> 	stw 5,24(31)
> 	stw 7,32(31)
> 	stw 8,36(31)
> 	stw 9,40(31)
> 	stb 0,8(31)
> 	lwz 11,0(1)
> 	lwz 31,-4(11)
> 	mr 1,11
> 	blr
> .LFE107:
> 	.size	arg_passing_test2, .-arg_passing_test2
> 
> The first 'lwz' is the first non-prologue instruction; the stores
> above it are argument spills.
> 
> You're referring to the catchall case that says that we scan past
> unfamiliar, non-branch instructions unless the frame is already set
> up, right?  This function leaves its return address in LR, so there's
> no need to wait for a return address save instruction.  And the frame
> pointer is set up by the time we see the 'mr 0,3' instruction.  So
> that catch-all case doesn't apply here.
> 
> According to the ABI, upon function entry, r0 is caller-saves, and not
> used for passing arguments, so its dead; a move into r0 can't destroy
> information.  Moving an argument register into it could at worst be an
> initialization of a variable that would be done sooner than expected.
> 
> But missing those spills is a serious problem.  Most users don't
> consider their frame 'set up' if GDB displays the arguments as
> garbage.  :)

Okay, I'm convinced.  This part is approved too.

Kevin