[PATCH, arm] Fix gcc.c-torture/execute/930628-1.c -O3 for Thumb

[Richard Earnshaw]

> On Wed, 30 Oct 2002 10:35:18 +0000
> Richard Earnshaw <rearnsha@arm.com> wrote:
> 
> > > Hi,
> > > 
> > > The testcase gcc.c-torture/execute/930628-1.c -O3 -funroll-loops
> > > fails to compile on Thumb.  The problem is that GCC generates
> > > instunctions like `str r6,[r2,sp]' which are not valid address
> > > combination on Thumb.
> > > 
> > > SP is a valid base register but only with an immediate index.  Since
> > > THUMB_REGNO_MODE_OK_FOR_BASE_P yields true for SP,
> > > find_reloads_address_1 does not reload this register.  The only
> > > solution I found was to manually reload either SP or the entire
> > > address in LEGITIMIZE_RELOAD_ADDRESS.
> > > 
> > > I reg-tested the patch with arm-sim/-mthumb.  The two failures with
> > > 930628-1.c have gone away.
> > > 
> > I've been thinking about this particular case, but the fix is quite 
> > complex and involves introducing a new macro that allows the index reg
> > class to be based both on the mode of access and the class of base 
> > register in use.  Unfortunately I don't have a patch available for
> > this yet.
> 
> Is it possible just to avoid using sp ?
> 
> ~velco

Not really, not using the stack pointer register at all would result in 
woefully inefficient code if we ever have to spill registers.

R.