Re: [again] switch() statements on ARM/Thumb

[Richard Earnshaw <rearnsha at arm dot com>]

> On Wed, 2002-01-09 at 12:51, Richard Earnshaw wrote:
> > > +; this works only if the label_ref is close enough (10 bit word aligned)
> > > +; but it should only be called from thumbpic_casesi anyway.
> > > +(define_insn "*thumb_adr"
> > > +  [(set (match_operand:SI 0 "s_register_operand" "=l")
> > > +        (label_ref (match_operand 1 "" "")))]
> > > +  "TARGET_THUMB"
> > > +  "adr\\t%0, %a1"
> > > +  [(set_attr "length" "2")]
> > > +)
> > > +
> > 
> > I'm very nervous about this.  In particular the assertion that this will 
> > only be generated by thumbpic_casesi and so will always be in range.  
> > Experience has shown that combine and CSE are very good at making use of 
> > such patterns in ways that are unexpected and that then lead to compiler 
> > aborts or invalid assembly code.
> 
> As I said this is the part I am least happy with, too. I'm not convinced
> that the (set (match_operand:SI 0 "s_register_operand" "=l") (label_ref
> (match_operand 1 "" ""))) construct is even legal, but I could not find
> the Official Way of loading the (pc relaive) address of a code label
> into a register.
> 
> Alternatively I could make this a named insn and emit it manually from
> the C code part of the casesi, but I though doing it this way would
> allow the optimizer to play around with it some more.
> 
> > I can't prove that it won't work, but 
> > have you run the entire testsuite (in thumb mode, of course) and shown 
> > that this doesn't cause regressions?
> 

The safest way to do it would be to use an unspec.   Create a constant, 
say THUMB_SWITCH_TABLE_ADDR at the top of the MD file, then define your 
insn as

(define_insn "*thumb_adr"
  [(set (match_operand:SI 0 "s_register_operand" "=l")
	(unspec [(label_ref (match_operand 1 "" "")))] 
		THUMB_SWITCH_TABLE_ADDR))]
  ...

Then combine and CSE will never create anything that will match this by 
accident.