Re: gcc vax bug in built-in ffs function/displacement indexed addressing

["Maciej W. Rozycki" <macro%linux-mips.org@localhost>]

On Fri, 6 Jan 2017, Paul Koning wrote:

> >> My VAX architecture handbook says that the operand type of that 
> >> operand is "$
> > 
> > Doh!  So does mine.
> > 
> >              opcode startpos.rl, size.rb, base.vb, {field.rv}, findpos.wl
> > 
> > I can only infer I was looking at the implicit field.rv operand.

 The {field.rv} specifier appears to be a notational convention only 
standing for the three first operands, and does not appear as an actual 
instruction operand in assembly or disassembly.  The document is lacking a 
description of this convention or at least I cannot find it either.

> >> So yes, that looks like a wrong code bug.  Perhaps the code generator 
> >> just h$
> > 
> > If so, it's probably worth checking the other bitfield instructions
> > too.
> 
> >From a quick look at the VAX stuff in gcc/config/vax, I see this in 
> >builtins.md:
> 
> (define_insn "ffssi2_internal"
>   [(set (match_operand:SI 0 "nonimmediate_operand" "=rQ")
> 	(ffs:SI (match_operand:SI 1 "general_operand" "nrmT")))
>    (set (cc0) (match_dup 0))]
>   ""
>   "ffs $0,$32,%1,%0")
> 
> I believe the issue is that the check for "can we use indexing" (in 
> vax.c) assumes that the answer is yes if the offset is a value 
> multiplied by sizeof(operand).  In this case, I assume that the operand 
> is a 32 bit integer.  The trouble is that FFS operates on values up to 
> 32 bits but indexes as if it were a byte operand.

 The operand size is specified by its mode being SI here, that is a 32-bit 
integer indeed.

> I'm not sure if constraint "Q" means what I hope it does, but if yes, 
> the fix might be to replace the nrmT in that second match_operand by 
> nrQT.

 Offhand the predicate might need to be changed accordingly as well so 
that it only passes through a memory operand expression which will later 
match one of the constraints given.  NB these days it can be easily done 
with the use of `satisfies_constraint_<name>' functions, where <name> is 
the constraint intended, so that code used for constraint matching does 
not have to be replicated in predicates.

  Maciej