Re: Re:

[Dale Johannesen <dalej@apple.com>]

On Sep 24, 2004, at 6:24 PM, Devang Patel wrote:
> On Sep 24, 2004, at 5:58 PM, Dale Johannesen wrote:
>> I'm tracking down a crash I'm getting locally.  It's not clear yet 
>> this is connected,
>> but I see a tree of this form in the Gimple and SSA dumps:
>>
>>   *pTmp4<D6472> = __NAGf90_dcdivdc (*pTmp2<D6470>, *pTmp3<D6471>);
>>
>> The Gimple doc does not allow a pointer deref in a function argument.
>> Is is supposed to?
>
> It says CALL_EXPR arguments are valid LHS. So it allows. I am reading 
> comment at the beginning of tree-gimple.c

Thanks, I was looking at tree-ssa.texi:

call-stmt: CALL_EXPR
   op0 -> _DECL | '&' _DECL
   op1 -> arglist
arglist:
   NULL_TREE
   | TREE_LIST
       op0 -> val
       op1 -> arglist
val : _DECL | CONST

There seem to be a number of differences.   Apparently the 
tree-ssa.texi version is older, as it's
closer (but not identical) to the summit paper, but some nits have 
crept in.   The version in
tree-gimple.c doesn't seem to allow ADDR_EXPR, but clearly it has to:

  __NAGf90_write (6, &FMT_129<D6540>, -1, 0B, 0, 0);

IMO it is highly desirable for future development to have a clean 
definition.  I expect not everyone
shares MO, so I'll volunteer to clean this up, but I'll need help 
determining what it should say.  I propose to
put a single copy of it one place or the other (I'd prefer the texi 
file), with a pointer to the master copy
in the other place.  Comments?