Re: IA64 HP-UX ILP32 mode patch

[Steve Ellcey <sje at cup dot hp dot com>]

Thanks for the feedback, I'll work on modifying the patch to see if I
can make it better.  Addressing some of the points:

> You don't seem to actually define MD_CONVERT_MEMORY_ADDRESS anywhere.
>
> - Geoffrey Keating <geoffk@geoffk.org>

> But you didn't document it.  Patches need documentation.  Is this called
> somewhere performance-sensitive, or should you be using the target
> structure instead?
>
> Joseph S. Myers

Both of these are right, I forgot the documentation and the definition
got lost from the patch, though both of these may be moot if I modify
convert_memory_address based on POINTERS_EXTEND_UNSIGNED being less then
zero instead of having a hook.

Moving on to Richards comments:

> FWIW, I'd rather this be called ptr_extend, to match
> sign_extend and zero_extend.

This is easy to do, I will make the change.

> Use gr_register_operand, not register_operand.

I fixed ptr_extend and added the new instructions as well but the new
instructions caused a couple of new optimization failures that I don't
yet understand.  I need to figure out what is happening with this and
fix it.

> I see two changes in your version.  One, you check cfun nonnull,
> which seems pointless to me -- why wouldn't we have run into this
> on other targets?  Two, you remove support for CONST, PLUS, and MULT.
> What examples lead you to this?  If it's really unavoidable, perhaps
> you should simply add POINTERS_EXTEND_UNSIGNED >= 0 to this cases.

If I build gcc with the standard convert_memory_address and compile the
following program:

	foo() { char * s = "hi"; }

I get:

	(insn 9 17 10 (set (reg/f:SI 340)
        	(symbol_ref/f:SI ("*.LC0"))) -1 (nil)
    	(expr_list:REG_EQUAL (symbol_ref/f:SI ("*.LC0"))
        	(nil)))

	bug1.c:1: Internal compiler error in extract_insn, at recog.c:2210

My thought was that I didn't want symbol_ref's to be in mode SI because
I always need to use a DI mode to access a symbol.  It looks like I
originally just took out the SYMBOL_REF and LABEL_REF cases completely
but I think I had to put them back in to deal with some global
initialization cases.  I just tried it again and with the SYMBOL_REF and
LABEL_REF cases completely gone I was able to compile and run the test
suite with no problems.  With the standard SYMBOL_REF and LABEL_REF
cases I get the error shown above.

Its possible I could just say:

	if (POINTERS_EXTEND_UNSIGNED < 0)
		return convert_modes (to_mode, from_mode, x,
				      POINTERS_EXTEND_UNSIGNED);

before the case statement.  I will experiment with this some more
and submit a new patch.

Steve Ellcey
sje@cup.hp.com