Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!bloom-beacon!bionet!ames!amdcad!weitek!dms!albaugh
From: albaugh@dms.UUCP (Mike Albaugh)
Newsgroups: gnu.gcc
Subject: Dumb questions on GCC
Keywords: define_insn, constraints
Message-ID: <616@dms.UUCP>
Date: 8 Feb 89 01:23:45 GMT
Organization: Atari Games Inc., Milpitas, CA
Lines: 51


	I am trying to port gcc. Since the only system I have
capable of compiling gcc is a vms vax, I am currently stuck at
version 1.22 (no flames please, unless accompanied by a check
sufficient to buy me a "real unix (tm)" machine). Anyway, I finally
got it to compile without errors, but now it aborts if I feed it
hello.c (of course that's the first program I tried :-), but not the
last). The abort is _reported_ in _fixuns_truncdfsi (or something of the
like) in gcclib, but appears to _really_ emanate from insn_extract()
called from regclass(). Tracing back a ways, this all seems to stem from
trying to figure out the "best class" of register for a given insn.
Since I don't have any floating point hardware, I don't have any define_insn's
for floating point. I presumed (apparently incorrectly) that gcc would
notice this and use libcalls. Anyway, the whole subject of "register class"
seems to key on the #define REGISTER_CONSTRAINTS which is generated by
gen-config when it sees _any_ constraint whatsoever (not just register
constraints). Since my machine is a load/store architecture, I can't
define movesi without _some_ constraints. I'm using something like:

	"general_operand" "=r,m" and
	"general_operand" "rmI,r"

and picking the right template. Note that there are no distinct flavors
of register, but at least one of the operands must be _some_ register.
I might be able to re-write most of my md without constraints, but some
of the pre-defined ones baffle me. While I'm at it, has anyone sleuthed
out a list of the required pre-defined insns. By required I mean you
either don't compile or dump core if you don't have all of them. The final
fix needed to get an error-free compile was to define "call_value",
which is not even mentioned as being defined, let alone essential.

	Wierd piece of evidence - this does not happen if I compile:

char *fred() { return "hello world\n"; }

so it seems this path is only exercised when a subroutine call is processed.
I know this is hardly an exhaustive test, but I found it interesting.

	Please don't think I'm just grousing. I like the structure and
much of the rest of gcc. I just need a little help across the more
obscure points. I apologize for the hand-waving in some of the above,
but I'm telnet'd two deep to get to a machine that can post news and
don't have real good access to my sources from here.

Thanks in advance for any pointers.
					Mike

| Mike Albaugh (albaugh@dms.UUCP || {...decwrl!turtlevax!}weitek!dms!albaugh)
| Atari Games Corp (Arcade Games, no relation to the makers of the ST)
| 675 Sycamore Dr. Milpitas, CA 95035		voice: (408)434-1709
| The opinions expressed are my own (Boy, are they ever)