i386.md bug + fix

[Jim Wilson]

	First of all some code taken from the output of compiling the
	function makeUtf8Const in kaffe/kaffevm/string.s using just
	the "-O -mcpu=pentium"---this code WORKS:

You should submit a testcase for this.  We can't fix the bug without a
testcase.

	(define_insn ""
	  [(set (match_operand:SI 0 "register_operand" "=&c")
	        (unspec:SI [(mem:BLK (match_operand:SI 1 "address_operand" "D"))
	                    (match_operand:QI 2 "immediate_operand" "a")
	                    (match_operand:SI 3 "immediate_operand" "i")] 0))
	   (clobber (match_dup 1))]

	  Where it goes wrong is, I beleive, that
	operand 2 is never actually used within any RTL generated by the
	define_expand.

Operand2 is clearly used by the UNSPEC rtl.  Adding a second use inside a
USE rtl does not make any sense.  I don't understand how that could be
in any way useful.

However, it is possible that changing the immediate_operand predicate to
a register_operand predicate did something useful.  I don't understand
why this fixes the problem though.  The real bug may be elsewhere, which
is why I would like to see a testcase.

I do see one obvious bug looking at this pattern.  Operand 1 is an
input/output operand, but is missing the required + constraint.
It is possible that this bug may be causing part of the trouble that
you are seeing.  I can't tell without a testcase though.

Jim