grokfield miscompiled with -O9

["Martin v. Loewis" <martin at loewis dot home dot cs dot tu-berlin dot de>]

> No.  Someone should sit down and figure out what the actual problem
> is.

Ok. Here is as far as I got. grokfield gets miscompiled under certain
circumstances, passing incorrect (out-of-range) decl_context values to
grokdeclarator. I've trimmed-down the source of c-decl.c to a minimum
size. I also attach the corresponding assembler code, obtained with
-O9, on i586-pc-linux-gnu, with 2.96 19991222.

Please note that the fragment

  push_obstacks_nochange ();
  value = grokdeclarator (declarator, declspecs, width ? BITFIELD : FIELD, 0);

is compiled to

	call	push_obstacks_nochange
	testl	%edi, %edi
	setne	%al
	pushl	$0
	addl	$3, %eax
	pushl	%eax
	pushl	%ebx
	pushl	%esi
	call	grokdeclarator

The optimization being done here is that FIELD+1 == BITFIELD, so if
edi (==width) is not null, 3+1 should be pushed, otherwise
3. Unfortunately, the high part of eax is not cleared, so it pushes
garbage. Under slight modification of the source, it produces the
fragment

	pushl	$0
	andl	$255, %eax
	addl	$3, %eax

instead, which is correct.

I could not convert this into a test case, because calling grokfield
from main results in inlining of grokfield (at -O9), which does not
trigger the bug.

If you need any further information, please let me know.

Regards,
Martin

bug.c

bug.s