make bootstrap segfaults on current cvs head (i686-pc-linux-gnu)

[Zack Weinberg]

On Fri, Oct 13, 2000 at 11:59:20PM +0100, Joseph S. Myers wrote:
> On Fri, 13 Oct 2000, Alan Modra wrote:
> 
> > On Fri, 13 Oct 2000, Mike Coleman wrote:
> > 
> > > I get a segfault when trying to build gcc from the current cvs
> > > head.
> > 
> > So do I.  This might help.
> > 
> > gcc/ChangeLog
> > * emit_rtl.c (init_emit_once): Create const_int_htab before using.
> 
> I think there's more to the problem than this.  gen_rtx_CONST_INT is at
> this point only called with values that shouldn't go to const_int_htab -
> arg == 0 at the crash.  It seems that

Here's an assembly dump.  The problematic sequence is starred.

gen_rtx_CONST_INT:
	push	%ebp
	mov	%esp, %ebp
	push	%ebx
	sub	$4, %esp
	mov	10(%ebp), %edx		; arg in edx
*	lea	64(%edx), %eax		; arg+MAX_SAVED_CONST_INT in eax
*	add	$-128, %eax		; subtract 2*MAX_SAVED_CONST_INT
*	ja	.L2			; if positive, not in const_int_rtx[]
	mov	const_int_rtx(%edx,4), %eax	; else fetch
.L1:
	mov	-4(%ebp), %ebx
	mov	%ebp, %esp
	pop	%ebp
	ret
.L2:
	<code to look stuff up in const_int_htab>

We're called with arg=0.  That becomes 64 in %eax, and then we
subtract 128 from that, giving -64.  The condition codes are set
'according to the result' - in this case, that means SF on, all others
off.  JA jumps if "above: CF=0 and ZF=0", which both are.

That add needs to be a compare, instead.  I think we decided to
convert <sub $128, %eax> into <add $-128, %eax> because it's smaller
that way (+128 doesn't fit into a signed byte, -128 does) without
considering that add and sub do different things to the flags.  From
the debugger, I adjusted the add instruction into a compare against
+127, which fits in the same 3 bytes, and cc1 successfully finished
processing tmp-dum.c.

zw