Re: middle-end/7245: ICE in find_reloads

[Michael Matz <matz@suse.de>]

The following reply was made to PR middle-end/7245; it has been noted by GNATS.

From: Michael Matz <matz@suse.de>
To: Andreas Jaeger <aj@suse.de>
Cc: <gcc-gnats@gcc.gnu.org>, Ulrich Hecht <uli@suse.de>
Subject: Re: middle-end/7245: ICE in find_reloads
Date: Tue, 9 Jul 2002 21:47:36 +0200 (CEST)

 Again for the bugtracking system.  Manually constructed answer mail ;(
 
 > (insn:QI 61 60 11 (parallel[
 >             (set (reg:CCZ 17 flags)
 >                 (compare:CCZ (ashiftrt:SI (reg/v:SI 0 eax [62])
 >                         (const_int -8 [0xfffffff8]))
 >                     (const_int 0 [0x0])))
 >             (set (reg/v:SI 1 edx [60])
 >                 (ashiftrt:SI (reg/v:SI 0 eax [62])
 >                     (const_int -8 [0xfffffff8])))
 >         ] ) 434 {*ashrsi3_cmp} (nil)
 
 The problem is, that this insn doesn't fulfill it's constraint on operand
 2, which is the -8.  It needs to be an immediate_operand() (which it is),
 with constraint "I", which means >=0 and <= 31, which it isn't.
 Constraints are first checked in reload, so only there we see, that the
 whole selection of the pattern shouldn't have happened.  I.e. I think in
 this case either the predicate immediate_operand is wrong, and shouldn't
 accept values <0, or we need another letter constraint for a range
 including some values <0.
 
 I'm not sure anymore, what happens on x86, if the shift-count is negative
 in the insn.  It's ANDed with 31, right?  In that case, we _might_ accept
 negative values in the constraint.
 
 The problem here is, that this operand can't be reloaded, so both, the
 constraint and the predicate should really match.
 
 
 Ciao,
 Michael.