From patchwork Thu Jul  1 03:17:48 2010
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: canonicalize_condition vs zero_extract
From: DJ Delorie <dj@redhat.com>
X-Patchwork-Id: 57462
Message-Id: <201007010317.o613HmhD016051@greed.delorie.com>
To: gcc-patches@gcc.gnu.org
Date: Wed, 30 Jun 2010 23:17:48 -0400

If you have an insn that does "test <const>,<reg>", gcc will look for
a pattern that uses zero_extract to generate the "condition" when
<const> is a power of two.  However, canincalize_condition gets passed
this condition, and only expects the RTL at this point to have two
subexpressions.  Zero_extract has three, so it blows up:

  return gen_rtx_fmt_ee (code, VOIDmode, op0, op1);

This is a simple check for this case, but is it the right solution?
Would it be better to:

* Check for a generic "this rtl has more than 2 xexp's"?
* Check for the codes we know, and return 0 for the rest?
* Find out why a zero_extract is getting this far in the first place?


Index: rtlanal.c
===================================================================
--- rtlanal.c      (revision 161641)
+++ rtlanal.c   (working copy)
@@ -4699,12 +4699,15 @@ canonicalize_condition (rtx insn, rtx co
 
   code = GET_CODE (cond);
   mode = GET_MODE (cond);
   op0 = XEXP (cond, 0);
   op1 = XEXP (cond, 1);
 
+  if (code == ZERO_EXTRACT)
+    return 0;
+
   if (reverse)
     code = reversed_comparison_code (cond, insn);
   if (code == UNKNOWN)
     return 0;
 
   if (earliest)