Re: PATCH [1/n]: Prepare x32: PR middle-end/47364: internal compiler error: in emit_move_insn, at expr.c:3355

["H.J. Lu" <hjl.tools@gmail.com>]

On Sun, Jun 12, 2011 at 7:00 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Sun, Jun 12, 2011 at 6:50 AM, Richard Guenther
> <richard.guenther@gmail.com> wrote:
>> On Sun, Jun 12, 2011 at 3:18 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>>> On Sun, Jun 12, 2011 at 3:48 AM, Richard Guenther
>>> <richard.guenther@gmail.com> wrote:
>>>> On Sat, Jun 11, 2011 at 5:09 PM, H.J. Lu <hongjiu.lu@intel.com> wrote:
>>>>> Hi,
>>>>>
>>>>> expand_builtin_strlen has
>>>>>
>>>>> src_reg = gen_reg_rtx (Pmode);
>>>>> ...
>>>>> pat = expand_expr (src, src_reg, ptr_mode, EXPAND_NORMAL);
>>>>> if (pat != src_reg)
>>>>>  emit_move_insn (src_reg, pat);
>>>>>
>>>>> But src_reg may be in ptr_mode, wich may not be the same as Pmode.
>>>>> This patch checks it.  OK for trunk?
>>>>>
>>>>> Thanks.
>>>>>
>>>>>
>>>>> H.J.
>>>>> ---
>>>>> 2011-06-11  H.J. Lu  <hongjiu.lu@intel.com>
>>>>>
>>>>>        PR middle-end/47364
>>>>>        * builtins.c (expand_builtin_strlen): Properly handle target
>>>>>        not in Pmode.
>>>>>
>>>>> diff --git a/gcc/builtins.c b/gcc/builtins.c
>>>>> index 7b24a0c..4e2cf31 100644
>>>>> --- a/gcc/builtins.c
>>>>> +++ b/gcc/builtins.c
>>>>> @@ -2941,7 +2941,11 @@ expand_builtin_strlen (tree exp, rtx target,
>>>>>       start_sequence ();
>>>>>       pat = expand_expr (src, src_reg, ptr_mode, EXPAND_NORMAL);
>>>>>       if (pat != src_reg)
>>>>> -       emit_move_insn (src_reg, pat);
>>>>> +       {
>>>>> +         if (GET_MODE (pat) != Pmode)
>>>>> +           pat = convert_to_mode (Pmode, pat, 1);
>>>>
>>>> Shouldn't this be POINTERS_EXTEND_UNSIGNED instead of 1?
>>>>
>>>>> +         emit_move_insn (src_reg, pat);
>>>>
>>>> Why not use convert_move unconditionally?
>>>>
>>>> Or, why not expand src in Pmode from the start?  After all, src_reg is
>>>> created as Pmode reg.
>>>>
>>>
>>> This patch works for my testcase.  OK for trunk?
>>
>> Ok if it passes bootstrap & regtest on a ptr_mode != Pmode target.
>>
>
> Only the following targets expand strlen:
>
> avr/avr.md:(define_expand "strlenhi"
> avr/avr.md:(define_insn "*strlenhi"
> i386/i386.md:(define_expand "strlen<mode>"
> i386/i386.md: if (ix86_expand_strlen (operands[0], operands[1],
> operands[2], operands[3]))
> i386/i386.md:(define_expand "strlenqi_1"
> i386/i386.md:(define_insn "*strlenqi_1"
> rs6000/rs6000.md:(define_expand "strlensi"
> s390/s390.md:; strlenM instruction pattern(s).
> s390/s390.md:(define_expand "strlen<mode>"
> s390/s390.md:(define_insn "*strlen<mode>"
>
> None of them, except for my x32 port, are ptr_mode != Pmode targets.
> I will bootstrap and test it on my x32 branch.
>

It doesn't work on x32. I got

/export/gnu/import/git/gcc-x32/libssp/gets-chk.c:74:14: internal
compiler error: in emit_move_insn, at expr.c:3319
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

How about this patch?

Thanks.

-- 
H.J.
---
2011-06-12  H.J. Lu  <hongjiu.lu@intel.com>

	PR middle-end/47364
	* builtins.c (expand_builtin_strlen): Properly handle target
	not in Pmode.

diff --git a/gcc/builtins.c b/gcc/builtins.c
index 7b24a0c..a2f175d 100644
--- a/gcc/builtins.c
+++ b/gcc/builtins.c
@@ -2941,7 +2941,14 @@ expand_builtin_strlen (tree exp, rtx target,
       start_sequence ();
       pat = expand_expr (src, src_reg, ptr_mode, EXPAND_NORMAL);
       if (pat != src_reg)
-	emit_move_insn (src_reg, pat);
+	{
+#ifdef POINTERS_EXTEND_UNSIGNED
+	  if (GET_MODE (pat) != Pmode)
+	    pat = convert_to_mode (Pmode, pat,
+				   POINTERS_EXTEND_UNSIGNED);
+#endif
+	  emit_move_insn (src_reg, pat);
+	}
       pat = get_insns ();
       end_sequence ();