Fix twolf -funroll-loops -O3 miscompilation (a semi-latent web.c bug)

[Paolo Bonzini]

Il 16/10/2012 12:35, Steven Bosscher ha scritto:
> On Tue, Oct 16, 2012 at 12:15 PM, Steven Bosscher wrote:
>> On Mon, Oct 15, 2012 at 3:26 PM, Steven Bosscher wrote:
>>> On Mon, Oct 15, 2012 at 3:21 PM, Paolo Bonzini wrote:
>>>> Il 15/10/2012 14:53, Steven Bosscher ha scritto:
>>>>> I think I've shown above that we're all looking at the wrong pass...
>>>>
>>>> I think you have... so we want a patch like this?
>>>
>>> I don't think so. df_kill_notes is already supposed to take care of this.
>>
>> But it doesn't because if the SET_DEST of an insn is the same as the
>> register dieing in the insn's REG_EQUAL note, the reg is live at the
>> end of the insn and so the note stays:
>>
>> Breakpoint 2, df_kill_notes (insn=0x7ffff5e3e7e0, live=0x7fffffffda90)
>> at ../../trunk/gcc/df-problems.c:2833
>> 2833      rtx *pprev = &REG_NOTES (insn);
>> 1: debug_rtx(insn) = (insn 79 50 52 8 (set (reg:DI 72 [ ivtmp.17D.1758 ])
>>         (reg:DI 103 [ ivtmp.17D.1758 ])) -1
>>      (expr_list:REG_DEAD (reg:DI 103 [ ivtmp.17D.1758 ])
>>         (expr_list:REG_EQUAL (plus:DI (reg:DI 72 [ ivtmp.17D.1758 ])
>>                 (const_int 24 [0x18]))
>>             (nil))))
>> void
>> (gdb) undisp 1
>> (gdb) p debug_bitmap(live)
>>
>> first = 0x1627200 current = 0x1627200 indx = 0
>>         0x1627200 next = (nil) prev = (nil) indx = 0
>>                 bits = { 6 7 16 20 72 82 85 87 }
>> $2 = void
>> (gdb)
>>
>>
>> So GCC should be looking at whether the reg in the REG_EQUAL note is
>> dead *before* the insn.
>>
>> Bottom line: This is a bug in df_kill_notes.

Yep, and it could cause wrong code generation, couldn't it?  Because the
new (reg:DI 72) is obviously not equal to (plus:DI (reg:DI 72)
(const_int 24)), but it could be propagated to subsequent insns.

So I think this patch should be backported to all release branches after
regstrapping.

> I think this should fix it. Can't test it right now, so help
> appreciated (Honza: hint hint! ;-)

Ok after bootstrap, regtest and checking that it actually fixes the bug.

Paolo