Hi, Sun
Would you please kindly furthur input to this issue? If we could not
find alternative better fix, can we check in the patch firstly? the
regression bugs are marked major priority, we should get it fixed quickly.
Thanks
Regards
Gang
On Thu, Dec 15, 2011 at 10:19 AM, Gang Yu <yugang...@gmail.com> wrote:
> Thanks for your comments.
> On Wed, Dec 14, 2011 at 10:06 PM, Sun Chan <sun.c...@gmail.com> wrote:
>
>> I don't understand the rationale for this fix. You are saying that the
>> dead branch is not dce'd, and lno does not like it. So, why not fix dce? A
>> dead branch is dead branch, why this workaround?
>> Sun
>>
>>
> This segmentation is in the preopt, but intermediate optimized result as:
> IF {line: 1/131}
> U4INTCONST 1 (0x1)
> THEN
> ...
> ELSE
> BLOCK
> END_BLOCK
> END_IF
> is acceptable, not 100% optimized but still in its own right and later
> mainopt can still eliminate the dead branch.
>
> for a long term, we should fix proactive_loop_transform, since we should
> not always assume the intermediate input in perfect form.
>
> Coding style in open64.net pointing this out:
>
> no assumption across
>
> - No assumption across different phases
>
>
> As a temporary workaround, we assume the previously code work quite well
> except for cases like bug912, so, we make more effort on limit the changes,
> i.e, more constraint on the change. if we add the MAINOPT condition, the
> branch get eliminted by later DCE in preopt, then proactive_loop_transform
> happy. we get things done.
>
> Regards
> Gang
>
>
>> On Wed, Dec 14, 2011 at 12:34 PM, Gang Yu <yugang...@gmail.com> wrote:
>>
>>> Hi,
>>>
>>> Would a gatekeeper please help review the fix for bug934 , regression
>>> caused by r3843?
>>>
>>> segment-fault function snippet of bug case:
>>>
>>> void LibRaw:: tiff_set (ushort *ntag,
>>> ushort tag, ushort type, int count, int val)
>>> {
>>> struct tiff_tag *tt;
>>> int c;
>>> if (type < 3 && count <= 4)
>>> for (c=0; c < 4; c++) tt->val.c[c] = val >> (c << 3);
>>> }
>>> void LibRaw:: tiff_head (struct tiff_hdr *th, int full)
>>> {
>>> if (full) {
>>> tiff_set (&th->ngps, 3, 2, 2, (imgdata.other.gpsdata)[30]);
>>> tiff_set (&th->ngps, 4, 5, 3, ((char *)(&(th->gps[6])) - (char
>>> *)th));
>>> tiff_set (&th->ngps, 5, 1, 1, (imgdata.other.gpsdata)[31]);
>>> }
>>> int c, row, col, soff, rstep, cstep;
>>> for (row=0; row < (imgdata.sizes.height); row++, soff += rstep) {
>>> }
>>> }
>>>
>>> at O3 , tiff_set is inlined into tiff_set, at PREOPT_LNO_PHASE,
>>> expression
>>> if (type <3 && count <=4), i.e, whirl expression
>>>
>>> U4U2LDID 0 <st 89> T<7,.predef_U2,2>
>>> U4INTCONST 2 (0x2)
>>> U4U4LE
>>> I4I4LDID 0 <st 90> T<4,.predef_I4,4>
>>> I4INTCONST 4 (0x4)
>>> U4I4LE
>>> I4CAND
>>> U4I4CVT
>>> after codemap, is transfered to
>>> > LDC U4 1 <u=2 cr19> flags:0x0 b=-1
>>> > U4I4CVT <u=2 cr33> isop_flags:0xc0000 flags:0x1 b=-1
>>> >OPR_FALSEBR 6914 b=-1 flags:0x2 pj2 Sid-1
>>> DCE does not delete the unreachable branch under this condition, which
>>> then emits back to H whirl
>>> as
>>> IF {line: 1/131}
>>> U4INTCONST 1 (0x1)
>>> THEN
>>> ...
>>> ELSE
>>> BLOCK
>>> END_BLOCK
>>> END_IF
>>>
>>> this will make proactive transformation phase in PREOPT_LNO1_PHASE
>>> uncomfort, then it asserts.
>>>
>>> Suggested patch, keep the orginal handling of U4I4CVT const , only
>>> change under the MAINOPT_phase.
>>>
>>>
>>> osprey/be/opt/opt_htable.cxx -- 148c4e2..6499b99 100644
>>> --- a/osprey/be/opt/opt_htable.cxx
>>> +++ b/osprey/be/opt/opt_htable.cxx
>>> @@ -2273,7 +2273,9 @@ CODEMAP::Canon_cvt(WN *wn,
>>> Get_mtype_class(OPCODE_desc(op))) != 0 &&
>>> MTYPE_size_min(OPCODE_rtype(op)) ==
>>> MTYPE_size_min(OPCODE_desc(op)) &&
>>> // bug912 open64.net. Do not delete U4I4CVT if his kid is a
>>> constant
>>> - (!(OPCODE_rtype(op) == MTYPE_U4 &&
>>> + // bug934 open64.net. Only do this at MAINOPT_PHASE.
>>> + (!(_phase == MAINOPT_PHASE &&
>>> + OPCODE_rtype(op) == MTYPE_U4 &&
>>> OPCODE_desc(op) == MTYPE_I4 &&
>>> ccr->Tree() == NULL)))
>>> return propagated;
>>>
>>> Would a gatekeeper help a review?
>>>
>>> BTW, thanks for Xiaojing for report the bug.
>>>
>>> Regards
>>> Gang
>>>
>>>
>>> ------------------------------------------------------------------------------
>>> Cloud Computing - Latest Buzzword or a Glimpse of the Future?
>>> This paper surveys cloud computing today: What are the benefits?
>>> Why are businesses embracing it? What are its payoffs and pitfalls?
>>> http://www.accelacomm.com/jaw/sdnl/114/51425149/
>>> _______________________________________________
>>> Open64-devel mailing list
>>> Open64-devel@lists.sourceforge.net
>>> https://lists.sourceforge.net/lists/listinfo/open64-devel
>>>
>>>
>>
>
------------------------------------------------------------------------------
Learn Windows Azure Live! Tuesday, Dec 13, 2011
Microsoft is holding a special Learn Windows Azure training event for
developers. It will provide a great way to learn Windows Azure and what it
provides. You can attend the event by watching it streamed LIVE online.
Learn more at http://p.sf.net/sfu/ms-windowsazure
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
