[Patch]New InstCombine pattern for Icmp

[Yi Jiang]

Ping ...
On Aug 11, 2014, at 10:19 AM, Yi Jiang <yjiang at apple.com> wrote:

> Hi Nuno, 
> Yes. it does. We have already extracted the test case to reflect the pattern we want to catch most and will go with the patch. We believe the most useful case is to optimize:
> ((c) >= 'A' && (c) <= 'Z') || ((c) >= 'a' && (c) <= 'z’).  Thank you.
> 
> -Yi
> 
> On Aug 11, 2014, at 9:17 AM, Nuno Lopes <nuno.lopes at ist.utl.pt> wrote:
> 
>> AFAICT, correctness wise the patch looks good. I've translated your optimization into Alive by hand and now it seems good.
>> Now I wonder whether the optimization still kicks in on your benchmarks? (given that now the constraints are much thigher)
>> 
>> Nuno
>> 
>> ----- Original Message ----- From: "Yi Jiang" <yjiang at apple.com>
>> Sent: Thursday, August 07, 2014 8:37 AM
>> Subject: Re: [Patch]New InstCombine pattern for Icmp
>> 
>> 
>>> With the help of Nuno and the Alive tool, I figured out the new condition. The original condition actually refers to the following two ranges:
>>> (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A + C2), C3) == [MAX_UINT-C1+1, MAX_UINT-C1+1+C3] and [MAX_UINT-C2+1, MAX_UINT-C2+1+C3]
>>> We can fold these two range sif:
>>> 1) C1 and C2 is unsigned greater than C3
>>> 2) The two ranges are separated.  abs(LowRange1-LowRange2) > C3
>>> 3) C1 ^ C2 is one-bit mask.
>>> 4) LowRange1 ^ LowRange2 and HighRange1 ^ HighRange2 are one-bit mask.
>>> 
>>> Here is the new patch:
>>> 
>>> 
>>> On Jul 30, 2014, at 2:24 PM, Nuno Lopes <nuno.lopes at ist.utl.pt> wrote:
>>> 
>>>> Hi Yi,
>>>> 
>>>> ALIVe still complains, I'm afraid:
>>>> 
>>>> Precondition: isPowerOf2(C1 ^ C2) && ugt(C1, C3) && ugt(C2, C3)
>>>> %x = add %A, C1
>>>> %i = icmp ult %x, C3
>>>> %y = add %A, C2
>>>> %j = icmp ult %y, C3
>>>> %r = or %i, %j
>>>> =>
>>>> %and = and %A, ~(C1 ^ C2)
>>>> %lhs = add %and, umax(C1, C2)
>>>> %r = icmp ult %lhs, C3
>>>> 
>>>> Done: 1
>>>> ERROR: Mismatch in values of i1 %r
>>>> 
>>>> Example:
>>>> %A i2 = 2 (0x2)
>>>> C1 i2 = 3 (0x3)
>>>> %x i2 = 1 (0x1)
>>>> C3 i2 = 1 (0x1)
>>>> %i i1 = 0 (0x0)
>>>> C2 i2 = 2 (0x2)
>>>> %y i2 = 0 (0x0)
>>>> %j i1 = 1 (0x1)
>>>> %and i2 = 2 (0x2)
>>>> %lhs i2 = 1 (0x1)
>>>> Source value: 1 (0x1)
>>>> Target value: 0 (0x0)
>>>> 
>>>> 
>>>> Nuno
>>>> 
>>>> ----- Original Message ----- From: "Yi Jiang" <yjiang at apple.com>
>>>> To: "Nuno Lopes" <nuno.lopes at ist.utl.pt>
>>>> Cc: "LLVM Commits" <llvm-commits at cs.uiuc.edu>
>>>> Sent: Wednesday, July 30, 2014 8:28 PM
>>>> Subject: Re: [Patch]New InstCombine pattern for Icmp
>>>> 
>>>> 
>>>> Hi Nuno,
>>>> 
>>>> Thank you for pointing out this! One condition is missing: C1 and C2 should be unsigned greater than C3.
>>>> Here is the new patch.
>>>> 
>>>> -Yi
>>>> 
>>>> On Jul 29, 2014, at 11:34 AM, Nuno Lopes <nuno.lopes at ist.utl.pt> wrote:
>>>> 
>>>>> Hi,
>>>>> 
>>>>> I sent you an email the other day, but it was rejected by Apple's email server.  Tyring now through another email account.
>>>>> Please take a look here: http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20140721/227575.html
>>>>> 
>>>>> Nuno
>>>>> 
>>>>> ----- Original Message -----
>>>>> 
>>>>> Ping...
>>>>> On Jul 25, 2014, at 4:11 PM, Yi Jiang <yjiang at apple.com> wrote:
>>>>> 
>>>>>> Hi Ben,
>>>>>> 
>>>>>> Thank you for your comments! Here is a new version.  In this version, I still kept some "if" thinking that some operand has been already extracted so I would like to use it directly.  Please let me know if any other comments.
>>>>>> <combinecmpV2.patch>
>>>>>> -Yi
>>>>>> On Jul 25, 2014, at 2:12 AM, Benjamin Kramer <benny.kra at gmail.com> wrote:
>>>>>> 
>>>>>>> On Fri, Jul 25, 2014 at 1:24 AM, Yi Jiang <yjiang at apple.com> wrote:
>>>>>>>> Hi,
>>>>>>>> 
>>>>>>>> his patch is trying to fold (icmp ult/ule (A + C1), C3) | (icmp ult/ule (A +
>>>>>>>> C2), C3)  to (icmp ult/ule ((A & ~(C1 ^ C2)) + max(C1, C2)), C3)  .
>>>>>>>> This transformation is legal if C1 ^ C2 is one-bit mask, in other word, C1
>>>>>>>> is only one bit different from C2. In this case, we can "mask" that bit and
>>>>>>>> do just one comparison.
>>>>>>>> A typical example is:
>>>>>>>> isALPHA(c)    ((c) >= 'A' && (c) <= 'Z') || ((c) >= 'a' && (c) <= 'z')
>>>>>>>> 
>>>>>>>> Now llvm will optimize it to ((c + 191) <=25) || ((c + 159) <=25)
>>>>>>>> 
>>>>>>>> With this patch, we can optimize it further to:
>>>>>>>> (c & 223) + 191 <= 25
>>>>>>>> 
>>>>>>>> The binary format of the constants are:
>>>>>>>> 191   10111111
>>>>>>>> 159   10011111
>>>>>>>> 223   11011111
>>>>>>>> 
>>>>>>>> Here is some experiment result on arm64:
>>>>>>>> The patch shows no regression and improve spec2000 perlbmk 3.8% in
>>>>>>>> test-suite under -O3.
>>>>>>>> We also test the spec2006 400.perlbench with ref size input, it will improve
>>>>>>>> 1% under -O3 and 1.2% under -O3+lto+pgo.
>>>>>>> 
>>>>>>> Wow, very nice.
>>>>>>> 
>>>>>>>> Any comments are appreciated.
>>>>>>> 
>>>>>>> The patch could be significantly simplified with the PatternMatch
>>>>>>> tool. Something like match(LHS, m_Add(m_OneUse(m_Value(A),
>>>>>>> m_ConstantInt(LAddCst)... could replace your if chain.
>>>>>>> 
>>>>>>> - Ben 
>> 
>