[PATCH] Make bitcast, extractelement, and insertelement considered cheap for speculation in SimplifyCFG.
Matt Arsenault
Matthew.Arsenault at amd.com
Fri Dec 13 17:30:59 PST 2013
On 12/13/2013 05:16 PM, Hal Finkel wrote:
> ----- Original Message -----
>> From: "Hal Finkel" <hfinkel at anl.gov>
>> To: "Matt Arsenault" <Matthew.Arsenault at amd.com>
>> Cc: "llvm commits" <llvm-commits at cs.uiuc.edu>, reviews+D2218+public+1b17721e58b4714d at llvm-reviews.chandlerc.com
>> Sent: Tuesday, November 19, 2013 1:23:41 PM
>> Subject: Re: [PATCH] Make bitcast, extractelement, and insertelement considered cheap for speculation in SimplifyCFG.
>>
>> ----- Original Message -----
>>> From: "Matt Arsenault" <Matthew.Arsenault at amd.com>
>>> To: "Nadav Rotem" <nrotem at apple.com>,
>>> reviews+D2218+public+1b17721e58b4714d at llvm-reviews.chandlerc.com
>>> Cc: "llvm commits" <llvm-commits at cs.uiuc.edu>
>>> Sent: Tuesday, November 19, 2013 1:12:01 PM
>>> Subject: Re: [PATCH] Make bitcast, extractelement, and
>>> insertelement considered cheap for speculation in SimplifyCFG.
>>>
>>> On 11/19/2013 08:42 AM, Nadav Rotem wrote:
>>>> Hi Matt,
>>>>
>>>> I don’t know this code very well but your patch looks okay to me.
>>>> Please benchmark the LLVM test suite and make sure that this
>>>> change does not introduce new regressions.
>>>>
>>>> SimplifyCFG should not use TTI because this pass is executed very
>>>> early in the optimization pipe. TTI should only be used by
>>>> “lowering” transformations in the late stages of the optimization
>>>> pipe.
>>> But it is included in SimplifyCFG, and it is used in a few places
>>> there.
>>> Are these supposed to be removed at some point?
>> No, TTI is used in a few other places in order to check for "free"
>> operations. We just don't use the costs until later in the pipeline.
> What ended up happening with this?
>
> -Hal
Nothing yet. I noticed a few more things I wanted to investigate here
and haven't gotten back to it. First it seems odd most of the other
instructions aren't also handled here (like all of the basic FP ops) and
was wondering why. This also didn't fully solve the problem I was
originally trying to solve. I was going to try running the benchmarks
but then I was failing to get LNT to work.
- Matt
>> -Hal
>>
>>>> Thanks,
>>>> Nadav
>>>>
>>>>
>>>> On Nov 18, 2013, at 7:00 PM, Matt Arsenault
>>>> <Matthew.Arsenault at amd.com> wrote:
>>>>
>>>>> Hi nadav,
>>>>>
>>>>> This help fold more branches into selects.
>>>>> On R600, vectors are cheap and anything that helps
>>>>> remove branches is very good.
>>>>>
>>>>> I don't know why this doesn't use the TTI for this cost
>>>>> calculation or if it should.
>>>>>
>>>>> http://llvm-reviews.chandlerc.com/D2218
>>>>>
>>>>> Files:
>>>>> lib/Transforms/Utils/SimplifyCFG.cpp
>>>>> test/Transforms/SimplifyCFG/speculate-vector-ops.ll
>>>>>
>>>>> Index: lib/Transforms/Utils/SimplifyCFG.cpp
>>>>> ===================================================================
>>>>> --- lib/Transforms/Utils/SimplifyCFG.cpp
>>>>> +++ lib/Transforms/Utils/SimplifyCFG.cpp
>>>>> @@ -224,6 +224,9 @@
>>>>> case Instruction::Trunc:
>>>>> case Instruction::ZExt:
>>>>> case Instruction::SExt:
>>>>> + case Instruction::BitCast:
>>>>> + case Instruction::ExtractElement:
>>>>> + case Instruction::InsertElement:
>>>>> return 1; // These are all cheap.
>>>>>
>>>>> case Instruction::Call:
>>>>> Index: test/Transforms/SimplifyCFG/speculate-vector-ops.ll
>>>>> ===================================================================
>>>>> --- /dev/null
>>>>> +++ test/Transforms/SimplifyCFG/speculate-vector-ops.ll
>>>>> @@ -0,0 +1,60 @@
>>>>> +; RUN: opt -S -simplifycfg < %s | FileCheck %s
>>>>> +
>>>>> +define i32 @speculate_vector_extract(i32 %d, <4 x i32> %v) #0 {
>>>>> +; CHECK-LABEL: @speculate_vector_extract(
>>>>> +; CHECK-NOT: br
>>>>> +entry:
>>>>> + %conv = insertelement <4 x i32> undef, i32 %d, i32 0
>>>>> + %conv2 = insertelement <4 x i32> %conv, i32 %d, i32 1
>>>>> + %conv3 = insertelement <4 x i32> %conv2, i32 %d, i32 2
>>>>> + %conv4 = insertelement <4 x i32> %conv3, i32 %d, i32 3
>>>>> + %tmp6 = add nsw <4 x i32> %conv4, <i32 0, i32 -1, i32 -2, i32
>>>>> -3>
>>>>> + %cmp = icmp eq <4 x i32> %tmp6, zeroinitializer
>>>>> + %cmp.ext = sext <4 x i1> %cmp to <4 x i32>
>>>>> + %tmp8 = extractelement <4 x i32> %cmp.ext, i32 0
>>>>> + %tobool = icmp eq i32 %tmp8, 0
>>>>> + br i1 %tobool, label %cond.else, label %cond.then
>>>>> +
>>>>> +return: ; preds =
>>>>> %cond.end28
>>>>> + ret i32 %cond32
>>>>> +
>>>>> +cond.then: ; preds =
>>>>> %entry
>>>>> + %tmp10 = extractelement <4 x i32> %v, i32 0
>>>>> + br label %cond.end
>>>>> +
>>>>> +cond.else: ; preds =
>>>>> %entry
>>>>> + %tmp12 = extractelement <4 x i32> %v, i32 3
>>>>> + br label %cond.end
>>>>> +
>>>>> +cond.end: ; preds =
>>>>> %cond.else, %cond.then
>>>>> + %cond = phi i32 [ %tmp10, %cond.then ], [ %tmp12, %cond.else
>>>>> ]
>>>>> + %tmp14 = extractelement <4 x i32> %cmp.ext, i32 1
>>>>> + %tobool15 = icmp eq i32 %tmp14, 0
>>>>> + br i1 %tobool15, label %cond.else17, label %cond.then16
>>>>> +
>>>>> +cond.then16: ; preds =
>>>>> %cond.end
>>>>> + %tmp20 = extractelement <4 x i32> %v, i32 1
>>>>> + br label %cond.end18
>>>>> +
>>>>> +cond.else17: ; preds =
>>>>> %cond.end
>>>>> + br label %cond.end18
>>>>> +
>>>>> +cond.end18: ; preds =
>>>>> %cond.else17, %cond.then16
>>>>> + %cond22 = phi i32 [ %tmp20, %cond.then16 ], [ %cond,
>>>>> %cond.else17 ]
>>>>> + %tmp24 = extractelement <4 x i32> %cmp.ext, i32 2
>>>>> + %tobool25 = icmp eq i32 %tmp24, 0
>>>>> + br i1 %tobool25, label %cond.else27, label %cond.then26
>>>>> +
>>>>> +cond.then26: ; preds =
>>>>> %cond.end18
>>>>> + %tmp30 = extractelement <4 x i32> %v, i32 2
>>>>> + br label %cond.end28
>>>>> +
>>>>> +cond.else27: ; preds =
>>>>> %cond.end18
>>>>> + br label %cond.end28
>>>>> +
>>>>> +cond.end28: ; preds =
>>>>> %cond.else27, %cond.then26
>>>>> + %cond32 = phi i32 [ %tmp30, %cond.then26 ], [ %cond22,
>>>>> %cond.else27 ]
>>>>> + br label %return
>>>>> +}
>>>>> +
>>>>> +attributes #0 = { nounwind }
>>>>> <D2218.1.patch>
>>>
>>>
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>>
>> --
>> Hal Finkel
>> Assistant Computational Scientist
>> Leadership Computing Facility
>> Argonne National Laboratory
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
More information about the llvm-commits
mailing list