[LLVMdev] [Review Request][PATCH] Add the function "vectorizeBasicBlock"

[Hongbin Zheng]

Hi Hal,

On Thu, Apr 5, 2012 at 12:07 AM, Hal Finkel <hfinkel at anl.gov> wrote:
> Ether,
>
> Sounds great! Please keep in mind that, eventually, we'll also want to
> configure those options from TLI (or something similar). The patch
Thanks a lot for reminding this and this is what i want, in fact i
have a private backend targeting FPGA and i only want to vectorize the
load/store instructions.
> looks good to me.
I am going to commit tomorrow.

best regards
ether
>
>  -Hal
>
> On Wed, 4 Apr 2012 23:54:18 +0800
> Hongbin Zheng <etherzhhb at gmail.com> wrote:
>
>> Hi Hal,
>>
>> I add a function named "vectorizeBasicBlock" which allow users to
>> perform basic block vectoirzation inside their pass. But i am not sure
>> whether i missed something as no one use the function right now (But
>> it will be used by Polly sometimes later[1]).
>>
>> In addition, we (tobi and me) also want to make the vectorizer being
>> configured command line flags. To achieve this, we are going to patch
>> the pass so that it do not read the options like ReqChainDepth,
>> SearchLimit, SplatBreaksChain, ... etc. from the command line option
>> directly, but read from an instance of the VectorizeConfig class,
>> where the VectorizeConfig class is designed to provide the vectorize
>> related parameters to to the BBVectorize pas.
>>
>> After this, people can pass target specific vectorize configuration to
>> the BBVectorize pass in tools like LLC. And original
>> createBBVectorizePass function can create a VectorizeConfig that read
>> the configuration from the commandline options and pass it to the
>> BBVectorizePass, hence users can still control the behaviour of the
>> pass with command like option.
>>
>> Any comment or suggestion is appreciated.
>>
>> best regards
>> ether
>>
>> [1]http://llvm.org/bugs/show_bug.cgi?id=12404
>> ---------- Forwarded message ----------
>> From: Hongbin Zheng <etherzhhb at gmail.com>
>> Date: Wed, Apr 4, 2012 at 9:55 PM
>> Subject: [Review Request][PATCH] Add a the function
>> "vectorizeBasicBlock" To: llvm-commits at cs.uiuc.edu
>> Cc: Tobias Grosser <tobias at grosser.es>
>>
>>
>> hi,
>>
>> This patch add a the function "vectorizeBasicBlock" which allow users
>> vectorize a  BasicBlock in other passes, e.g. we can call
>> vectorizeBasicBlock in the  loop unroll pass right after the loop is
>> unrolled.
>>
>> The function is currently declared in
>> "include/llvm/Transforms/Vectorize.h",  I have no idea in which header
>> the function should be declared though.
>>
>> best regards
>> ether
>>
>> ---
>>  include/llvm/Transforms/Vectorize.h      |   13 ++++++++++++-
>>  lib/Transforms/Vectorize/BBVectorize.cpp |   25
>> ++++++++++++++++++++----- 2 files changed, 32 insertions(+), 6
>> deletions(-)
>>
>> diff --git a/include/llvm/Transforms/Vectorize.h
>> b/include/llvm/Transforms/Vectorize.h
>> index dfc099d..373d501 100644
>> --- a/include/llvm/Transforms/Vectorize.h
>> +++ b/include/llvm/Transforms/Vectorize.h
>> @@ -16,7 +16,7 @@
>>  #define LLVM_TRANSFORMS_VECTORIZE_H
>>
>>  namespace llvm {
>> -
>> +class BasicBlock;
>>  class BasicBlockPass;
>>
>>  //===----------------------------------------------------------------------===//
>> @@ -25,6 +25,17 @@ class BasicBlockPass;
>>  //
>>  BasicBlockPass *createBBVectorizePass();
>>
>> +//===----------------------------------------------------------------------===//
>> +/// @brief Vectorize the BasicBlock.
>> +///
>> +/// @param BB The BasicBlock to be vectorized
>> +/// @param P  The current running pass, should require AliasAnalysis
>> and +///           ScalarEvolution. After the vectorization,
>> AliasAnalysis, +///           ScalarEvolution and CFG are preserved.
>> +///
>> +/// @returns True if the BB is changed, false otherwise.
>> +bool vectorizeBasicBlock(Pass *P, BasicBlock &BB);
>> +
>>  } // End llvm namespace
>>
>>  #endif
>> diff --git a/lib/Transforms/Vectorize/BBVectorize.cpp
>> b/lib/Transforms/Vectorize/BBVectorize.cpp
>> index 28eb634..d65cdd8 100644
>> --- a/lib/Transforms/Vectorize/BBVectorize.cpp
>> +++ b/lib/Transforms/Vectorize/BBVectorize.cpp
>> @@ -143,6 +143,12 @@ namespace {
>>     BBVectorize() : BasicBlockPass(ID) {
>>       initializeBBVectorizePass(*PassRegistry::getPassRegistry());
>>     }
>> +
>> +    BBVectorize(Pass *P) : BasicBlockPass(ID) {
>> +      AA = &P->getAnalysis<AliasAnalysis>();
>> +      SE = &P->getAnalysis<ScalarEvolution>();
>> +      TD = P->getAnalysisIfAvailable<TargetData>();
>> +    }
>>
>>     typedef std::pair<Value *, Value *> ValuePair;
>>     typedef std::pair<ValuePair, size_t> ValuePairWithDepth;
>> @@ -280,11 +286,7 @@ namespace {
>>                      Instruction *&InsertionPt,
>>                      Instruction *I, Instruction *J);
>>
>> -    virtual bool runOnBasicBlock(BasicBlock &BB) {
>> -      AA = &getAnalysis<AliasAnalysis>();
>> -      SE = &getAnalysis<ScalarEvolution>();
>> -      TD = getAnalysisIfAvailable<TargetData>();
>> -
>> +    bool vectorizeBB(BasicBlock &BB) {
>>       bool changed = false;
>>       // Iterate a sufficient number of times to merge types of size
>> 1 bit, // then 2 bits, then 4, etc. up to half of the target vector
>> width of the @@ -303,6 +305,14 @@ namespace {
>>       DEBUG(dbgs() << "BBV: done!\n");
>>       return changed;
>>     }
>> +
>> +    virtual bool runOnBasicBlock(BasicBlock &BB) {
>> +      AA = &getAnalysis<AliasAnalysis>();
>> +      SE = &getAnalysis<ScalarEvolution>();
>> +      TD = getAnalysisIfAvailable<TargetData>();
>> +
>> +      return vectorizeBB(BB);
>> +    }
>>
>>     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
>>       BasicBlockPass::getAnalysisUsage(AU);
>> @@ -1861,3 +1871,8 @@ BasicBlockPass *llvm::createBBVectorizePass() {
>>   return new BBVectorize();
>>  }
>>
>> +bool llvm::vectorizeBasicBlock(Pass *P, BasicBlock &BB) {
>> +  BBVectorize BBVectorizer(P);
>> +  return BBVectorizer.vectorizeBB(BB);
>> +}
>> +
>> --
>> 1.7.5.4
>
>
>
> --
> Hal Finkel
> Postdoctoral Appointee
> Leadership Computing Facility
> Argonne National Laboratory