[PATCH][polly] Support for generating vectors for loads with -1 stride
Chris Jenneisch
chrisj at codeaurora.org
Tue Mar 18 17:45:18 PDT 2014
Hi Tobi,
Thanks for the comments, I have attached the updated patch.
Chris
> On 03/10/2014 04:46 PM, chrisj at codeaurora.org wrote:
>> Hi all,
>>
>> This patch enables vectorization of loops containing backward array
>> traversal (array stride is -1).
>
> Hi Chris,
>
> this patch looks generally fine. Just some minor nits.
>
>
>> 0001-Support-for-generating-vectors-for-loads-with-1-stri.patch
>>
>>
>> From 1d8a62768085a53f5a4bff2c80ba9c8b5930460a Mon Sep 17 00:00:00 2001
>> From: Chris Jenneisch<chrisj at codeaurora.org>
>> Date: Tue, 4 Mar 2014 12:31:57 -0800
>> Subject: [PATCH] Support for generating vectors for loads with -1 stride
>>
>> This patch enables vectorization of loops containing backward array
>> traversal (array stride is -1).
>> ---
>> include/polly/CodeGen/BlockGenerators.h | 9 +++-
>> lib/CodeGen/BlockGenerators.cpp | 27 +++++++++--
>> test/Isl/CodeGen/simple_vec_stride_negative_one.ll | 49
>> ++++++++++++++++++++
>> 3 files changed, 79 insertions(+), 6 deletions(-)
>> create mode 100644 test/Isl/CodeGen/simple_vec_stride_negative_one.ll
>>
>> diff --git a/include/polly/CodeGen/BlockGenerators.h
>> b/include/polly/CodeGen/BlockGenerators.h
>> index 077236d..487521f 100644
>> --- a/include/polly/CodeGen/BlockGenerators.h
>> +++ b/include/polly/CodeGen/BlockGenerators.h
>> @@ -265,7 +265,14 @@ private:
>> /// %vector_ptr= bitcast double* %p to <4 x double>*
>> /// %vec_full = load <4 x double>* %vector_ptr
>> ///
>> - Value *generateStrideOneLoad(const LoadInst *Load, ValueMapT &BBMap);
>> + /// @param Direction This is used to indicate a -1 stride. In such
>> + /// a case we load the end of a base address and
>> + /// shuffle the accesses in reverse order into the
>> + /// vector. By default we would do only positive
>> + /// strides.
>
> Maybe use 'bool NegativeStride' instead of Direction.
>
>> + Value *generateStrideOneLoad(const LoadInst *Load,
>> + VectorValueMapT &ScalarMaps, int
>> Direction);
>>
>> /// @brief Load a vector initialized from a single scalar in memory
>> ///
>> diff --git a/lib/CodeGen/BlockGenerators.cpp
>> b/lib/CodeGen/BlockGenerators.cpp
>> index a5d1f0f..4ce8ca1 100644
>> --- a/lib/CodeGen/BlockGenerators.cpp
>> +++ b/lib/CodeGen/BlockGenerators.cpp
>> @@ -440,11 +440,16 @@ Type *VectorBlockGenerator::getVectorPtrTy(const
>> Value *Val, int Width) {
>> }
>>
>> Value *VectorBlockGenerator::generateStrideOneLoad(const LoadInst
>> *Load,
>> - ValueMapT &BBMap) {
>> + VectorValueMapT
>> &ScalarMaps,
>> + int Direction = 0) {
>> + unsigned VectorWidth = getVectorWidth();
>> const Value *Pointer = Load->getPointerOperand();
>> - Type *VectorPtrType = getVectorPtrTy(Pointer, getVectorWidth());
>> - Value *NewPointer =
>> - getNewValue(Pointer, BBMap, GlobalMaps[0], VLTS[0],
>> getLoopForInst(Load));
>> + Type *VectorPtrType = getVectorPtrTy(Pointer, VectorWidth);
>> + unsigned Offset = Direction ? VectorWidth - 1 : 0;
>> +
>> + Value *NewPointer = NULL;
>> + NewPointer = getNewValue(Pointer, ScalarMaps[Offset],
>> GlobalMaps[Offset],
>> + VLTS[Offset], getLoopForInst(Load));
>> Value *VectorPtr =
>> Builder.CreateBitCast(NewPointer, VectorPtrType, "vector_ptr");
>> LoadInst *VecLoad =
>> @@ -452,6 +457,16 @@ Value
>> *VectorBlockGenerator::generateStrideOneLoad(const LoadInst *Load,
>> if (!Aligned)
>> VecLoad->setAlignment(8);
>>
>> + if (Direction) {
>> + SmallVector<Constant *, 16> Indices;
>> + for (int i = VectorWidth - 1; i >= 0; i--)
>> + Indices.push_back(ConstantInt::get(Builder.getInt32Ty(), i));
>> + Constant *SV = llvm::ConstantVector::get(Indices);
>> + Value *RevVecLoad = Builder.CreateShuffleVector(
>> + VecLoad, VecLoad, SV, Load->getName() + "reverse");
>> + return RevVecLoad;
>> + }
>> +
>> return VecLoad;
>> }
>>
>> @@ -516,7 +531,9 @@ void VectorBlockGenerator::generateLoad(const
>> LoadInst *Load,
>> if (Access.isStrideZero(isl_map_copy(Schedule)))
>> NewLoad = generateStrideZeroLoad(Load, ScalarMaps[0]);
>> else if (Access.isStrideOne(isl_map_copy(Schedule)))
>> - NewLoad = generateStrideOneLoad(Load, ScalarMaps[0]);
>> + NewLoad = generateStrideOneLoad(Load, ScalarMaps);
>> + else if (Access.isStrideX(isl_map_copy(Schedule), -1))
>> + NewLoad = generateStrideOneLoad(Load, ScalarMaps, 1);
>> else
>> NewLoad = generateUnknownStrideLoad(Load, ScalarMaps);
>>
>> diff --git a/test/Isl/CodeGen/simple_vec_stride_negative_one.ll
>> b/test/Isl/CodeGen/simple_vec_stride_negative_one.ll
>> new file mode 100644
>> index 0000000..37a160e
>> --- /dev/null
>> +++ b/test/Isl/CodeGen/simple_vec_stride_negative_one.ll
>> @@ -0,0 +1,49 @@
>> +; RUN: opt %loadPolly %defaultOpts -polly-codegen-isl
>> -polly-codegen-scev %vector-opt -dce -polly -O3 -S < %s | FileCheck %s
>
> %defaultOps does not seem to be necessary.
>
> Also, why are you running '-dce -polly -O3' at the end?
>
> I would expect something like:
>
> opt %loadPolly -polly-codegen-isl -polly-codegen-scev %vector-opt -S
>
> Also, if the test case is not immediately vectorized, it may be
> necessary to reduce the number of loop iterations to a smaller number (4
> e.g.).
>
>> +; ModuleID = 'reverse.c'
>> +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
>> +target triple = "x86_64-unknown-linux-gnu"
>> +
>> +;int A[100];
>> +;void foo() {
>> +; for (int i=999; i >= 0; i--)
>> +; A[i]+=1;
>> +;}
>> +
>> +
>> + at A = common global [100 x i32] zeroinitializer, align 16
>> +
>> +; Function Attrs: nounwind uwtable
>> +define void @foo() #0 {
>> +entry:
>> + br label %for.body
>> +
>> +for.body: ; preds = %entry,
>> %for.body
>> + %indvars.iv = phi i64 [ 999, %entry ], [ %indvars.iv.next, %for.body
>> ]
>> + %arrayidx = getelementptr inbounds [100 x i32]* @A, i64 0, i64
>> %indvars.iv
>> + %0 = load i32* %arrayidx, align 4, !tbaa !1
>> + %add = add nsw i32 %0, 1
>> + store i32 %add, i32* %arrayidx, align 4, !tbaa !1
>> + %indvars.iv.next = add nsw i64 %indvars.iv, -1
>> + %1 = trunc i64 %indvars.iv to i32
>> + %cmp = icmp sgt i32 %1, 0
>> + br i1 %cmp, label %for.body, label %for.end
>> +
>> +for.end: ; preds = %for.body
>> + ret void
>> +}
>> +
>> +attributes #0 = { nounwind uwtable "less-precise-fpmad"="false"
>> "no-frame-pointer-elim"="false" "no-infs-fp-math"="false"
>> "no-nans-fp-math"="false" "stack-protector-buffer-size"="8"
>> "unsafe-fp-math"="false" "use-soft-float"="false" }
>
> Can you remove unnecessary attributes.
>
>> +!llvm.ident = !{!0}
>> +
>> +!0 = metadata !{metadata !"clang version 3.5.0
>> (http://llvm.org/git/clang.git
>> b6b24026464c7c5053e1903a27f4dd1f95faf5c8) (http://llvm.org/git/llvm.git
>> 824dfb1c5690ad1494572c95909c0d9dc27338c8)"}
>> +!1 = metadata !{metadata !2, metadata !2, i64 0}
>> +!2 = metadata !{metadata !"int", metadata !3, i64 0}
>> +!3 = metadata !{metadata !"omnipotent char", metadata !4, i64 0}
>> +!4 = metadata !{metadata !"Simple C/C++ TBAA"}
>
> Can you remove unnecessary metadata.
>
>
>> +
>> +; CHECK: @foo
>> +; CHECK: [[LOAD:%[a-zA-Z0-9_]+]] = load <4 x i32>*
>> +; CHECK: [[REVERSE_LOAD:%[a-zA-Z0-9_]+reverse]] = shufflevector <4 x
>> i32> [[LOAD]], <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
>> -- 1.7.8.3
>
>
--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
hosted by The Linux Foundation
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Support-for-generating-vectors-for-loads-with-1-stri.patch
Type: application/octet-stream
Size: 5572 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140319/e61f6a79/attachment.obj>
More information about the llvm-commits
mailing list