maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

Support for generating vectors for loads with -1 stride 

This patch enables vectorization of loops containing backward array
traversal (array stride is -1).

Contributed-by: Chris Jenneisch <chrisj@codeaurora.org>
llvm-svn: 204257
This commit is contained in:
Tobias Grosser 2014-03-19 19:27:24 +00:00
parent f88287fd7a
commit 0dd463facf
3 changed files with 71 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
polly/include/polly/CodeGen/BlockGenerators.h
View File

@ -265,7 +265,15 @@ 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 NegativeStride 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.
///
Value *generateStrideOneLoad(const LoadInst *Load,
VectorValueMapT &ScalarMaps,
bool NegativeStride);
/// @brief Load a vector initialized from a single scalar in memory
///

30
polly/lib/CodeGen/BlockGenerators.cpp
View File

@ -439,12 +439,18 @@ Type *VectorBlockGenerator::getVectorPtrTy(const Value *Val, int Width) {
return PointerType::getUnqual(VectorType);
}
Value *VectorBlockGenerator::generateStrideOneLoad(const LoadInst *Load,
ValueMapT &BBMap) {
Value *
VectorBlockGenerator::generateStrideOneLoad(const LoadInst *Load,
VectorValueMapT &ScalarMaps,
bool NegativeStride = false) {
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 = NegativeStride ? 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 +458,16 @@ Value *VectorBlockGenerator::generateStrideOneLoad(const LoadInst *Load,
if (!Aligned)
VecLoad->setAlignment(8);
if (NegativeStride) {
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 +532,9 @@ 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, true);
else
NewLoad = generateUnknownStrideLoad(Load, ScalarMaps);

38
polly/test/Isl/CodeGen/simple_vec_stride_negative_one.ll Normal file
View File

@ -0,0 +1,38 @@
; RUN: opt %loadPolly -polly-codegen-isl -polly-codegen-scev %vector-opt -S < %s | FileCheck %s
; 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=3; i >= 0; i--)
; A[i]+=1;
;}
@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 [ 3, %entry ], [ %indvars.iv.next, %for.body ]
%arrayidx = getelementptr inbounds [100 x i32]* @A, i64 0, i64 %indvars.iv
%0 = load i32* %arrayidx, align 4
%add = add nsw i32 %0, 1
store i32 %add, i32* %arrayidx, align 4
%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
}
; 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> [[LOAD]], <4 x i32> <i32 3, i32 2, i32 1, i32 0>