diff options
author | Andrea Di Biagio <Andrea_DiBiagio@sn.scee.net> | 2014-02-12 23:43:47 +0000 |
---|---|---|
committer | Andrea Di Biagio <Andrea_DiBiagio@sn.scee.net> | 2014-02-12 23:43:47 +0000 |
commit | b7882b3bd11b46cca650e318bb7b921a1bfc8e21 (patch) | |
tree | 97a7fdb6b7e8ffe0337e7c9260cc03cb39f3c1b9 /llvm/lib/Transforms/Vectorize/BBVectorize.cpp | |
parent | 386d566395d90b01ddddf3d0ab219097cca8c37c (diff) | |
download | bcm5719-llvm-b7882b3bd11b46cca650e318bb7b921a1bfc8e21.tar.gz bcm5719-llvm-b7882b3bd11b46cca650e318bb7b921a1bfc8e21.zip |
[Vectorizer] Add a new 'OperandValueKind' in TargetTransformInfo called
'OK_NonUniformConstValue' to identify operands which are constants but
not constant splats.
The cost model now allows returning 'OK_NonUniformConstValue'
for non splat operands that are instances of ConstantVector or
ConstantDataVector.
With this change, targets are now able to compute different costs
for instructions with non-uniform constant operands.
For example, On X86 the cost of a vector shift may vary depending on whether
the second operand is a uniform or non-uniform constant.
This patch applies the following changes:
- The cost model computation now takes into account non-uniform constants;
- The cost of vector shift instructions has been improved in
X86TargetTransformInfo analysis pass;
- BBVectorize, SLPVectorizer and LoopVectorize now know how to distinguish
between non-uniform and uniform constant operands.
Added a new test to verify that the output of opt
'-cost-model -analyze' is valid in the following configurations: SSE2,
SSE4.1, AVX, AVX2.
llvm-svn: 201272
Diffstat (limited to 'llvm/lib/Transforms/Vectorize/BBVectorize.cpp')
-rw-r--r-- | llvm/lib/Transforms/Vectorize/BBVectorize.cpp | 55 |
1 files changed, 52 insertions, 3 deletions
diff --git a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp index 0cc1f3962a4..f59dd2160a9 100644 --- a/llvm/lib/Transforms/Vectorize/BBVectorize.cpp +++ b/llvm/lib/Transforms/Vectorize/BBVectorize.cpp @@ -532,7 +532,11 @@ namespace { // Returns the cost of the provided instruction using TTI. // This does not handle loads and stores. - unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2) { + unsigned getInstrCost(unsigned Opcode, Type *T1, Type *T2, + TargetTransformInfo::OperandValueKind Op1VK = + TargetTransformInfo::OK_AnyValue, + TargetTransformInfo::OperandValueKind Op2VK = + TargetTransformInfo::OK_AnyValue) { switch (Opcode) { default: break; case Instruction::GetElementPtr: @@ -562,7 +566,7 @@ namespace { case Instruction::And: case Instruction::Or: case Instruction::Xor: - return TTI->getArithmeticInstrCost(Opcode, T1); + return TTI->getArithmeticInstrCost(Opcode, T1, Op1VK, Op2VK); case Instruction::Select: case Instruction::ICmp: case Instruction::FCmp: @@ -1013,13 +1017,58 @@ namespace { unsigned JCost = getInstrCost(J->getOpcode(), JT1, JT2); Type *VT1 = getVecTypeForPair(IT1, JT1), *VT2 = getVecTypeForPair(IT2, JT2); + TargetTransformInfo::OperandValueKind Op1VK = + TargetTransformInfo::OK_AnyValue; + TargetTransformInfo::OperandValueKind Op2VK = + TargetTransformInfo::OK_AnyValue; + + // On some targets (example X86) the cost of a vector shift may vary + // depending on whether the second operand is a Uniform or + // NonUniform Constant. + switch (I->getOpcode()) { + default : break; + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + + // If both I and J are scalar shifts by constant, then the + // merged vector shift count would be either a constant splat value + // or a non-uniform vector of constants. + if (ConstantInt *CII = dyn_cast<ConstantInt>(I->getOperand(1))) { + if (ConstantInt *CIJ = dyn_cast<ConstantInt>(J->getOperand(1))) + Op2VK = CII == CIJ ? TargetTransformInfo::OK_UniformConstantValue : + TargetTransformInfo::OK_NonUniformConstantValue; + } else { + // Check for a splat of a constant or for a non uniform vector + // of constants. + Value *IOp = I->getOperand(1); + Value *JOp = J->getOperand(1); + if (ConstantDataVector *CDVI = dyn_cast<ConstantDataVector>(IOp)) { + if (ConstantDataVector *CDVJ = dyn_cast<ConstantDataVector>(JOp)) { + Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; + Constant *SplatValue = CDVI->getSplatValue(); + if (SplatValue != NULL && SplatValue == CDVJ->getSplatValue()) + Op2VK = TargetTransformInfo::OK_UniformConstantValue; + } + } + + if (ConstantVector *CVI = dyn_cast<ConstantVector>(IOp)) { + if (ConstantVector *CVJ = dyn_cast<ConstantVector>(JOp)) { + Op2VK = TargetTransformInfo::OK_NonUniformConstantValue; + Constant *SplatValue = CVI->getSplatValue(); + if (SplatValue != NULL && SplatValue == CVJ->getSplatValue()) + Op2VK = TargetTransformInfo::OK_UniformConstantValue; + } + } + } + } // Note that this procedure is incorrect for insert and extract element // instructions (because combining these often results in a shuffle), // but this cost is ignored (because insert and extract element // instructions are assigned a zero depth factor and are not really // fused in general). - unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2); + unsigned VCost = getInstrCost(I->getOpcode(), VT1, VT2, Op1VK, Op2VK); if (VCost > ICost + JCost) return false; |