[llvm-commits] [llvm] r170995 - in /llvm/trunk: include/llvm/Target/TargetTransformImpl.h include/llvm/TargetTransformInfo.h lib/Target/TargetTransformImpl.cpp lib/Target/X86/X86ISelLowering.cpp lib/Transforms/Vectorize/LoopVectorize.cpp test/Transforms/LoopVectorize/X86/cost-model.ll
Nadav Rotem
nrotem at apple.com
Sat Dec 22 23:23:55 PST 2012
Author: nadav
Date: Sun Dec 23 01:23:55 2012
New Revision: 170995
URL: http://llvm.org/viewvc/llvm-project?rev=170995&view=rev
Log:
Loop Vectorizer: Update the cost model of scatter/gather operations and make
them more expensive.
Modified:
llvm/trunk/include/llvm/Target/TargetTransformImpl.h
llvm/trunk/include/llvm/TargetTransformInfo.h
llvm/trunk/lib/Target/TargetTransformImpl.cpp
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp
llvm/trunk/test/Transforms/LoopVectorize/X86/cost-model.ll
Modified: llvm/trunk/include/llvm/Target/TargetTransformImpl.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Target/TargetTransformImpl.h?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/include/llvm/Target/TargetTransformImpl.h (original)
+++ llvm/trunk/include/llvm/Target/TargetTransformImpl.h Sun Dec 23 01:23:55 2012
@@ -69,8 +69,6 @@
virtual ~VectorTargetTransformImpl() {}
- virtual unsigned getInstrCost(unsigned Opcode, Type *Ty1, Type *Ty2) const;
-
virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const;
virtual unsigned getBroadcastCost(Type *Tp) const;
Modified: llvm/trunk/include/llvm/TargetTransformInfo.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/TargetTransformInfo.h?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/include/llvm/TargetTransformInfo.h (original)
+++ llvm/trunk/include/llvm/TargetTransformInfo.h Sun Dec 23 01:23:55 2012
@@ -135,44 +135,28 @@
virtual bool shouldBuildLookupTables() const {
return true;
}
-
/// getPopcntHwSupport - Return hardware support for population count.
virtual PopcntHwSupport getPopcntHwSupport(unsigned IntTyWidthInBit) const {
return None;
}
-
/// getIntImmCost - Return the expected cost of materializing the given
/// integer immediate of the specified type.
virtual unsigned getIntImmCost(const APInt&, Type*) const {
- // Default assumption is immediate is cheap.
+ // The default assumption is that the immediate is cheap.
return 1;
}
};
/// VectorTargetTransformInfo - This interface is used by the vectorizers
/// to estimate the profitability of vectorization for different instructions.
+/// This interface provides the cost of different IR instructions. The cost
+/// is unit-less and represents the estimated throughput of the instruction
+/// (not the latency!) assuming that all branches are predicted, cache is hit,
+/// etc.
class VectorTargetTransformInfo {
public:
virtual ~VectorTargetTransformInfo() {}
- /// Returns the expected cost of the instruction opcode. The opcode is one of
- /// the enums like Instruction::Add. The type arguments are the type of the
- /// operation.
- /// Most instructions only use the first type and in that case the second
- /// operand is ignored.
- ///
- /// Exceptions:
- /// * Br instructions do not use any of the types.
- /// * Select instructions pass the return type as Ty1 and the selector as Ty2.
- /// * Cast instructions pass the destination as Ty1 and the source as Ty2.
- /// * Insert/Extract element pass only the vector type as Ty1.
- /// * ShuffleVector, Load, Store do not use this call.
- virtual unsigned getInstrCost(unsigned Opcode,
- Type *Ty1 = 0,
- Type *Ty2 = 0) const {
- return 1;
- }
-
/// Returns the expected cost of arithmetic ops, such as mul, xor, fsub, etc.
virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const {
return 1;
Modified: llvm/trunk/lib/Target/TargetTransformImpl.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/TargetTransformImpl.cpp?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/lib/Target/TargetTransformImpl.cpp (original)
+++ llvm/trunk/lib/Target/TargetTransformImpl.cpp Sun Dec 23 01:23:55 2012
@@ -132,7 +132,6 @@
std::pair<unsigned, MVT>
VectorTargetTransformImpl::getTypeLegalizationCost(Type *Ty) const {
-
LLVMContext &C = Ty->getContext();
EVT MTy = TLI->getValueType(Ty);
@@ -271,7 +270,7 @@
return getScalarizationOverhead(Dst, true, true) + Num * Cost;
}
- // We already handled vector-to-vector and scalar-to-scalar conversions. This
+ // We already handled vector-to-vector and scalar-to-scalar conversions. This
// is where we handle bitcast between vectors and scalars. We need to assume
// that the conversion is scalarized in one way or another.
if (Opcode == Instruction::BitCast)
@@ -283,6 +282,7 @@
}
unsigned VectorTargetTransformImpl::getCFInstrCost(unsigned Opcode) const {
+ // Branches are assumed to be predicted.
return 0;
}
@@ -331,12 +331,6 @@
}
unsigned
-VectorTargetTransformImpl::getInstrCost(unsigned Opcode, Type *Ty1,
- Type *Ty2) const {
- return 1;
-}
-
-unsigned
VectorTargetTransformImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
unsigned Alignment,
unsigned AddressSpace) const {
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Sun Dec 23 01:23:55 2012
@@ -17988,7 +17988,6 @@
return VectorTargetTransformImpl::getArithmeticInstrCost(Opcode, Ty);
}
-
unsigned
X86VectorTargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
unsigned Alignment,
Modified: llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp (original)
+++ llvm/trunk/lib/Transforms/Vectorize/LoopVectorize.cpp Sun Dec 23 01:23:55 2012
@@ -2080,17 +2080,23 @@
VectorTy = ToVectorTy(ValTy, VF);
if (VF == 1)
- return VTTI->getMemoryOpCost(I->getOpcode(), ValTy,
+ return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy,
SI->getAlignment(),
SI->getPointerAddressSpace());
// Scalarized stores.
if (!Legal->isConsecutivePtr(SI->getPointerOperand())) {
unsigned Cost = 0;
- unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
- ValTy);
- // The cost of extracting from the value vector.
- Cost += VF * (ExtCost);
+
+ // The cost of extracting from the value vector and pointer vector.
+ Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
+ for (unsigned i = 0; i < VF; ++i) {
+ Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
+ VectorTy, i);
+ Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
+ PtrTy, i);
+ }
+
// The cost of the scalar stores.
Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
ValTy->getScalarType(),
@@ -2107,16 +2113,25 @@
LoadInst *LI = cast<LoadInst>(I);
if (VF == 1)
- return VTTI->getMemoryOpCost(I->getOpcode(), RetTy,
+ return VTTI->getMemoryOpCost(I->getOpcode(), VectorTy,
LI->getAlignment(),
LI->getPointerAddressSpace());
// Scalarized loads.
if (!Legal->isConsecutivePtr(LI->getPointerOperand())) {
unsigned Cost = 0;
- unsigned InCost = VTTI->getInstrCost(Instruction::InsertElement, RetTy);
- // The cost of inserting the loaded value into the result vector.
- Cost += VF * (InCost);
+ Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
+
+ // The cost of extracting from the pointer vector.
+ for (unsigned i = 0; i < VF; ++i)
+ Cost += VTTI->getVectorInstrCost(Instruction::ExtractElement,
+ PtrTy, i);
+
+ // The cost of inserting data to the result vector.
+ for (unsigned i = 0; i < VF; ++i)
+ Cost += VTTI->getVectorInstrCost(Instruction::InsertElement,
+ VectorTy, i);
+
// The cost of the scalar stores.
Cost += VF * VTTI->getMemoryOpCost(I->getOpcode(),
RetTy->getScalarType(),
@@ -2169,18 +2184,19 @@
bool IsVoid = RetTy->isVoidTy();
unsigned InsCost = (IsVoid ? 0 :
- VTTI->getInstrCost(Instruction::InsertElement,
+ VTTI->getVectorInstrCost(Instruction::InsertElement,
VectorTy));
- unsigned ExtCost = VTTI->getInstrCost(Instruction::ExtractElement,
+ unsigned ExtCost = VTTI->getVectorInstrCost(Instruction::ExtractElement,
VectorTy);
// The cost of inserting the results plus extracting each one of the
// operands.
Cost += VF * (InsCost + ExtCost * I->getNumOperands());
- // The cost of executing VF copies of the scalar instruction.
- Cost += VF * VTTI->getInstrCost(I->getOpcode(), RetTy);
+ // The cost of executing VF copies of the scalar instruction. This opcode
+ // is unknown. Assume that it is the same as 'mul'.
+ Cost += VF * VTTI->getArithmeticInstrCost(Instruction::Mul, VectorTy);
return Cost;
}
}// end of switch.
Modified: llvm/trunk/test/Transforms/LoopVectorize/X86/cost-model.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopVectorize/X86/cost-model.ll?rev=170995&r1=170994&r2=170995&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/LoopVectorize/X86/cost-model.ll (original)
+++ llvm/trunk/test/Transforms/LoopVectorize/X86/cost-model.ll Sun Dec 23 01:23:55 2012
@@ -8,8 +8,11 @@
@d = common global [2048 x i32] zeroinitializer, align 16
@a = common global [2048 x i32] zeroinitializer, align 16
+; The program below gathers and scatters data. We better not vectorize it.
;CHECK: cost_model_1
-;CHECK: <4 x i32>
+;CHECK-NOT: <2 x i32>
+;CHECK-NOT: <4 x i32>
+;CHECK-NOT: <8 x i32>
;CHECK: ret void
define void @cost_model_1() nounwind uwtable noinline ssp {
entry:
More information about the llvm-commits
mailing list