Given the following C code llvm currently generates suboptimal code for
x86-64:
__m128 bss4( const __m128 *ptr, size_t i, size_t j )
{
float f = ptr[i][j];
return (__m128) { f, f, f, f };
}
=================================================
define <4 x float> @_Z4bss4PKDv4_fmm(<4 x float>* nocapture readonly %ptr, i64 %i, i64 %j) #0 {
%a1 = getelementptr inbounds <4 x float>* %ptr, i64 %i
%a2 = load <4 x float>* %a1, align 16, !tbaa !1
%a3 = trunc i64 %j to i32
%a4 = extractelement <4 x float> %a2, i32 %a3
%a5 = insertelement <4 x float> undef, float %a4, i32 0
%a6 = insertelement <4 x float> %a5, float %a4, i32 1
%a7 = insertelement <4 x float> %a6, float %a4, i32 2
%a8 = insertelement <4 x float> %a7, float %a4, i32 3
ret <4 x float> %a8
}
=================================================
shlq $4, %rsi
addq %rdi, %rsi
movslq %edx, %rax
vbroadcastss (%rsi,%rax,4), %xmm0
retq
=================================================
The movslq is uneeded, but is present because of the trunc to i32 and then
sext back to i64 that the backend adds for vbroadcastss.
We can't remove it because it changes the meaning. The IR that clang
generates is already suboptimal. What clang really should emit is:
%a4 = extractelement <4 x float> %a2, i64 %j
This patch makes that legal. A separate patch will teach clang to do it.
Differential Revision: http://reviews.llvm.org/D3519
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@207801
91177308-0d34-0410-b5e6-
96231b3b80d8
::
- <result> = extractelement <n x <ty>> <val>, i32 <idx> ; yields <ty>
+ <result> = extractelement <n x <ty>> <val>, <ty2> <idx> ; yields <ty>
Overview:
"""""""""
The first operand of an '``extractelement``' instruction is a value of
:ref:`vector <t_vector>` type. The second operand is an index indicating
the position from which to extract the element. The index may be a
-variable.
+variable of any integer type.
Semantics:
""""""""""
::
- <result> = insertelement <n x <ty>> <val>, <ty> <elt>, i32 <idx> ; yields <n x <ty>>
+ <result> = insertelement <n x <ty>> <val>, <ty> <elt>, <ty2> <idx> ; yields <n x <ty>>
Overview:
"""""""""
:ref:`vector <t_vector>` type. The second operand is a scalar value whose
type must equal the element type of the first operand. The third operand
is an index indicating the position at which to insert the value. The
-index may be a variable.
+index may be a variable of any integer type.
Semantics:
""""""""""
ValueList.getConstantFwdRef(Record[2],CurTy));
break;
}
- case bitc::CST_CODE_CE_EXTRACTELT: { // CE_EXTRACTELT: [opty, opval, opval]
+ case bitc::CST_CODE_CE_EXTRACTELT
+ : { // CE_EXTRACTELT: [opty, opval, opty, opval]
if (Record.size() < 3)
return Error(InvalidRecord);
VectorType *OpTy =
if (!OpTy)
return Error(InvalidRecord);
Constant *Op0 = ValueList.getConstantFwdRef(Record[1], OpTy);
- Constant *Op1 = ValueList.getConstantFwdRef(Record[2],
- Type::getInt32Ty(Context));
+ Constant *Op1 = nullptr;
+ if (Record.size() == 4) {
+ Type *IdxTy = getTypeByID(Record[2]);
+ if (!IdxTy)
+ return Error(InvalidRecord);
+ Op1 = ValueList.getConstantFwdRef(Record[3], IdxTy);
+ } else // TODO: Remove with llvm 4.0
+ Op1 = ValueList.getConstantFwdRef(Record[2], Type::getInt32Ty(Context));
+ if (!Op1)
+ return Error(InvalidRecord);
V = ConstantExpr::getExtractElement(Op0, Op1);
break;
}
- case bitc::CST_CODE_CE_INSERTELT: { // CE_INSERTELT: [opval, opval, opval]
+ case bitc::CST_CODE_CE_INSERTELT
+ : { // CE_INSERTELT: [opval, opval, opty, opval]
VectorType *OpTy = dyn_cast<VectorType>(CurTy);
if (Record.size() < 3 || !OpTy)
return Error(InvalidRecord);
Constant *Op0 = ValueList.getConstantFwdRef(Record[0], OpTy);
Constant *Op1 = ValueList.getConstantFwdRef(Record[1],
OpTy->getElementType());
- Constant *Op2 = ValueList.getConstantFwdRef(Record[2],
- Type::getInt32Ty(Context));
+ Constant *Op2 = nullptr;
+ if (Record.size() == 4) {
+ Type *IdxTy = getTypeByID(Record[2]);
+ if (!IdxTy)
+ return Error(InvalidRecord);
+ Op2 = ValueList.getConstantFwdRef(Record[3], IdxTy);
+ } else // TODO: Remove with llvm 4.0
+ Op2 = ValueList.getConstantFwdRef(Record[2], Type::getInt32Ty(Context));
+ if (!Op2)
+ return Error(InvalidRecord);
V = ConstantExpr::getInsertElement(Op0, Op1, Op2);
break;
}
unsigned OpNum = 0;
Value *Vec, *Idx;
if (getValueTypePair(Record, OpNum, NextValueNo, Vec) ||
- popValue(Record, OpNum, NextValueNo, Type::getInt32Ty(Context), Idx))
+ getValueTypePair(Record, OpNum, NextValueNo, Idx))
return Error(InvalidRecord);
I = ExtractElementInst::Create(Vec, Idx);
InstructionList.push_back(I);
if (getValueTypePair(Record, OpNum, NextValueNo, Vec) ||
popValue(Record, OpNum, NextValueNo,
cast<VectorType>(Vec->getType())->getElementType(), Elt) ||
- popValue(Record, OpNum, NextValueNo, Type::getInt32Ty(Context), Idx))
+ getValueTypePair(Record, OpNum, NextValueNo, Idx))
return Error(InvalidRecord);
I = InsertElementInst::Create(Vec, Elt, Idx);
InstructionList.push_back(I);
Code = bitc::CST_CODE_CE_EXTRACTELT;
Record.push_back(VE.getTypeID(C->getOperand(0)->getType()));
Record.push_back(VE.getValueID(C->getOperand(0)));
+ Record.push_back(VE.getTypeID(C->getOperand(1)->getType()));
Record.push_back(VE.getValueID(C->getOperand(1)));
break;
case Instruction::InsertElement:
Code = bitc::CST_CODE_CE_INSERTELT;
Record.push_back(VE.getValueID(C->getOperand(0)));
Record.push_back(VE.getValueID(C->getOperand(1)));
+ Record.push_back(VE.getTypeID(C->getOperand(2)->getType()));
Record.push_back(VE.getValueID(C->getOperand(2)));
break;
case Instruction::ShuffleVector:
case Instruction::ExtractElement:
Code = bitc::FUNC_CODE_INST_EXTRACTELT;
PushValueAndType(I.getOperand(0), InstID, Vals, VE);
- pushValue(I.getOperand(1), InstID, Vals, VE);
+ PushValueAndType(I.getOperand(1), InstID, Vals, VE);
break;
case Instruction::InsertElement:
Code = bitc::FUNC_CODE_INST_INSERTELT;
PushValueAndType(I.getOperand(0), InstID, Vals, VE);
pushValue(I.getOperand(1), InstID, Vals, VE);
- pushValue(I.getOperand(2), InstID, Vals, VE);
+ PushValueAndType(I.getOperand(2), InstID, Vals, VE);
break;
case Instruction::ShuffleVector:
Code = bitc::FUNC_CODE_INST_SHUFFLEVEC;
Constant *ConstantExpr::getExtractElement(Constant *Val, Constant *Idx) {
assert(Val->getType()->isVectorTy() &&
"Tried to create extractelement operation on non-vector type!");
- assert(Idx->getType()->isIntegerTy(32) &&
- "Extractelement index must be i32 type!");
+ assert(Idx->getType()->isIntegerTy() &&
+ "Extractelement index must be an integer type!");
if (Constant *FC = ConstantFoldExtractElementInstruction(Val, Idx))
return FC; // Fold a few common cases.
"Tried to create insertelement operation on non-vector type!");
assert(Elt->getType() == Val->getType()->getVectorElementType() &&
"Insertelement types must match!");
- assert(Idx->getType()->isIntegerTy(32) &&
+ assert(Idx->getType()->isIntegerTy() &&
"Insertelement index must be i32 type!");
if (Constant *FC = ConstantFoldInsertElementInstruction(Val, Elt, Idx))
bool ExtractElementInst::isValidOperands(const Value *Val, const Value *Index) {
- if (!Val->getType()->isVectorTy() || !Index->getType()->isIntegerTy(32))
+ if (!Val->getType()->isVectorTy() || !Index->getType()->isIntegerTy())
return false;
return true;
}
if (Elt->getType() != cast<VectorType>(Vec->getType())->getElementType())
return false;// Second operand of insertelement must be vector element type.
- if (!Index->getType()->isIntegerTy(32))
+ if (!Index->getType()->isIntegerTy())
return false; // Third operand of insertelement must be i32.
return true;
}
; SSE3-LABEL: test_v2sd:
; SSE3: movddup
}
+
+; Fold extract of a load into the load's address computation. This avoids spilling to the stack.
+define <4 x float> @load_extract_splat(<4 x float>* nocapture readonly %ptr, i64 %i, i64 %j) nounwind {
+ %1 = getelementptr inbounds <4 x float>* %ptr, i64 %i
+ %2 = load <4 x float>* %1, align 16
+ %3 = extractelement <4 x float> %2, i64 %j
+ %4 = insertelement <4 x float> undef, float %3, i32 0
+ %5 = insertelement <4 x float> %4, float %3, i32 1
+ %6 = insertelement <4 x float> %5, float %3, i32 2
+ %7 = insertelement <4 x float> %6, float %3, i32 3
+ ret <4 x float> %7
+
+; AVX-LABEL: load_extract_splat
+; AVX-NOT: movs
+; AVX: vbroadcastss
+}
define i32 @test_extractelement(<4 x i32> %V) {
%R = extractelement <4 x i32> %V, i32 1 ; <i32> [#uses=1]
+ %S = extractelement <4 x i32> %V, i64 1 ; <i32> [#uses=0]
ret i32 %R
}
define <4 x i32> @test_insertelement(<4 x i32> %V) {
%R = insertelement <4 x i32> %V, i32 0, i32 0 ; <<4 x i32>> [#uses=1]
+ %S = insertelement <4 x i32> %V, i32 0, i64 0 ; <<4 x i32>> [#uses=0]
ret <4 x i32> %R
}