Instead of a vector load, shuffle and then extract an element. Load the element from address with an offset.
pshufd $1, (%rdi), %xmm0
movd %xmm0, %eax
=>
movl 4(%rdi), %eax
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@51026 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 684b2f6..8fe6eb7 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4682,49 +4682,82 @@
}
SDOperand DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) {
+ // (vextract (v4f32 load $addr), c) -> (f32 load $addr+c*size)
+ // (vextract (v4f32 s2v (f32 load $addr)), c) -> (f32 load $addr+c*size)
+ // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), 0) -> (f32 load $addr)
+
+ // Perform only after legalization to ensure build_vector / vector_shuffle
+ // optimizations have already been done.
+ if (!AfterLegalize) return SDOperand();
+
SDOperand InVec = N->getOperand(0);
SDOperand EltNo = N->getOperand(1);
- // (vextract (v4f32 s2v (f32 load $addr)), 0) -> (f32 load $addr)
- // (vextract (v4i32 bc (v4f32 s2v (f32 load $addr))), 0) -> (i32 load $addr)
if (isa<ConstantSDNode>(EltNo)) {
unsigned Elt = cast<ConstantSDNode>(EltNo)->getValue();
bool NewLoad = false;
- if (Elt == 0) {
- MVT::ValueType VT = InVec.getValueType();
- MVT::ValueType EVT = MVT::getVectorElementType(VT);
- MVT::ValueType LVT = EVT;
- unsigned NumElts = MVT::getVectorNumElements(VT);
- if (InVec.getOpcode() == ISD::BIT_CONVERT) {
- MVT::ValueType BCVT = InVec.getOperand(0).getValueType();
- if (!MVT::isVector(BCVT) ||
- NumElts != MVT::getVectorNumElements(BCVT))
- return SDOperand();
- InVec = InVec.getOperand(0);
- EVT = MVT::getVectorElementType(BCVT);
- NewLoad = true;
- }
- if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
- InVec.getOperand(0).getValueType() == EVT &&
- ISD::isNormalLoad(InVec.getOperand(0).Val) &&
- InVec.getOperand(0).hasOneUse()) {
- LoadSDNode *LN0 = cast<LoadSDNode>(InVec.getOperand(0));
- unsigned Align = LN0->getAlignment();
- if (NewLoad) {
- // Check the resultant load doesn't need a higher alignment than the
- // original load.
- unsigned NewAlign = TLI.getTargetMachine().getTargetData()->
- getABITypeAlignment(MVT::getTypeForValueType(LVT));
- if (!TLI.isOperationLegal(ISD::LOAD, LVT) || NewAlign > Align)
- return SDOperand();
- Align = NewAlign;
- }
+ MVT::ValueType VT = InVec.getValueType();
+ MVT::ValueType EVT = MVT::getVectorElementType(VT);
+ MVT::ValueType LVT = EVT;
+ if (InVec.getOpcode() == ISD::BIT_CONVERT) {
+ MVT::ValueType BCVT = InVec.getOperand(0).getValueType();
+ if (!MVT::isVector(BCVT)
+ || (MVT::getSizeInBits(EVT) >
+ MVT::getSizeInBits(MVT::getVectorElementType(BCVT))))
+ return SDOperand();
+ InVec = InVec.getOperand(0);
+ EVT = MVT::getVectorElementType(BCVT);
+ NewLoad = true;
+ }
- return DAG.getLoad(LVT, LN0->getChain(), LN0->getBasePtr(),
- LN0->getSrcValue(), LN0->getSrcValueOffset(),
- LN0->isVolatile(), Align);
+ LoadSDNode *LN0 = NULL;
+ if (ISD::isNormalLoad(InVec.Val))
+ LN0 = cast<LoadSDNode>(InVec);
+ else if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR &&
+ InVec.getOperand(0).getValueType() == EVT &&
+ ISD::isNormalLoad(InVec.getOperand(0).Val)) {
+ LN0 = cast<LoadSDNode>(InVec.getOperand(0));
+ } else if (InVec.getOpcode() == ISD::VECTOR_SHUFFLE) {
+ // (vextract (vector_shuffle (load $addr), v2, <1, u, u, u>), 1)
+ // =>
+ // (load $addr+1*size)
+ unsigned Idx = cast<ConstantSDNode>(InVec.getOperand(2).
+ getOperand(Elt))->getValue();
+ unsigned NumElems = InVec.getOperand(2).getNumOperands();
+ InVec = (Idx < NumElems) ? InVec.getOperand(0) : InVec.getOperand(1);
+ if (InVec.getOpcode() == ISD::BIT_CONVERT)
+ InVec = InVec.getOperand(0);
+ if (ISD::isNormalLoad(InVec.Val)) {
+ LN0 = cast<LoadSDNode>(InVec);
+ Elt = (Idx < NumElems) ? Idx : Idx - NumElems;
}
}
+ if (!LN0 || !LN0->hasOneUse())
+ return SDOperand();
+
+ unsigned Align = LN0->getAlignment();
+ if (NewLoad) {
+ // Check the resultant load doesn't need a higher alignment than the
+ // original load.
+ unsigned NewAlign = TLI.getTargetMachine().getTargetData()->
+ getABITypeAlignment(MVT::getTypeForValueType(LVT));
+ if (!TLI.isOperationLegal(ISD::LOAD, LVT) || NewAlign > Align)
+ return SDOperand();
+ Align = NewAlign;
+ }
+
+ SDOperand NewPtr = LN0->getBasePtr();
+ if (Elt) {
+ unsigned PtrOff = MVT::getSizeInBits(LVT) * Elt / 8;
+ MVT::ValueType PtrType = NewPtr.getValueType();
+ if (TLI.isBigEndian())
+ PtrOff = MVT::getSizeInBits(VT) / 8 - PtrOff;
+ NewPtr = DAG.getNode(ISD::ADD, PtrType, NewPtr,
+ DAG.getConstant(PtrOff, PtrType));
+ }
+ return DAG.getLoad(LVT, LN0->getChain(), NewPtr,
+ LN0->getSrcValue(), LN0->getSrcValueOffset(),
+ LN0->isVolatile(), Align);
}
return SDOperand();
}
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index f05f444..058e60f 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -1838,6 +1838,28 @@
}
+/// getShuffleScalarElt - Returns the scalar element that will make up the ith
+/// element of the result of the vector shuffle.
+SDOperand SelectionDAG::getShuffleScalarElt(const SDNode *N, unsigned Idx) {
+ MVT::ValueType VT = N->getValueType(0);
+ SDOperand PermMask = N->getOperand(2);
+ unsigned NumElems = PermMask.getNumOperands();
+ SDOperand V = (Idx < NumElems) ? N->getOperand(0) : N->getOperand(1);
+ Idx %= NumElems;
+ if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ return (Idx == 0)
+ ? V.getOperand(0) : getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
+ }
+ if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
+ SDOperand Elt = PermMask.getOperand(Idx);
+ if (Elt.getOpcode() == ISD::UNDEF)
+ return getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
+ return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Elt)->getValue());
+ }
+ return SDOperand();
+}
+
+
/// getNode - Gets or creates the specified node.
///
SDOperand SelectionDAG::getNode(unsigned Opcode, MVT::ValueType VT) {
diff --git a/lib/Target/X86/README-SSE.txt b/lib/Target/X86/README-SSE.txt
index c78e13b..5c681e0 100644
--- a/lib/Target/X86/README-SSE.txt
+++ b/lib/Target/X86/README-SSE.txt
@@ -545,35 +545,6 @@
//===---------------------------------------------------------------------===//
-These functions should produce the same code:
-
-#include <emmintrin.h>
-
-typedef long long __m128i __attribute__ ((__vector_size__ (16)));
-
-int foo(__m128i* val) {
- return __builtin_ia32_vec_ext_v4si(*val, 1);
-}
-int bar(__m128i* val) {
- union vs {
- __m128i *_v;
- int* _s;
- } v = {val};
- return v._s[1];
-}
-
-We currently produce (with -m64):
-
-_foo:
- pshufd $1, (%rdi), %xmm0
- movd %xmm0, %eax
- ret
-_bar:
- movl 4(%rdi), %eax
- ret
-
-//===---------------------------------------------------------------------===//
-
We should materialize vector constants like "all ones" and "signbit" with
code like:
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 806b626..4cc3f27 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -6182,26 +6182,6 @@
}
}
-/// getShuffleScalarElt - Returns the scalar element that will make up the ith
-/// element of the result of the vector shuffle.
-static SDOperand getShuffleScalarElt(SDNode *N, unsigned i, SelectionDAG &DAG) {
- MVT::ValueType VT = N->getValueType(0);
- SDOperand PermMask = N->getOperand(2);
- unsigned NumElems = PermMask.getNumOperands();
- SDOperand V = (i < NumElems) ? N->getOperand(0) : N->getOperand(1);
- i %= NumElems;
- if (V.getOpcode() == ISD::SCALAR_TO_VECTOR) {
- return (i == 0)
- ? V.getOperand(0) : DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
- } else if (V.getOpcode() == ISD::VECTOR_SHUFFLE) {
- SDOperand Idx = PermMask.getOperand(i);
- if (Idx.getOpcode() == ISD::UNDEF)
- return DAG.getNode(ISD::UNDEF, MVT::getVectorElementType(VT));
- return getShuffleScalarElt(V.Val,cast<ConstantSDNode>(Idx)->getValue(),DAG);
- }
- return SDOperand();
-}
-
/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
/// node is a GlobalAddress + offset.
bool X86TargetLowering::isGAPlusOffset(SDNode *N,
@@ -6240,7 +6220,7 @@
}
unsigned Index = cast<ConstantSDNode>(Idx)->getValue();
- SDOperand Elt = getShuffleScalarElt(N, Index, DAG);
+ SDOperand Elt = DAG.getShuffleScalarElt(N, Index);
if (!Elt.Val ||
(Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.Val)))
return false;