Improved widening loads by adding support for wider loads if
the alignment allows. Fixed a bug where we didn't use a
vector load/store for PR5626.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@94338 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 11e07df..ab51369 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -747,6 +747,12 @@
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
+ setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
+
// Custom lower build_vector, vector_shuffle, and extract_vector_elt.
for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
EVT VT = (MVT::SimpleValueType)i;
@@ -3686,6 +3692,33 @@
return SDValue();
}
+SDValue
+X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+ // We support concatenate two MMX registers and place them in a MMX
+ // register. This is better than doing a stack convert.
+ DebugLoc dl = Op.getDebugLoc();
+ EVT ResVT = Op.getValueType();
+ assert(Op.getNumOperands() == 2);
+ assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
+ ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
+ int Mask[2];
+ SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
+ SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
+ InVec = Op.getOperand(1);
+ if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ unsigned NumElts = ResVT.getVectorNumElements();
+ VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+ VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
+ InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
+ } else {
+ InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
+ SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
+ Mask[0] = 0; Mask[1] = 2;
+ VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
+ }
+ return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+}
+
// v8i16 shuffles - Prefer shuffles in the following order:
// 1. [all] pshuflw, pshufhw, optional move
// 2. [ssse3] 1 x pshufb
@@ -7238,6 +7271,7 @@
case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
+ case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 64bc70c..11b5225 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -156,6 +156,11 @@
/// relative displacements.
WrapperRIP,
+ /// MOVQ2DQ - Copies a 64-bit value from a vector to another vector.
+ /// Can be used to move a vector value from a MMX register to a XMM
+ /// register.
+ MOVQ2DQ,
+
/// PEXTRB - Extract an 8-bit value from a vector and zero extend it to
/// i32, corresponds to X86::PEXTRB.
PEXTRB,
@@ -634,6 +639,7 @@
SDValue LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
SelectionDAG &DAG);
SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG);
+ SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG);
SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG);
SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG);
SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG);
diff --git a/lib/Target/X86/X86InstrMMX.td b/lib/Target/X86/X86InstrMMX.td
index fc40c9a..ab169ac 100644
--- a/lib/Target/X86/X86InstrMMX.td
+++ b/lib/Target/X86/X86InstrMMX.td
@@ -501,6 +501,20 @@
(iPTR imm:$src3))))]>;
}
+// MMX to XMM for vector types
+def MMX_X86movq2dq : SDNode<"X86ISD::MOVQ2DQ", SDTypeProfile<1, 1,
+ [SDTCisVT<0, v2i64>, SDTCisVT<1, v1i64>]>>;
+
+def : Pat<(v2i64 (MMX_X86movq2dq VR64:$src)),
+ (v2i64 (MMX_MOVQ2DQrr VR64:$src))>;
+
+def : Pat<(v2i64 (MMX_X86movq2dq (load_mmx addr:$src))),
+ (v2i64 (MOVQI2PQIrm addr:$src))>;
+
+def : Pat<(v2i64 (MMX_X86movq2dq (v1i64 (bitconvert
+ (v2i32 (scalar_to_vector (loadi32 addr:$src))))))),
+ (v2i64 (MOVDI2PDIrm addr:$src))>;
+
// Mask creation
def MMX_PMOVMSKBrr : MMXI<0xD7, MRMSrcReg, (outs GR32:$dst), (ins VR64:$src),
"pmovmskb\t{$src, $dst|$dst, $src}",