summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorBob Wilson <bob.wilson@apple.com>2010-11-23 19:38:38 +0000
committerBob Wilson <bob.wilson@apple.com>2010-11-23 19:38:38 +0000
commit626613d5e8805db1f40e956faf47c87bea3aee44 (patch)
treefbec1eec6cfd3337dc2ea1afc3c43bf90302cd96
parentb055f740bd25317fabd1065f2c95bc4f03cc1b5f (diff)
downloadllvm-626613d5e8805db1f40e956faf47c87bea3aee44.tar.gz
llvm-626613d5e8805db1f40e956faf47c87bea3aee44.tar.bz2
llvm-626613d5e8805db1f40e956faf47c87bea3aee44.tar.xz
Recognize sign/zero-extended constant BUILD_VECTORs for VMULL operations.
We need to check if the individual vector elements are sign/zero-extended values. For now this only handles constants values. Radar 8687140. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@120034 91177308-0d34-0410-b5e6-96231b3b80d8
-rw-r--r--lib/Target/ARM/ARMISelLowering.cpp122
-rw-r--r--test/CodeGen/ARM/vmul.ll72
2 files changed, 181 insertions, 13 deletions
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 37d7fccbd7..507356a768 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -3860,15 +3860,114 @@ static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
}
-/// SkipExtension - For a node that is either a SIGN_EXTEND, ZERO_EXTEND, or
-/// an extending load, return the unextended value.
+/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
+/// element has been zero/sign-extended, depending on the isSigned parameter,
+/// from an integer type half its size.
+static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
+ bool isSigned) {
+ // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
+ EVT VT = N->getValueType(0);
+ if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
+ SDNode *BVN = N->getOperand(0).getNode();
+ if (BVN->getValueType(0) != MVT::v4i32 ||
+ BVN->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+ unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
+ unsigned HiElt = 1 - LoElt;
+ ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
+ ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
+ ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
+ ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
+ if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
+ return false;
+ if (isSigned) {
+ if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
+ Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
+ return true;
+ } else {
+ if (Hi0->isNullValue() && Hi1->isNullValue())
+ return true;
+ }
+ return false;
+ }
+
+ if (N->getOpcode() != ISD::BUILD_VECTOR)
+ return false;
+
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ SDNode *Elt = N->getOperand(i).getNode();
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+ unsigned HalfSize = EltSize / 2;
+ if (isSigned) {
+ int64_t SExtVal = C->getSExtValue();
+ if ((SExtVal >> HalfSize) != (SExtVal >> EltSize))
+ return false;
+ } else {
+ if ((C->getZExtValue() >> HalfSize) != 0)
+ return false;
+ }
+ continue;
+ }
+ return false;
+ }
+
+ return true;
+}
+
+/// isSignExtended - Check if a node is a vector value that is sign-extended
+/// or a constant BUILD_VECTOR with sign-extended elements.
+static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, true))
+ return true;
+ return false;
+}
+
+/// isZeroExtended - Check if a node is a vector value that is zero-extended
+/// or a constant BUILD_VECTOR with zero-extended elements.
+static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
+ if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
+ return true;
+ if (isExtendedBUILD_VECTOR(N, DAG, false))
+ return true;
+ return false;
+}
+
+/// SkipExtension - For a node that is a SIGN_EXTEND, ZERO_EXTEND, extending
+/// load, or BUILD_VECTOR with extended elements, return the unextended value.
static SDValue SkipExtension(SDNode *N, SelectionDAG &DAG) {
if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
return N->getOperand(0);
- LoadSDNode *LD = cast<LoadSDNode>(N);
- return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
- LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
- LD->isNonTemporal(), LD->getAlignment());
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
+ return DAG.getLoad(LD->getMemoryVT(), N->getDebugLoc(), LD->getChain(),
+ LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
+ LD->isNonTemporal(), LD->getAlignment());
+ // Otherwise, the value must be a BUILD_VECTOR. For v2i64, it will
+ // have been legalized as a BITCAST from v4i32.
+ if (N->getOpcode() == ISD::BITCAST) {
+ SDNode *BVN = N->getOperand(0).getNode();
+ assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
+ BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
+ unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
+ BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
+ }
+ // Construct a new BUILD_VECTOR with elements truncated to half the size.
+ assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
+ EVT VT = N->getValueType(0);
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
+ unsigned NumElts = VT.getVectorNumElements();
+ MVT TruncVT = MVT::getIntegerVT(EltSize);
+ SmallVector<SDValue, 8> Ops;
+ for (unsigned i = 0; i != NumElts; ++i) {
+ ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
+ const APInt &CInt = C->getAPIntValue();
+ Ops.push_back(DAG.getConstant(APInt(CInt).trunc(EltSize), TruncVT));
+ }
+ return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
+ MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
}
static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
@@ -3879,19 +3978,16 @@ static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
SDNode *N0 = Op.getOperand(0).getNode();
SDNode *N1 = Op.getOperand(1).getNode();
unsigned NewOpc = 0;
- if ((N0->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N0)) &&
- (N1->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N1))) {
+ if (isSignExtended(N0, DAG) && isSignExtended(N1, DAG))
NewOpc = ARMISD::VMULLs;
- } else if ((N0->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N0)) &&
- (N1->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N1))) {
+ else if (isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG))
NewOpc = ARMISD::VMULLu;
- } else if (VT == MVT::v2i64) {
+ else if (VT == MVT::v2i64)
// Fall through to expand this. It is not legal.
return SDValue();
- } else {
+ else
// Other vector multiplications are legal.
return Op;
- }
// Legalize to a VMULL instruction.
DebugLoc DL = Op.getDebugLoc();
diff --git a/test/CodeGen/ARM/vmul.ll b/test/CodeGen/ARM/vmul.ll
index 5383425018..ee033caa00 100644
--- a/test/CodeGen/ARM/vmul.ll
+++ b/test/CodeGen/ARM/vmul.ll
@@ -267,3 +267,75 @@ entry:
}
declare <8 x i16> @llvm.arm.neon.vmullp.v8i16(<8 x i8>, <8 x i8>) nounwind readnone
+
+
+; Radar 8687140
+; VMULL needs to recognize BUILD_VECTORs with sign/zero-extended elements.
+
+define <8 x i16> @vmull_extvec_s8(<8 x i8> %arg) nounwind {
+; CHECK: vmull_extvec_s8
+; CHECK: vmull.s8
+ %tmp3 = sext <8 x i8> %arg to <8 x i16>
+ %tmp4 = mul <8 x i16> %tmp3, <i16 -12, i16 -12, i16 -12, i16 -12, i16 -12, i16 -12, i16 -12, i16 -12>
+ ret <8 x i16> %tmp4
+}
+
+define <8 x i16> @vmull_extvec_u8(<8 x i8> %arg) nounwind {
+; CHECK: vmull_extvec_u8
+; CHECK: vmull.u8
+ %tmp3 = zext <8 x i8> %arg to <8 x i16>
+ %tmp4 = mul <8 x i16> %tmp3, <i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12, i16 12>
+ ret <8 x i16> %tmp4
+}
+
+define <8 x i16> @vmull_noextvec_s8(<8 x i8> %arg) nounwind {
+; Do not use VMULL if the BUILD_VECTOR element values are too big.
+; CHECK: vmull_noextvec_s8
+; CHECK: vmovl.s8
+; CHECK: vmul.i16
+ %tmp3 = sext <8 x i8> %arg to <8 x i16>
+ %tmp4 = mul <8 x i16> %tmp3, <i16 -999, i16 -999, i16 -999, i16 -999, i16 -999, i16 -999, i16 -999, i16 -999>
+ ret <8 x i16> %tmp4
+}
+
+define <8 x i16> @vmull_noextvec_u8(<8 x i8> %arg) nounwind {
+; Do not use VMULL if the BUILD_VECTOR element values are too big.
+; CHECK: vmull_noextvec_u8
+; CHECK: vmovl.u8
+; CHECK: vmul.i16
+ %tmp3 = zext <8 x i8> %arg to <8 x i16>
+ %tmp4 = mul <8 x i16> %tmp3, <i16 999, i16 999, i16 999, i16 999, i16 999, i16 999, i16 999, i16 999>
+ ret <8 x i16> %tmp4
+}
+
+define <4 x i32> @vmull_extvec_s16(<4 x i16> %arg) nounwind {
+; CHECK: vmull_extvec_s16
+; CHECK: vmull.s16
+ %tmp3 = sext <4 x i16> %arg to <4 x i32>
+ %tmp4 = mul <4 x i32> %tmp3, <i32 -12, i32 -12, i32 -12, i32 -12>
+ ret <4 x i32> %tmp4
+}
+
+define <4 x i32> @vmull_extvec_u16(<4 x i16> %arg) nounwind {
+; CHECK: vmull_extvec_u16
+; CHECK: vmull.u16
+ %tmp3 = zext <4 x i16> %arg to <4 x i32>
+ %tmp4 = mul <4 x i32> %tmp3, <i32 1234, i32 1234, i32 1234, i32 1234>
+ ret <4 x i32> %tmp4
+}
+
+define <2 x i64> @vmull_extvec_s32(<2 x i32> %arg) nounwind {
+; CHECK: vmull_extvec_s32
+; CHECK: vmull.s32
+ %tmp3 = sext <2 x i32> %arg to <2 x i64>
+ %tmp4 = mul <2 x i64> %tmp3, <i64 -1234, i64 -1234>
+ ret <2 x i64> %tmp4
+}
+
+define <2 x i64> @vmull_extvec_u32(<2 x i32> %arg) nounwind {
+; CHECK: vmull_extvec_u32
+; CHECK: vmull.u32
+ %tmp3 = zext <2 x i32> %arg to <2 x i64>
+ %tmp4 = mul <2 x i64> %tmp3, <i64 1234, i64 1234>
+ ret <2 x i64> %tmp4
+}