[llvm] r204071 - Make DAGCombiner work on vector bitshifts with constant splat vectors.
Andrea Di Biagio
andrea.dibiagio at gmail.com
Fri Mar 21 17:22:21 PDT 2014
Hi Rafael and Matt,
I have already sent a patch for review to fix this problem.
(Search for an email with subject: "[PATCH][DAG] Fix assertion failure
caused by an invalid cast in method
'BuildVectorSDNode::isConstantSplat'").
The patch I have sent for review is this one btw.
On Fri, Mar 21, 2014 at 11:15 PM, Rafael Espíndola
<rafael.espindola at gmail.com> wrote:
> Testcase attached.
>
> Matt, do you want me to open a bug with it?
>
> On 18 March 2014 02:13, Patrik Hägglund H
> <patrik.h.hagglund at ericsson.com> wrote:
>> Hi Matt,
>>
>> This commit makes the following test to fail on x86-64:
>>
>> bin/llvm-stress -size 185 -seed 3115 | bin/llc -mcpu=generic -o /dev/null
>>
>> llc: ../include/llvm/Support/Casting.h:230: typename cast_retty<X, Y>::ret_type llvm::cast(Y &) [X = llvm::ConstantSDNode, Y = llvm::SDValue]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
>> 0 llc 0x0000000000fada05 llvm::sys::PrintStackTrace(_IO_FILE*) + 37
>> 1 llc 0x0000000000fade43
>> 2 libpthread.so.0 0x00007f258b67e7c0
>> 3 libc.so.6 0x00007f258a982b35 gsignal + 53
>> 4 libc.so.6 0x00007f258a984111 abort + 385
>> 5 libc.so.6 0x00007f258a97b9f0 __assert_fail + 240
>> 6 llc 0x0000000000e6a14b
>> 7 llc 0x0000000000dd4d93
>> 8 llc 0x0000000000dcb28c
>> 9 llc 0x0000000000dcac2b llvm::SelectionDAG::Combine(llvm::CombineLevel, llvm::AliasAnalysis&, llvm::CodeGenOpt::Level) + 827
>> 10 llc 0x0000000000eb20bf llvm::SelectionDAGISel::CodeGenAndEmitDAG() + 1791
>> 11 llc 0x0000000000eb1078 llvm::SelectionDAGISel::SelectAllBasicBlocks(llvm::Function const&) + 6504
>> 12 llc 0x0000000000eae974 llvm::SelectionDAGISel::runOnMachineFunction(llvm::MachineFunction&) + 1332
>> 13 llc 0x0000000000b46e6c llvm::MachineFunctionPass::runOnFunction(llvm::Function&) + 124
>> 14 llc 0x0000000000d28eda llvm::FPPassManager::runOnFunction(llvm::Function&) + 362
>> 15 llc 0x0000000000d2916b llvm::FPPassManager::runOnModule(llvm::Module&) + 43
>> 16 llc 0x0000000000d29707 llvm::legacy::PassManagerImpl::run(llvm::Module&) + 999
>> 17 llc 0x000000000055903b main + 6331
>> 18 libc.so.6 0x00007f258a96ec16 __libc_start_main + 230
>> 19 llc 0x0000000000557699
>> Stack dump:
>> 0. Program arguments: bin/llc -mcpu=generic -o /dev/null
>> 1. Running pass 'Function Pass Manager' on module '<stdin>'.
>> 2. Running pass 'X86 DAG->DAG Instruction Selection' on function '@autogen_SD3115'
>>
>> /Patrik Hägglund
>>
>> -----Original Message-----
>> From: llvm-commits-bounces at cs.uiuc.edu [mailto:llvm-commits-bounces at cs.uiuc.edu] On Behalf Of Matt Arsenault
>> Sent: den 17 mars 2014 19:58
>> To: llvm-commits at cs.uiuc.edu
>> Subject: [llvm] r204071 - Make DAGCombiner work on vector bitshifts with constant splat vectors.
>>
>> Author: arsenm
>> Date: Mon Mar 17 13:58:01 2014
>> New Revision: 204071
>>
>> URL: http://llvm.org/viewvc/llvm-project?rev=204071&view=rev
>> Log:
>> Make DAGCombiner work on vector bitshifts with constant splat vectors.
>>
>> Modified:
>> llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
>> llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
>> llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
>> llvm/trunk/test/CodeGen/AArch64/neon-shl-ashr-lshr.ll
>> llvm/trunk/test/CodeGen/X86/avx2-vector-shifts.ll
>> llvm/trunk/test/CodeGen/X86/sse2-vector-shifts.ll
>>
>> Modified: llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h (original)
>> +++ llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h Mon Mar 17 13:58:01 2014
>> @@ -1522,6 +1522,11 @@ public:
>> unsigned MinSplatBits = 0,
>> bool isBigEndian = false) const;
>>
>> + /// isConstantSplat - Simpler form of isConstantSplat. Get the constant splat
>> + /// when you only care about the value. Returns nullptr if this isn't a
>> + /// constant splat vector.
>> + ConstantSDNode *isConstantSplat() const;
>> +
>> bool isConstant() const;
>>
>> static inline bool classof(const SDNode *N) {
>>
>> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (original)
>> +++ llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp Mon Mar 17 13:58:01 2014
>> @@ -280,7 +280,7 @@ namespace {
>> SDValue XformToShuffleWithZero(SDNode *N);
>> SDValue ReassociateOps(unsigned Opc, SDLoc DL, SDValue LHS, SDValue RHS);
>>
>> - SDValue visitShiftByConstant(SDNode *N, unsigned Amt);
>> + SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
>>
>> bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
>> SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);
>> @@ -634,7 +634,23 @@ static bool isOneUseSetCC(SDValue N) {
>> return false;
>> }
>>
>> -// \brief Returns the SDNode if it is a constant BuildVector or constant int.
>> +/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
>> +/// elements are all the same constant or undefined.
>> +static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
>> + BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
>> + if (!C)
>> + return false;
>> +
>> + APInt SplatUndef;
>> + unsigned SplatBitSize;
>> + bool HasAnyUndefs;
>> + EVT EltVT = N->getValueType(0).getVectorElementType();
>> + return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
>> + HasAnyUndefs) &&
>> + EltVT.getSizeInBits() >= SplatBitSize);
>> +}
>> +
>> +// \brief Returns the SDNode if it is a constant BuildVector or constant.
>> static SDNode *isConstantBuildVectorOrConstantInt(SDValue N) {
>> if (isa<ConstantSDNode>(N))
>> return N.getNode();
>> @@ -644,6 +660,18 @@ static SDNode *isConstantBuildVectorOrCo
>> return NULL;
>> }
>>
>> +// \brief Returns the SDNode if it is a constant splat BuildVector or constant
>> +// int.
>> +static ConstantSDNode *isConstOrConstSplat(SDValue N) {
>> + if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N))
>> + return CN;
>> +
>> + if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N))
>> + return BV->isConstantSplat();
>> +
>> + return nullptr;
>> +}
>> +
>> SDValue DAGCombiner::ReassociateOps(unsigned Opc, SDLoc DL,
>> SDValue N0, SDValue N1) {
>> EVT VT = N0.getValueType();
>> @@ -1830,22 +1858,6 @@ SDValue DAGCombiner::visitSUBE(SDNode *N
>> return SDValue();
>> }
>>
>> -/// isConstantSplatVector - Returns true if N is a BUILD_VECTOR node whose
>> -/// elements are all the same constant or undefined.
>> -static bool isConstantSplatVector(SDNode *N, APInt& SplatValue) {
>> - BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(N);
>> - if (!C)
>> - return false;
>> -
>> - APInt SplatUndef;
>> - unsigned SplatBitSize;
>> - bool HasAnyUndefs;
>> - EVT EltVT = N->getValueType(0).getVectorElementType();
>> - return (C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
>> - HasAnyUndefs) &&
>> - EltVT.getSizeInBits() >= SplatBitSize);
>> -}
>> -
>> SDValue DAGCombiner::visitMUL(SDNode *N) {
>> SDValue N0 = N->getOperand(0);
>> SDValue N1 = N->getOperand(1);
>> @@ -3805,11 +3817,9 @@ SDValue DAGCombiner::visitXOR(SDNode *N)
>>
>> /// visitShiftByConstant - Handle transforms common to the three shifts, when
>> /// the shift amount is a constant.
>> -SDValue DAGCombiner::visitShiftByConstant(SDNode *N, unsigned Amt) {
>> - assert(isa<ConstantSDNode>(N->getOperand(1)) &&
>> - "Expected an ConstantSDNode operand.");
>> +SDValue DAGCombiner::visitShiftByConstant(SDNode *N, ConstantSDNode *Amt) {
>> // We can't and shouldn't fold opaque constants.
>> - if (cast<ConstantSDNode>(N->getOperand(1))->isOpaque())
>> + if (Amt->isOpaque())
>> return SDValue();
>>
>> SDNode *LHS = N->getOperand(0).getNode();
>> @@ -3888,11 +3898,11 @@ SDValue DAGCombiner::distributeTruncateT
>> if (N->hasOneUse() && N->getOperand(0).hasOneUse()) {
>> SDValue N01 = N->getOperand(0).getOperand(1);
>>
>> - if (ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N01)) {
>> + if (ConstantSDNode *N01C = isConstOrConstSplat(N01)) {
>> EVT TruncVT = N->getValueType(0);
>> SDValue N00 = N->getOperand(0).getOperand(0);
>> APInt TruncC = N01C->getAPIntValue();
>> - TruncC = TruncC.trunc(TruncVT.getScalarType().getSizeInBits());
>> + TruncC = TruncC.trunc(TruncVT.getScalarSizeInBits());
>>
>> return DAG.getNode(ISD::AND, SDLoc(N), TruncVT,
>> DAG.getNode(ISD::TRUNCATE, SDLoc(N), TruncVT, N00),
>> @@ -3921,7 +3931,7 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0);
>> ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
>> EVT VT = N0.getValueType();
>> - unsigned OpSizeInBits = VT.getScalarType().getSizeInBits();
>> + unsigned OpSizeInBits = VT.getScalarSizeInBits();
>>
>> // fold vector ops
>> if (VT.isVector()) {
>> @@ -3931,18 +3941,21 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> BuildVectorSDNode *N1CV = dyn_cast<BuildVectorSDNode>(N1);
>> // If setcc produces all-one true value then:
>> // (shl (and (setcc) N01CV) N1CV) -> (and (setcc) N01CV<<N1CV)
>> - if (N1CV && N1CV->isConstant() &&
>> - TLI.getBooleanContents(true) ==
>> - TargetLowering::ZeroOrNegativeOneBooleanContent &&
>> - N0.getOpcode() == ISD::AND) {
>> - SDValue N00 = N0->getOperand(0);
>> - SDValue N01 = N0->getOperand(1);
>> - BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
>> -
>> - if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC) {
>> - SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
>> - if (C.getNode())
>> - return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
>> + if (N1CV && N1CV->isConstant()) {
>> + if (N0.getOpcode() == ISD::AND &&
>> + TLI.getBooleanContents(true) ==
>> + TargetLowering::ZeroOrNegativeOneBooleanContent) {
>> + SDValue N00 = N0->getOperand(0);
>> + SDValue N01 = N0->getOperand(1);
>> + BuildVectorSDNode *N01CV = dyn_cast<BuildVectorSDNode>(N01);
>> +
>> + if (N01CV && N01CV->isConstant() && N00.getOpcode() == ISD::SETCC) {
>> + SDValue C = DAG.FoldConstantArithmetic(ISD::SHL, VT, N01CV, N1CV);
>> + if (C.getNode())
>> + return DAG.getNode(ISD::AND, SDLoc(N), VT, N00, C);
>> + }
>> + } else {
>> + N1C = isConstOrConstSplat(N1);
>> }
>> }
>> }
>> @@ -3978,14 +3991,15 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> return SDValue(N, 0);
>>
>> // fold (shl (shl x, c1), c2) -> 0 or (shl x, (add c1, c2))
>> - if (N1C && N0.getOpcode() == ISD::SHL &&
>> - N0.getOperand(1).getOpcode() == ISD::Constant) {
>> - uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
>> - uint64_t c2 = N1C->getZExtValue();
>> - if (c1 + c2 >= OpSizeInBits)
>> - return DAG.getConstant(0, VT);
>> - return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(c1 + c2, N1.getValueType()));
>> + if (N1C && N0.getOpcode() == ISD::SHL) {
>> + if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
>> + uint64_t c1 = N0C1->getZExtValue();
>> + uint64_t c2 = N1C->getZExtValue();
>> + if (c1 + c2 >= OpSizeInBits)
>> + return DAG.getConstant(0, VT);
>> + return DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
>> + DAG.getConstant(c1 + c2, N1.getValueType()));
>> + }
>> }
>>
>> // fold (shl (ext (shl x, c1)), c2) -> (ext (shl x, (add c1, c2)))
>> @@ -3996,20 +4010,21 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> if (N1C && (N0.getOpcode() == ISD::ZERO_EXTEND ||
>> N0.getOpcode() == ISD::ANY_EXTEND ||
>> N0.getOpcode() == ISD::SIGN_EXTEND) &&
>> - N0.getOperand(0).getOpcode() == ISD::SHL &&
>> - isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
>> - uint64_t c1 =
>> - cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
>> - uint64_t c2 = N1C->getZExtValue();
>> - EVT InnerShiftVT = N0.getOperand(0).getValueType();
>> - uint64_t InnerShiftSize = InnerShiftVT.getScalarType().getSizeInBits();
>> - if (c2 >= OpSizeInBits - InnerShiftSize) {
>> - if (c1 + c2 >= OpSizeInBits)
>> - return DAG.getConstant(0, VT);
>> - return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
>> - DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
>> - N0.getOperand(0)->getOperand(0)),
>> - DAG.getConstant(c1 + c2, N1.getValueType()));
>> + N0.getOperand(0).getOpcode() == ISD::SHL) {
>> + SDValue N0Op0 = N0.getOperand(0);
>> + if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
>> + uint64_t c1 = N0Op0C1->getZExtValue();
>> + uint64_t c2 = N1C->getZExtValue();
>> + EVT InnerShiftVT = N0Op0.getValueType();
>> + uint64_t InnerShiftSize = InnerShiftVT.getScalarSizeInBits();
>> + if (c2 >= OpSizeInBits - InnerShiftSize) {
>> + if (c1 + c2 >= OpSizeInBits)
>> + return DAG.getConstant(0, VT);
>> + return DAG.getNode(ISD::SHL, SDLoc(N0), VT,
>> + DAG.getNode(N0.getOpcode(), SDLoc(N0), VT,
>> + N0Op0->getOperand(0)),
>> + DAG.getConstant(c1 + c2, N1.getValueType()));
>> + }
>> }
>> }
>>
>> @@ -4017,19 +4032,20 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> // Only fold this if the inner zext has no other uses to avoid increasing
>> // the total number of instructions.
>> if (N1C && N0.getOpcode() == ISD::ZERO_EXTEND && N0.hasOneUse() &&
>> - N0.getOperand(0).getOpcode() == ISD::SRL &&
>> - isa<ConstantSDNode>(N0.getOperand(0)->getOperand(1))) {
>> - uint64_t c1 =
>> - cast<ConstantSDNode>(N0.getOperand(0)->getOperand(1))->getZExtValue();
>> - if (c1 < VT.getSizeInBits()) {
>> - uint64_t c2 = N1C->getZExtValue();
>> - if (c1 == c2) {
>> - SDValue NewOp0 = N0.getOperand(0);
>> - EVT CountVT = NewOp0.getOperand(1).getValueType();
>> - SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
>> - NewOp0, DAG.getConstant(c2, CountVT));
>> - AddToWorkList(NewSHL.getNode());
>> - return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
>> + N0.getOperand(0).getOpcode() == ISD::SRL) {
>> + SDValue N0Op0 = N0.getOperand(0);
>> + if (ConstantSDNode *N0Op0C1 = isConstOrConstSplat(N0Op0.getOperand(1))) {
>> + uint64_t c1 = N0Op0C1->getZExtValue();
>> + if (c1 < VT.getScalarSizeInBits()) {
>> + uint64_t c2 = N1C->getZExtValue();
>> + if (c1 == c2) {
>> + SDValue NewOp0 = N0.getOperand(0);
>> + EVT CountVT = NewOp0.getOperand(1).getValueType();
>> + SDValue NewSHL = DAG.getNode(ISD::SHL, SDLoc(N), NewOp0.getValueType(),
>> + NewOp0, DAG.getConstant(c2, CountVT));
>> + AddToWorkList(NewSHL.getNode());
>> + return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N0), VT, NewSHL);
>> + }
>> }
>> }
>> }
>> @@ -4038,40 +4054,39 @@ SDValue DAGCombiner::visitSHL(SDNode *N)
>> // (and (srl x, (sub c1, c2), MASK)
>> // Only fold this if the inner shift has no other uses -- if it does, folding
>> // this will increase the total number of instructions.
>> - if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse() &&
>> - N0.getOperand(1).getOpcode() == ISD::Constant) {
>> - uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
>> - if (c1 < VT.getSizeInBits()) {
>> - uint64_t c2 = N1C->getZExtValue();
>> - APInt Mask = APInt::getHighBitsSet(VT.getSizeInBits(),
>> - VT.getSizeInBits() - c1);
>> - SDValue Shift;
>> - if (c2 > c1) {
>> - Mask = Mask.shl(c2-c1);
>> - Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(c2-c1, N1.getValueType()));
>> - } else {
>> - Mask = Mask.lshr(c1-c2);
>> - Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(c1-c2, N1.getValueType()));
>> + if (N1C && N0.getOpcode() == ISD::SRL && N0.hasOneUse()) {
>> + if (ConstantSDNode *N0C1 = isConstOrConstSplat(N0.getOperand(1))) {
>> + uint64_t c1 = N0C1->getZExtValue();
>> + if (c1 < OpSizeInBits) {
>> + uint64_t c2 = N1C->getZExtValue();
>> + APInt Mask = APInt::getHighBitsSet(OpSizeInBits, OpSizeInBits - c1);
>> + SDValue Shift;
>> + if (c2 > c1) {
>> + Mask = Mask.shl(c2 - c1);
>> + Shift = DAG.getNode(ISD::SHL, SDLoc(N), VT, N0.getOperand(0),
>> + DAG.getConstant(c2 - c1, N1.getValueType()));
>> + } else {
>> + Mask = Mask.lshr(c1 - c2);
>> + Shift = DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
>> + DAG.getConstant(c1 - c2, N1.getValueType()));
>> + }
>> + return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
>> + DAG.getConstant(Mask, VT));
>> }
>> - return DAG.getNode(ISD::AND, SDLoc(N0), VT, Shift,
>> - DAG.getConstant(Mask, VT));
>> }
>> }
>> // fold (shl (sra x, c1), c1) -> (and x, (shl -1, c1))
>> if (N1C && N0.getOpcode() == ISD::SRA && N1 == N0.getOperand(1)) {
>> + unsigned BitSize = VT.getScalarSizeInBits();
>> SDValue HiBitsMask =
>> - DAG.getConstant(APInt::getHighBitsSet(VT.getSizeInBits(),
>> - VT.getSizeInBits() -
>> - N1C->getZExtValue()),
>> - VT);
>> + DAG.getConstant(APInt::getHighBitsSet(BitSize,
>> + BitSize - N1C->getZExtValue()), VT);
>> return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
>> HiBitsMask);
>> }
>>
>> if (N1C) {
>> - SDValue NewSHL = visitShiftByConstant(N, N1C->getZExtValue());
>> + SDValue NewSHL = visitShiftByConstant(N, N1C);
>> if (NewSHL.getNode())
>> return NewSHL;
>> }
>> @@ -4091,6 +4106,8 @@ SDValue DAGCombiner::visitSRA(SDNode *N)
>> if (VT.isVector()) {
>> SDValue FoldedVOp = SimplifyVBinOp(N);
>> if (FoldedVOp.getNode()) return FoldedVOp;
>> +
>> + N1C = isConstOrConstSplat(N1);
>> }
>>
>> // fold (sra c1, c2) -> (sra c1, c2)
>> @@ -4124,11 +4141,12 @@ SDValue DAGCombiner::visitSRA(SDNode *N)
>>
>> // fold (sra (sra x, c1), c2) -> (sra x, (add c1, c2))
>> if (N1C && N0.getOpcode() == ISD::SRA) {
>> - if (ConstantSDNode *C1 = dyn_cast<ConstantSDNode>(N0.getOperand(1))) {
>> + if (ConstantSDNode *C1 = isConstOrConstSplat(N0.getOperand(1))) {
>> unsigned Sum = N1C->getZExtValue() + C1->getZExtValue();
>> - if (Sum >= OpSizeInBits) Sum = OpSizeInBits-1;
>> + if (Sum >= OpSizeInBits)
>> + Sum = OpSizeInBits - 1;
>> return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(Sum, N1C->getValueType(0)));
>> + DAG.getConstant(Sum, N1.getValueType()));
>> }
>> }
>>
>> @@ -4137,14 +4155,17 @@ SDValue DAGCombiner::visitSRA(SDNode *N)
>> // result_size - n != m.
>> // If truncate is free for the target sext(shl) is likely to result in better
>> // code.
>> - if (N0.getOpcode() == ISD::SHL) {
>> + if (N0.getOpcode() == ISD::SHL && N1C) {
>> // Get the two constanst of the shifts, CN0 = m, CN = n.
>> - const ConstantSDNode *N01C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
>> - if (N01C && N1C) {
>> + const ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1));
>> + if (N01C) {
>> + LLVMContext &Ctx = *DAG.getContext();
>> // Determine what the truncate's result bitsize and type would be.
>> - EVT TruncVT =
>> - EVT::getIntegerVT(*DAG.getContext(),
>> - OpSizeInBits - N1C->getZExtValue());
>> + EVT TruncVT = EVT::getIntegerVT(Ctx, OpSizeInBits - N1C->getZExtValue());
>> +
>> + if (VT.isVector())
>> + TruncVT = EVT::getVectorVT(Ctx, TruncVT, VT.getVectorNumElements());
>> +
>> // Determine the residual right-shift amount.
>> signed ShiftAmt = N1C->getZExtValue() - N01C->getZExtValue();
>>
>> @@ -4177,26 +4198,27 @@ SDValue DAGCombiner::visitSRA(SDNode *N)
>> return DAG.getNode(ISD::SRA, SDLoc(N), VT, N0, NewOp1);
>> }
>>
>> - // fold (sra (trunc (sr x, c1)), c2) -> (trunc (sra x, c1+c2))
>> + // fold (sra (trunc (srl x, c1)), c2) -> (trunc (sra x, c1 + c2))
>> // if c1 is equal to the number of bits the trunc removes
>> if (N0.getOpcode() == ISD::TRUNCATE &&
>> (N0.getOperand(0).getOpcode() == ISD::SRL ||
>> N0.getOperand(0).getOpcode() == ISD::SRA) &&
>> N0.getOperand(0).hasOneUse() &&
>> N0.getOperand(0).getOperand(1).hasOneUse() &&
>> - N1C && isa<ConstantSDNode>(N0.getOperand(0).getOperand(1))) {
>> - EVT LargeVT = N0.getOperand(0).getValueType();
>> - ConstantSDNode *LargeShiftAmt =
>> - cast<ConstantSDNode>(N0.getOperand(0).getOperand(1));
>> -
>> - if (LargeVT.getScalarType().getSizeInBits() - OpSizeInBits ==
>> - LargeShiftAmt->getZExtValue()) {
>> - SDValue Amt =
>> - DAG.getConstant(LargeShiftAmt->getZExtValue() + N1C->getZExtValue(),
>> - getShiftAmountTy(N0.getOperand(0).getOperand(0).getValueType()));
>> - SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
>> - N0.getOperand(0).getOperand(0), Amt);
>> - return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
>> + N1C) {
>> + SDValue N0Op0 = N0.getOperand(0);
>> + if (ConstantSDNode *LargeShift = isConstOrConstSplat(N0Op0.getOperand(1))) {
>> + unsigned LargeShiftVal = LargeShift->getZExtValue();
>> + EVT LargeVT = N0Op0.getValueType();
>> +
>> + if (LargeVT.getScalarSizeInBits() - OpSizeInBits == LargeShiftVal) {
>> + SDValue Amt =
>> + DAG.getConstant(LargeShiftVal + N1C->getZExtValue(),
>> + getShiftAmountTy(N0Op0.getOperand(0).getValueType()));
>> + SDValue SRA = DAG.getNode(ISD::SRA, SDLoc(N), LargeVT,
>> + N0Op0.getOperand(0), Amt);
>> + return DAG.getNode(ISD::TRUNCATE, SDLoc(N), VT, SRA);
>> + }
>> }
>> }
>>
>> @@ -4210,7 +4232,7 @@ SDValue DAGCombiner::visitSRA(SDNode *N)
>> return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0, N1);
>>
>> if (N1C) {
>> - SDValue NewSRA = visitShiftByConstant(N, N1C->getZExtValue());
>> + SDValue NewSRA = visitShiftByConstant(N, N1C);
>> if (NewSRA.getNode())
>> return NewSRA;
>> }
>> @@ -4230,6 +4252,8 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>> if (VT.isVector()) {
>> SDValue FoldedVOp = SimplifyVBinOp(N);
>> if (FoldedVOp.getNode()) return FoldedVOp;
>> +
>> + N1C = isConstOrConstSplat(N1);
>> }
>>
>> // fold (srl c1, c2) -> c1 >>u c2
>> @@ -4250,14 +4274,15 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>> return DAG.getConstant(0, VT);
>>
>> // fold (srl (srl x, c1), c2) -> 0 or (srl x, (add c1, c2))
>> - if (N1C && N0.getOpcode() == ISD::SRL &&
>> - N0.getOperand(1).getOpcode() == ISD::Constant) {
>> - uint64_t c1 = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue();
>> - uint64_t c2 = N1C->getZExtValue();
>> - if (c1 + c2 >= OpSizeInBits)
>> - return DAG.getConstant(0, VT);
>> - return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(c1 + c2, N1.getValueType()));
>> + if (N1C && N0.getOpcode() == ISD::SRL) {
>> + if (ConstantSDNode *N01C = isConstOrConstSplat(N0.getOperand(1))) {
>> + uint64_t c1 = N01C->getZExtValue();
>> + uint64_t c2 = N1C->getZExtValue();
>> + if (c1 + c2 >= OpSizeInBits)
>> + return DAG.getConstant(0, VT);
>> + return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0),
>> + DAG.getConstant(c1 + c2, N1.getValueType()));
>> + }
>> }
>>
>> // fold (srl (trunc (srl x, c1)), c2) -> 0 or (trunc (srl x, (add c1, c2)))
>> @@ -4282,18 +4307,21 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>> }
>>
>> // fold (srl (shl x, c), c) -> (and x, cst2)
>> - if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1 &&
>> - N0.getValueSizeInBits() <= 64) {
>> - uint64_t ShAmt = N1C->getZExtValue()+64-N0.getValueSizeInBits();
>> - return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
>> - DAG.getConstant(~0ULL >> ShAmt, VT));
>> + if (N1C && N0.getOpcode() == ISD::SHL && N0.getOperand(1) == N1) {
>> + unsigned BitSize = N0.getScalarValueSizeInBits();
>> + if (BitSize <= 64) {
>> + uint64_t ShAmt = N1C->getZExtValue() + 64 - BitSize;
>> + return DAG.getNode(ISD::AND, SDLoc(N), VT, N0.getOperand(0),
>> + DAG.getConstant(~0ULL >> ShAmt, VT));
>> + }
>> }
>>
>> // fold (srl (anyextend x), c) -> (and (anyextend (srl x, c)), mask)
>> if (N1C && N0.getOpcode() == ISD::ANY_EXTEND) {
>> // Shifting in all undef bits?
>> EVT SmallVT = N0.getOperand(0).getValueType();
>> - if (N1C->getZExtValue() >= SmallVT.getSizeInBits())
>> + unsigned BitSize = SmallVT.getScalarSizeInBits();
>> + if (N1C->getZExtValue() >= BitSize)
>> return DAG.getUNDEF(VT);
>>
>> if (!LegalTypes || TLI.isTypeDesirableForOp(ISD::SRL, SmallVT)) {
>> @@ -4302,7 +4330,7 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>> N0.getOperand(0),
>> DAG.getConstant(ShiftAmt, getShiftAmountTy(SmallVT)));
>> AddToWorkList(SmallShift.getNode());
>> - APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()).lshr(ShiftAmt);
>> + APInt Mask = APInt::getAllOnesValue(OpSizeInBits).lshr(ShiftAmt);
>> return DAG.getNode(ISD::AND, SDLoc(N), VT,
>> DAG.getNode(ISD::ANY_EXTEND, SDLoc(N), VT, SmallShift),
>> DAG.getConstant(Mask, VT));
>> @@ -4311,14 +4339,14 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>>
>> // fold (srl (sra X, Y), 31) -> (srl X, 31). This srl only looks at the sign
>> // bit, which is unmodified by sra.
>> - if (N1C && N1C->getZExtValue() + 1 == VT.getSizeInBits()) {
>> + if (N1C && N1C->getZExtValue() + 1 == OpSizeInBits) {
>> if (N0.getOpcode() == ISD::SRA)
>> return DAG.getNode(ISD::SRL, SDLoc(N), VT, N0.getOperand(0), N1);
>> }
>>
>> // fold (srl (ctlz x), "5") -> x iff x has one bit set (the low bit).
>> if (N1C && N0.getOpcode() == ISD::CTLZ &&
>> - N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) {
>> + N1C->getAPIntValue() == Log2_32(OpSizeInBits)) {
>> APInt KnownZero, KnownOne;
>> DAG.ComputeMaskedBits(N0.getOperand(0), KnownZero, KnownOne);
>>
>> @@ -4365,7 +4393,7 @@ SDValue DAGCombiner::visitSRL(SDNode *N)
>> return SDValue(N, 0);
>>
>> if (N1C) {
>> - SDValue NewSRL = visitShiftByConstant(N, N1C->getZExtValue());
>> + SDValue NewSRL = visitShiftByConstant(N, N1C);
>> if (NewSRL.getNode())
>> return NewSRL;
>> }
>>
>> Modified: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (original)
>> +++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp Mon Mar 17 13:58:01 2014
>> @@ -6573,6 +6573,19 @@ bool BuildVectorSDNode::isConstantSplat(
>> return true;
>> }
>>
>> +ConstantSDNode *BuildVectorSDNode::isConstantSplat() const {
>> + SDValue Op0 = getOperand(0);
>> + for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
>> + SDValue Opi = getOperand(i);
>> + unsigned Opc = Opi.getOpcode();
>> + if ((Opc != ISD::UNDEF && Opc != ISD::Constant && Opc != ISD::ConstantFP) ||
>> + Opi != Op0)
>> + return nullptr;
>> + }
>> +
>> + return cast<ConstantSDNode>(Op0);
>> +}
>> +
>> bool BuildVectorSDNode::isConstant() const {
>> for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
>> unsigned Opc = getOperand(i).getOpcode();
>>
>> Modified: llvm/trunk/test/CodeGen/AArch64/neon-shl-ashr-lshr.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/AArch64/neon-shl-ashr-lshr.ll?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/AArch64/neon-shl-ashr-lshr.ll (original)
>> +++ llvm/trunk/test/CodeGen/AArch64/neon-shl-ashr-lshr.ll Mon Mar 17 13:58:01 2014
>> @@ -186,14 +186,14 @@ define <2 x i64> @ashr.v2i64(<2 x i64> %
>>
>> define <1 x i64> @shl.v1i64.0(<1 x i64> %a) {
>> ; CHECK-LABEL: shl.v1i64.0:
>> -; CHECK: shl d{{[0-9]+}}, d{{[0-9]+}}, #0
>> +; CHECK-NOT: shl d{{[0-9]+}}, d{{[0-9]+}}, #0
>> %c = shl <1 x i64> %a, zeroinitializer
>> ret <1 x i64> %c
>> }
>>
>> define <2 x i32> @shl.v2i32.0(<2 x i32> %a) {
>> ; CHECK-LABEL: shl.v2i32.0:
>> -; CHECK: shl v{{[0-9]+}}.2s, v{{[0-9]+}}.2s, #0
>> +; CHECK-NOT: shl v{{[0-9]+}}.2s, v{{[0-9]+}}.2s, #0
>> %c = shl <2 x i32> %a, zeroinitializer
>> ret <2 x i32> %c
>> }
>> @@ -285,7 +285,7 @@ define <1 x i16> @shl.v1i16.imm(<1 x i16
>>
>> define <1 x i32> @shl.v1i32.imm(<1 x i32> %a) {
>> ; CHECK-LABEL: shl.v1i32.imm:
>> -; CHECK: shl v{{[0-9]+}}.2s, v{{[0-9]+}}.2s, #0
>> +; CHECK-NOT: shl v{{[0-9]+}}.2s, v{{[0-9]+}}.2s, #0
>> %c = shl <1 x i32> %a, zeroinitializer
>> ret <1 x i32> %c
>> }
>>
>> Modified: llvm/trunk/test/CodeGen/X86/avx2-vector-shifts.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/avx2-vector-shifts.ll?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/X86/avx2-vector-shifts.ll (original)
>> +++ llvm/trunk/test/CodeGen/X86/avx2-vector-shifts.ll Mon Mar 17 13:58:01 2014
>> @@ -9,7 +9,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sllw_1:
>> -; CHECK: vpsllw $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsllw $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <16 x i16> @test_sllw_2(<16 x i16> %InVec) {
>> @@ -39,7 +39,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_slld_1:
>> -; CHECK: vpslld $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpslld $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <8 x i32> @test_slld_2(<8 x i32> %InVec) {
>> @@ -69,7 +69,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sllq_1:
>> -; CHECK: vpsllq $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsllq $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <4 x i64> @test_sllq_2(<4 x i64> %InVec) {
>> @@ -101,7 +101,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sraw_1:
>> -; CHECK: vpsraw $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsraw $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <16 x i16> @test_sraw_2(<16 x i16> %InVec) {
>> @@ -131,7 +131,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srad_1:
>> -; CHECK: vpsrad $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsrad $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <8 x i32> @test_srad_2(<8 x i32> %InVec) {
>> @@ -163,7 +163,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srlw_1:
>> -; CHECK: vpsrlw $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsrlw $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <16 x i16> @test_srlw_2(<16 x i16> %InVec) {
>> @@ -193,7 +193,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srld_1:
>> -; CHECK: vpsrld $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsrld $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <8 x i32> @test_srld_2(<8 x i32> %InVec) {
>> @@ -223,7 +223,7 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srlq_1:
>> -; CHECK: vpsrlq $0, %ymm0, %ymm0
>> +; CHECK-NOT: vpsrlq $0, %ymm0, %ymm0
>> ; CHECK: ret
>>
>> define <4 x i64> @test_srlq_2(<4 x i64> %InVec) {
>> @@ -245,3 +245,14 @@ entry:
>> ; CHECK-LABEL: test_srlq_3:
>> ; CHECK: vpsrlq $63, %ymm0, %ymm0
>> ; CHECK: ret
>> +
>> +; CHECK-LABEL: @srl_trunc_and_v4i64
>> +; CHECK: vpand
>> +; CHECK-NEXT: vpsrlvd
>> +; CHECK: ret
>> +define <4 x i32> @srl_trunc_and_v4i64(<4 x i32> %x, <4 x i64> %y) nounwind {
>> + %and = and <4 x i64> %y, <i64 8, i64 8, i64 8, i64 8>
>> + %trunc = trunc <4 x i64> %and to <4 x i32>
>> + %sra = lshr <4 x i32> %x, %trunc
>> + ret <4 x i32> %sra
>> +}
>>
>> Modified: llvm/trunk/test/CodeGen/X86/sse2-vector-shifts.ll
>> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/sse2-vector-shifts.ll?rev=204071&r1=204070&r2=204071&view=diff
>> ==============================================================================
>> --- llvm/trunk/test/CodeGen/X86/sse2-vector-shifts.ll (original)
>> +++ llvm/trunk/test/CodeGen/X86/sse2-vector-shifts.ll Mon Mar 17 13:58:01 2014
>> @@ -9,8 +9,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sllw_1:
>> -; CHECK: psllw $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psllw $0, %xmm0
>> +; CHECK: ret
>>
>> define <8 x i16> @test_sllw_2(<8 x i16> %InVec) {
>> entry:
>> @@ -39,8 +39,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_slld_1:
>> -; CHECK: pslld $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: pslld $0, %xmm0
>> +; CHECK: ret
>>
>> define <4 x i32> @test_slld_2(<4 x i32> %InVec) {
>> entry:
>> @@ -69,8 +69,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sllq_1:
>> -; CHECK: psllq $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psllq $0, %xmm0
>> +; CHECK: ret
>>
>> define <2 x i64> @test_sllq_2(<2 x i64> %InVec) {
>> entry:
>> @@ -101,8 +101,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_sraw_1:
>> -; CHECK: psraw $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psraw $0, %xmm0
>> +; CHECK: ret
>>
>> define <8 x i16> @test_sraw_2(<8 x i16> %InVec) {
>> entry:
>> @@ -131,8 +131,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srad_1:
>> -; CHECK: psrad $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psrad $0, %xmm0
>> +; CHECK: ret
>>
>> define <4 x i32> @test_srad_2(<4 x i32> %InVec) {
>> entry:
>> @@ -163,8 +163,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srlw_1:
>> -; CHECK: psrlw $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psrlw $0, %xmm0
>> +; CHECK: ret
>>
>> define <8 x i16> @test_srlw_2(<8 x i16> %InVec) {
>> entry:
>> @@ -193,8 +193,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srld_1:
>> -; CHECK: psrld $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psrld $0, %xmm0
>> +; CHECK: ret
>>
>> define <4 x i32> @test_srld_2(<4 x i32> %InVec) {
>> entry:
>> @@ -223,8 +223,8 @@ entry:
>> }
>>
>> ; CHECK-LABEL: test_srlq_1:
>> -; CHECK: psrlq $0, %xmm0
>> -; CHECK-NEXT: ret
>> +; CHECK-NOT: psrlq $0, %xmm0
>> +; CHECK: ret
>>
>> define <2 x i64> @test_srlq_2(<2 x i64> %InVec) {
>> entry:
>> @@ -245,3 +245,123 @@ entry:
>> ; CHECK-LABEL: test_srlq_3:
>> ; CHECK: psrlq $63, %xmm0
>> ; CHECK-NEXT: ret
>> +
>> +
>> +; CHECK-LABEL: sra_sra_v4i32:
>> +; CHECK: psrad $6, %xmm0
>> +; CHECK-NEXT: retq
>> +define <4 x i32> @sra_sra_v4i32(<4 x i32> %x) nounwind {
>> + %sra0 = ashr <4 x i32> %x, <i32 2, i32 2, i32 2, i32 2>
>> + %sra1 = ashr <4 x i32> %sra0, <i32 4, i32 4, i32 4, i32 4>
>> + ret <4 x i32> %sra1
>> +}
>> +
>> +; CHECK-LABEL: @srl_srl_v4i32
>> +; CHECK: psrld $6, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @srl_srl_v4i32(<4 x i32> %x) nounwind {
>> + %srl0 = lshr <4 x i32> %x, <i32 2, i32 2, i32 2, i32 2>
>> + %srl1 = lshr <4 x i32> %srl0, <i32 4, i32 4, i32 4, i32 4>
>> + ret <4 x i32> %srl1
>> +}
>> +
>> +; CHECK-LABEL: @srl_shl_v4i32
>> +; CHECK: andps
>> +; CHECK-NEXT: retq
>> +define <4 x i32> @srl_shl_v4i32(<4 x i32> %x) nounwind {
>> + %srl0 = shl <4 x i32> %x, <i32 4, i32 4, i32 4, i32 4>
>> + %srl1 = lshr <4 x i32> %srl0, <i32 4, i32 4, i32 4, i32 4>
>> + ret <4 x i32> %srl1
>> +}
>> +
>> +; CHECK-LABEL: @srl_sra_31_v4i32
>> +; CHECK: psrld $31, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @srl_sra_31_v4i32(<4 x i32> %x, <4 x i32> %y) nounwind {
>> + %sra = ashr <4 x i32> %x, %y
>> + %srl1 = lshr <4 x i32> %sra, <i32 31, i32 31, i32 31, i32 31>
>> + ret <4 x i32> %srl1
>> +}
>> +
>> +; CHECK-LABEL: @shl_shl_v4i32
>> +; CHECK: pslld $6, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_shl_v4i32(<4 x i32> %x) nounwind {
>> + %shl0 = shl <4 x i32> %x, <i32 2, i32 2, i32 2, i32 2>
>> + %shl1 = shl <4 x i32> %shl0, <i32 4, i32 4, i32 4, i32 4>
>> + ret <4 x i32> %shl1
>> +}
>> +
>> +; CHECK-LABEL: @shl_sra_v4i32
>> +; CHECK: andps
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_sra_v4i32(<4 x i32> %x) nounwind {
>> + %shl0 = ashr <4 x i32> %x, <i32 4, i32 4, i32 4, i32 4>
>> + %shl1 = shl <4 x i32> %shl0, <i32 4, i32 4, i32 4, i32 4>
>> + ret <4 x i32> %shl1
>> +}
>> +
>> +; CHECK-LABEL: @shl_srl_v4i32
>> +; CHECK: pslld $3, %xmm0
>> +; CHECK-NEXT: pand
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_srl_v4i32(<4 x i32> %x) nounwind {
>> + %shl0 = lshr <4 x i32> %x, <i32 2, i32 2, i32 2, i32 2>
>> + %shl1 = shl <4 x i32> %shl0, <i32 5, i32 5, i32 5, i32 5>
>> + ret <4 x i32> %shl1
>> +}
>> +
>> +; CHECK-LABEL: @shl_zext_srl_v4i32
>> +; CHECK: andps
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_zext_srl_v4i32(<4 x i16> %x) nounwind {
>> + %srl = lshr <4 x i16> %x, <i16 2, i16 2, i16 2, i16 2>
>> + %zext = zext <4 x i16> %srl to <4 x i32>
>> + %shl = shl <4 x i32> %zext, <i32 2, i32 2, i32 2, i32 2>
>> + ret <4 x i32> %shl
>> +}
>> +
>> +; CHECK: @sra_trunc_srl_v4i32
>> +; CHECK: psrad $19, %xmm0
>> +; CHECK-NEXT: retq
>> +define <4 x i16> @sra_trunc_srl_v4i32(<4 x i32> %x) nounwind {
>> + %srl = lshr <4 x i32> %x, <i32 16, i32 16, i32 16, i32 16>
>> + %trunc = trunc <4 x i32> %srl to <4 x i16>
>> + %sra = ashr <4 x i16> %trunc, <i16 3, i16 3, i16 3, i16 3>
>> + ret <4 x i16> %sra
>> +}
>> +
>> +; CHECK-LABEL: @shl_zext_shl_v4i32
>> +; CHECK: pand
>> +; CHECK-NEXT: pslld $19, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_zext_shl_v4i32(<4 x i16> %x) nounwind {
>> + %shl0 = shl <4 x i16> %x, <i16 2, i16 2, i16 2, i16 2>
>> + %ext = zext <4 x i16> %shl0 to <4 x i32>
>> + %shl1 = shl <4 x i32> %ext, <i32 17, i32 17, i32 17, i32 17>
>> + ret <4 x i32> %shl1
>> +}
>> +
>> +; CHECK-LABEL: @sra_v4i32
>> +; CHECK: psrad $3, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @sra_v4i32(<4 x i32> %x) nounwind {
>> + %sra = ashr <4 x i32> %x, <i32 3, i32 3, i32 3, i32 3>
>> + ret <4 x i32> %sra
>> +}
>> +
>> +; CHECK-LABEL: @srl_v4i32
>> +; CHECK: psrld $3, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @srl_v4i32(<4 x i32> %x) nounwind {
>> + %sra = lshr <4 x i32> %x, <i32 3, i32 3, i32 3, i32 3>
>> + ret <4 x i32> %sra
>> +}
>> +
>> +; CHECK-LABEL: @shl_v4i32
>> +; CHECK: pslld $3, %xmm0
>> +; CHECK-NEXT: ret
>> +define <4 x i32> @shl_v4i32(<4 x i32> %x) nounwind {
>> + %sra = shl <4 x i32> %x, <i32 3, i32 3, i32 3, i32 3>
>> + ret <4 x i32> %sra
>> +}
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
-------------- next part --------------
Index: lib/CodeGen/SelectionDAG/SelectionDAG.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/SelectionDAG.cpp (revision 204238)
+++ lib/CodeGen/SelectionDAG/SelectionDAG.cpp (working copy)
@@ -6573,15 +6573,14 @@
return true;
}
-ConstantSDNode *BuildVectorSDNode::isConstantSplat() const {
+ConstantSDNode *BuildVectorSDNode::getConstantSplatValue() const {
SDValue Op0 = getOperand(0);
- for (unsigned i = 0, e = getNumOperands(); i != e; ++i) {
- SDValue Opi = getOperand(i);
- unsigned Opc = Opi.getOpcode();
- if ((Opc != ISD::UNDEF && Opc != ISD::Constant && Opc != ISD::ConstantFP) ||
- Opi != Op0)
+ if (Op0.getOpcode() != ISD::Constant)
+ return nullptr;
+
+ for (unsigned i = 1, e = getNumOperands(); i != e; ++i)
+ if (getOperand(i) != Op0)
return nullptr;
- }
return cast<ConstantSDNode>(Op0);
}
Index: lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- lib/CodeGen/SelectionDAG/DAGCombiner.cpp (revision 204238)
+++ lib/CodeGen/SelectionDAG/DAGCombiner.cpp (working copy)
@@ -667,7 +667,7 @@
return CN;
if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N))
- return BV->isConstantSplat();
+ return BV->getConstantSplatValue();
return nullptr;
}
Index: test/CodeGen/X86/shift-combine-crash.ll
===================================================================
--- test/CodeGen/X86/shift-combine-crash.ll (revision 0)
+++ test/CodeGen/X86/shift-combine-crash.ll (working copy)
@@ -0,0 +1,18 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -mcpu=corei7 > /dev/null
+
+; Verify that DAGCombiner doesn't crash with an assertion failure in the
+; attempt to cast a ISD::UNDEF node to a ConstantSDNode.
+
+; During type legalization, the vector shift operation in function @foo is
+; split into two legal shifts that work on <2 x i64> elements.
+; The first shift of the legalized sequence would be a shift by all undefs.
+; DAGCombiner will then try to simplify the vector shift and check if the
+; vector of shift counts is a splat. Make sure that llc doesn't crash
+; at that stage.
+
+
+define <4 x i64> @foo(<4 x i64> %A) {
+ %shl = shl <4 x i64> %A, <i64 undef, i64 undef, i64 1, i64 2>
+ ret <4 x i64> %shl
+}
+
Index: include/llvm/CodeGen/SelectionDAGNodes.h
===================================================================
--- include/llvm/CodeGen/SelectionDAGNodes.h (revision 204238)
+++ include/llvm/CodeGen/SelectionDAGNodes.h (working copy)
@@ -1522,10 +1522,11 @@
unsigned MinSplatBits = 0,
bool isBigEndian = false) const;
- /// isConstantSplat - Simpler form of isConstantSplat. Get the constant splat
- /// when you only care about the value. Returns nullptr if this isn't a
- /// constant splat vector.
- ConstantSDNode *isConstantSplat() const;
+ /// getConstantSplatValue - Check if this is a constant splat, and if so,
+ /// return the splat value only if it is a ConstantSDNode. Otherwise
+ /// return nullptr. This is a simpler form of isConstantSplat.
+ /// Get the constant splat only if you care about the splat value.
+ ConstantSDNode *getConstantSplatValue() const;
bool isConstant() const;
More information about the llvm-commits
mailing list