//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "systemz-lower"
-
#include "SystemZISelLowering.h"
#include "SystemZCallingConv.h"
#include "SystemZConstantPoolValue.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
-
#include <cctype>
using namespace llvm;
+#define DEBUG_TYPE "systemz-lower"
+
+namespace {
+// Represents a sequence for extracting a 0/1 value from an IPM result:
+// (((X ^ XORValue) + AddValue) >> Bit)
+struct IPMConversion {
+ IPMConversion(unsigned xorValue, int64_t addValue, unsigned bit)
+ : XORValue(xorValue), AddValue(addValue), Bit(bit) {}
+
+ int64_t XORValue;
+ int64_t AddValue;
+ unsigned Bit;
+};
+
+// Represents information about a comparison.
+struct Comparison {
+ Comparison(SDValue Op0In, SDValue Op1In)
+ : Op0(Op0In), Op1(Op1In), Opcode(0), ICmpType(0), CCValid(0), CCMask(0) {}
+
+ // The operands to the comparison.
+ SDValue Op0, Op1;
+
+ // The opcode that should be used to compare Op0 and Op1.
+ unsigned Opcode;
+
+ // A SystemZICMP value. Only used for integer comparisons.
+ unsigned ICmpType;
+
+ // The mask of CC values that Opcode can produce.
+ unsigned CCValid;
+
+ // The mask of CC values for which the original condition is true.
+ unsigned CCMask;
+};
+} // end anonymous namespace
+
// Classify VT as either 32 or 64 bit.
static bool is32Bit(EVT VT) {
switch (VT.getSimpleVT().SimpleTy) {
return Op;
}
-SystemZTargetLowering::SystemZTargetLowering(SystemZTargetMachine &tm)
- : TargetLowering(tm, new TargetLoweringObjectFileELF()),
- Subtarget(*tm.getSubtargetImpl()), TM(tm) {
+SystemZTargetLowering::SystemZTargetLowering(const TargetMachine &tm,
+ const SystemZSubtarget &STI)
+ : TargetLowering(tm), Subtarget(STI) {
MVT PtrVT = getPointerTy();
// Set up the register classes.
addRegisterClass(MVT::f128, &SystemZ::FP128BitRegClass);
// Compute derived properties from the register classes
- computeRegisterProperties();
+ computeRegisterProperties(Subtarget.getRegisterInfo());
// Set up special registers.
setExceptionPointerRegister(SystemZ::R6D);
++I) {
MVT VT = MVT::SimpleValueType(I);
if (isTypeLegal(VT)) {
- // Expand SETCC(X, Y, COND) into SELECT_CC(X, Y, 1, 0, COND).
- setOperationAction(ISD::SETCC, VT, Expand);
+ // Lower SET_CC into an IPM-based sequence.
+ setOperationAction(ISD::SETCC, VT, Custom);
// Expand SELECT(C, A, B) into SELECT_CC(X, 0, A, B, NE).
setOperationAction(ISD::SELECT, VT, Expand);
setOperationAction(ISD::SDIVREM, VT, Custom);
setOperationAction(ISD::UDIVREM, VT, Custom);
- // Expand ATOMIC_LOAD and ATOMIC_STORE using ATOMIC_CMP_SWAP.
- // FIXME: probably much too conservative.
- setOperationAction(ISD::ATOMIC_LOAD, VT, Expand);
- setOperationAction(ISD::ATOMIC_STORE, VT, Expand);
+ // Lower ATOMIC_LOAD and ATOMIC_STORE into normal volatile loads and
+ // stores, putting a serialization instruction after the stores.
+ setOperationAction(ISD::ATOMIC_LOAD, VT, Custom);
+ setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
+
+ // Lower ATOMIC_LOAD_SUB into ATOMIC_LOAD_ADD if LAA and LAAG are
+ // available, or if the operand is constant.
+ setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
// No special instructions for these.
setOperationAction(ISD::CTPOP, VT, Expand);
setOperationAction(ISD::SMUL_LOHI, VT, Custom);
setOperationAction(ISD::UMUL_LOHI, VT, Custom);
- // We have instructions for signed but not unsigned FP conversion.
- setOperationAction(ISD::FP_TO_UINT, VT, Expand);
+ // Only z196 and above have native support for conversions to unsigned.
+ if (!Subtarget.hasFPExtension())
+ setOperationAction(ISD::FP_TO_UINT, VT, Expand);
}
}
setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Custom);
setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
- // We have instructions for signed but not unsigned FP conversion.
+ // z10 has instructions for signed but not unsigned FP conversion.
// Handle unsigned 32-bit types as signed 64-bit types.
- setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
- setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
+ if (!Subtarget.hasFPExtension()) {
+ setOperationAction(ISD::UINT_TO_FP, MVT::i32, Promote);
+ setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
+ }
// We have native support for a 64-bit CTLZ, via FLOGR.
setOperationAction(ISD::CTLZ, MVT::i32, Promote);
setOperationAction(ISD::SRA_PARTS, MVT::i64, Expand);
// We have native instructions for i8, i16 and i32 extensions, but not i1.
- setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
- setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
- setLoadExtAction(ISD::EXTLOAD, MVT::i1, Promote);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
+ for (MVT VT : MVT::integer_valuetypes()) {
+ setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
+ setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
+ setLoadExtAction(ISD::EXTLOAD, VT, MVT::i1, Promote);
+ }
// Handle the various types of symbolic address.
setOperationAction(ISD::ConstantPool, PtrVT, Custom);
// Needed so that we don't try to implement f128 constant loads using
// a load-and-extend of a f80 constant (in cases where the constant
// would fit in an f80).
- setLoadExtAction(ISD::EXTLOAD, MVT::f80, Expand);
+ for (MVT VT : MVT::fp_valuetypes())
+ setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
// Floating-point truncation and stores need to be done separately.
setTruncStoreAction(MVT::f64, MVT::f32, Expand);
setOperationAction(ISD::VACOPY, MVT::Other, Custom);
setOperationAction(ISD::VAEND, MVT::Other, Expand);
+ // Codes for which we want to perform some z-specific combinations.
+ setTargetDAGCombine(ISD::SIGN_EXTEND);
+
// We want to use MVC in preference to even a single load/store pair.
MaxStoresPerMemcpy = 0;
MaxStoresPerMemcpyOptSize = 0;
MaxStoresPerMemsetOptSize = 0;
}
-bool
-SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
+EVT SystemZTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
+ if (!VT.isVector())
+ return MVT::i32;
+ return VT.changeVectorElementTypeToInteger();
+}
+
+bool SystemZTargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
VT = VT.getScalarType();
if (!VT.isSimple())
return Imm.isZero() || Imm.isNegZero();
}
-bool SystemZTargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
- bool *Fast) const {
+bool SystemZTargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
+ unsigned,
+ unsigned,
+ bool *Fast) const {
// Unaligned accesses should never be slower than the expanded version.
// We check specifically for aligned accesses in the few cases where
// they are required.
Value *CallOperandVal = info.CallOperandVal;
// If we don't have a value, we can't do a match,
// but allow it at the lowest weight.
- if (CallOperandVal == NULL)
+ if (!CallOperandVal)
return CW_Default;
Type *type = CallOperandVal->getType();
// Look at the constraint type.
break;
case 'I': // Unsigned 8-bit constant
- if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
+ if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
if (isUInt<8>(C->getZExtValue()))
weight = CW_Constant;
break;
case 'J': // Unsigned 12-bit constant
- if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
+ if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
if (isUInt<12>(C->getZExtValue()))
weight = CW_Constant;
break;
case 'K': // Signed 16-bit constant
- if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
+ if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
if (isInt<16>(C->getSExtValue()))
weight = CW_Constant;
break;
case 'L': // Signed 20-bit displacement (on all targets we support)
- if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
+ if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
if (isInt<20>(C->getSExtValue()))
weight = CW_Constant;
break;
case 'M': // 0x7fffffff
- if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal))
+ if (auto *C = dyn_cast<ConstantInt>(CallOperandVal))
if (C->getZExtValue() == 0x7fffffff)
weight = CW_Constant;
break;
if (Index < 16 && Map[Index])
return std::make_pair(Map[Index], RC);
}
- return std::make_pair(0u, static_cast<TargetRegisterClass*>(0));
+ return std::make_pair(0U, nullptr);
}
-std::pair<unsigned, const TargetRegisterClass *> SystemZTargetLowering::
-getRegForInlineAsmConstraint(const std::string &Constraint, MVT VT) const {
+std::pair<unsigned, const TargetRegisterClass *>
+SystemZTargetLowering::getRegForInlineAsmConstraint(
+ const TargetRegisterInfo *TRI, const std::string &Constraint,
+ MVT VT) const {
if (Constraint.size() == 1) {
// GCC Constraint Letters
switch (Constraint[0]) {
SystemZMC::FP64Regs);
}
}
- return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
+ return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
}
void SystemZTargetLowering::
if (Constraint.length() == 1) {
switch (Constraint[0]) {
case 'I': // Unsigned 8-bit constant
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (isUInt<8>(C->getZExtValue()))
Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
Op.getValueType()));
return;
case 'J': // Unsigned 12-bit constant
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (isUInt<12>(C->getZExtValue()))
Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
Op.getValueType()));
return;
case 'K': // Signed 16-bit constant
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (isInt<16>(C->getSExtValue()))
Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
Op.getValueType()));
return;
case 'L': // Signed 20-bit displacement (on all targets we support)
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (isInt<20>(C->getSExtValue()))
Ops.push_back(DAG.getTargetConstant(C->getSExtValue(),
Op.getValueType()));
return;
case 'M': // 0x7fffffff
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
+ if (auto *C = dyn_cast<ConstantSDNode>(Op))
if (C->getZExtValue() == 0x7fffffff)
Ops.push_back(DAG.getTargetConstant(C->getZExtValue(),
Op.getValueType()));
MachineFrameInfo *MFI = MF.getFrameInfo();
MachineRegisterInfo &MRI = MF.getRegInfo();
SystemZMachineFunctionInfo *FuncInfo =
- MF.getInfo<SystemZMachineFunctionInfo>();
- const SystemZFrameLowering *TFL =
- static_cast<const SystemZFrameLowering *>(TM.getFrameLowering());
+ MF.getInfo<SystemZMachineFunctionInfo>();
+ auto *TFL =
+ static_cast<const SystemZFrameLowering *>(Subtarget.getFrameLowering());
// Assign locations to all of the incoming arguments.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
+ CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
CCInfo.AnalyzeFormalArguments(Ins, CC_SystemZ);
unsigned NumFixedGPRs = 0;
}
// Join the stores, which are independent of one another.
Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
- &MemOps[NumFixedFPRs],
- SystemZ::NumArgFPRs - NumFixedFPRs);
+ makeArrayRef(&MemOps[NumFixedFPRs],
+ SystemZ::NumArgFPRs-NumFixedFPRs));
}
}
return Chain;
}
-static bool canUseSiblingCall(CCState ArgCCInfo,
+static bool canUseSiblingCall(const CCState &ArgCCInfo,
SmallVectorImpl<CCValAssign> &ArgLocs) {
// Punt if there are any indirect or stack arguments, or if the call
// needs the call-saved argument register R6.
// Analyze the operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState ArgCCInfo(CallConv, IsVarArg, MF, TM, ArgLocs, *DAG.getContext());
+ CCState ArgCCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
ArgCCInfo.AnalyzeCallOperands(Outs, CC_SystemZ);
// We don't support GuaranteedTailCallOpt, only automatically-detected
// Join the stores, which are independent of one another.
if (!MemOpChains.empty())
- Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
- &MemOpChains[0], MemOpChains.size());
+ Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);
// Accept direct calls by converting symbolic call addresses to the
// associated Target* opcodes. Force %r1 to be used for indirect
// tail calls.
SDValue Glue;
- if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
+ if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
Callee = DAG.getTargetGlobalAddress(G->getGlobal(), DL, PtrVT);
Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
- } else if (ExternalSymbolSDNode *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
+ } else if (auto *E = dyn_cast<ExternalSymbolSDNode>(Callee)) {
Callee = DAG.getTargetExternalSymbol(E->getSymbol(), PtrVT);
Callee = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Callee);
} else if (IsTailCall) {
Ops.push_back(DAG.getRegister(RegsToPass[I].first,
RegsToPass[I].second.getValueType()));
+ // Add a register mask operand representing the call-preserved registers.
+ const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
+ const uint32_t *Mask = TRI->getCallPreservedMask(MF, CallConv);
+ assert(Mask && "Missing call preserved mask for calling convention");
+ Ops.push_back(DAG.getRegisterMask(Mask));
+
// Glue the call to the argument copies, if any.
if (Glue.getNode())
Ops.push_back(Glue);
// Emit the call.
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
if (IsTailCall)
- return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, &Ops[0], Ops.size());
- Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, &Ops[0], Ops.size());
+ return DAG.getNode(SystemZISD::SIBCALL, DL, NodeTys, Ops);
+ Chain = DAG.getNode(SystemZISD::CALL, DL, NodeTys, Ops);
Glue = Chain.getValue(1);
// Mark the end of the call, which is glued to the call itself.
// Assign locations to each value returned by this call.
SmallVector<CCValAssign, 16> RetLocs;
- CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
+ CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
RetCCInfo.AnalyzeCallResult(Ins, RetCC_SystemZ);
// Copy all of the result registers out of their specified physreg.
// Assign locations to each returned value.
SmallVector<CCValAssign, 16> RetLocs;
- CCState RetCCInfo(CallConv, IsVarArg, MF, TM, RetLocs, *DAG.getContext());
+ CCState RetCCInfo(CallConv, IsVarArg, MF, RetLocs, *DAG.getContext());
RetCCInfo.AnalyzeReturn(Outs, RetCC_SystemZ);
// Quick exit for void returns
if (Glue.getNode())
RetOps.push_back(Glue);
- return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other,
- RetOps.data(), RetOps.size());
+ return DAG.getNode(SystemZISD::RET_FLAG, DL, MVT::Other, RetOps);
+}
+
+SDValue SystemZTargetLowering::
+prepareVolatileOrAtomicLoad(SDValue Chain, SDLoc DL, SelectionDAG &DAG) const {
+ return DAG.getNode(SystemZISD::SERIALIZE, DL, MVT::Other, Chain);
}
// CC is a comparison that will be implemented using an integer or
#undef CONV
}
-// If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
-// can be converted to a comparison against zero, adjust the operands
+// Return a sequence for getting a 1 from an IPM result when CC has a
+// value in CCMask and a 0 when CC has a value in CCValid & ~CCMask.
+// The handling of CC values outside CCValid doesn't matter.
+static IPMConversion getIPMConversion(unsigned CCValid, unsigned CCMask) {
+ // Deal with cases where the result can be taken directly from a bit
+ // of the IPM result.
+ if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_3)))
+ return IPMConversion(0, 0, SystemZ::IPM_CC);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_2 | SystemZ::CCMASK_3)))
+ return IPMConversion(0, 0, SystemZ::IPM_CC + 1);
+
+ // Deal with cases where we can add a value to force the sign bit
+ // to contain the right value. Putting the bit in 31 means we can
+ // use SRL rather than RISBG(L), and also makes it easier to get a
+ // 0/-1 value, so it has priority over the other tests below.
+ //
+ // These sequences rely on the fact that the upper two bits of the
+ // IPM result are zero.
+ uint64_t TopBit = uint64_t(1) << 31;
+ if (CCMask == (CCValid & SystemZ::CCMASK_0))
+ return IPMConversion(0, -(1 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_1)))
+ return IPMConversion(0, -(2 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0
+ | SystemZ::CCMASK_1
+ | SystemZ::CCMASK_2)))
+ return IPMConversion(0, -(3 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & SystemZ::CCMASK_3))
+ return IPMConversion(0, TopBit - (3 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_1
+ | SystemZ::CCMASK_2
+ | SystemZ::CCMASK_3)))
+ return IPMConversion(0, TopBit - (1 << SystemZ::IPM_CC), 31);
+
+ // Next try inverting the value and testing a bit. 0/1 could be
+ // handled this way too, but we dealt with that case above.
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_2)))
+ return IPMConversion(-1, 0, SystemZ::IPM_CC);
+
+ // Handle cases where adding a value forces a non-sign bit to contain
+ // the right value.
+ if (CCMask == (CCValid & (SystemZ::CCMASK_1 | SystemZ::CCMASK_2)))
+ return IPMConversion(0, 1 << SystemZ::IPM_CC, SystemZ::IPM_CC + 1);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0 | SystemZ::CCMASK_3)))
+ return IPMConversion(0, -(1 << SystemZ::IPM_CC), SystemZ::IPM_CC + 1);
+
+ // The remaining cases are 1, 2, 0/1/3 and 0/2/3. All these are
+ // can be done by inverting the low CC bit and applying one of the
+ // sign-based extractions above.
+ if (CCMask == (CCValid & SystemZ::CCMASK_1))
+ return IPMConversion(1 << SystemZ::IPM_CC, -(1 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & SystemZ::CCMASK_2))
+ return IPMConversion(1 << SystemZ::IPM_CC,
+ TopBit - (3 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0
+ | SystemZ::CCMASK_1
+ | SystemZ::CCMASK_3)))
+ return IPMConversion(1 << SystemZ::IPM_CC, -(3 << SystemZ::IPM_CC), 31);
+ if (CCMask == (CCValid & (SystemZ::CCMASK_0
+ | SystemZ::CCMASK_2
+ | SystemZ::CCMASK_3)))
+ return IPMConversion(1 << SystemZ::IPM_CC,
+ TopBit - (1 << SystemZ::IPM_CC), 31);
+
+ llvm_unreachable("Unexpected CC combination");
+}
+
+// If C can be converted to a comparison against zero, adjust the operands
// as necessary.
-static void adjustZeroCmp(SelectionDAG &DAG, bool &IsUnsigned,
- SDValue &CmpOp0, SDValue &CmpOp1,
- unsigned &CCMask) {
- if (IsUnsigned)
+static void adjustZeroCmp(SelectionDAG &DAG, Comparison &C) {
+ if (C.ICmpType == SystemZICMP::UnsignedOnly)
return;
- ConstantSDNode *ConstOp1 = dyn_cast<ConstantSDNode>(CmpOp1.getNode());
+ auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1.getNode());
if (!ConstOp1)
return;
int64_t Value = ConstOp1->getSExtValue();
- if ((Value == -1 && CCMask == SystemZ::CCMASK_CMP_GT) ||
- (Value == -1 && CCMask == SystemZ::CCMASK_CMP_LE) ||
- (Value == 1 && CCMask == SystemZ::CCMASK_CMP_LT) ||
- (Value == 1 && CCMask == SystemZ::CCMASK_CMP_GE)) {
- CCMask ^= SystemZ::CCMASK_CMP_EQ;
- CmpOp1 = DAG.getConstant(0, CmpOp1.getValueType());
+ if ((Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_GT) ||
+ (Value == -1 && C.CCMask == SystemZ::CCMASK_CMP_LE) ||
+ (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_LT) ||
+ (Value == 1 && C.CCMask == SystemZ::CCMASK_CMP_GE)) {
+ C.CCMask ^= SystemZ::CCMASK_CMP_EQ;
+ C.Op1 = DAG.getConstant(0, C.Op1.getValueType());
}
}
-// If a comparison described by IsUnsigned, CCMask, CmpOp0 and CmpOp1
-// is suitable for CLI(Y), CHHSI or CLHHSI, adjust the operands as necessary.
-static void adjustSubwordCmp(SelectionDAG &DAG, bool &IsUnsigned,
- SDValue &CmpOp0, SDValue &CmpOp1,
- unsigned &CCMask) {
+// If a comparison described by C is suitable for CLI(Y), CHHSI or CLHHSI,
+// adjust the operands as necessary.
+static void adjustSubwordCmp(SelectionDAG &DAG, Comparison &C) {
// For us to make any changes, it must a comparison between a single-use
// load and a constant.
- if (!CmpOp0.hasOneUse() ||
- CmpOp0.getOpcode() != ISD::LOAD ||
- CmpOp1.getOpcode() != ISD::Constant)
+ if (!C.Op0.hasOneUse() ||
+ C.Op0.getOpcode() != ISD::LOAD ||
+ C.Op1.getOpcode() != ISD::Constant)
return;
// We must have an 8- or 16-bit load.
- LoadSDNode *Load = cast<LoadSDNode>(CmpOp0);
+ auto *Load = cast<LoadSDNode>(C.Op0);
unsigned NumBits = Load->getMemoryVT().getStoreSizeInBits();
if (NumBits != 8 && NumBits != 16)
return;
// The load must be an extending one and the constant must be within the
// range of the unextended value.
- ConstantSDNode *Constant = cast<ConstantSDNode>(CmpOp1);
- uint64_t Value = Constant->getZExtValue();
+ auto *ConstOp1 = cast<ConstantSDNode>(C.Op1);
+ uint64_t Value = ConstOp1->getZExtValue();
uint64_t Mask = (1 << NumBits) - 1;
if (Load->getExtensionType() == ISD::SEXTLOAD) {
- int64_t SignedValue = Constant->getSExtValue();
- if (uint64_t(SignedValue) + (1ULL << (NumBits - 1)) > Mask)
+ // Make sure that ConstOp1 is in range of C.Op0.
+ int64_t SignedValue = ConstOp1->getSExtValue();
+ if (uint64_t(SignedValue) + (uint64_t(1) << (NumBits - 1)) > Mask)
return;
- // Unsigned comparison between two sign-extended values is equivalent
- // to unsigned comparison between two zero-extended values.
- if (IsUnsigned)
+ if (C.ICmpType != SystemZICMP::SignedOnly) {
+ // Unsigned comparison between two sign-extended values is equivalent
+ // to unsigned comparison between two zero-extended values.
Value &= Mask;
- else if (CCMask == SystemZ::CCMASK_CMP_EQ ||
- CCMask == SystemZ::CCMASK_CMP_NE)
- // Any choice of IsUnsigned is OK for equality comparisons.
- // We could use either CHHSI or CLHHSI for 16-bit comparisons,
- // but since we use CLHHSI for zero extensions, it seems better
- // to be consistent and do the same here.
- Value &= Mask, IsUnsigned = true;
- else if (NumBits == 8) {
+ } else if (NumBits == 8) {
// Try to treat the comparison as unsigned, so that we can use CLI.
// Adjust CCMask and Value as necessary.
- if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_LT)
+ if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_LT)
// Test whether the high bit of the byte is set.
- Value = 127, CCMask = SystemZ::CCMASK_CMP_GT, IsUnsigned = true;
- else if (Value == 0 && CCMask == SystemZ::CCMASK_CMP_GE)
+ Value = 127, C.CCMask = SystemZ::CCMASK_CMP_GT;
+ else if (Value == 0 && C.CCMask == SystemZ::CCMASK_CMP_GE)
// Test whether the high bit of the byte is clear.
- Value = 128, CCMask = SystemZ::CCMASK_CMP_LT, IsUnsigned = true;
+ Value = 128, C.CCMask = SystemZ::CCMASK_CMP_LT;
else
// No instruction exists for this combination.
return;
+ C.ICmpType = SystemZICMP::UnsignedOnly;
}
} else if (Load->getExtensionType() == ISD::ZEXTLOAD) {
if (Value > Mask)
return;
- // Signed comparison between two zero-extended values is equivalent
- // to unsigned comparison.
- IsUnsigned = true;
+ assert(C.ICmpType == SystemZICMP::Any &&
+ "Signedness shouldn't matter here.");
} else
return;
// Make sure that the first operand is an i32 of the right extension type.
- ISD::LoadExtType ExtType = IsUnsigned ? ISD::ZEXTLOAD : ISD::SEXTLOAD;
- if (CmpOp0.getValueType() != MVT::i32 ||
+ ISD::LoadExtType ExtType = (C.ICmpType == SystemZICMP::SignedOnly ?
+ ISD::SEXTLOAD :
+ ISD::ZEXTLOAD);
+ if (C.Op0.getValueType() != MVT::i32 ||
Load->getExtensionType() != ExtType)
- CmpOp0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
- Load->getChain(), Load->getBasePtr(),
- Load->getPointerInfo(), Load->getMemoryVT(),
- Load->isVolatile(), Load->isNonTemporal(),
- Load->getAlignment());
+ C.Op0 = DAG.getExtLoad(ExtType, SDLoc(Load), MVT::i32,
+ Load->getChain(), Load->getBasePtr(),
+ Load->getPointerInfo(), Load->getMemoryVT(),
+ Load->isVolatile(), Load->isNonTemporal(),
+ Load->isInvariant(), Load->getAlignment());
// Make sure that the second operand is an i32 with the right value.
- if (CmpOp1.getValueType() != MVT::i32 ||
- Value != Constant->getZExtValue())
- CmpOp1 = DAG.getConstant(Value, MVT::i32);
+ if (C.Op1.getValueType() != MVT::i32 ||
+ Value != ConstOp1->getZExtValue())
+ C.Op1 = DAG.getConstant(Value, MVT::i32);
}
// Return true if Op is either an unextended load, or a load suitable
// for integer register-memory comparisons of type ICmpType.
static bool isNaturalMemoryOperand(SDValue Op, unsigned ICmpType) {
- LoadSDNode *Load = dyn_cast<LoadSDNode>(Op.getNode());
+ auto *Load = dyn_cast<LoadSDNode>(Op.getNode());
if (Load) {
// There are no instructions to compare a register with a memory byte.
if (Load->getMemoryVT() == MVT::i8)
return false;
}
-// Return true if it is better to swap comparison operands Op0 and Op1.
-// ICmpType is the type of an integer comparison.
-static bool shouldSwapCmpOperands(SDValue Op0, SDValue Op1,
- unsigned ICmpType) {
+// Return true if it is better to swap the operands of C.
+static bool shouldSwapCmpOperands(const Comparison &C) {
// Leave f128 comparisons alone, since they have no memory forms.
- if (Op0.getValueType() == MVT::f128)
+ if (C.Op0.getValueType() == MVT::f128)
return false;
// Always keep a floating-point constant second, since comparisons with
// zero can use LOAD TEST and comparisons with other constants make a
// natural memory operand.
- if (isa<ConstantFPSDNode>(Op1))
+ if (isa<ConstantFPSDNode>(C.Op1))
return false;
// Never swap comparisons with zero since there are many ways to optimize
// those later.
- ConstantSDNode *COp1 = dyn_cast<ConstantSDNode>(Op1);
- if (COp1 && COp1->getZExtValue() == 0)
+ auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
+ if (ConstOp1 && ConstOp1->getZExtValue() == 0)
+ return false;
+
+ // Also keep natural memory operands second if the loaded value is
+ // only used here. Several comparisons have memory forms.
+ if (isNaturalMemoryOperand(C.Op1, C.ICmpType) && C.Op1.hasOneUse())
return false;
// Look for cases where Cmp0 is a single-use load and Cmp1 isn't.
// In that case we generally prefer the memory to be second.
- if ((isNaturalMemoryOperand(Op0, ICmpType) && Op0.hasOneUse()) &&
- !(isNaturalMemoryOperand(Op1, ICmpType) && Op1.hasOneUse())) {
+ if (isNaturalMemoryOperand(C.Op0, C.ICmpType) && C.Op0.hasOneUse()) {
// The only exceptions are when the second operand is a constant and
// we can use things like CHHSI.
- if (!COp1)
+ if (!ConstOp1)
return true;
// The unsigned memory-immediate instructions can handle 16-bit
// unsigned integers.
- if (ICmpType != SystemZICMP::SignedOnly &&
- isUInt<16>(COp1->getZExtValue()))
+ if (C.ICmpType != SystemZICMP::SignedOnly &&
+ isUInt<16>(ConstOp1->getZExtValue()))
return false;
// The signed memory-immediate instructions can handle 16-bit
// signed integers.
- if (ICmpType != SystemZICMP::UnsignedOnly &&
- isInt<16>(COp1->getSExtValue()))
+ if (C.ICmpType != SystemZICMP::UnsignedOnly &&
+ isInt<16>(ConstOp1->getSExtValue()))
return false;
return true;
}
+
+ // Try to promote the use of CGFR and CLGFR.
+ unsigned Opcode0 = C.Op0.getOpcode();
+ if (C.ICmpType != SystemZICMP::UnsignedOnly && Opcode0 == ISD::SIGN_EXTEND)
+ return true;
+ if (C.ICmpType != SystemZICMP::SignedOnly && Opcode0 == ISD::ZERO_EXTEND)
+ return true;
+ if (C.ICmpType != SystemZICMP::SignedOnly &&
+ Opcode0 == ISD::AND &&
+ C.Op0.getOperand(1).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(C.Op0.getOperand(1))->getZExtValue() == 0xffffffff)
+ return true;
+
return false;
}
+// Return a version of comparison CC mask CCMask in which the LT and GT
+// actions are swapped.
+static unsigned reverseCCMask(unsigned CCMask) {
+ return ((CCMask & SystemZ::CCMASK_CMP_EQ) |
+ (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
+ (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
+ (CCMask & SystemZ::CCMASK_CMP_UO));
+}
+
+// Check whether C tests for equality between X and Y and whether X - Y
+// or Y - X is also computed. In that case it's better to compare the
+// result of the subtraction against zero.
+static void adjustForSubtraction(SelectionDAG &DAG, Comparison &C) {
+ if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
+ C.CCMask == SystemZ::CCMASK_CMP_NE) {
+ for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
+ SDNode *N = *I;
+ if (N->getOpcode() == ISD::SUB &&
+ ((N->getOperand(0) == C.Op0 && N->getOperand(1) == C.Op1) ||
+ (N->getOperand(0) == C.Op1 && N->getOperand(1) == C.Op0))) {
+ C.Op0 = SDValue(N, 0);
+ C.Op1 = DAG.getConstant(0, N->getValueType(0));
+ return;
+ }
+ }
+ }
+}
+
+// Check whether C compares a floating-point value with zero and if that
+// floating-point value is also negated. In this case we can use the
+// negation to set CC, so avoiding separate LOAD AND TEST and
+// LOAD (NEGATIVE/COMPLEMENT) instructions.
+static void adjustForFNeg(Comparison &C) {
+ auto *C1 = dyn_cast<ConstantFPSDNode>(C.Op1);
+ if (C1 && C1->isZero()) {
+ for (auto I = C.Op0->use_begin(), E = C.Op0->use_end(); I != E; ++I) {
+ SDNode *N = *I;
+ if (N->getOpcode() == ISD::FNEG) {
+ C.Op0 = SDValue(N, 0);
+ C.CCMask = reverseCCMask(C.CCMask);
+ return;
+ }
+ }
+ }
+}
+
+// Check whether C compares (shl X, 32) with 0 and whether X is
+// also sign-extended. In that case it is better to test the result
+// of the sign extension using LTGFR.
+//
+// This case is important because InstCombine transforms a comparison
+// with (sext (trunc X)) into a comparison with (shl X, 32).
+static void adjustForLTGFR(Comparison &C) {
+ // Check for a comparison between (shl X, 32) and 0.
+ if (C.Op0.getOpcode() == ISD::SHL &&
+ C.Op0.getValueType() == MVT::i64 &&
+ C.Op1.getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
+ auto *C1 = dyn_cast<ConstantSDNode>(C.Op0.getOperand(1));
+ if (C1 && C1->getZExtValue() == 32) {
+ SDValue ShlOp0 = C.Op0.getOperand(0);
+ // See whether X has any SIGN_EXTEND_INREG uses.
+ for (auto I = ShlOp0->use_begin(), E = ShlOp0->use_end(); I != E; ++I) {
+ SDNode *N = *I;
+ if (N->getOpcode() == ISD::SIGN_EXTEND_INREG &&
+ cast<VTSDNode>(N->getOperand(1))->getVT() == MVT::i32) {
+ C.Op0 = SDValue(N, 0);
+ return;
+ }
+ }
+ }
+ }
+}
+
+// If C compares the truncation of an extending load, try to compare
+// the untruncated value instead. This exposes more opportunities to
+// reuse CC.
+static void adjustICmpTruncate(SelectionDAG &DAG, Comparison &C) {
+ if (C.Op0.getOpcode() == ISD::TRUNCATE &&
+ C.Op0.getOperand(0).getOpcode() == ISD::LOAD &&
+ C.Op1.getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
+ auto *L = cast<LoadSDNode>(C.Op0.getOperand(0));
+ if (L->getMemoryVT().getStoreSizeInBits()
+ <= C.Op0.getValueType().getSizeInBits()) {
+ unsigned Type = L->getExtensionType();
+ if ((Type == ISD::ZEXTLOAD && C.ICmpType != SystemZICMP::SignedOnly) ||
+ (Type == ISD::SEXTLOAD && C.ICmpType != SystemZICMP::UnsignedOnly)) {
+ C.Op0 = C.Op0.getOperand(0);
+ C.Op1 = DAG.getConstant(0, C.Op0.getValueType());
+ }
+ }
+ }
+}
+
// Return true if shift operation N has an in-range constant shift value.
// Store it in ShiftVal if so.
static bool isSimpleShift(SDValue N, unsigned &ShiftVal) {
- ConstantSDNode *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
+ auto *Shift = dyn_cast<ConstantSDNode>(N.getOperand(1));
if (!Shift)
return false;
return 0;
}
-// See whether the comparison (Opcode CmpOp0, CmpOp1, ICmpType) can be
-// implemented as a TEST UNDER MASK instruction when the condition being
-// tested is as described by CCValid and CCMask. Update the arguments
-// with the TM version if so.
-static void adjustForTestUnderMask(SelectionDAG &DAG, unsigned &Opcode,
- SDValue &CmpOp0, SDValue &CmpOp1,
- unsigned &CCValid, unsigned &CCMask,
- unsigned &ICmpType) {
+// See whether C can be implemented as a TEST UNDER MASK instruction.
+// Update the arguments with the TM version if so.
+static void adjustForTestUnderMask(SelectionDAG &DAG, Comparison &C) {
// Check that we have a comparison with a constant.
- ConstantSDNode *ConstCmpOp1 = dyn_cast<ConstantSDNode>(CmpOp1);
- if (!ConstCmpOp1)
+ auto *ConstOp1 = dyn_cast<ConstantSDNode>(C.Op1);
+ if (!ConstOp1)
return;
- uint64_t CmpVal = ConstCmpOp1->getZExtValue();
+ uint64_t CmpVal = ConstOp1->getZExtValue();
// Check whether the nonconstant input is an AND with a constant mask.
- if (CmpOp0.getOpcode() != ISD::AND)
- return;
- SDValue AndOp0 = CmpOp0.getOperand(0);
- SDValue AndOp1 = CmpOp0.getOperand(1);
- ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(AndOp1.getNode());
- if (!Mask)
- return;
- uint64_t MaskVal = Mask->getZExtValue();
+ Comparison NewC(C);
+ uint64_t MaskVal;
+ ConstantSDNode *Mask = nullptr;
+ if (C.Op0.getOpcode() == ISD::AND) {
+ NewC.Op0 = C.Op0.getOperand(0);
+ NewC.Op1 = C.Op0.getOperand(1);
+ Mask = dyn_cast<ConstantSDNode>(NewC.Op1);
+ if (!Mask)
+ return;
+ MaskVal = Mask->getZExtValue();
+ } else {
+ // There is no instruction to compare with a 64-bit immediate
+ // so use TMHH instead if possible. We need an unsigned ordered
+ // comparison with an i64 immediate.
+ if (NewC.Op0.getValueType() != MVT::i64 ||
+ NewC.CCMask == SystemZ::CCMASK_CMP_EQ ||
+ NewC.CCMask == SystemZ::CCMASK_CMP_NE ||
+ NewC.ICmpType == SystemZICMP::SignedOnly)
+ return;
+ // Convert LE and GT comparisons into LT and GE.
+ if (NewC.CCMask == SystemZ::CCMASK_CMP_LE ||
+ NewC.CCMask == SystemZ::CCMASK_CMP_GT) {
+ if (CmpVal == uint64_t(-1))
+ return;
+ CmpVal += 1;
+ NewC.CCMask ^= SystemZ::CCMASK_CMP_EQ;
+ }
+ // If the low N bits of Op1 are zero than the low N bits of Op0 can
+ // be masked off without changing the result.
+ MaskVal = -(CmpVal & -CmpVal);
+ NewC.ICmpType = SystemZICMP::UnsignedOnly;
+ }
// Check whether the combination of mask, comparison value and comparison
// type are suitable.
- unsigned BitSize = CmpOp0.getValueType().getSizeInBits();
+ unsigned BitSize = NewC.Op0.getValueType().getSizeInBits();
unsigned NewCCMask, ShiftVal;
- if (ICmpType != SystemZICMP::SignedOnly &&
- AndOp0.getOpcode() == ISD::SHL &&
- isSimpleShift(AndOp0, ShiftVal) &&
- (NewCCMask = getTestUnderMaskCond(BitSize, CCMask, MaskVal >> ShiftVal,
+ if (NewC.ICmpType != SystemZICMP::SignedOnly &&
+ NewC.Op0.getOpcode() == ISD::SHL &&
+ isSimpleShift(NewC.Op0, ShiftVal) &&
+ (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
+ MaskVal >> ShiftVal,
CmpVal >> ShiftVal,
SystemZICMP::Any))) {
- AndOp0 = AndOp0.getOperand(0);
- AndOp1 = DAG.getConstant(MaskVal >> ShiftVal, AndOp0.getValueType());
- } else if (ICmpType != SystemZICMP::SignedOnly &&
- AndOp0.getOpcode() == ISD::SRL &&
- isSimpleShift(AndOp0, ShiftVal) &&
- (NewCCMask = getTestUnderMaskCond(BitSize, CCMask,
+ NewC.Op0 = NewC.Op0.getOperand(0);
+ MaskVal >>= ShiftVal;
+ } else if (NewC.ICmpType != SystemZICMP::SignedOnly &&
+ NewC.Op0.getOpcode() == ISD::SRL &&
+ isSimpleShift(NewC.Op0, ShiftVal) &&
+ (NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask,
MaskVal << ShiftVal,
CmpVal << ShiftVal,
SystemZICMP::UnsignedOnly))) {
- AndOp0 = AndOp0.getOperand(0);
- AndOp1 = DAG.getConstant(MaskVal << ShiftVal, AndOp0.getValueType());
+ NewC.Op0 = NewC.Op0.getOperand(0);
+ MaskVal <<= ShiftVal;
} else {
- NewCCMask = getTestUnderMaskCond(BitSize, CCMask, MaskVal, CmpVal,
- ICmpType);
+ NewCCMask = getTestUnderMaskCond(BitSize, NewC.CCMask, MaskVal, CmpVal,
+ NewC.ICmpType);
if (!NewCCMask)
return;
}
// Go ahead and make the change.
- Opcode = SystemZISD::TM;
- CmpOp0 = AndOp0;
- CmpOp1 = AndOp1;
- ICmpType = (bool(NewCCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
- bool(NewCCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
- CCValid = SystemZ::CCMASK_TM;
- CCMask = NewCCMask;
-}
-
-// Return a target node that compares CmpOp0 with CmpOp1 and stores a
-// 2-bit result in CC. Set CCValid to the CCMASK_* of all possible
-// 2-bit results and CCMask to the subset of those results that are
-// associated with Cond.
-static SDValue emitCmp(const SystemZTargetMachine &TM, SelectionDAG &DAG,
- SDLoc DL, SDValue CmpOp0, SDValue CmpOp1,
- ISD::CondCode Cond, unsigned &CCValid,
- unsigned &CCMask) {
- bool IsUnsigned = false;
- CCMask = CCMaskForCondCode(Cond);
- unsigned Opcode, ICmpType = 0;
- if (CmpOp0.getValueType().isFloatingPoint()) {
- CCValid = SystemZ::CCMASK_FCMP;
- Opcode = SystemZISD::FCMP;
+ C.Opcode = SystemZISD::TM;
+ C.Op0 = NewC.Op0;
+ if (Mask && Mask->getZExtValue() == MaskVal)
+ C.Op1 = SDValue(Mask, 0);
+ else
+ C.Op1 = DAG.getConstant(MaskVal, C.Op0.getValueType());
+ C.CCValid = SystemZ::CCMASK_TM;
+ C.CCMask = NewCCMask;
+}
+
+// Decide how to implement a comparison of type Cond between CmpOp0 with CmpOp1.
+static Comparison getCmp(SelectionDAG &DAG, SDValue CmpOp0, SDValue CmpOp1,
+ ISD::CondCode Cond) {
+ Comparison C(CmpOp0, CmpOp1);
+ C.CCMask = CCMaskForCondCode(Cond);
+ if (C.Op0.getValueType().isFloatingPoint()) {
+ C.CCValid = SystemZ::CCMASK_FCMP;
+ C.Opcode = SystemZISD::FCMP;
+ adjustForFNeg(C);
} else {
- IsUnsigned = CCMask & SystemZ::CCMASK_CMP_UO;
- CCValid = SystemZ::CCMASK_ICMP;
- CCMask &= CCValid;
- adjustZeroCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask);
- adjustSubwordCmp(DAG, IsUnsigned, CmpOp0, CmpOp1, CCMask);
- Opcode = SystemZISD::ICMP;
+ C.CCValid = SystemZ::CCMASK_ICMP;
+ C.Opcode = SystemZISD::ICMP;
// Choose the type of comparison. Equality and inequality tests can
// use either signed or unsigned comparisons. The choice also doesn't
// matter if both sign bits are known to be clear. In those cases we
// want to give the main isel code the freedom to choose whichever
// form fits best.
- if (CCMask == SystemZ::CCMASK_CMP_EQ ||
- CCMask == SystemZ::CCMASK_CMP_NE ||
- (DAG.SignBitIsZero(CmpOp0) && DAG.SignBitIsZero(CmpOp1)))
- ICmpType = SystemZICMP::Any;
- else if (IsUnsigned)
- ICmpType = SystemZICMP::UnsignedOnly;
+ if (C.CCMask == SystemZ::CCMASK_CMP_EQ ||
+ C.CCMask == SystemZ::CCMASK_CMP_NE ||
+ (DAG.SignBitIsZero(C.Op0) && DAG.SignBitIsZero(C.Op1)))
+ C.ICmpType = SystemZICMP::Any;
+ else if (C.CCMask & SystemZ::CCMASK_CMP_UO)
+ C.ICmpType = SystemZICMP::UnsignedOnly;
else
- ICmpType = SystemZICMP::SignedOnly;
+ C.ICmpType = SystemZICMP::SignedOnly;
+ C.CCMask &= ~SystemZ::CCMASK_CMP_UO;
+ adjustZeroCmp(DAG, C);
+ adjustSubwordCmp(DAG, C);
+ adjustForSubtraction(DAG, C);
+ adjustForLTGFR(C);
+ adjustICmpTruncate(DAG, C);
}
- if (shouldSwapCmpOperands(CmpOp0, CmpOp1, ICmpType)) {
- std::swap(CmpOp0, CmpOp1);
- CCMask = ((CCMask & SystemZ::CCMASK_CMP_EQ) |
- (CCMask & SystemZ::CCMASK_CMP_GT ? SystemZ::CCMASK_CMP_LT : 0) |
- (CCMask & SystemZ::CCMASK_CMP_LT ? SystemZ::CCMASK_CMP_GT : 0) |
- (CCMask & SystemZ::CCMASK_CMP_UO));
+ if (shouldSwapCmpOperands(C)) {
+ std::swap(C.Op0, C.Op1);
+ C.CCMask = reverseCCMask(C.CCMask);
}
- adjustForTestUnderMask(DAG, Opcode, CmpOp0, CmpOp1, CCValid, CCMask,
- ICmpType);
- if (Opcode == SystemZISD::ICMP || Opcode == SystemZISD::TM)
- return DAG.getNode(Opcode, DL, MVT::Glue, CmpOp0, CmpOp1,
- DAG.getConstant(ICmpType, MVT::i32));
- return DAG.getNode(Opcode, DL, MVT::Glue, CmpOp0, CmpOp1);
+ adjustForTestUnderMask(DAG, C);
+ return C;
+}
+
+// Emit the comparison instruction described by C.
+static SDValue emitCmp(SelectionDAG &DAG, SDLoc DL, Comparison &C) {
+ if (C.Opcode == SystemZISD::ICMP)
+ return DAG.getNode(SystemZISD::ICMP, DL, MVT::Glue, C.Op0, C.Op1,
+ DAG.getConstant(C.ICmpType, MVT::i32));
+ if (C.Opcode == SystemZISD::TM) {
+ bool RegisterOnly = (bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_0) !=
+ bool(C.CCMask & SystemZ::CCMASK_TM_MIXED_MSB_1));
+ return DAG.getNode(SystemZISD::TM, DL, MVT::Glue, C.Op0, C.Op1,
+ DAG.getConstant(RegisterOnly, MVT::i32));
+ }
+ return DAG.getNode(C.Opcode, DL, MVT::Glue, C.Op0, C.Op1);
}
// Implement a 32-bit *MUL_LOHI operation by extending both operands to
Odd = DAG.getTargetExtractSubreg(SystemZ::odd128(Is32Bit), DL, VT, Result);
}
+// Return an i32 value that is 1 if the CC value produced by Glue is
+// in the mask CCMask and 0 otherwise. CC is known to have a value
+// in CCValid, so other values can be ignored.
+static SDValue emitSETCC(SelectionDAG &DAG, SDLoc DL, SDValue Glue,
+ unsigned CCValid, unsigned CCMask) {
+ IPMConversion Conversion = getIPMConversion(CCValid, CCMask);
+ SDValue Result = DAG.getNode(SystemZISD::IPM, DL, MVT::i32, Glue);
+
+ if (Conversion.XORValue)
+ Result = DAG.getNode(ISD::XOR, DL, MVT::i32, Result,
+ DAG.getConstant(Conversion.XORValue, MVT::i32));
+
+ if (Conversion.AddValue)
+ Result = DAG.getNode(ISD::ADD, DL, MVT::i32, Result,
+ DAG.getConstant(Conversion.AddValue, MVT::i32));
+
+ // The SHR/AND sequence should get optimized to an RISBG.
+ Result = DAG.getNode(ISD::SRL, DL, MVT::i32, Result,
+ DAG.getConstant(Conversion.Bit, MVT::i32));
+ if (Conversion.Bit != 31)
+ Result = DAG.getNode(ISD::AND, DL, MVT::i32, Result,
+ DAG.getConstant(1, MVT::i32));
+ return Result;
+}
+
+SDValue SystemZTargetLowering::lowerSETCC(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDValue CmpOp0 = Op.getOperand(0);
+ SDValue CmpOp1 = Op.getOperand(1);
+ ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
+ SDLoc DL(Op);
+
+ Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
+ SDValue Glue = emitCmp(DAG, DL, C);
+ return emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
+}
+
SDValue SystemZTargetLowering::lowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
SDValue Chain = Op.getOperand(0);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
SDValue Dest = Op.getOperand(4);
SDLoc DL(Op);
- unsigned CCValid, CCMask;
- SDValue Flags = emitCmp(TM, DAG, DL, CmpOp0, CmpOp1, CC, CCValid, CCMask);
+ Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
+ SDValue Glue = emitCmp(DAG, DL, C);
return DAG.getNode(SystemZISD::BR_CCMASK, DL, Op.getValueType(),
- Chain, DAG.getConstant(CCValid, MVT::i32),
- DAG.getConstant(CCMask, MVT::i32), Dest, Flags);
+ Chain, DAG.getConstant(C.CCValid, MVT::i32),
+ DAG.getConstant(C.CCMask, MVT::i32), Dest, Glue);
+}
+
+// Return true if Pos is CmpOp and Neg is the negative of CmpOp,
+// allowing Pos and Neg to be wider than CmpOp.
+static bool isAbsolute(SDValue CmpOp, SDValue Pos, SDValue Neg) {
+ return (Neg.getOpcode() == ISD::SUB &&
+ Neg.getOperand(0).getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(Neg.getOperand(0))->getZExtValue() == 0 &&
+ Neg.getOperand(1) == Pos &&
+ (Pos == CmpOp ||
+ (Pos.getOpcode() == ISD::SIGN_EXTEND &&
+ Pos.getOperand(0) == CmpOp)));
+}
+
+// Return the absolute or negative absolute of Op; IsNegative decides which.
+static SDValue getAbsolute(SelectionDAG &DAG, SDLoc DL, SDValue Op,
+ bool IsNegative) {
+ Op = DAG.getNode(SystemZISD::IABS, DL, Op.getValueType(), Op);
+ if (IsNegative)
+ Op = DAG.getNode(ISD::SUB, DL, Op.getValueType(),
+ DAG.getConstant(0, Op.getValueType()), Op);
+ return Op;
}
SDValue SystemZTargetLowering::lowerSELECT_CC(SDValue Op,
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
SDLoc DL(Op);
- unsigned CCValid, CCMask;
- SDValue Flags = emitCmp(TM, DAG, DL, CmpOp0, CmpOp1, CC, CCValid, CCMask);
+ Comparison C(getCmp(DAG, CmpOp0, CmpOp1, CC));
+
+ // Check for absolute and negative-absolute selections, including those
+ // where the comparison value is sign-extended (for LPGFR and LNGFR).
+ // This check supplements the one in DAGCombiner.
+ if (C.Opcode == SystemZISD::ICMP &&
+ C.CCMask != SystemZ::CCMASK_CMP_EQ &&
+ C.CCMask != SystemZ::CCMASK_CMP_NE &&
+ C.Op1.getOpcode() == ISD::Constant &&
+ cast<ConstantSDNode>(C.Op1)->getZExtValue() == 0) {
+ if (isAbsolute(C.Op0, TrueOp, FalseOp))
+ return getAbsolute(DAG, DL, TrueOp, C.CCMask & SystemZ::CCMASK_CMP_LT);
+ if (isAbsolute(C.Op0, FalseOp, TrueOp))
+ return getAbsolute(DAG, DL, FalseOp, C.CCMask & SystemZ::CCMASK_CMP_GT);
+ }
+
+ SDValue Glue = emitCmp(DAG, DL, C);
+
+ // Special case for handling -1/0 results. The shifts we use here
+ // should get optimized with the IPM conversion sequence.
+ auto *TrueC = dyn_cast<ConstantSDNode>(TrueOp);
+ auto *FalseC = dyn_cast<ConstantSDNode>(FalseOp);
+ if (TrueC && FalseC) {
+ int64_t TrueVal = TrueC->getSExtValue();
+ int64_t FalseVal = FalseC->getSExtValue();
+ if ((TrueVal == -1 && FalseVal == 0) || (TrueVal == 0 && FalseVal == -1)) {
+ // Invert the condition if we want -1 on false.
+ if (TrueVal == 0)
+ C.CCMask ^= C.CCValid;
+ SDValue Result = emitSETCC(DAG, DL, Glue, C.CCValid, C.CCMask);
+ EVT VT = Op.getValueType();
+ // Extend the result to VT. Upper bits are ignored.
+ if (!is32Bit(VT))
+ Result = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Result);
+ // Sign-extend from the low bit.
+ SDValue ShAmt = DAG.getConstant(VT.getSizeInBits() - 1, MVT::i32);
+ SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Result, ShAmt);
+ return DAG.getNode(ISD::SRA, DL, VT, Shl, ShAmt);
+ }
+ }
- SmallVector<SDValue, 5> Ops;
- Ops.push_back(TrueOp);
- Ops.push_back(FalseOp);
- Ops.push_back(DAG.getConstant(CCValid, MVT::i32));
- Ops.push_back(DAG.getConstant(CCMask, MVT::i32));
- Ops.push_back(Flags);
+ SDValue Ops[] = {TrueOp, FalseOp, DAG.getConstant(C.CCValid, MVT::i32),
+ DAG.getConstant(C.CCMask, MVT::i32), Glue};
SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
- return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, &Ops[0], Ops.size());
+ return DAG.getNode(SystemZISD::SELECT_CCMASK, DL, VTs, Ops);
}
SDValue SystemZTargetLowering::lowerGlobalAddress(GlobalAddressSDNode *Node,
const GlobalValue *GV = Node->getGlobal();
int64_t Offset = Node->getOffset();
EVT PtrVT = getPointerTy();
- Reloc::Model RM = TM.getRelocationModel();
- CodeModel::Model CM = TM.getCodeModel();
+ Reloc::Model RM = DAG.getTarget().getRelocationModel();
+ CodeModel::Model CM = DAG.getTarget().getCodeModel();
SDValue Result;
if (Subtarget.isPC32DBLSymbol(GV, RM, CM)) {
return Result;
}
+SDValue SystemZTargetLowering::lowerTLSGetOffset(GlobalAddressSDNode *Node,
+ SelectionDAG &DAG,
+ unsigned Opcode,
+ SDValue GOTOffset) const {
+ SDLoc DL(Node);
+ EVT PtrVT = getPointerTy();
+ SDValue Chain = DAG.getEntryNode();
+ SDValue Glue;
+
+ // __tls_get_offset takes the GOT offset in %r2 and the GOT in %r12.
+ SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
+ Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R12D, GOT, Glue);
+ Glue = Chain.getValue(1);
+ Chain = DAG.getCopyToReg(Chain, DL, SystemZ::R2D, GOTOffset, Glue);
+ Glue = Chain.getValue(1);
+
+ // The first call operand is the chain and the second is the TLS symbol.
+ SmallVector<SDValue, 8> Ops;
+ Ops.push_back(Chain);
+ Ops.push_back(DAG.getTargetGlobalAddress(Node->getGlobal(), DL,
+ Node->getValueType(0),
+ 0, 0));
+
+ // Add argument registers to the end of the list so that they are
+ // known live into the call.
+ Ops.push_back(DAG.getRegister(SystemZ::R2D, PtrVT));
+ Ops.push_back(DAG.getRegister(SystemZ::R12D, PtrVT));
+
+ // Add a register mask operand representing the call-preserved registers.
+ const TargetRegisterInfo *TRI = Subtarget.getRegisterInfo();
+ const uint32_t *Mask =
+ TRI->getCallPreservedMask(DAG.getMachineFunction(), CallingConv::C);
+ assert(Mask && "Missing call preserved mask for calling convention");
+ Ops.push_back(DAG.getRegisterMask(Mask));
+
+ // Glue the call to the argument copies.
+ Ops.push_back(Glue);
+
+ // Emit the call.
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
+ Chain = DAG.getNode(Opcode, DL, NodeTys, Ops);
+ Glue = Chain.getValue(1);
+
+ // Copy the return value from %r2.
+ return DAG.getCopyFromReg(Chain, DL, SystemZ::R2D, PtrVT, Glue);
+}
+
SDValue SystemZTargetLowering::lowerGlobalTLSAddress(GlobalAddressSDNode *Node,
SelectionDAG &DAG) const {
SDLoc DL(Node);
const GlobalValue *GV = Node->getGlobal();
EVT PtrVT = getPointerTy();
- TLSModel::Model model = TM.getTLSModel(GV);
-
- if (model != TLSModel::LocalExec)
- llvm_unreachable("only local-exec TLS mode supported");
+ TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
// The high part of the thread pointer is in access register 0.
SDValue TPHi = DAG.getNode(SystemZISD::EXTRACT_ACCESS, DL, MVT::i32,
DAG.getConstant(32, PtrVT));
SDValue TP = DAG.getNode(ISD::OR, DL, PtrVT, TPHiShifted, TPLo);
- // Get the offset of GA from the thread pointer.
- SystemZConstantPoolValue *CPV =
- SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
+ // Get the offset of GA from the thread pointer, based on the TLS model.
+ SDValue Offset;
+ switch (model) {
+ case TLSModel::GeneralDynamic: {
+ // Load the GOT offset of the tls_index (module ID / per-symbol offset).
+ SystemZConstantPoolValue *CPV =
+ SystemZConstantPoolValue::Create(GV, SystemZCP::TLSGD);
+
+ Offset = DAG.getConstantPool(CPV, PtrVT, 8);
+ Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
+ Offset, MachinePointerInfo::getConstantPool(),
+ false, false, false, 0);
+
+ // Call __tls_get_offset to retrieve the offset.
+ Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_GDCALL, Offset);
+ break;
+ }
+
+ case TLSModel::LocalDynamic: {
+ // Load the GOT offset of the module ID.
+ SystemZConstantPoolValue *CPV =
+ SystemZConstantPoolValue::Create(GV, SystemZCP::TLSLDM);
- // Force the offset into the constant pool and load it from there.
- SDValue CPAddr = DAG.getConstantPool(CPV, PtrVT, 8);
- SDValue Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
- CPAddr, MachinePointerInfo::getConstantPool(),
- false, false, false, 0);
+ Offset = DAG.getConstantPool(CPV, PtrVT, 8);
+ Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
+ Offset, MachinePointerInfo::getConstantPool(),
+ false, false, false, 0);
+
+ // Call __tls_get_offset to retrieve the module base offset.
+ Offset = lowerTLSGetOffset(Node, DAG, SystemZISD::TLS_LDCALL, Offset);
+
+ // Note: The SystemZLDCleanupPass will remove redundant computations
+ // of the module base offset. Count total number of local-dynamic
+ // accesses to trigger execution of that pass.
+ SystemZMachineFunctionInfo* MFI =
+ DAG.getMachineFunction().getInfo<SystemZMachineFunctionInfo>();
+ MFI->incNumLocalDynamicTLSAccesses();
+
+ // Add the per-symbol offset.
+ CPV = SystemZConstantPoolValue::Create(GV, SystemZCP::DTPOFF);
+
+ SDValue DTPOffset = DAG.getConstantPool(CPV, PtrVT, 8);
+ DTPOffset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
+ DTPOffset, MachinePointerInfo::getConstantPool(),
+ false, false, false, 0);
+
+ Offset = DAG.getNode(ISD::ADD, DL, PtrVT, Offset, DTPOffset);
+ break;
+ }
+
+ case TLSModel::InitialExec: {
+ // Load the offset from the GOT.
+ Offset = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
+ SystemZII::MO_INDNTPOFF);
+ Offset = DAG.getNode(SystemZISD::PCREL_WRAPPER, DL, PtrVT, Offset);
+ Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
+ Offset, MachinePointerInfo::getGOT(),
+ false, false, false, 0);
+ break;
+ }
+
+ case TLSModel::LocalExec: {
+ // Force the offset into the constant pool and load it from there.
+ SystemZConstantPoolValue *CPV =
+ SystemZConstantPoolValue::Create(GV, SystemZCP::NTPOFF);
+
+ Offset = DAG.getConstantPool(CPV, PtrVT, 8);
+ Offset = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(),
+ Offset, MachinePointerInfo::getConstantPool(),
+ false, false, false, 0);
+ break;
+ }
+ }
// Add the base and offset together.
return DAG.getNode(ISD::ADD, DL, PtrVT, TP, Offset);
false, false, 0);
Offset += 8;
}
- return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps, NumFields);
+ return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
}
SDValue SystemZTargetLowering::lowerVACOPY(SDValue Op,
SDValue Result = DAG.getNode(ISD::ADD, DL, MVT::i64, NewSP, ArgAdjust);
SDValue Ops[2] = { Result, Chain };
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue SystemZTargetLowering::lowerSMUL_LOHI(SDValue Op,
SDValue NegSum = DAG.getNode(ISD::ADD, DL, VT, NegLLTimesRH, NegLHTimesRL);
Ops[1] = DAG.getNode(ISD::SUB, DL, VT, Ops[1], NegSum);
}
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue SystemZTargetLowering::lowerUMUL_LOHI(SDValue Op,
// low half first, so the results are in reverse order.
lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, SystemZISD::UMUL_LOHI64,
Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue SystemZTargetLowering::lowerSDIVREM(SDValue Op,
SDValue Ops[2];
lowerGR128Binary(DAG, DL, VT, SystemZ::AEXT128_64, Opcode,
Op0, Op1, Ops[1], Ops[0]);
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue SystemZTargetLowering::lowerUDIVREM(SDValue Op,
else
lowerGR128Binary(DAG, DL, VT, SystemZ::ZEXT128_64, SystemZISD::UDIVREM64,
Op.getOperand(0), Op.getOperand(1), Ops[1], Ops[0]);
- return DAG.getMergeValues(Ops, 2, DL);
+ return DAG.getMergeValues(Ops, DL);
}
SDValue SystemZTargetLowering::lowerOR(SDValue Op, SelectionDAG &DAG) const {
// Get the known-zero masks for each operand.
SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1) };
APInt KnownZero[2], KnownOne[2];
- DAG.ComputeMaskedBits(Ops[0], KnownZero[0], KnownOne[0]);
- DAG.ComputeMaskedBits(Ops[1], KnownZero[1], KnownOne[1]);
+ DAG.computeKnownBits(Ops[0], KnownZero[0], KnownOne[0]);
+ DAG.computeKnownBits(Ops[1], KnownZero[1], KnownOne[1]);
// See if the upper 32 bits of one operand and the lower 32 bits of the
// other are known zero. They are the low and high operands respectively.
// high 32 bits and just masks out low bits. We can skip it if so.
if (HighOp.getOpcode() == ISD::AND &&
HighOp.getOperand(1).getOpcode() == ISD::Constant) {
- ConstantSDNode *MaskNode = cast<ConstantSDNode>(HighOp.getOperand(1));
- uint64_t Mask = MaskNode->getZExtValue() | Masks[High];
- if ((Mask >> 32) == 0xffffffff)
- HighOp = HighOp.getOperand(0);
+ SDValue HighOp0 = HighOp.getOperand(0);
+ uint64_t Mask = cast<ConstantSDNode>(HighOp.getOperand(1))->getZExtValue();
+ if (DAG.MaskedValueIsZero(HighOp0, APInt(64, ~(Mask | 0xffffffff))))
+ HighOp = HighOp0;
}
// Take advantage of the fact that all GR32 operations only change the
MVT::i64, HighOp, Low32);
}
+// Op is an atomic load. Lower it into a normal volatile load.
+SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
+ SelectionDAG &DAG) const {
+ auto *Node = cast<AtomicSDNode>(Op.getNode());
+ return DAG.getExtLoad(ISD::EXTLOAD, SDLoc(Op), Op.getValueType(),
+ Node->getChain(), Node->getBasePtr(),
+ Node->getMemoryVT(), Node->getMemOperand());
+}
+
+// Op is an atomic store. Lower it into a normal volatile store followed
+// by a serialization.
+SDValue SystemZTargetLowering::lowerATOMIC_STORE(SDValue Op,
+ SelectionDAG &DAG) const {
+ auto *Node = cast<AtomicSDNode>(Op.getNode());
+ SDValue Chain = DAG.getTruncStore(Node->getChain(), SDLoc(Op), Node->getVal(),
+ Node->getBasePtr(), Node->getMemoryVT(),
+ Node->getMemOperand());
+ return SDValue(DAG.getMachineNode(SystemZ::Serialize, SDLoc(Op), MVT::Other,
+ Chain), 0);
+}
+
// Op is an 8-, 16-bit or 32-bit ATOMIC_LOAD_* operation. Lower the first
// two into the fullword ATOMIC_LOADW_* operation given by Opcode.
-SDValue SystemZTargetLowering::lowerATOMIC_LOAD(SDValue Op,
- SelectionDAG &DAG,
- unsigned Opcode) const {
- AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
+SDValue SystemZTargetLowering::lowerATOMIC_LOAD_OP(SDValue Op,
+ SelectionDAG &DAG,
+ unsigned Opcode) const {
+ auto *Node = cast<AtomicSDNode>(Op.getNode());
// 32-bit operations need no code outside the main loop.
EVT NarrowVT = Node->getMemoryVT();
// Convert atomic subtracts of constants into additions.
if (Opcode == SystemZISD::ATOMIC_LOADW_SUB)
- if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(Src2)) {
+ if (auto *Const = dyn_cast<ConstantSDNode>(Src2)) {
Opcode = SystemZISD::ATOMIC_LOADW_ADD;
Src2 = DAG.getConstant(-Const->getSExtValue(), Src2.getValueType());
}
SDValue Ops[] = { ChainIn, AlignedAddr, Src2, BitShift, NegBitShift,
DAG.getConstant(BitSize, WideVT) };
SDValue AtomicOp = DAG.getMemIntrinsicNode(Opcode, DL, VTList, Ops,
- array_lengthof(Ops),
NarrowVT, MMO);
// Rotate the result of the final CS so that the field is in the lower
SDValue Result = DAG.getNode(ISD::ROTL, DL, WideVT, AtomicOp, ResultShift);
SDValue RetOps[2] = { Result, AtomicOp.getValue(1) };
- return DAG.getMergeValues(RetOps, 2, DL);
+ return DAG.getMergeValues(RetOps, DL);
+}
+
+// Op is an ATOMIC_LOAD_SUB operation. Lower 8- and 16-bit operations
+// into ATOMIC_LOADW_SUBs and decide whether to convert 32- and 64-bit
+// operations into additions.
+SDValue SystemZTargetLowering::lowerATOMIC_LOAD_SUB(SDValue Op,
+ SelectionDAG &DAG) const {
+ auto *Node = cast<AtomicSDNode>(Op.getNode());
+ EVT MemVT = Node->getMemoryVT();
+ if (MemVT == MVT::i32 || MemVT == MVT::i64) {
+ // A full-width operation.
+ assert(Op.getValueType() == MemVT && "Mismatched VTs");
+ SDValue Src2 = Node->getVal();
+ SDValue NegSrc2;
+ SDLoc DL(Src2);
+
+ if (auto *Op2 = dyn_cast<ConstantSDNode>(Src2)) {
+ // Use an addition if the operand is constant and either LAA(G) is
+ // available or the negative value is in the range of A(G)FHI.
+ int64_t Value = (-Op2->getAPIntValue()).getSExtValue();
+ if (isInt<32>(Value) || Subtarget.hasInterlockedAccess1())
+ NegSrc2 = DAG.getConstant(Value, MemVT);
+ } else if (Subtarget.hasInterlockedAccess1())
+ // Use LAA(G) if available.
+ NegSrc2 = DAG.getNode(ISD::SUB, DL, MemVT, DAG.getConstant(0, MemVT),
+ Src2);
+
+ if (NegSrc2.getNode())
+ return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, DL, MemVT,
+ Node->getChain(), Node->getBasePtr(), NegSrc2,
+ Node->getMemOperand(), Node->getOrdering(),
+ Node->getSynchScope());
+
+ // Use the node as-is.
+ return Op;
+ }
+
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
}
// Node is an 8- or 16-bit ATOMIC_CMP_SWAP operation. Lower the first two
// into a fullword ATOMIC_CMP_SWAPW operation.
SDValue SystemZTargetLowering::lowerATOMIC_CMP_SWAP(SDValue Op,
SelectionDAG &DAG) const {
- AtomicSDNode *Node = cast<AtomicSDNode>(Op.getNode());
+ auto *Node = cast<AtomicSDNode>(Op.getNode());
// We have native support for 32-bit compare and swap.
EVT NarrowVT = Node->getMemoryVT();
SDValue Ops[] = { ChainIn, AlignedAddr, CmpVal, SwapVal, BitShift,
NegBitShift, DAG.getConstant(BitSize, WideVT) };
SDValue AtomicOp = DAG.getMemIntrinsicNode(SystemZISD::ATOMIC_CMP_SWAPW, DL,
- VTList, Ops, array_lengthof(Ops),
- NarrowVT, MMO);
+ VTList, Ops, NarrowVT, MMO);
return AtomicOp;
}
bool IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
unsigned Code = IsWrite ? SystemZ::PFD_WRITE : SystemZ::PFD_READ;
- MemIntrinsicSDNode *Node = cast<MemIntrinsicSDNode>(Op.getNode());
+ auto *Node = cast<MemIntrinsicSDNode>(Op.getNode());
SDValue Ops[] = {
Op.getOperand(0),
DAG.getConstant(Code, MVT::i32),
Op.getOperand(1)
};
return DAG.getMemIntrinsicNode(SystemZISD::PREFETCH, SDLoc(Op),
- Node->getVTList(), Ops, array_lengthof(Ops),
+ Node->getVTList(), Ops,
Node->getMemoryVT(), Node->getMemOperand());
}
return lowerBR_CC(Op, DAG);
case ISD::SELECT_CC:
return lowerSELECT_CC(Op, DAG);
+ case ISD::SETCC:
+ return lowerSETCC(Op, DAG);
case ISD::GlobalAddress:
return lowerGlobalAddress(cast<GlobalAddressSDNode>(Op), DAG);
case ISD::GlobalTLSAddress:
case ISD::OR:
return lowerOR(Op, DAG);
case ISD::ATOMIC_SWAP:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_SWAPW);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_SWAPW);
+ case ISD::ATOMIC_STORE:
+ return lowerATOMIC_STORE(Op, DAG);
+ case ISD::ATOMIC_LOAD:
+ return lowerATOMIC_LOAD(Op, DAG);
case ISD::ATOMIC_LOAD_ADD:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_ADD);
case ISD::ATOMIC_LOAD_SUB:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_SUB);
+ return lowerATOMIC_LOAD_SUB(Op, DAG);
case ISD::ATOMIC_LOAD_AND:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_AND);
case ISD::ATOMIC_LOAD_OR:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_OR);
case ISD::ATOMIC_LOAD_XOR:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_XOR);
case ISD::ATOMIC_LOAD_NAND:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_NAND);
case ISD::ATOMIC_LOAD_MIN:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MIN);
case ISD::ATOMIC_LOAD_MAX:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_MAX);
case ISD::ATOMIC_LOAD_UMIN:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMIN);
case ISD::ATOMIC_LOAD_UMAX:
- return lowerATOMIC_LOAD(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
+ return lowerATOMIC_LOAD_OP(Op, DAG, SystemZISD::ATOMIC_LOADW_UMAX);
case ISD::ATOMIC_CMP_SWAP:
return lowerATOMIC_CMP_SWAP(Op, DAG);
case ISD::STACKSAVE:
OPCODE(SIBCALL);
OPCODE(PCREL_WRAPPER);
OPCODE(PCREL_OFFSET);
+ OPCODE(IABS);
OPCODE(ICMP);
OPCODE(FCMP);
OPCODE(TM);
OPCODE(STPCPY);
OPCODE(SEARCH_STRING);
OPCODE(IPM);
+ OPCODE(SERIALIZE);
OPCODE(ATOMIC_SWAPW);
OPCODE(ATOMIC_LOADW_ADD);
OPCODE(ATOMIC_LOADW_SUB);
OPCODE(ATOMIC_CMP_SWAPW);
OPCODE(PREFETCH);
}
- return NULL;
+ return nullptr;
#undef OPCODE
}
+SDValue SystemZTargetLowering::PerformDAGCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ SelectionDAG &DAG = DCI.DAG;
+ unsigned Opcode = N->getOpcode();
+ if (Opcode == ISD::SIGN_EXTEND) {
+ // Convert (sext (ashr (shl X, C1), C2)) to
+ // (ashr (shl (anyext X), C1'), C2')), since wider shifts are as
+ // cheap as narrower ones.
+ SDValue N0 = N->getOperand(0);
+ EVT VT = N->getValueType(0);
+ if (N0.hasOneUse() && N0.getOpcode() == ISD::SRA) {
+ auto *SraAmt = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ SDValue Inner = N0.getOperand(0);
+ if (SraAmt && Inner.hasOneUse() && Inner.getOpcode() == ISD::SHL) {
+ if (auto *ShlAmt = dyn_cast<ConstantSDNode>(Inner.getOperand(1))) {
+ unsigned Extra = (VT.getSizeInBits() -
+ N0.getValueType().getSizeInBits());
+ unsigned NewShlAmt = ShlAmt->getZExtValue() + Extra;
+ unsigned NewSraAmt = SraAmt->getZExtValue() + Extra;
+ EVT ShiftVT = N0.getOperand(1).getValueType();
+ SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, SDLoc(Inner), VT,
+ Inner.getOperand(0));
+ SDValue Shl = DAG.getNode(ISD::SHL, SDLoc(Inner), VT, Ext,
+ DAG.getConstant(NewShlAmt, ShiftVT));
+ return DAG.getNode(ISD::SRA, SDLoc(N0), VT, Shl,
+ DAG.getConstant(NewSraAmt, ShiftVT));
+ }
+ }
+ }
+ }
+ return SDValue();
+}
+
//===----------------------------------------------------------------------===//
// Custom insertion
//===----------------------------------------------------------------------===//
static MachineBasicBlock *emitBlockAfter(MachineBasicBlock *MBB) {
MachineFunction &MF = *MBB->getParent();
MachineBasicBlock *NewMBB = MF.CreateMachineBasicBlock(MBB->getBasicBlock());
- MF.insert(llvm::next(MachineFunction::iterator(MBB)), NewMBB);
+ MF.insert(std::next(MachineFunction::iterator(MBB)), NewMBB);
return NewMBB;
}
MachineBasicBlock *MBB) {
MachineBasicBlock *NewMBB = emitBlockAfter(MBB);
NewMBB->splice(NewMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
NewMBB->transferSuccessorsAndUpdatePHIs(MBB);
return NewMBB;
}
MachineBasicBlock *
SystemZTargetLowering::emitSelect(MachineInstr *MI,
MachineBasicBlock *MBB) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
unsigned DestReg = MI->getOperand(0).getReg();
unsigned TrueReg = MI->getOperand(1).getReg();
MachineBasicBlock *MBB,
unsigned StoreOpcode, unsigned STOCOpcode,
bool Invert) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
unsigned SrcReg = MI->getOperand(0).getReg();
MachineOperand Base = MI->getOperand(1);
// Use STOCOpcode if possible. We could use different store patterns in
// order to avoid matching the index register, but the performance trade-offs
// might be more complicated in that case.
- if (STOCOpcode && !IndexReg && TM.getSubtargetImpl()->hasLoadStoreOnCond()) {
+ if (STOCOpcode && !IndexReg && Subtarget.hasLoadStoreOnCond()) {
if (Invert)
CCMask ^= CCValid;
BuildMI(*MBB, MI, DL, TII->get(STOCOpcode))
unsigned BinOpcode,
unsigned BitSize,
bool Invert) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
bool IsSubWord = (BitSize < 32);
unsigned Tmp = MRI.createVirtualRegister(RC);
BuildMI(MBB, DL, TII->get(BinOpcode), Tmp)
.addReg(RotatedOldVal).addOperand(Src2);
- if (BitSize < 32)
+ if (BitSize <= 32)
// XILF with the upper BitSize bits set.
BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
- .addReg(Tmp).addImm(uint32_t(~0 << (32 - BitSize)));
- else if (BitSize == 32)
- // XILF with every bit set.
- BuildMI(MBB, DL, TII->get(SystemZ::XILF), RotatedNewVal)
- .addReg(Tmp).addImm(~uint32_t(0));
+ .addReg(Tmp).addImm(-1U << (32 - BitSize));
else {
// Use LCGR and add -1 to the result, which is more compact than
// an XILF, XILH pair.
unsigned CompareOpcode,
unsigned KeepOldMask,
unsigned BitSize) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
bool IsSubWord = (BitSize < 32);
MachineBasicBlock *
SystemZTargetLowering::emitAtomicCmpSwapW(MachineInstr *MI,
MachineBasicBlock *MBB) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
// Extract the operands. Base can be a register or a frame index.
SystemZTargetLowering::emitExt128(MachineInstr *MI,
MachineBasicBlock *MBB,
bool ClearEven, unsigned SubReg) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
DebugLoc DL = MI->getDebugLoc();
SystemZTargetLowering::emitMemMemWrapper(MachineInstr *MI,
MachineBasicBlock *MBB,
unsigned Opcode) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
DebugLoc DL = MI->getDebugLoc();
// When generating more than one CLC, all but the last will need to
// branch to the end when a difference is found.
MachineBasicBlock *EndMBB = (Length > 256 && Opcode == SystemZ::CLC ?
- splitBlockAfter(MI, MBB) : 0);
+ splitBlockAfter(MI, MBB) : nullptr);
// Check for the loop form, in which operand 5 is the trip count.
if (MI->getNumExplicitOperands() > 5) {
SystemZTargetLowering::emitStringWrapper(MachineInstr *MI,
MachineBasicBlock *MBB,
unsigned Opcode) const {
- const SystemZInstrInfo *TII = TM.getInstrInfo();
MachineFunction &MF = *MBB->getParent();
+ const SystemZInstrInfo *TII =
+ static_cast<const SystemZInstrInfo *>(Subtarget.getInstrInfo());
MachineRegisterInfo &MRI = MF.getRegInfo();
DebugLoc DL = MI->getDebugLoc();