#include "PPCISelLowering.h"
#include "PPCTargetMachine.h"
#include "llvm/ADT/VectorExtras.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
+#include "llvm/Intrinsics.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;
setOperationAction(ISD::FSQRT, MVT::f32, Expand);
}
+ setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
+
// PowerPC does not have BSWAP, CTPOP or CTTZ
setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
// PowerPC wants to optimize integer setcc a bit
setOperationAction(ISD::SETCC, MVT::i32, Custom);
- // PowerPC does not have BRCOND* which requires SetCC
- setOperationAction(ISD::BRCOND, MVT::Other, Expand);
- setOperationAction(ISD::BRCONDTWOWAY, MVT::Other, Expand);
+ // PowerPC does not have BRCOND which requires SetCC
+ setOperationAction(ISD::BRCOND, MVT::Other, Expand);
// PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
+ // We want to custom lower some of our intrinsics.
+ setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
+
if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
// They also have instructions for converting between i64 and fp.
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
+
+ // FIXME: disable this lowered code. This generates 64-bit register values,
+ // and we don't model the fact that the top part is clobbered by calls. We
+ // need to flag these together so that the value isn't live across a call.
+ //setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
+
// To take advantage of the above i64 FP_TO_SINT, promote i32 FP_TO_UINT
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Promote);
} else {
setOperationAction(ISD::SRL, MVT::i64, Custom);
setOperationAction(ISD::SRA, MVT::i64, Custom);
}
-
- // First set operation action for all vector types to expand. Then we
- // will selectively turn on ones that can be effectively codegen'd.
- for (unsigned VT = (unsigned)MVT::Vector + 1;
- VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
- setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
- setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
- setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
- setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
- }
if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
+ // First set operation action for all vector types to expand. Then we
+ // will selectively turn on ones that can be effectively codegen'd.
+ for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
+ // add/sub/and/or/xor are legal for all supported vector VT's.
+ setOperationAction(ISD::ADD , (MVT::ValueType)VT, Legal);
+ setOperationAction(ISD::SUB , (MVT::ValueType)VT, Legal);
+ setOperationAction(ISD::AND , (MVT::ValueType)VT, Legal);
+ setOperationAction(ISD::OR , (MVT::ValueType)VT, Legal);
+ setOperationAction(ISD::XOR , (MVT::ValueType)VT, Legal);
+
+ // We can custom expand all VECTOR_SHUFFLEs to VPERM.
+ setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::ValueType)VT, Custom);
+
+ setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::SDIV, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::SREM, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::UDIV, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::UREM, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::EXTRACT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, (MVT::ValueType)VT, Expand);
+ setOperationAction(ISD::BUILD_VECTOR, (MVT::ValueType)VT, Expand);
+ }
+
addRegisterClass(MVT::v4f32, PPC::VRRCRegisterClass);
addRegisterClass(MVT::v4i32, PPC::VRRCRegisterClass);
+ addRegisterClass(MVT::v8i16, PPC::VRRCRegisterClass);
+ addRegisterClass(MVT::v16i8, PPC::VRRCRegisterClass);
- setOperationAction(ISD::ADD , MVT::v4f32, Legal);
- setOperationAction(ISD::SUB , MVT::v4f32, Legal);
- setOperationAction(ISD::MUL , MVT::v4f32, Legal);
- setOperationAction(ISD::LOAD , MVT::v4f32, Legal);
- setOperationAction(ISD::ADD , MVT::v4i32, Legal);
- setOperationAction(ISD::LOAD , MVT::v4i32, Legal);
- // FIXME: We don't support any ConstantVec's yet. We should custom expand
- // the ones we do!
- setOperationAction(ISD::ConstantVec, MVT::v4f32, Expand);
- setOperationAction(ISD::ConstantVec, MVT::v4i32, Expand);
+ setOperationAction(ISD::MUL, MVT::v4f32, Legal);
+
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4f32, Custom);
+ setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i32, Custom);
+
+ setOperationAction(ISD::BUILD_VECTOR, MVT::v16i8, Custom);
+ setOperationAction(ISD::BUILD_VECTOR, MVT::v8i16, Custom);
+ setOperationAction(ISD::BUILD_VECTOR, MVT::v4i32, Custom);
+ setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
}
setSetCCResultContents(ZeroOrOneSetCCResult);
setStackPointerRegisterToSaveRestore(PPC::R1);
+ // We have target-specific dag combine patterns for the following nodes:
+ setTargetDAGCombine(ISD::SINT_TO_FP);
+ setTargetDAGCombine(ISD::STORE);
+
computeRegisterProperties();
}
case PPCISD::FCFID: return "PPCISD::FCFID";
case PPCISD::FCTIDZ: return "PPCISD::FCTIDZ";
case PPCISD::FCTIWZ: return "PPCISD::FCTIWZ";
+ case PPCISD::STFIWX: return "PPCISD::STFIWX";
case PPCISD::VMADDFP: return "PPCISD::VMADDFP";
case PPCISD::VNMSUBFP: return "PPCISD::VNMSUBFP";
+ case PPCISD::VPERM: return "PPCISD::VPERM";
case PPCISD::Hi: return "PPCISD::Hi";
case PPCISD::Lo: return "PPCISD::Lo";
case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
case PPCISD::SRL: return "PPCISD::SRL";
case PPCISD::SRA: return "PPCISD::SRA";
case PPCISD::SHL: return "PPCISD::SHL";
+ case PPCISD::EXTSW_32: return "PPCISD::EXTSW_32";
+ case PPCISD::STD_32: return "PPCISD::STD_32";
case PPCISD::CALL: return "PPCISD::CALL";
case PPCISD::RET_FLAG: return "PPCISD::RET_FLAG";
+ case PPCISD::MFCR: return "PPCISD::MFCR";
+ case PPCISD::VCMP: return "PPCISD::VCMP";
+ case PPCISD::VCMPo: return "PPCISD::VCMPo";
}
}
return false;
}
+
+/// isSplatShuffleMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a splat of a single element that is suitable for input to
+/// VSPLTB/VSPLTH/VSPLTW.
+bool PPC::isSplatShuffleMask(SDNode *N) {
+ assert(N->getOpcode() == ISD::BUILD_VECTOR);
+
+ // We can only splat 8-bit, 16-bit, and 32-bit quantities.
+ if (N->getNumOperands() != 4 && N->getNumOperands() != 8 &&
+ N->getNumOperands() != 16)
+ return false;
+
+ // This is a splat operation if each element of the permute is the same, and
+ // if the value doesn't reference the second vector.
+ SDOperand Elt = N->getOperand(0);
+ assert(isa<ConstantSDNode>(Elt) && "Invalid VECTOR_SHUFFLE mask!");
+ for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i) {
+ assert(isa<ConstantSDNode>(N->getOperand(i)) &&
+ "Invalid VECTOR_SHUFFLE mask!");
+ if (N->getOperand(i) != Elt) return false;
+ }
+
+ // Make sure it is a splat of the first vector operand.
+ return cast<ConstantSDNode>(Elt)->getValue() < N->getNumOperands();
+}
+
+/// getVSPLTImmediate - Return the appropriate VSPLT* immediate to splat the
+/// specified isSplatShuffleMask VECTOR_SHUFFLE mask.
+unsigned PPC::getVSPLTImmediate(SDNode *N) {
+ assert(isSplatShuffleMask(N));
+ return cast<ConstantSDNode>(N->getOperand(0))->getValue();
+}
+
+/// isVecSplatImm - Return true if this is a build_vector of constants which
+/// can be formed by using a vspltis[bhw] instruction. The ByteSize field
+/// indicates the number of bytes of each element [124] -> [bhw].
+bool PPC::isVecSplatImm(SDNode *N, unsigned ByteSize, char *Val) {
+ SDOperand OpVal(0, 0);
+ // Check to see if this buildvec has a single non-undef value in its elements.
+ for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+ if (N->getOperand(i).getOpcode() == ISD::UNDEF) continue;
+ if (OpVal.Val == 0)
+ OpVal = N->getOperand(i);
+ else if (OpVal != N->getOperand(i))
+ return false;
+ }
+
+ if (OpVal.Val == 0) return false; // All UNDEF: use implicit def.
+
+ unsigned ValSizeInBytes = 0;
+ uint64_t Value = 0;
+ if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(OpVal)) {
+ Value = CN->getValue();
+ ValSizeInBytes = MVT::getSizeInBits(CN->getValueType(0))/8;
+ } else if (ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(OpVal)) {
+ assert(CN->getValueType(0) == MVT::f32 && "Only one legal FP vector type!");
+ Value = FloatToBits(CN->getValue());
+ ValSizeInBytes = 4;
+ }
+
+ // If the splat value is larger than the element value, then we can never do
+ // this splat. The only case that we could fit the replicated bits into our
+ // immediate field for would be zero, and we prefer to use vxor for it.
+ if (ValSizeInBytes < ByteSize) return false;
+
+ // If the element value is larger than the splat value, cut it in half and
+ // check to see if the two halves are equal. Continue doing this until we
+ // get to ByteSize. This allows us to handle 0x01010101 as 0x01.
+ while (ValSizeInBytes > ByteSize) {
+ ValSizeInBytes >>= 1;
+
+ // If the top half equals the bottom half, we're still ok.
+ if (((Value >> (ValSizeInBytes*8)) & ((8 << ValSizeInBytes)-1)) !=
+ (Value & ((8 << ValSizeInBytes)-1)))
+ return false;
+ }
+
+ // Properly sign extend the value.
+ int ShAmt = (4-ByteSize)*8;
+ int MaskVal = ((int)Value << ShAmt) >> ShAmt;
+
+ // If this is zero, don't match, zero matches ISD::isBuildVectorAllZeros.
+ if (MaskVal == 0) return false;
+
+ if (Val) *Val = MaskVal;
+
+ // Finally, if this value fits in a 5 bit sext field, return true.
+ return ((MaskVal << (32-5)) >> (32-5)) == MaskVal;
+}
+
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDOperand PPCTargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
Bits = DAG.getNode(ISD::TRUNCATE, MVT::i32, Bits);
return Bits;
}
- case ISD::SINT_TO_FP: {
- assert(MVT::i64 == Op.getOperand(0).getValueType() &&
- "Unhandled SINT_TO_FP type in custom expander!");
- SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
- SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
- if (MVT::f32 == Op.getValueType())
- FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
- return FP;
- }
+ case ISD::SINT_TO_FP:
+ if (Op.getOperand(0).getValueType() == MVT::i64) {
+ SDOperand Bits = DAG.getNode(ISD::BIT_CONVERT, MVT::f64, Op.getOperand(0));
+ SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Bits);
+ if (Op.getValueType() == MVT::f32)
+ FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
+ return FP;
+ } else {
+ assert(Op.getOperand(0).getValueType() == MVT::i32 &&
+ "Unhandled SINT_TO_FP type in custom expander!");
+ // Since we only generate this in 64-bit mode, we can take advantage of
+ // 64-bit registers. In particular, sign extend the input value into the
+ // 64-bit register with extsw, store the WHOLE 64-bit value into the stack
+ // then lfd it and fcfid it.
+ MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
+ int FrameIdx = FrameInfo->CreateStackObject(8, 8);
+ SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
+
+ SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
+ Op.getOperand(0));
+
+ // STD the extended value into the stack slot.
+ SDOperand Store = DAG.getNode(PPCISD::STD_32, MVT::Other,
+ DAG.getEntryNode(), Ext64, FIdx,
+ DAG.getSrcValue(NULL));
+ // Load the value as a double.
+ SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
+
+ // FCFID it and return it.
+ SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
+ if (Op.getValueType() == MVT::f32)
+ FP = DAG.getNode(ISD::FP_ROUND, MVT::f32, FP);
+ return FP;
+ }
+ break;
+
case ISD::SELECT_CC: {
// Turn FP only select_cc's into fsel instructions.
if (!MVT::isFloatingPoint(Op.getOperand(0).getValueType()) ||
}
return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
}
+ case ISD::SCALAR_TO_VECTOR: {
+ // Create a stack slot that is 16-byte aligned.
+ MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
+ int FrameIdx = FrameInfo->CreateStackObject(16, 16);
+ SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
+
+ // Store the input value into Value#0 of the stack slot.
+ SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
+ Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
+ // Load it out.
+ return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
+ }
+ case ISD::BUILD_VECTOR:
+ // If this is a case we can't handle, return null and let the default
+ // expansion code take care of it. If we CAN select this case, return Op.
+
+ // See if this is all zeros.
+ // FIXME: We should handle splat(-0.0), and other cases here.
+ if (ISD::isBuildVectorAllZeros(Op.Val))
+ return Op;
+
+ if (PPC::isVecSplatImm(Op.Val, 1) || // vspltisb
+ PPC::isVecSplatImm(Op.Val, 2) || // vspltish
+ PPC::isVecSplatImm(Op.Val, 4)) // vspltisw
+ return Op;
+
+ return SDOperand();
+
+ case ISD::VECTOR_SHUFFLE: {
+ SDOperand V1 = Op.getOperand(0);
+ SDOperand V2 = Op.getOperand(1);
+ SDOperand PermMask = Op.getOperand(2);
+
+ // Cases that are handled by instructions that take permute immediates
+ // (such as vsplt*) should be left as VECTOR_SHUFFLE nodes so they can be
+ // selected by the instruction selector.
+ if (PPC::isSplatShuffleMask(PermMask.Val) && V2.getOpcode() == ISD::UNDEF)
+ break;
+
+ // TODO: Handle more cases, and also handle cases that are cheaper to do as
+ // multiple such instructions than as a constant pool load/vperm pair.
+
+ // Lower this to a VPERM(V1, V2, V3) expression, where V3 is a constant
+ // vector that will get spilled to the constant pool.
+ if (V2.getOpcode() == ISD::UNDEF) V2 = V1;
+
+ // The SHUFFLE_VECTOR mask is almost exactly what we want for vperm, except
+ // that it is in input element units, not in bytes. Convert now.
+ MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
+ unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
+
+ std::vector<SDOperand> ResultMask;
+ for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
+ unsigned SrcElt =cast<ConstantSDNode>(PermMask.getOperand(i))->getValue();
+
+ for (unsigned j = 0; j != BytesPerElement; ++j)
+ ResultMask.push_back(DAG.getConstant(SrcElt*BytesPerElement+j,
+ MVT::i8));
+ }
+
+ SDOperand VPermMask =DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
+ return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
+ }
+ case ISD::INTRINSIC_WO_CHAIN: {
+ unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getValue();
+
+ // If this is a lowered altivec predicate compare, CompareOpc is set to the
+ // opcode number of the comparison.
+ int CompareOpc = -1;
+ bool isDot = false;
+ switch (IntNo) {
+ default: return SDOperand(); // Don't custom lower most intrinsics.
+ // Comparison predicates.
+ case Intrinsic::ppc_altivec_vcmpbfp_p: CompareOpc = 966; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpeqfp_p: CompareOpc = 198; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpequb_p: CompareOpc = 6; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpequh_p: CompareOpc = 70; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpequw_p: CompareOpc = 134; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgefp_p: CompareOpc = 454; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtfp_p: CompareOpc = 710; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtsb_p: CompareOpc = 774; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtsh_p: CompareOpc = 838; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtsw_p: CompareOpc = 902; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtub_p: CompareOpc = 518; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtuh_p: CompareOpc = 582; isDot = 1; break;
+ case Intrinsic::ppc_altivec_vcmpgtuw_p: CompareOpc = 646; isDot = 1; break;
+
+ // Normal Comparisons.
+ case Intrinsic::ppc_altivec_vcmpbfp: CompareOpc = 966; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpeqfp: CompareOpc = 198; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpequb: CompareOpc = 6; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpequh: CompareOpc = 70; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpequw: CompareOpc = 134; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgefp: CompareOpc = 454; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtfp: CompareOpc = 710; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtsb: CompareOpc = 774; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtsh: CompareOpc = 838; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtsw: CompareOpc = 902; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtub: CompareOpc = 518; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtuh: CompareOpc = 582; isDot = 0; break;
+ case Intrinsic::ppc_altivec_vcmpgtuw: CompareOpc = 646; isDot = 0; break;
+ }
+
+ assert(CompareOpc>0 && "We only lower altivec predicate compares so far!");
+
+ // If this is a non-dot comparison, make the VCMP node.
+ if (!isDot)
+ return DAG.getNode(PPCISD::VCMP, Op.getOperand(2).getValueType(),
+ Op.getOperand(1), Op.getOperand(2),
+ DAG.getConstant(CompareOpc, MVT::i32));
+
+ // Create the PPCISD altivec 'dot' comparison node.
+ std::vector<SDOperand> Ops;
+ std::vector<MVT::ValueType> VTs;
+ Ops.push_back(Op.getOperand(2)); // LHS
+ Ops.push_back(Op.getOperand(3)); // RHS
+ Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
+ VTs.push_back(Op.getOperand(2).getValueType());
+ VTs.push_back(MVT::Flag);
+ SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
+
+ // Now that we have the comparison, emit a copy from the CR to a GPR.
+ // This is flagged to the above dot comparison.
+ SDOperand Flags = DAG.getNode(PPCISD::MFCR, MVT::i32,
+ DAG.getRegister(PPC::CR6, MVT::i32),
+ CompNode.getValue(1));
+
+ // Unpack the result based on how the target uses it.
+ unsigned BitNo; // Bit # of CR6.
+ bool InvertBit; // Invert result?
+ switch (cast<ConstantSDNode>(Op.getOperand(1))->getValue()) {
+ default: // Can't happen, don't crash on invalid number though.
+ case 0: // Return the value of the EQ bit of CR6.
+ BitNo = 0; InvertBit = false;
+ break;
+ case 1: // Return the inverted value of the EQ bit of CR6.
+ BitNo = 0; InvertBit = true;
+ break;
+ case 2: // Return the value of the LT bit of CR6.
+ BitNo = 2; InvertBit = false;
+ break;
+ case 3: // Return the inverted value of the LT bit of CR6.
+ BitNo = 2; InvertBit = true;
+ break;
+ }
+
+ // Shift the bit into the low position.
+ Flags = DAG.getNode(ISD::SRL, MVT::i32, Flags,
+ DAG.getConstant(8-(3-BitNo), MVT::i32));
+ // Isolate the bit.
+ Flags = DAG.getNode(ISD::AND, MVT::i32, Flags,
+ DAG.getConstant(1, MVT::i32));
+
+ // If we are supposed to, toggle the bit.
+ if (InvertBit)
+ Flags = DAG.getNode(ISD::XOR, MVT::i32, Flags,
+ DAG.getConstant(1, MVT::i32));
+ return Flags;
+ }
}
return SDOperand();
}
MachineFunction *F = BB->getParent();
F->getBasicBlockList().insert(It, copy0MBB);
F->getBasicBlockList().insert(It, sinkMBB);
- // Update machine-CFG edges
+ // Update machine-CFG edges by first adding all successors of the current
+ // block to the new block which will contain the Phi node for the select.
+ for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
+ e = BB->succ_end(); i != e; ++i)
+ sinkMBB->addSuccessor(*i);
+ // Next, remove all successors of the current block, and add the true
+ // and fallthrough blocks as its successors.
+ while(!BB->succ_empty())
+ BB->removeSuccessor(BB->succ_begin());
BB->addSuccessor(copy0MBB);
BB->addSuccessor(sinkMBB);
return BB;
}
+SDOperand PPCTargetLowering::PerformDAGCombine(SDNode *N,
+ DAGCombinerInfo &DCI) const {
+ TargetMachine &TM = getTargetMachine();
+ SelectionDAG &DAG = DCI.DAG;
+ switch (N->getOpcode()) {
+ default: break;
+ case ISD::SINT_TO_FP:
+ if (TM.getSubtarget<PPCSubtarget>().is64Bit()) {
+ if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
+ // Turn (sint_to_fp (fp_to_sint X)) -> fctidz/fcfid without load/stores.
+ // We allow the src/dst to be either f32/f64, but the intermediate
+ // type must be i64.
+ if (N->getOperand(0).getValueType() == MVT::i64) {
+ SDOperand Val = N->getOperand(0).getOperand(0);
+ if (Val.getValueType() == MVT::f32) {
+ Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+ }
+
+ Val = DAG.getNode(PPCISD::FCTIDZ, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+ Val = DAG.getNode(PPCISD::FCFID, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+ if (N->getValueType(0) == MVT::f32) {
+ Val = DAG.getNode(ISD::FP_ROUND, MVT::f32, Val);
+ DCI.AddToWorklist(Val.Val);
+ }
+ return Val;
+ } else if (N->getOperand(0).getValueType() == MVT::i32) {
+ // If the intermediate type is i32, we can avoid the load/store here
+ // too.
+ }
+ }
+ }
+ break;
+ case ISD::STORE:
+ // Turn STORE (FP_TO_SINT F) -> STFIWX(FCTIWZ(F)).
+ if (TM.getSubtarget<PPCSubtarget>().hasSTFIWX() &&
+ N->getOperand(1).getOpcode() == ISD::FP_TO_SINT &&
+ N->getOperand(1).getValueType() == MVT::i32) {
+ SDOperand Val = N->getOperand(1).getOperand(0);
+ if (Val.getValueType() == MVT::f32) {
+ Val = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+ }
+ Val = DAG.getNode(PPCISD::FCTIWZ, MVT::f64, Val);
+ DCI.AddToWorklist(Val.Val);
+
+ Val = DAG.getNode(PPCISD::STFIWX, MVT::Other, N->getOperand(0), Val,
+ N->getOperand(2), N->getOperand(3));
+ DCI.AddToWorklist(Val.Val);
+ return Val;
+ }
+ break;
+ case PPCISD::VCMP: {
+ // If a VCMPo node already exists with exactly the same operands as this
+ // node, use its result instead of this node (VCMPo computes both a CR6 and
+ // a normal output).
+ //
+ if (!N->getOperand(0).hasOneUse() &&
+ !N->getOperand(1).hasOneUse() &&
+ !N->getOperand(2).hasOneUse()) {
+
+ // Scan all of the users of the LHS, looking for VCMPo's that match.
+ SDNode *VCMPoNode = 0;
+
+ SDNode *LHSN = N->getOperand(0).Val;
+ for (SDNode::use_iterator UI = LHSN->use_begin(), E = LHSN->use_end();
+ UI != E; ++UI)
+ if ((*UI)->getOpcode() == PPCISD::VCMPo &&
+ (*UI)->getOperand(1) == N->getOperand(1) &&
+ (*UI)->getOperand(2) == N->getOperand(2) &&
+ (*UI)->getOperand(0) == N->getOperand(0)) {
+ VCMPoNode = *UI;
+ break;
+ }
+
+ // If there are non-zero uses of the flag value, use the VCMPo node!
+ if (VCMPoNode && !VCMPoNode->hasNUsesOfValue(0, 1))
+ return SDOperand(VCMPoNode, 0);
+ }
+ break;
+ }
+ }
+
+ return SDOperand();
+}
+
+void PPCTargetLowering::computeMaskedBitsForTargetNode(const SDOperand Op,
+ uint64_t Mask,
+ uint64_t &KnownZero,
+ uint64_t &KnownOne,
+ unsigned Depth) const {
+ KnownZero = 0;
+ KnownOne = 0;
+ switch (Op.getOpcode()) {
+ default: break;
+ case ISD::INTRINSIC_WO_CHAIN: {
+ switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
+ default: break;
+ case Intrinsic::ppc_altivec_vcmpbfp_p:
+ case Intrinsic::ppc_altivec_vcmpeqfp_p:
+ case Intrinsic::ppc_altivec_vcmpequb_p:
+ case Intrinsic::ppc_altivec_vcmpequh_p:
+ case Intrinsic::ppc_altivec_vcmpequw_p:
+ case Intrinsic::ppc_altivec_vcmpgefp_p:
+ case Intrinsic::ppc_altivec_vcmpgtfp_p:
+ case Intrinsic::ppc_altivec_vcmpgtsb_p:
+ case Intrinsic::ppc_altivec_vcmpgtsh_p:
+ case Intrinsic::ppc_altivec_vcmpgtsw_p:
+ case Intrinsic::ppc_altivec_vcmpgtub_p:
+ case Intrinsic::ppc_altivec_vcmpgtuh_p:
+ case Intrinsic::ppc_altivec_vcmpgtuw_p:
+ KnownZero = ~1U; // All bits but the low one are known to be zero.
+ break;
+ }
+ }
+ }
+}
+
+
/// getConstraintType - Given a constraint letter, return the type of
/// constraint it is for this target.
PPCTargetLowering::ConstraintType
// Handle standard constraint letters.
return TargetLowering::isOperandValidForConstraint(Op, Letter);
}
+
+/// isLegalAddressImmediate - Return true if the integer value can be used
+/// as the offset of the target addressing mode.
+bool PPCTargetLowering::isLegalAddressImmediate(int64_t V) const {
+ // PPC allows a sign-extended 16-bit immediate field.
+ return (V > -(1 << 16) && V < (1 << 16)-1);
+}