#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
+#include "llvm/IR/Function.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
+#include <stdint.h>
using namespace llvm;
#define DEBUG_TYPE "x86-isel"
inline void getAddressOperands(X86ISelAddressMode &AM, SDValue &Base,
SDValue &Scale, SDValue &Index,
SDValue &Disp, SDValue &Segment) {
- Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase) ?
- CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
- getTargetLowering()->getPointerTy()) :
- AM.Base_Reg;
+ Base = (AM.BaseType == X86ISelAddressMode::FrameIndexBase)
+ ? CurDAG->getTargetFrameIndex(AM.Base_FrameIndex,
+ TLI->getPointerTy())
+ : AM.Base_Reg;
Scale = getI8Imm(AM.Scale);
Index = AM.IndexReg;
// These are 32-bit even in 64-bit mode since RIP relative offset
const X86InstrInfo *getInstrInfo() const {
return getTargetMachine().getSubtargetImpl()->getInstrInfo();
}
+
+ /// \brief Address-mode matching performs shift-of-and to and-of-shift
+ /// reassociation in order to expose more scaled addressing
+ /// opportunities.
+ bool ComplexPatternFuncMutatesDAG() const override {
+ return true;
+ }
};
}
// If the source and destination are SSE registers, then this is a legal
// conversion that should not be lowered.
const X86TargetLowering *X86Lowering =
- static_cast<const X86TargetLowering *>(getTargetLowering());
+ static_cast<const X86TargetLowering *>(TLI);
bool SrcIsSSE = X86Lowering->isScalarFPTypeInSSEReg(SrcVT);
bool DstIsSSE = X86Lowering->isScalarFPTypeInSSEReg(DstVT);
if (SrcIsSSE && DstIsSSE)
}
}
-// Transform "(X >> (8-C1)) & C2" to "(X >> 8) & 0xff)" if safe. This
-// allows us to convert the shift and and into an h-register extract and
-// a scaled index. Returns false if the simplification is performed.
+// Transform "(X >> (8-C1)) & (0xff << C1)" to "((X >> 8) & 0xff) << C1" if
+// safe. This allows us to convert the shift and and into an h-register
+// extract and a scaled index. Returns false if the simplification is
+// performed.
static bool FoldMaskAndShiftToExtract(SelectionDAG &DAG, SDValue N,
uint64_t Mask,
SDValue Shift, SDValue X,
RegisterSDNode *RN = dyn_cast<RegisterSDNode>(Base);
if (RN && RN->getReg() == 0)
Base = CurDAG->getRegister(0, MVT::i64);
- else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(N)) {
+ else if (Base.getValueType() == MVT::i32 && !dyn_cast<FrameIndexSDNode>(Base)) {
// Base could already be %rip, particularly in the x32 ABI.
Base = SDValue(CurDAG->getMachineNode(
TargetOpcode::SUBREG_TO_REG, DL, MVT::i64,
///
SDNode *X86DAGToDAGISel::getGlobalBaseReg() {
unsigned GlobalBaseReg = getInstrInfo()->getGlobalBaseReg(MF);
- return CurDAG->getRegister(GlobalBaseReg,
- getTargetLowering()->getPointerTy()).getNode();
+ return CurDAG->getRegister(GlobalBaseReg, TLI->getPointerTy()).getNode();
}
/// Atomic opcode table
static SDValue getAtomicLoadArithTargetConstant(SelectionDAG *CurDAG,
SDLoc dl,
enum AtomicOpc &Op, MVT NVT,
- SDValue Val) {
+ SDValue Val,
+ const X86Subtarget *Subtarget) {
if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Val)) {
int64_t CNVal = CN->getSExtValue();
// Quit if not 32-bit imm.
if ((int32_t)CNVal != CNVal)
return Val;
+ // Quit if INT32_MIN: it would be negated as it is negative and overflow,
+ // producing an immediate that does not fit in the 32 bits available for
+ // an immediate operand to sub. However, it still fits in 32 bits for the
+ // add (since it is not negated) so we can return target-constant.
+ if (CNVal == INT32_MIN)
+ return CurDAG->getTargetConstant(CNVal, NVT);
// For atomic-load-add, we could do some optimizations.
if (Op == ADD) {
// Translate to INC/DEC if ADD by 1 or -1.
- if ((CNVal == 1) || (CNVal == -1)) {
+ if (((CNVal == 1) || (CNVal == -1)) && !Subtarget->slowIncDec()) {
Op = (CNVal == 1) ? INC : DEC;
// No more constant operand after being translated into INC/DEC.
return SDValue();
SDValue Chain = Node->getOperand(0);
SDValue Ptr = Node->getOperand(1);
SDValue Val = Node->getOperand(2);
- SDValue Tmp0, Tmp1, Tmp2, Tmp3, Tmp4;
- if (!SelectAddr(Node, Ptr, Tmp0, Tmp1, Tmp2, Tmp3, Tmp4))
+ SDValue Base, Scale, Index, Disp, Segment;
+ if (!SelectAddr(Node, Ptr, Base, Scale, Index, Disp, Segment))
return nullptr;
// Which index into the table.
break;
}
- Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val);
+ Val = getAtomicLoadArithTargetConstant(CurDAG, dl, Op, NVT, Val, Subtarget);
bool isUnOp = !Val.getNode();
bool isCN = Val.getNode() && (Val.getOpcode() == ISD::TargetConstant);
Opc = AtomicOpcTbl[Op][I32];
break;
case MVT::i64:
- Opc = AtomicOpcTbl[Op][I64];
if (isCN) {
if (immSext8(Val.getNode()))
Opc = AtomicOpcTbl[Op][SextConstantI64];
else if (i64immSExt32(Val.getNode()))
Opc = AtomicOpcTbl[Op][ConstantI64];
- }
+ else
+ llvm_unreachable("True 64 bits constant in SelectAtomicLoadArith");
+ } else
+ Opc = AtomicOpcTbl[Op][I64];
break;
}
assert(Opc != 0 && "Invalid arith lock transform!");
+ // Building the new node.
SDValue Ret;
- SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
- dl, NVT), 0);
- MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
- MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
if (isUnOp) {
- SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Chain };
+ SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Chain };
Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
} else {
- SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, Val, Chain };
+ SDValue Ops[] = { Base, Scale, Index, Disp, Segment, Val, Chain };
Ret = SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops), 0);
}
+
+ // Copying the MachineMemOperand.
+ MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+ MemOp[0] = cast<MemSDNode>(Node)->getMemOperand();
cast<MachineSDNode>(Ret)->setMemRefs(MemOp, MemOp + 1);
+
+ // We need to have two outputs as that is what the original instruction had.
+ // So we add a dummy, undefined output. This is safe as we checked first
+ // that no-one uses our output anyway.
+ SDValue Undef = SDValue(CurDAG->getMachineNode(TargetOpcode::IMPLICIT_DEF,
+ dl, NVT), 0);
SDValue RetVals[] = { Undef, Ret };
return CurDAG->getMergeValues(RetVals, dl).getNode();
}
case X86ISD::GlobalBaseReg:
return getGlobalBaseReg();
+ case X86ISD::SHRUNKBLEND: {
+ // SHRUNKBLEND selects like a regular VSELECT.
+ SDValue VSelect = CurDAG->getNode(
+ ISD::VSELECT, SDLoc(Node), Node->getValueType(0), Node->getOperand(0),
+ Node->getOperand(1), Node->getOperand(2));
+ ReplaceUses(SDValue(Node, 0), VSelect);
+ SelectCode(VSelect.getNode());
+ // We already called ReplaceUses.
+ return nullptr;
+ }
case ISD::ATOMIC_LOAD_XOR:
case ISD::ATOMIC_LOAD_AND:
return CurDAG->SelectNodeTo(Node, ShlOp, NVT, SDValue(New, 0),
getI8Imm(ShlVal));
}
+ case X86ISD::UMUL8:
+ case X86ISD::SMUL8: {
+ SDValue N0 = Node->getOperand(0);
+ SDValue N1 = Node->getOperand(1);
+
+ Opc = (Opcode == X86ISD::SMUL8 ? X86::IMUL8r : X86::MUL8r);
+
+ SDValue InFlag = CurDAG->getCopyToReg(CurDAG->getEntryNode(), dl, X86::AL,
+ N0, SDValue()).getValue(1);
+
+ SDVTList VTs = CurDAG->getVTList(NVT, MVT::i32);
+ SDValue Ops[] = {N1, InFlag};
+ SDNode *CNode = CurDAG->getMachineNode(Opc, dl, VTs, Ops);
+
+ ReplaceUses(SDValue(Node, 0), SDValue(CNode, 0));
+ ReplaceUses(SDValue(Node, 1), SDValue(CNode, 1));
+ return nullptr;
+ }
+
case X86ISD::UMUL: {
SDValue N0 = Node->getOperand(0);
SDValue N1 = Node->getOperand(1);
}
case ISD::SDIVREM:
- case ISD::UDIVREM: {
+ case ISD::UDIVREM:
+ case X86ISD::SDIVREM8_SEXT_HREG:
+ case X86ISD::UDIVREM8_ZEXT_HREG: {
SDValue N0 = Node->getOperand(0);
SDValue N1 = Node->getOperand(1);
- bool isSigned = Opcode == ISD::SDIVREM;
+ bool isSigned = (Opcode == ISD::SDIVREM ||
+ Opcode == X86ISD::SDIVREM8_SEXT_HREG);
if (!isSigned) {
switch (NVT.SimpleTy) {
default: llvm_unreachable("Unsupported VT!");
SDValue(CurDAG->getMachineNode(SExtOpcode, dl, MVT::Glue, InFlag),0);
} else {
// Zero out the high part, effectively zero extending the input.
- SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
+ SDValue ClrNode = SDValue(CurDAG->getMachineNode(X86::MOV32r0, dl, NVT), 0);
switch (NVT.SimpleTy) {
case MVT::i16:
ClrNode =
SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
}
- // Prevent use of AH in a REX instruction by referencing AX instead.
- // Shift it down 8 bits.
+ // Prevent use of AH in a REX instruction by explicitly copying it to
+ // an ABCD_L register.
//
// The current assumption of the register allocator is that isel
- // won't generate explicit references to the GPR8_NOREX registers. If
+ // won't generate explicit references to the GR8_ABCD_H registers. If
// the allocator and/or the backend get enhanced to be more robust in
// that regard, this can be, and should be, removed.
- if (HiReg == X86::AH && Subtarget->is64Bit() &&
- !SDValue(Node, 1).use_empty()) {
- SDValue Result = CurDAG->getCopyFromReg(CurDAG->getEntryNode(), dl,
- X86::AX, MVT::i16, InFlag);
- InFlag = Result.getValue(2);
-
- // If we also need AL (the quotient), get it by extracting a subreg from
- // Result. The fast register allocator does not like multiple CopyFromReg
- // nodes using aliasing registers.
- if (!SDValue(Node, 0).use_empty())
- ReplaceUses(SDValue(Node, 0),
- CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
-
- // Shift AX right by 8 bits instead of using AH.
- Result = SDValue(CurDAG->getMachineNode(X86::SHR16ri, dl, MVT::i16,
- Result,
- CurDAG->getTargetConstant(8, MVT::i8)),
- 0);
- ReplaceUses(SDValue(Node, 1),
- CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result));
+ if (HiReg == X86::AH && !SDValue(Node, 1).use_empty()) {
+ SDValue AHCopy = CurDAG->getRegister(X86::AH, MVT::i8);
+ unsigned AHExtOpcode =
+ isSigned ? X86::MOVSX32_NOREXrr8 : X86::MOVZX32_NOREXrr8;
+
+ SDNode *RNode = CurDAG->getMachineNode(AHExtOpcode, dl, MVT::i32,
+ MVT::Glue, AHCopy, InFlag);
+ SDValue Result(RNode, 0);
+ InFlag = SDValue(RNode, 1);
+
+ if (Opcode == X86ISD::UDIVREM8_ZEXT_HREG ||
+ Opcode == X86ISD::SDIVREM8_SEXT_HREG) {
+ if (Node->getValueType(1) == MVT::i64) {
+ // It's not possible to directly movsx AH to a 64bit register, because
+ // the latter needs the REX prefix, but the former can't have it.
+ assert(Opcode != X86ISD::SDIVREM8_SEXT_HREG &&
+ "Unexpected i64 sext of h-register");
+ Result =
+ SDValue(CurDAG->getMachineNode(
+ TargetOpcode::SUBREG_TO_REG, dl, MVT::i64,
+ CurDAG->getTargetConstant(0, MVT::i64), Result,
+ CurDAG->getTargetConstant(X86::sub_32bit, MVT::i32)),
+ 0);
+ }
+ } else {
+ Result =
+ CurDAG->getTargetExtractSubreg(X86::sub_8bit, dl, MVT::i8, Result);
+ }
+ ReplaceUses(SDValue(Node, 1), Result);
+ DEBUG(dbgs() << "=> "; Result.getNode()->dump(CurDAG); dbgs() << '\n');
}
// Copy the division (low) result, if it is needed.
if (!SDValue(Node, 0).use_empty()) {
projects / oota-llvm.git / blobdiff