-//===- SPUInstrInfo.cpp - Cell SPU Instruction Information ----------------===//
+//===-- SPUInstrInfo.cpp - Cell SPU Instruction Information ---------------===//
//
// The LLVM Compiler Infrastructure
//
//
//===----------------------------------------------------------------------===//
-#include "SPURegisterNames.h"
#include "SPUInstrInfo.h"
#include "SPUInstrBuilder.h"
#include "SPUTargetMachine.h"
-#include "SPUGenInstrInfo.inc"
+#include "SPUHazardRecognizers.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
-#include "llvm/Support/Streams.h"
+#include "llvm/MC/MCContext.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/TargetRegistry.h"
+#include "llvm/Support/raw_ostream.h"
+
+#define GET_INSTRINFO_CTOR
+#include "SPUGenInstrInfo.inc"
using namespace llvm;
unsigned opc = I->getOpcode();
return (opc == SPU::BR
- || opc == SPU::BRA
- || opc == SPU::BI);
+ || opc == SPU::BRA
+ || opc == SPU::BI);
}
//! Predicate for a conditional branch instruction
return (opc == SPU::BRNZr32
|| opc == SPU::BRNZv4i32
- || opc == SPU::BRZr32
- || opc == SPU::BRZv4i32
- || opc == SPU::BRHNZr16
- || opc == SPU::BRHNZv8i16
- || opc == SPU::BRHZr16
- || opc == SPU::BRHZv8i16);
+ || opc == SPU::BRZr32
+ || opc == SPU::BRZv4i32
+ || opc == SPU::BRHNZr16
+ || opc == SPU::BRHNZv8i16
+ || opc == SPU::BRHZr16
+ || opc == SPU::BRHZv8i16);
}
}
SPUInstrInfo::SPUInstrInfo(SPUTargetMachine &tm)
- : TargetInstrInfoImpl(SPUInsts, sizeof(SPUInsts)/sizeof(SPUInsts[0])),
+ : SPUGenInstrInfo(SPU::ADJCALLSTACKDOWN, SPU::ADJCALLSTACKUP),
TM(tm),
RI(*TM.getSubtargetImpl(), *this)
{ /* NOP */ }
-/// getPointerRegClass - Return the register class to use to hold pointers.
-/// This is used for addressing modes.
-const TargetRegisterClass *
-SPUInstrInfo::getPointerRegClass() const
-{
- return &SPU::R32CRegClass;
-}
-
-bool
-SPUInstrInfo::isMoveInstr(const MachineInstr& MI,
- unsigned& sourceReg,
- unsigned& destReg,
- unsigned& SrcSR, unsigned& DstSR) const {
- SrcSR = DstSR = 0; // No sub-registers.
-
- // Primarily, ORI and OR are generated by copyRegToReg. But, there are other
- // cases where we can safely say that what's being done is really a move
- // (see how PowerPC does this -- it's the model for this code too.)
- switch (MI.getOpcode()) {
- default:
- break;
- case SPU::ORIv4i32:
- case SPU::ORIr32:
- case SPU::ORHIv8i16:
- case SPU::ORHIr16:
- case SPU::ORHIi8i16:
- case SPU::ORBIv16i8:
- case SPU::ORBIr8:
- case SPU::ORIi16i32:
- case SPU::ORIi8i32:
- case SPU::AHIvec:
- case SPU::AHIr16:
- case SPU::AIv4i32:
- assert(MI.getNumOperands() == 3 &&
- MI.getOperand(0).isReg() &&
- MI.getOperand(1).isReg() &&
- MI.getOperand(2).isImm() &&
- "invalid SPU ORI/ORHI/ORBI/AHI/AI/SFI/SFHI instruction!");
- if (MI.getOperand(2).getImm() == 0) {
- sourceReg = MI.getOperand(1).getReg();
- destReg = MI.getOperand(0).getReg();
- return true;
- }
- break;
- case SPU::AIr32:
- assert(MI.getNumOperands() == 3 &&
- "wrong number of operands to AIr32");
- if (MI.getOperand(0).isReg() &&
- MI.getOperand(1).isReg() &&
- (MI.getOperand(2).isImm() &&
- MI.getOperand(2).getImm() == 0)) {
- sourceReg = MI.getOperand(1).getReg();
- destReg = MI.getOperand(0).getReg();
- return true;
- }
- break;
- case SPU::LRr8:
- case SPU::LRr16:
- case SPU::LRr32:
- case SPU::LRf32:
- case SPU::LRr64:
- case SPU::LRf64:
- case SPU::LRr128:
- case SPU::LRv16i8:
- case SPU::LRv8i16:
- case SPU::LRv4i32:
- case SPU::LRv4f32:
- case SPU::LRv2i64:
- case SPU::LRv2f64:
- case SPU::ORv16i8_i8:
- case SPU::ORv8i16_i16:
- case SPU::ORv4i32_i32:
- case SPU::ORv2i64_i64:
- case SPU::ORv4f32_f32:
- case SPU::ORv2f64_f64:
- case SPU::ORi8_v16i8:
- case SPU::ORi16_v8i16:
- case SPU::ORi32_v4i32:
- case SPU::ORi64_v2i64:
- case SPU::ORf32_v4f32:
- case SPU::ORf64_v2f64:
- case SPU::ORi128_r64:
- case SPU::ORi128_f64:
- case SPU::ORi128_r32:
- case SPU::ORi128_f32:
- case SPU::ORi128_r16:
- case SPU::ORi128_r8:
- case SPU::ORi128_vec:
- case SPU::ORr64_i128:
- case SPU::ORf64_i128:
- case SPU::ORr32_i128:
- case SPU::ORf32_i128:
- case SPU::ORr16_i128:
- case SPU::ORr8_i128:
- case SPU::ORvec_i128:
- case SPU::ORr16_r32:
- case SPU::ORr8_r32:
- case SPU::ORr32_r16:
- case SPU::ORr32_r8:
- case SPU::ORr32_r64:
- case SPU::ORr16_r64:
- case SPU::ORr8_r64:
- case SPU::ORr64_r32:
- case SPU::ORr64_r16:
- case SPU::ORr64_r8:
- {
- assert(MI.getNumOperands() == 2 &&
- MI.getOperand(0).isReg() &&
- MI.getOperand(1).isReg() &&
- "invalid SPU OR<type>_<vec> or LR instruction!");
- if (MI.getOperand(0).getReg() == MI.getOperand(1).getReg()) {
- sourceReg = MI.getOperand(0).getReg();
- destReg = MI.getOperand(0).getReg();
- return true;
- }
- break;
- }
- case SPU::ORv16i8:
- case SPU::ORv8i16:
- case SPU::ORv4i32:
- case SPU::ORv2i64:
- case SPU::ORr8:
- case SPU::ORr16:
- case SPU::ORr32:
- case SPU::ORr64:
- case SPU::ORf32:
- case SPU::ORf64:
- assert(MI.getNumOperands() == 3 &&
- MI.getOperand(0).isReg() &&
- MI.getOperand(1).isReg() &&
- MI.getOperand(2).isReg() &&
- "invalid SPU OR(vec|r32|r64|gprc) instruction!");
- if (MI.getOperand(1).getReg() == MI.getOperand(2).getReg()) {
- sourceReg = MI.getOperand(1).getReg();
- destReg = MI.getOperand(0).getReg();
- return true;
- }
- break;
- }
-
- return false;
+/// CreateTargetHazardRecognizer - Return the hazard recognizer to use for
+/// this target when scheduling the DAG.
+ScheduleHazardRecognizer *SPUInstrInfo::CreateTargetHazardRecognizer(
+ const TargetMachine *TM,
+ const ScheduleDAG *DAG) const {
+ const TargetInstrInfo *TII = TM->getInstrInfo();
+ assert(TII && "No InstrInfo?");
+ return new SPUHazardRecognizer(*TII);
}
unsigned
return 0;
}
-bool SPUInstrInfo::copyRegToReg(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- unsigned DestReg, unsigned SrcReg,
- const TargetRegisterClass *DestRC,
- const TargetRegisterClass *SrcRC) const
+void SPUInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I, DebugLoc DL,
+ unsigned DestReg, unsigned SrcReg,
+ bool KillSrc) const
{
// We support cross register class moves for our aliases, such as R3 in any
// reg class to any other reg class containing R3. This is required because
// we instruction select bitconvert i64 -> f64 as a noop for example, so our
// types have no specific meaning.
- if (DestRC == SPU::R8CRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRr8), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::R16CRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRr16), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::R32CRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRr32), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::R32FPRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRf32), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::R64CRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRr64), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::R64FPRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRf64), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::GPRCRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRr128), DestReg).addReg(SrcReg);
- } else if (DestRC == SPU::VECREGRegisterClass) {
- BuildMI(MBB, MI, get(SPU::LRv16i8), DestReg).addReg(SrcReg);
- } else {
- // Attempt to copy unknown/unsupported register class!
- return false;
- }
-
- return true;
+ BuildMI(MBB, I, DL, get(SPU::LRr128), DestReg)
+ .addReg(SrcReg, getKillRegState(KillSrc));
}
void
SPUInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- unsigned SrcReg, bool isKill, int FrameIdx,
- const TargetRegisterClass *RC) const
-{
+ MachineBasicBlock::iterator MI,
+ unsigned SrcReg, bool isKill, int FrameIdx,
+ const TargetRegisterClass *RC,
+ const TargetRegisterInfo *TRI) const {
unsigned opc;
- bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
- if (RC == SPU::GPRCRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr128 : SPU::STQXr128);
- } else if (RC == SPU::R64CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
- } else if (RC == SPU::R64FPRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64);
- } else if (RC == SPU::R32CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
- } else if (RC == SPU::R32FPRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32);
- } else if (RC == SPU::R16CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr16 : SPU::STQXr16);
- } else if (RC == SPU::R8CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::STQDr8 : SPU::STQXr8);
- } else if (RC == SPU::VECREGRegisterClass) {
- opc = (isValidFrameIdx) ? SPU::STQDv16i8 : SPU::STQXv16i8;
- } else {
- assert(0 && "Unknown regclass!");
- abort();
- }
-
- addFrameReference(BuildMI(MBB, MI, get(opc))
- .addReg(SrcReg, false, false, isKill), FrameIdx);
-}
-
-void SPUInstrInfo::storeRegToAddr(MachineFunction &MF, unsigned SrcReg,
- bool isKill,
- SmallVectorImpl<MachineOperand> &Addr,
- const TargetRegisterClass *RC,
- SmallVectorImpl<MachineInstr*> &NewMIs) const {
- cerr << "storeRegToAddr() invoked!\n";
- abort();
-
- if (Addr[0].isFI()) {
- /* do what storeRegToStackSlot does here */
- } else {
- unsigned Opc = 0;
- if (RC == SPU::GPRCRegisterClass) {
- /* Opc = PPC::STW; */
- } else if (RC == SPU::R16CRegisterClass) {
- /* Opc = PPC::STD; */
- } else if (RC == SPU::R32CRegisterClass) {
- /* Opc = PPC::STFD; */
- } else if (RC == SPU::R32FPRegisterClass) {
- /* Opc = PPC::STFD; */
- } else if (RC == SPU::R64FPRegisterClass) {
- /* Opc = PPC::STFS; */
- } else if (RC == SPU::VECREGRegisterClass) {
- /* Opc = PPC::STVX; */
- } else {
- assert(0 && "Unknown regclass!");
- abort();
- }
- MachineInstrBuilder MIB = BuildMI(MF, get(Opc))
- .addReg(SrcReg, false, false, isKill);
- for (unsigned i = 0, e = Addr.size(); i != e; ++i) {
- MachineOperand &MO = Addr[i];
- if (MO.isReg())
- MIB.addReg(MO.getReg());
- else if (MO.isImm())
- MIB.addImm(MO.getImm());
- else
- MIB.addFrameIndex(MO.getIndex());
- }
- NewMIs.push_back(MIB);
- }
+ bool isValidFrameIdx = (FrameIdx < SPUFrameLowering::maxFrameOffset());
+ if (RC == &SPU::GPRCRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr128 : SPU::STQXr128;
+ else if (RC == &SPU::R64CRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64;
+ else if (RC == &SPU::R64FPRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr64 : SPU::STQXr64;
+ else if (RC == &SPU::R32CRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32;
+ else if (RC == &SPU::R32FPRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr32 : SPU::STQXr32;
+ else if (RC == &SPU::R16CRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr16 : SPU::STQXr16;
+ else if (RC == &SPU::R8CRegClass)
+ opc = isValidFrameIdx ? SPU::STQDr8 : SPU::STQXr8;
+ else if (RC == &SPU::VECREGRegClass)
+ opc = isValidFrameIdx ? SPU::STQDv16i8 : SPU::STQXv16i8;
+ else
+ llvm_unreachable("Unknown regclass!");
+
+ DebugLoc DL;
+ if (MI != MBB.end()) DL = MI->getDebugLoc();
+ addFrameReference(BuildMI(MBB, MI, DL, get(opc))
+ .addReg(SrcReg, getKillRegState(isKill)), FrameIdx);
}
void
SPUInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
- MachineBasicBlock::iterator MI,
- unsigned DestReg, int FrameIdx,
- const TargetRegisterClass *RC) const
-{
- unsigned opc;
- bool isValidFrameIdx = (FrameIdx < SPUFrameInfo::maxFrameOffset());
- if (RC == SPU::GPRCRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr128 : SPU::LQXr128);
- } else if (RC == SPU::R64CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
- } else if (RC == SPU::R64FPRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64);
- } else if (RC == SPU::R32CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
- } else if (RC == SPU::R32FPRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32);
- } else if (RC == SPU::R16CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr16 : SPU::LQXr16);
- } else if (RC == SPU::R8CRegisterClass) {
- opc = (isValidFrameIdx ? SPU::LQDr8 : SPU::LQXr8);
- } else if (RC == SPU::VECREGRegisterClass) {
- opc = (isValidFrameIdx) ? SPU::LQDv16i8 : SPU::LQXv16i8;
- } else {
- assert(0 && "Unknown regclass in loadRegFromStackSlot!");
- abort();
- }
-
- addFrameReference(BuildMI(MBB, MI, get(opc)).addReg(DestReg), FrameIdx);
-}
-
-/*!
- \note We are really pessimistic here about what kind of a load we're doing.
- */
-void SPUInstrInfo::loadRegFromAddr(MachineFunction &MF, unsigned DestReg,
- SmallVectorImpl<MachineOperand> &Addr,
+ MachineBasicBlock::iterator MI,
+ unsigned DestReg, int FrameIdx,
const TargetRegisterClass *RC,
- SmallVectorImpl<MachineInstr*> &NewMIs)
- const {
- cerr << "loadRegToAddr() invoked!\n";
- abort();
-
- if (Addr[0].isFI()) {
- /* do what loadRegFromStackSlot does here... */
- } else {
- unsigned Opc = 0;
- if (RC == SPU::R8CRegisterClass) {
- /* do brilliance here */
- } else if (RC == SPU::R16CRegisterClass) {
- /* Opc = PPC::LWZ; */
- } else if (RC == SPU::R32CRegisterClass) {
- /* Opc = PPC::LD; */
- } else if (RC == SPU::R32FPRegisterClass) {
- /* Opc = PPC::LFD; */
- } else if (RC == SPU::R64FPRegisterClass) {
- /* Opc = PPC::LFS; */
- } else if (RC == SPU::VECREGRegisterClass) {
- /* Opc = PPC::LVX; */
- } else if (RC == SPU::GPRCRegisterClass) {
- /* Opc = something else! */
- } else {
- assert(0 && "Unknown regclass!");
- abort();
- }
- MachineInstrBuilder MIB = BuildMI(MF, get(Opc), DestReg);
- for (unsigned i = 0, e = Addr.size(); i != e; ++i) {
- MachineOperand &MO = Addr[i];
- if (MO.isReg())
- MIB.addReg(MO.getReg());
- else if (MO.isImm())
- MIB.addImm(MO.getImm());
- else
- MIB.addFrameIndex(MO.getIndex());
- }
- NewMIs.push_back(MIB);
- }
-}
-
-//! Return true if the specified load or store can be folded
-bool
-SPUInstrInfo::canFoldMemoryOperand(const MachineInstr *MI,
- const SmallVectorImpl<unsigned> &Ops) const {
- if (Ops.size() != 1) return false;
-
- // Make sure this is a reg-reg copy.
- unsigned Opc = MI->getOpcode();
-
- switch (Opc) {
- case SPU::ORv16i8:
- case SPU::ORv8i16:
- case SPU::ORv4i32:
- case SPU::ORv2i64:
- case SPU::ORr8:
- case SPU::ORr16:
- case SPU::ORr32:
- case SPU::ORr64:
- case SPU::ORf32:
- case SPU::ORf64:
- if (MI->getOperand(1).getReg() == MI->getOperand(2).getReg())
- return true;
- break;
- }
-
- return false;
-}
-
-/// foldMemoryOperand - SPU, like PPC, can only fold spills into
-/// copy instructions, turning them into load/store instructions.
-MachineInstr *
-SPUInstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
- MachineInstr *MI,
- const SmallVectorImpl<unsigned> &Ops,
- int FrameIndex) const
-{
- if (Ops.size() != 1) return 0;
-
- unsigned OpNum = Ops[0];
- unsigned Opc = MI->getOpcode();
- MachineInstr *NewMI = 0;
-
- switch (Opc) {
- case SPU::ORv16i8:
- case SPU::ORv8i16:
- case SPU::ORv4i32:
- case SPU::ORv2i64:
- case SPU::ORr8:
- case SPU::ORr16:
- case SPU::ORr32:
- case SPU::ORr64:
- case SPU::ORf32:
- case SPU::ORf64:
- if (OpNum == 0) { // move -> store
- unsigned InReg = MI->getOperand(1).getReg();
- bool isKill = MI->getOperand(1).isKill();
- if (FrameIndex < SPUFrameInfo::maxFrameOffset()) {
- MachineInstrBuilder MIB = BuildMI(MF, get(SPU::STQDr32));
-
- MIB.addReg(InReg, false, false, isKill);
- NewMI = addFrameReference(MIB, FrameIndex);
- }
- } else { // move -> load
- unsigned OutReg = MI->getOperand(0).getReg();
- bool isDead = MI->getOperand(0).isDead();
- MachineInstrBuilder MIB = BuildMI(MF, get(Opc));
-
- MIB.addReg(OutReg, true, false, false, isDead);
- Opc = (FrameIndex < SPUFrameInfo::maxFrameOffset())
- ? SPU::STQDr32 : SPU::STQXr32;
- NewMI = addFrameReference(MIB, FrameIndex);
- break;
- }
- }
-
- return NewMI;
+ const TargetRegisterInfo *TRI) const {
+ unsigned opc;
+ bool isValidFrameIdx = (FrameIdx < SPUFrameLowering::maxFrameOffset());
+ if (RC == &SPU::GPRCRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr128 : SPU::LQXr128;
+ else if (RC == &SPU::R64CRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64;
+ else if (RC == &SPU::R64FPRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr64 : SPU::LQXr64;
+ else if (RC == &SPU::R32CRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32;
+ else if (RC == &SPU::R32FPRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr32 : SPU::LQXr32;
+ else if (RC == &SPU::R16CRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr16 : SPU::LQXr16;
+ else if (RC == &SPU::R8CRegClass)
+ opc = isValidFrameIdx ? SPU::LQDr8 : SPU::LQXr8;
+ else if (RC == &SPU::VECREGRegClass)
+ opc = isValidFrameIdx ? SPU::LQDv16i8 : SPU::LQXv16i8;
+ else
+ llvm_unreachable("Unknown regclass in loadRegFromStackSlot!");
+
+ DebugLoc DL;
+ if (MI != MBB.end()) DL = MI->getDebugLoc();
+ addFrameReference(BuildMI(MBB, MI, DL, get(opc), DestReg), FrameIdx);
}
//! Branch analysis
*/
bool
SPUInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
- MachineBasicBlock *&FBB,
- SmallVectorImpl<MachineOperand> &Cond) const {
+ MachineBasicBlock *&FBB,
+ SmallVectorImpl<MachineOperand> &Cond,
+ bool AllowModify) const {
// If the block has no terminators, it just falls into the block after it.
MachineBasicBlock::iterator I = MBB.end();
- if (I == MBB.begin() || !isUnpredicatedTerminator(--I))
+ if (I == MBB.begin())
+ return false;
+ --I;
+ while (I->isDebugValue()) {
+ if (I == MBB.begin())
+ return false;
+ --I;
+ }
+ if (!isUnpredicatedTerminator(I))
return false;
// Get the last instruction in the block.
// If there is only one terminator instruction, process it.
if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) {
if (isUncondBranch(LastInst)) {
+ // Check for jump tables
+ if (!LastInst->getOperand(0).isMBB())
+ return true;
TBB = LastInst->getOperand(0).getMBB();
return false;
} else if (isCondBranch(LastInst)) {
// Block ends with fall-through condbranch.
TBB = LastInst->getOperand(1).getMBB();
- DEBUG(cerr << "Pushing LastInst: ");
+ DEBUG(errs() << "Pushing LastInst: ");
DEBUG(LastInst->dump());
Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
Cond.push_back(LastInst->getOperand(0));
// If the block ends with a conditional and unconditional branch, handle it.
if (isCondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
TBB = SecondLastInst->getOperand(1).getMBB();
- DEBUG(cerr << "Pushing SecondLastInst: ");
+ DEBUG(errs() << "Pushing SecondLastInst: ");
DEBUG(SecondLastInst->dump());
Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
Cond.push_back(SecondLastInst->getOperand(0));
if (isUncondBranch(SecondLastInst) && isUncondBranch(LastInst)) {
TBB = SecondLastInst->getOperand(0).getMBB();
I = LastInst;
- I->eraseFromParent();
+ if (AllowModify)
+ I->eraseFromParent();
return false;
}
return true;
}
+// search MBB for branch hint labels and branch hit ops
+static void removeHBR( MachineBasicBlock &MBB) {
+ for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I){
+ if (I->getOpcode() == SPU::HBRA ||
+ I->getOpcode() == SPU::HBR_LABEL){
+ I=MBB.erase(I);
+ if (I == MBB.end())
+ break;
+ }
+ }
+}
+
unsigned
SPUInstrInfo::RemoveBranch(MachineBasicBlock &MBB) const {
MachineBasicBlock::iterator I = MBB.end();
+ removeHBR(MBB);
if (I == MBB.begin())
return 0;
--I;
+ while (I->isDebugValue()) {
+ if (I == MBB.begin())
+ return 0;
+ --I;
+ }
if (!isCondBranch(I) && !isUncondBranch(I))
return 0;
// Remove the first branch.
- DEBUG(cerr << "Removing branch: ");
+ DEBUG(errs() << "Removing branch: ");
DEBUG(I->dump());
I->eraseFromParent();
I = MBB.end();
return 1;
// Remove the second branch.
- DEBUG(cerr << "Removing second branch: ");
+ DEBUG(errs() << "Removing second branch: ");
DEBUG(I->dump());
I->eraseFromParent();
return 2;
}
+/** Find the optimal position for a hint branch instruction in a basic block.
+ * This should take into account:
+ * -the branch hint delays
+ * -congestion of the memory bus
+ * -dual-issue scheduling (i.e. avoid insertion of nops)
+ * Current implementation is rather simplistic.
+ */
+static MachineBasicBlock::iterator findHBRPosition(MachineBasicBlock &MBB)
+{
+ MachineBasicBlock::iterator J = MBB.end();
+ for( int i=0; i<8; i++) {
+ if( J == MBB.begin() ) return J;
+ J--;
+ }
+ return J;
+}
+
unsigned
SPUInstrInfo::InsertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
- MachineBasicBlock *FBB,
- const SmallVectorImpl<MachineOperand> &Cond) const {
+ MachineBasicBlock *FBB,
+ const SmallVectorImpl<MachineOperand> &Cond,
+ DebugLoc DL) const {
// Shouldn't be a fall through.
assert(TBB && "InsertBranch must not be told to insert a fallthrough");
assert((Cond.size() == 2 || Cond.size() == 0) &&
"SPU branch conditions have two components!");
+ MachineInstrBuilder MIB;
+ //TODO: make a more accurate algorithm.
+ bool haveHBR = MBB.size()>8;
+
+ removeHBR(MBB);
+ MCSymbol *branchLabel = MBB.getParent()->getContext().CreateTempSymbol();
+ // Add a label just before the branch
+ if (haveHBR)
+ MIB = BuildMI(&MBB, DL, get(SPU::HBR_LABEL)).addSym(branchLabel);
+
// One-way branch.
if (FBB == 0) {
if (Cond.empty()) {
// Unconditional branch
- MachineInstrBuilder MIB = BuildMI(&MBB, get(SPU::BR));
+ MIB = BuildMI(&MBB, DL, get(SPU::BR));
MIB.addMBB(TBB);
- DEBUG(cerr << "Inserted one-way uncond branch: ");
+ DEBUG(errs() << "Inserted one-way uncond branch: ");
DEBUG((*MIB).dump());
+
+ // basic blocks have just one branch so it is safe to add the hint a its
+ if (haveHBR) {
+ MIB = BuildMI( MBB, findHBRPosition(MBB), DL, get(SPU::HBRA));
+ MIB.addSym(branchLabel);
+ MIB.addMBB(TBB);
+ }
} else {
// Conditional branch
- MachineInstrBuilder MIB = BuildMI(&MBB, get(Cond[0].getImm()));
+ MIB = BuildMI(&MBB, DL, get(Cond[0].getImm()));
MIB.addReg(Cond[1].getReg()).addMBB(TBB);
- DEBUG(cerr << "Inserted one-way cond branch: ");
+ if (haveHBR) {
+ MIB = BuildMI(MBB, findHBRPosition(MBB), DL, get(SPU::HBRA));
+ MIB.addSym(branchLabel);
+ MIB.addMBB(TBB);
+ }
+
+ DEBUG(errs() << "Inserted one-way cond branch: ");
DEBUG((*MIB).dump());
}
return 1;
} else {
- MachineInstrBuilder MIB = BuildMI(&MBB, get(Cond[0].getImm()));
- MachineInstrBuilder MIB2 = BuildMI(&MBB, get(SPU::BR));
+ MIB = BuildMI(&MBB, DL, get(Cond[0].getImm()));
+ MachineInstrBuilder MIB2 = BuildMI(&MBB, DL, get(SPU::BR));
// Two-way Conditional Branch.
MIB.addReg(Cond[1].getReg()).addMBB(TBB);
MIB2.addMBB(FBB);
- DEBUG(cerr << "Inserted conditional branch: ");
+ if (haveHBR) {
+ MIB = BuildMI( MBB, findHBRPosition(MBB), DL, get(SPU::HBRA));
+ MIB.addSym(branchLabel);
+ MIB.addMBB(FBB);
+ }
+
+ DEBUG(errs() << "Inserted conditional branch: ");
DEBUG((*MIB).dump());
- DEBUG(cerr << "part 2: ");
+ DEBUG(errs() << "part 2: ");
DEBUG((*MIB2).dump());
return 2;
}
}
-bool
-SPUInstrInfo::BlockHasNoFallThrough(const MachineBasicBlock &MBB) const {
- return (!MBB.empty() && isUncondBranch(&MBB.back()));
-}
//! Reverses a branch's condition, returning false on success.
bool
SPUInstrInfo::ReverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond)
projects / oota-llvm.git / blobdiff