}
}
-bool PPCFrameLowering::findScratchRegister(MachineBasicBlock *MBB,
- bool UseAtEnd,
- unsigned *ScratchRegister) const {
+/* This function will do the following:
+ - If MBB is an entry or exit block, set SR1 and SR2 to R0 and R12
+ respectively (defaults recommended by the ABI) and return true
+ - If MBB is not an entry block, initialize the register scavenger and look
+ for available registers.
+ - If the defaults (R0/R12) are available, return true
+ - If TwoUniqueRegsRequired is set to true, it looks for two unique
+ registers. Otherwise, look for a single available register.
+ - If the required registers are found, set SR1 and SR2 and return true.
+ - If the required registers are not found, set SR2 or both SR1 and SR2 to
+ PPC::NoRegister and return false.
+
+ Note that if both SR1 and SR2 are valid parameters and TwoUniqueRegsRequired
+ is not set, this function will attempt to find two different registers, but
+ still return true if only one register is available (and set SR1 == SR2).
+*/
+bool
+PPCFrameLowering::findScratchRegister(MachineBasicBlock *MBB,
+ bool UseAtEnd,
+ bool TwoUniqueRegsRequired,
+ unsigned *SR1,
+ unsigned *SR2) const {
RegScavenger RS;
- unsigned R0 = Subtarget.isPPC64() ? PPC::X0 : PPC::R0;
+ unsigned R0 = Subtarget.isPPC64() ? PPC::X0 : PPC::R0;
+ unsigned R12 = Subtarget.isPPC64() ? PPC::X12 : PPC::R12;
+
+ // Set the defaults for the two scratch registers.
+ if (SR1)
+ *SR1 = R0;
- if (ScratchRegister)
- *ScratchRegister = R0;
+ if (SR2) {
+ assert (SR1 && "Asking for the second scratch register but not the first?");
+ *SR2 = R12;
+ }
- // If MBB is an entry or exit block, use R0 as the scratch register
+ // If MBB is an entry or exit block, use R0 and R12 as the scratch registers.
if ((UseAtEnd && MBB->isReturnBlock()) ||
(!UseAtEnd && (&MBB->getParent()->front() == MBB)))
return true;
RS.enterBasicBlock(MBB);
if (UseAtEnd && !MBB->empty()) {
- // The scratch register will be used at the end of the block, so must consider
- // all registers used within the block
+ // The scratch register will be used at the end of the block, so must
+ // consider all registers used within the block
MachineBasicBlock::iterator MBBI = MBB->getFirstTerminator();
// If no terminator, back iterator up to previous instruction.
if (MBBI != MBB->begin())
RS.forward(MBBI);
}
-
- if (!RS.isRegUsed(R0))
+
+ // If the two registers are available, we're all good.
+ // Note that we only return here if both R0 and R12 are available because
+ // although the function may not require two unique registers, it may benefit
+ // from having two so we should try to provide them.
+ if (!RS.isRegUsed(R0) && !RS.isRegUsed(R12))
return true;
- unsigned Reg = RS.FindUnusedReg(Subtarget.isPPC64() ? &PPC::G8RCRegClass
- : &PPC::GPRCRegClass);
-
- // Make sure the register scavenger was able to find an available register
- // If not, use R0 but return false to indicate no register was available and
- // R0 must be used (as recommended by the ABI)
- if (Reg == 0)
- return false;
+ // Get the list of callee-saved registers for the target.
+ const PPCRegisterInfo *RegInfo =
+ static_cast<const PPCRegisterInfo *>(Subtarget.getRegisterInfo());
+ const MCPhysReg *CSRegs = RegInfo->getCalleeSavedRegs(MBB->getParent());
+
+ // Get all the available registers in the block.
+ BitVector BV = RS.getRegsAvailable(Subtarget.isPPC64() ? &PPC::G8RCRegClass :
+ &PPC::GPRCRegClass);
+
+ // We shouldn't use callee-saved registers as scratch registers as they may be
+ // available when looking for a candidate block for shrink wrapping but not
+ // available when the actual prologue/epilogue is being emitted because they
+ // were added as live-in to the prologue block by PrologueEpilogueInserter.
+ for (int i = 0; CSRegs[i]; ++i)
+ BV.reset(CSRegs[i]);
+
+ // Set the first scratch register to the first available one.
+ if (SR1) {
+ int FirstScratchReg = BV.find_first();
+ *SR1 = FirstScratchReg == -1 ? (unsigned)PPC::NoRegister : FirstScratchReg;
+ }
- if (ScratchRegister)
- *ScratchRegister = Reg;
+ // If there is another one available, set the second scratch register to that.
+ // Otherwise, set it to either PPC::NoRegister if this function requires two
+ // or to whatever SR1 is set to if this function doesn't require two.
+ if (SR2) {
+ int SecondScratchReg = BV.find_next(*SR1);
+ if (SecondScratchReg != -1)
+ *SR2 = SecondScratchReg;
+ else
+ *SR2 = TwoUniqueRegsRequired ? (unsigned)PPC::NoRegister : *SR1;
+ }
+
+ // Now that we've done our best to provide both registers, double check
+ // whether we were unable to provide enough.
+ if (BV.count() < (TwoUniqueRegsRequired ? 2 : 1))
+ return false;
return true;
}
+// We need a scratch register for spilling LR and for spilling CR. By default,
+// we use two scratch registers to hide latency. However, if only one scratch
+// register is available, we can adjust for that by not overlapping the spill
+// code. However, if we need to realign the stack (i.e. have a base pointer)
+// and the stack frame is large, we need two scratch registers.
+bool
+PPCFrameLowering::twoUniqueScratchRegsRequired(MachineBasicBlock *MBB) const {
+ const PPCRegisterInfo *RegInfo =
+ static_cast<const PPCRegisterInfo *>(Subtarget.getRegisterInfo());
+ MachineFunction &MF = *(MBB->getParent());
+ bool HasBP = RegInfo->hasBasePointer(MF);
+ unsigned FrameSize = determineFrameLayout(MF, false);
+ int NegFrameSize = -FrameSize;
+ bool IsLargeFrame = !isInt<16>(NegFrameSize);
+ MachineFrameInfo *MFI = MF.getFrameInfo();
+ unsigned MaxAlign = MFI->getMaxAlignment();
+
+ return IsLargeFrame && HasBP && MaxAlign > 1;
+}
+
bool PPCFrameLowering::canUseAsPrologue(const MachineBasicBlock &MBB) const {
MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB);
- return findScratchRegister(TmpMBB, false, nullptr);
+ return findScratchRegister(TmpMBB, false,
+ twoUniqueScratchRegsRequired(TmpMBB));
}
bool PPCFrameLowering::canUseAsEpilogue(const MachineBasicBlock &MBB) const {
MachineBasicBlock *TmpMBB = const_cast<MachineBasicBlock *>(&MBB);
- return findScratchRegister(TmpMBB, true, nullptr);
+ return findScratchRegister(TmpMBB, true);
}
void PPCFrameLowering::emitPrologue(MachineFunction &MF,
PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
bool MustSaveLR = FI->mustSaveLR();
const SmallVectorImpl<unsigned> &MustSaveCRs = FI->getMustSaveCRs();
+ bool MustSaveCR = !MustSaveCRs.empty();
// Do we have a frame pointer and/or base pointer for this function?
bool HasFP = hasFP(MF);
bool HasBP = RegInfo->hasBasePointer(MF);
assert((isPPC64 || !isSVR4ABI || !(!FrameSize && (MustSaveLR || HasFP))) &&
"FrameSize must be >0 to save/restore the FP or LR for 32-bit SVR4.");
- findScratchRegister(&MBB, false, &ScratchReg);
- assert(ScratchReg && "No scratch register!");
-
+ // Using the same bool variable as below to supress compiler warnings.
+ bool SingleScratchReg =
+ findScratchRegister(&MBB, false, twoUniqueScratchRegsRequired(&MBB),
+ &ScratchReg, &TempReg);
+ assert(SingleScratchReg &&
+ "Required number of registers not available in this block");
+
+ SingleScratchReg = ScratchReg == TempReg;
+
int LROffset = getReturnSaveOffset();
int FPOffset = 0;
// indexed into with a simple STDU/STWU/STD/STW immediate offset operand.
bool isLargeFrame = !isInt<16>(NegFrameSize);
+ assert((isPPC64 || !MustSaveCR) &&
+ "Prologue CR saving supported only in 64-bit mode");
+
+ // If we need to spill the CR and the LR but we don't have two separate
+ // registers available, we must spill them one at a time
+ if (MustSaveCR && SingleScratchReg && MustSaveLR) {
+ // FIXME: In the ELFv2 ABI, we are not required to save all CR fields.
+ // If only one or two CR fields are clobbered, it could be more
+ // efficient to use mfocrf to selectively save just those fields.
+ MachineInstrBuilder MIB =
+ BuildMI(MBB, MBBI, dl, TII.get(PPC::MFCR8), TempReg);
+ for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
+ MIB.addReg(MustSaveCRs[i], RegState::ImplicitKill);
+ BuildMI(MBB, MBBI, dl, TII.get(PPC::STW8))
+ .addReg(TempReg, getKillRegState(true))
+ .addImm(8)
+ .addReg(SPReg);
+ }
+
if (MustSaveLR)
BuildMI(MBB, MBBI, dl, MFLRInst, ScratchReg);
- assert((isPPC64 || MustSaveCRs.empty()) &&
- "Prologue CR saving supported only in 64-bit mode");
-
- if (!MustSaveCRs.empty()) { // will only occur for PPC64
+ if (MustSaveCR &&
+ !(SingleScratchReg && MustSaveLR)) { // will only occur for PPC64
// FIXME: In the ELFv2 ABI, we are not required to save all CR fields.
// If only one or two CR fields are clobbered, it could be more
// efficient to use mfocrf to selectively save just those fields.
.addImm(LROffset)
.addReg(SPReg);
- if (!MustSaveCRs.empty()) // will only occur for PPC64
+ if (MustSaveCR &&
+ !(SingleScratchReg && MustSaveLR)) // will only occur for PPC64
BuildMI(MBB, MBBI, dl, TII.get(PPC::STW8))
.addReg(TempReg, getKillRegState(true))
.addImm(8)
.addReg(SPReg);
}
+ // This condition must be kept in sync with canUseAsPrologue.
if (HasBP && MaxAlign > 1) {
if (isPPC64)
BuildMI(MBB, MBBI, dl, TII.get(PPC::RLDICL), ScratchReg)
.addReg(ScratchReg, RegState::Kill)
.addImm(NegFrameSize);
} else {
+ assert(!SingleScratchReg && "Only a single scratch reg available");
BuildMI(MBB, MBBI, dl, LoadImmShiftedInst, TempReg)
.addImm(NegFrameSize >> 16);
BuildMI(MBB, MBBI, dl, OrImmInst, TempReg)
// For SVR4, don't emit a move for the CR spill slot if we haven't
// spilled CRs.
if (isSVR4ABI && (PPC::CR2 <= Reg && Reg <= PPC::CR4)
- && MustSaveCRs.empty())
+ && !MustSaveCR)
continue;
// For 64-bit SVR4 when we have spilled CRs, the spill location
PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
bool MustSaveLR = FI->mustSaveLR();
const SmallVectorImpl<unsigned> &MustSaveCRs = FI->getMustSaveCRs();
+ bool MustSaveCR = !MustSaveCRs.empty();
// Do we have a frame pointer and/or base pointer for this function?
bool HasFP = hasFP(MF);
bool HasBP = RegInfo->hasBasePointer(MF);
: PPC::ADDI );
const MCInstrDesc& AddInst = TII.get( isPPC64 ? PPC::ADD8
: PPC::ADD4 );
-
+
int LROffset = getReturnSaveOffset();
int FPOffset = 0;
- findScratchRegister(&MBB, true, &ScratchReg);
- assert(ScratchReg && "No scratch register!");
-
+ // Using the same bool variable as below to supress compiler warnings.
+ bool SingleScratchReg = findScratchRegister(&MBB, true, false, &ScratchReg,
+ &TempReg);
+ assert(SingleScratchReg &&
+ "Could not find an available scratch register");
+
+ SingleScratchReg = ScratchReg == TempReg;
+
if (HasFP) {
if (isSVR4ABI) {
MachineFrameInfo *FFI = MF.getFrameInfo();
}
}
+ assert((isPPC64 || !MustSaveCR) &&
+ "Epilogue CR restoring supported only in 64-bit mode");
+
+ // If we need to save both the LR and the CR and we only have one available
+ // scratch register, we must do them one at a time.
+ if (MustSaveCR && SingleScratchReg && MustSaveLR) {
+ BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ8), TempReg)
+ .addImm(8)
+ .addReg(SPReg);
+ for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
+ BuildMI(MBB, MBBI, dl, TII.get(PPC::MTOCRF8), MustSaveCRs[i])
+ .addReg(TempReg, getKillRegState(i == e-1));
+ }
+
if (MustSaveLR)
BuildMI(MBB, MBBI, dl, LoadInst, ScratchReg)
.addImm(LROffset)
.addReg(SPReg);
- assert((isPPC64 || MustSaveCRs.empty()) &&
- "Epilogue CR restoring supported only in 64-bit mode");
-
- if (!MustSaveCRs.empty()) // will only occur for PPC64
+ if (MustSaveCR &&
+ !(SingleScratchReg && MustSaveLR)) // will only occur for PPC64
BuildMI(MBB, MBBI, dl, TII.get(PPC::LWZ8), TempReg)
.addImm(8)
.addReg(SPReg);
.addImm(BPOffset)
.addReg(SPReg);
- if (!MustSaveCRs.empty()) // will only occur for PPC64
+ if (MustSaveCR &&
+ !(SingleScratchReg && MustSaveLR)) // will only occur for PPC64
for (unsigned i = 0, e = MustSaveCRs.size(); i != e; ++i)
BuildMI(MBB, MBBI, dl, TII.get(PPC::MTOCRF8), MustSaveCRs[i])
.addReg(TempReg, getKillRegState(i == e-1));
const unsigned BasePointerSaveOffset;
/**
- * \brief Find a register that can be used in function prologue and epilogue
+ * \brief Find register[s] that can be used in function prologue and epilogue
*
- * Find a register that can be use as the scratch register in function
+ * Find register[s] that can be use as scratch register[s] in function
* prologue and epilogue to save various registers (Link Register, Base
- * Pointer, etc.). Prefer R0, if it is available. If it is not available,
- * then choose a different register.
+ * Pointer, etc.). Prefer R0/R12, if available. Otherwise choose whatever
+ * register[s] are available.
*
- * This method will return true if an available register was found (including
- * R0). If no available registers are found, the method returns false and sets
- * ScratchRegister to R0, as per the recommendation in the ABI.
+ * This method will return true if it is able to find enough unique scratch
+ * registers (1 or 2 depending on the requirement). If it is unable to find
+ * enough available registers in the block, it will return false and set
+ * any passed output parameter that corresponds to a required unique register
+ * to PPC::NoRegister.
*
* \param[in] MBB The machine basic block to find an available register for
* \param[in] UseAtEnd Specify whether the scratch register will be used at
* the end of the basic block (i.e., will the scratch
* register kill a register defined in the basic block)
- * \param[out] ScratchRegister The scratch register to use
- * \return true if a scratch register was found. false of a scratch register
- * was not found and R0 is being used as the default.
+ * \param[in] TwoUniqueRegsRequired Specify whether this basic block will
+ * require two unique scratch registers.
+ * \param[out] SR1 The scratch register to use
+ * \param[out] SR2 The second scratch register. If this pointer is not null
+ * the function will attempt to set it to an available
+ * register regardless of whether there is a hard requirement
+ * for two unique scratch registers.
+ * \return true if the required number of registers was found.
+ * false if the required number of scratch register weren't available.
+ * If either output parameter refers to a required scratch register
+ * that isn't available, it will be set to an invalid value.
*/
bool findScratchRegister(MachineBasicBlock *MBB,
bool UseAtEnd,
- unsigned *ScratchRegister) const;
+ bool TwoUniqueRegsRequired = false,
+ unsigned *SR1 = nullptr,
+ unsigned *SR2 = nullptr) const;
+ bool twoUniqueScratchRegsRequired(MachineBasicBlock *MBB) const;
public:
PPCFrameLowering(const PPCSubtarget &STI);
--- /dev/null
+; RUN: llc -mcpu=pwr8 -mtriple=powerpc64le-unknown-linux-gnu < %s | FileCheck %s
+
+%struct.anon = type { %struct.anon.0, %struct.anon.1 }
+%struct.anon.0 = type { i32 }
+%struct.anon.1 = type { i32 }
+
+@i = common global i32 0, align 4
+@b = common global i32* null, align 8
+@c = common global i32 0, align 4
+@a = common global i32 0, align 4
+@h = common global i32 0, align 4
+@g = common global i32 0, align 4
+@j = common global i32 0, align 4
+@f = common global %struct.anon zeroinitializer, align 4
+@d = common global i32 0, align 4
+@e = common global i32 0, align 4
+
+; Function Attrs: norecurse nounwind
+define signext i32 @fn1(i32* nocapture %p1, i32 signext %p2, i32* nocapture %p3) {
+entry:
+ %0 = load i32, i32* @i, align 4, !tbaa !1
+ %cond = icmp eq i32 %0, 8
+ br i1 %cond, label %if.end16, label %while.cond.preheader
+
+while.cond.preheader: ; preds = %entry
+ %1 = load i32*, i32** @b, align 8, !tbaa !5
+ %2 = load i32, i32* %1, align 4, !tbaa !1
+ %tobool18 = icmp eq i32 %2, 0
+ br i1 %tobool18, label %while.end, label %while.body.lr.ph
+
+while.body.lr.ph: ; preds = %while.cond.preheader
+ %.pre = load i32, i32* @c, align 4, !tbaa !1
+ br label %while.body
+
+while.body: ; preds = %while.body.backedge, %while.body.lr.ph
+ switch i32 %.pre, label %while.body.backedge [
+ i32 0, label %sw.bb1
+ i32 8, label %sw.bb1
+ i32 6, label %sw.bb1
+ i32 24, label %while.cond.backedge
+ ]
+
+while.body.backedge: ; preds = %while.body, %while.cond.backedge
+ br label %while.body
+
+sw.bb1: ; preds = %while.body, %while.body, %while.body
+ store i32 2, i32* @a, align 4, !tbaa !1
+ br label %while.cond.backedge
+
+while.cond.backedge: ; preds = %while.body, %sw.bb1
+ store i32 4, i32* @a, align 4, !tbaa !1
+ %.pre19 = load i32, i32* %1, align 4, !tbaa !1
+ %tobool = icmp eq i32 %.pre19, 0
+ br i1 %tobool, label %while.end.loopexit, label %while.body.backedge
+
+while.end.loopexit: ; preds = %while.cond.backedge
+ br label %while.end
+
+while.end: ; preds = %while.end.loopexit, %while.cond.preheader
+ %3 = load i32, i32* @h, align 4, !tbaa !1
+ %mul = mul nsw i32 %0, %3
+ %4 = load i32, i32* @g, align 4, !tbaa !1
+ %mul4 = mul nsw i32 %mul, %4
+ store i32 %mul4, i32* @j, align 4, !tbaa !1
+ %5 = load i32, i32* getelementptr inbounds (%struct.anon, %struct.anon* @f, i64 0, i32 0, i32 0), align 4, !tbaa !7
+ %tobool5 = icmp eq i32 %5, 0
+ br i1 %tobool5, label %if.end, label %if.then
+
+if.then: ; preds = %while.end
+ %div = sdiv i32 %5, %mul
+ store i32 %div, i32* @g, align 4, !tbaa !1
+ br label %if.end
+
+if.end: ; preds = %while.end, %if.then
+ %6 = phi i32 [ %4, %while.end ], [ %div, %if.then ]
+ %7 = load i32, i32* getelementptr inbounds (%struct.anon, %struct.anon* @f, i64 0, i32 1, i32 0), align 4, !tbaa !10
+ %tobool7 = icmp ne i32 %7, 0
+ %tobool8 = icmp ne i32 %mul4, 0
+ %or.cond = and i1 %tobool7, %tobool8
+ %tobool10 = icmp ne i32 %0, 0
+ %or.cond17 = and i1 %or.cond, %tobool10
+ br i1 %or.cond17, label %if.then11, label %if.end13
+
+if.then11: ; preds = %if.end
+ store i32 %3, i32* @d, align 4, !tbaa !1
+ %8 = load i32, i32* @e, align 4, !tbaa !1
+ store i32 %8, i32* %p3, align 4, !tbaa !1
+ %.pre20 = load i32, i32* @g, align 4, !tbaa !1
+ br label %if.end13
+
+if.end13: ; preds = %if.then11, %if.end
+ %9 = phi i32 [ %.pre20, %if.then11 ], [ %6, %if.end ]
+ %tobool14 = icmp eq i32 %9, 0
+ br i1 %tobool14, label %if.end16, label %if.then15
+
+if.then15: ; preds = %if.end13
+ store i32 %p2, i32* %p1, align 4, !tbaa !1
+ br label %if.end16
+
+if.end16: ; preds = %entry, %if.end13, %if.then15
+ ret i32 2
+}
+
+; CHECK: mfcr {{[0-9]+}}
+
+!llvm.ident = !{!0}
+
+!0 = !{!"clang version 3.9.0 (trunk 261520)"}
+!1 = !{!2, !2, i64 0}
+!2 = !{!"int", !3, i64 0}
+!3 = !{!"omnipotent char", !4, i64 0}
+!4 = !{!"Simple C/C++ TBAA"}
+!5 = !{!6, !6, i64 0}
+!6 = !{!"any pointer", !3, i64 0}
+!7 = !{!8, !2, i64 0}
+!8 = !{!"", !9, i64 0, !9, i64 4}
+!9 = !{!"", !2, i64 0}
+!10 = !{!8, !2, i64 4}