//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "x86-codegen"
#include "X86.h"
#include "X86InstrInfo.h"
#include "llvm/ADT/DepthFirstIterator.h"
-#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/STLExtras.h"
#include "llvm/CodeGen/EdgeBundles.h"
+#include "llvm/CodeGen/LivePhysRegs.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/Passes.h"
-#include "llvm/InlineAsm.h"
+#include "llvm/IR/InlineAsm.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
#include <algorithm>
+#include <bitset>
using namespace llvm;
+#define DEBUG_TYPE "x86-codegen"
+
STATISTIC(NumFXCH, "Number of fxch instructions inserted");
STATISTIC(NumFP , "Number of floating point instructions");
namespace {
+ const unsigned ScratchFPReg = 7;
+
struct FPS : public MachineFunctionPass {
static char ID;
FPS() : MachineFunctionPass(ID) {
memset(RegMap, 0, sizeof(RegMap));
}
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
AU.addRequired<EdgeBundles>();
AU.addPreservedID(MachineLoopInfoID);
MachineFunctionPass::getAnalysisUsage(AU);
}
- virtual bool runOnMachineFunction(MachineFunction &MF);
+ bool runOnMachineFunction(MachineFunction &MF) override;
- virtual const char *getPassName() const { return "X86 FP Stackifier"; }
+ const char *getPassName() const override { return "X86 FP Stackifier"; }
private:
const TargetInstrInfo *TII; // Machine instruction info.
EdgeBundles *Bundles;
// Return a bitmask of FP registers in block's live-in list.
- unsigned calcLiveInMask(MachineBasicBlock *MBB) {
+ static unsigned calcLiveInMask(MachineBasicBlock *MBB) {
unsigned Mask = 0;
- for (MachineBasicBlock::livein_iterator I = MBB->livein_begin(),
- E = MBB->livein_end(); I != E; ++I) {
- unsigned Reg = *I - X86::FP0;
- if (Reg < 8)
- Mask |= 1 << Reg;
+ for (const auto &LI : MBB->liveins()) {
+ if (LI.PhysReg < X86::FP0 || LI.PhysReg > X86::FP6)
+ continue;
+ Mask |= 1 << (LI.PhysReg - X86::FP0);
}
return Mask;
}
// The hardware keeps track of how many FP registers are live, so we have
// to model that exactly. Usually, each live register corresponds to an
// FP<n> register, but when dealing with calls, returns, and inline
- // assembly, it is sometimes neccesary to have live scratch registers.
+ // assembly, it is sometimes necessary to have live scratch registers.
unsigned Stack[8]; // FP<n> Registers in each stack slot...
unsigned StackTop; // The current top of the FP stack.
enum {
- NumFPRegs = 16 // Including scratch pseudo-registers.
+ NumFPRegs = 8 // Including scratch pseudo-registers.
};
// For each live FP<n> register, point to its Stack[] entry.
// register allocator thinks.
unsigned RegMap[NumFPRegs];
- // Pending fixed registers - Inline assembly needs FP registers to appear
- // in fixed stack slot positions. This is handled by copying FP registers
- // to ST registers before the instruction, and copying back after the
- // instruction.
- //
- // This is modeled with pending ST registers. NumPendingSTs is the number
- // of ST registers (ST0-STn) we are tracking. PendingST[n] points to an FP
- // register that holds the ST value. The ST registers are not moved into
- // place until immediately before the instruction that needs them.
- //
- // It can happen that we need an ST register to be live when no FP register
- // holds the value:
- //
- // %ST0 = COPY %FP4<kill>
- //
- // When that happens, we allocate a scratch FP register to hold the ST
- // value. That means every register in PendingST must be live.
-
- unsigned NumPendingSTs;
- unsigned char PendingST[8];
-
// Set up our stack model to match the incoming registers to MBB.
void setupBlockStack();
// Shuffle live registers to match the expectations of successor blocks.
void finishBlockStack();
+#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
void dumpStack() const {
dbgs() << "Stack contents:";
for (unsigned i = 0; i != StackTop; ++i) {
dbgs() << " FP" << Stack[i];
assert(RegMap[Stack[i]] == i && "Stack[] doesn't match RegMap[]!");
}
- for (unsigned i = 0; i != NumPendingSTs; ++i)
- dbgs() << ", ST" << i << " in FP" << unsigned(PendingST[i]);
- dbgs() << "\n";
}
+#endif
/// getSlot - Return the stack slot number a particular register number is
/// in.
return Slot < StackTop && Stack[Slot] == RegNo;
}
- /// getScratchReg - Return an FP register that is not currently in use.
- unsigned getScratchReg() {
- for (int i = NumFPRegs - 1; i >= 8; --i)
- if (!isLive(i))
- return i;
- llvm_unreachable("Ran out of scratch FP registers");
- }
-
- /// isScratchReg - Returns trus if RegNo is a scratch FP register.
- bool isScratchReg(unsigned RegNo) {
- return RegNo > 8 && RegNo < NumFPRegs;
- }
-
/// getStackEntry - Return the X86::FP<n> register in register ST(i).
unsigned getStackEntry(unsigned STi) const {
if (STi >= StackTop)
/// getSTReg - Return the X86::ST(i) register which contains the specified
/// FP<RegNo> register.
unsigned getSTReg(unsigned RegNo) const {
- return StackTop - 1 - getSlot(RegNo) + llvm::X86::ST0;
+ return StackTop - 1 - getSlot(RegNo) + X86::ST0;
}
// pushReg - Push the specified FP<n> register onto the stack.
bool processBasicBlock(MachineFunction &MF, MachineBasicBlock &MBB);
+ void handleCall(MachineBasicBlock::iterator &I);
void handleZeroArgFP(MachineBasicBlock::iterator &I);
void handleOneArgFP(MachineBasicBlock::iterator &I);
void handleOneArgFPRW(MachineBasicBlock::iterator &I);
void handleSpecialFP(MachineBasicBlock::iterator &I);
// Check if a COPY instruction is using FP registers.
- bool isFPCopy(MachineInstr *MI) {
+ static bool isFPCopy(MachineInstr *MI) {
unsigned DstReg = MI->getOperand(0).getReg();
unsigned SrcReg = MI->getOperand(1).getReg();
return X86::RFP80RegClass.contains(DstReg) ||
X86::RFP80RegClass.contains(SrcReg);
}
+
+ void setKillFlags(MachineBasicBlock &MBB) const;
};
char FPS::ID = 0;
}
// function. If it is all integer, there is nothing for us to do!
bool FPIsUsed = false;
- assert(X86::FP6 == X86::FP0+6 && "Register enums aren't sorted right!");
+ static_assert(X86::FP6 == X86::FP0+6, "Register enums aren't sorted right!");
+ const MachineRegisterInfo &MRI = MF.getRegInfo();
for (unsigned i = 0; i <= 6; ++i)
- if (MF.getRegInfo().isPhysRegUsed(X86::FP0+i)) {
+ if (!MRI.reg_nodbg_empty(X86::FP0 + i)) {
FPIsUsed = true;
break;
}
if (!FPIsUsed) return false;
Bundles = &getAnalysis<EdgeBundles>();
- TII = MF.getTarget().getInstrInfo();
+ TII = MF.getSubtarget().getInstrInfo();
// Prepare cross-MBB liveness.
bundleCFG(MF);
// Process the function in depth first order so that we process at least one
// of the predecessors for every reachable block in the function.
SmallPtrSet<MachineBasicBlock*, 8> Processed;
- MachineBasicBlock *Entry = MF.begin();
+ MachineBasicBlock *Entry = &MF.front();
bool Changed = false;
- for (df_ext_iterator<MachineBasicBlock*, SmallPtrSet<MachineBasicBlock*, 8> >
- I = df_ext_begin(Entry, Processed), E = df_ext_end(Entry, Processed);
- I != E; ++I)
- Changed |= processBasicBlock(MF, **I);
+ for (MachineBasicBlock *BB : depth_first_ext(Entry, Processed))
+ Changed |= processBasicBlock(MF, *BB);
// Process any unreachable blocks in arbitrary order now.
if (MF.size() != Processed.size())
- for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
- if (Processed.insert(BB))
- Changed |= processBasicBlock(MF, *BB);
+ for (MachineBasicBlock &BB : MF)
+ if (Processed.insert(&BB).second)
+ Changed |= processBasicBlock(MF, BB);
LiveBundles.clear();
LiveBundles.resize(Bundles->getNumBundles());
// Gather the actual live-in masks for all MBBs.
- for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
- MachineBasicBlock *MBB = I;
- const unsigned Mask = calcLiveInMask(MBB);
+ for (MachineBasicBlock &MBB : MF) {
+ const unsigned Mask = calcLiveInMask(&MBB);
if (!Mask)
continue;
// Update MBB ingoing bundle mask.
- LiveBundles[Bundles->getBundle(MBB->getNumber(), false)].Mask |= Mask;
+ LiveBundles[Bundles->getBundle(MBB.getNumber(), false)].Mask |= Mask;
}
}
bool FPS::processBasicBlock(MachineFunction &MF, MachineBasicBlock &BB) {
bool Changed = false;
MBB = &BB;
- NumPendingSTs = 0;
+ setKillFlags(BB);
setupBlockStack();
for (MachineBasicBlock::iterator I = BB.begin(); I != BB.end(); ++I) {
X86::RFP80RegClass.contains(MI->getOperand(0).getReg()))
FPInstClass = X86II::SpecialFP;
+ if (MI->isCall())
+ FPInstClass = X86II::SpecialFP;
+
if (FPInstClass == X86II::NotFP)
continue; // Efficiently ignore non-fp insts!
- MachineInstr *PrevMI = 0;
+ MachineInstr *PrevMI = nullptr;
if (I != BB.begin())
- PrevMI = prior(I);
+ PrevMI = std::prev(I);
++NumFP; // Keep track of # of pseudo instrs
DEBUG(dbgs() << "\nFPInst:\t" << *MI);
// after definition. If so, pop them.
for (unsigned i = 0, e = DeadRegs.size(); i != e; ++i) {
unsigned Reg = DeadRegs[i];
- if (Reg >= X86::FP0 && Reg <= X86::FP6) {
+ // Check if Reg is live on the stack. An inline-asm register operand that
+ // is in the clobber list and marked dead might not be live on the stack.
+ if (Reg >= X86::FP0 && Reg <= X86::FP6 && isLive(Reg-X86::FP0)) {
DEBUG(dbgs() << "Register FP#" << Reg-X86::FP0 << " is dead!\n");
freeStackSlotAfter(I, Reg-X86::FP0);
}
} else {
MachineBasicBlock::iterator Start = I;
// Rewind to first instruction newly inserted.
- while (Start != BB.begin() && prior(Start) != PrevI) --Start;
+ while (Start != BB.begin() && std::prev(Start) != PrevI) --Start;
dbgs() << "Inserted instructions:\n\t";
- Start->print(dbgs(), &MF.getTarget());
- while (++Start != llvm::next(I)) {}
+ Start->print(dbgs());
+ while (++Start != std::next(I)) {}
}
dumpStack();
);
+ (void)PrevMI;
Changed = true;
}
namespace {
struct TableEntry {
- unsigned from;
- unsigned to;
+ uint16_t from;
+ uint16_t to;
bool operator<(const TableEntry &TE) const { return from < TE.from; }
friend bool operator<(const TableEntry &TE, unsigned V) {
return TE.from < V;
}
- friend bool LLVM_ATTRIBUTE_USED operator<(unsigned V,
- const TableEntry &TE) {
+ friend bool LLVM_ATTRIBUTE_UNUSED operator<(unsigned V,
+ const TableEntry &TE) {
return V < TE.from;
}
};
}
-#ifndef NDEBUG
-static bool TableIsSorted(const TableEntry *Table, unsigned NumEntries) {
- for (unsigned i = 0; i != NumEntries-1; ++i)
- if (!(Table[i] < Table[i+1])) return false;
- return true;
-}
-#endif
-
-static int Lookup(const TableEntry *Table, unsigned N, unsigned Opcode) {
- const TableEntry *I = std::lower_bound(Table, Table+N, Opcode);
- if (I != Table+N && I->from == Opcode)
+static int Lookup(ArrayRef<TableEntry> Table, unsigned Opcode) {
+ const TableEntry *I = std::lower_bound(Table.begin(), Table.end(), Opcode);
+ if (I != Table.end() && I->from == Opcode)
return I->to;
return -1;
}
#define ASSERT_SORTED(TABLE) \
{ static bool TABLE##Checked = false; \
if (!TABLE##Checked) { \
- assert(TableIsSorted(TABLE, array_lengthof(TABLE)) && \
+ assert(std::is_sorted(std::begin(TABLE), std::end(TABLE)) && \
"All lookup tables must be sorted for efficient access!"); \
TABLE##Checked = true; \
} \
static unsigned getConcreteOpcode(unsigned Opcode) {
ASSERT_SORTED(OpcodeTable);
- int Opc = Lookup(OpcodeTable, array_lengthof(OpcodeTable), Opcode);
+ int Opc = Lookup(OpcodeTable, Opcode);
assert(Opc != -1 && "FP Stack instruction not in OpcodeTable!");
return Opc;
}
RegMap[Stack[--StackTop]] = ~0; // Update state
// Check to see if there is a popping version of this instruction...
- int Opcode = Lookup(PopTable, array_lengthof(PopTable), I->getOpcode());
+ int Opcode = Lookup(PopTable, I->getOpcode());
if (Opcode != -1) {
I->setDesc(TII->get(Opcode));
if (Opcode == X86::UCOM_FPPr)
RegMap[TopReg] = OldSlot;
RegMap[FPRegNo] = ~0;
Stack[--StackTop] = ~0;
- return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr)).addReg(STReg);
+ return BuildMI(*MBB, I, DebugLoc(), TII->get(X86::ST_FPrr))
+ .addReg(STReg)
+ .getInstr();
}
/// adjustLiveRegs - Kill and revive registers such that exactly the FP
// Produce implicit-defs for free by using killed registers.
while (Kills && Defs) {
- unsigned KReg = CountTrailingZeros_32(Kills);
- unsigned DReg = CountTrailingZeros_32(Defs);
+ unsigned KReg = countTrailingZeros(Kills);
+ unsigned DReg = countTrailingZeros(Defs);
DEBUG(dbgs() << "Renaming %FP" << KReg << " as imp %FP" << DReg << "\n");
std::swap(Stack[getSlot(KReg)], Stack[getSlot(DReg)]);
std::swap(RegMap[KReg], RegMap[DReg]);
// Kill registers by popping.
if (Kills && I != MBB->begin()) {
- MachineBasicBlock::iterator I2 = llvm::prior(I);
+ MachineBasicBlock::iterator I2 = std::prev(I);
while (StackTop) {
unsigned KReg = getStackEntry(0);
if (!(Kills & (1 << KReg)))
// Manually kill the rest.
while (Kills) {
- unsigned KReg = CountTrailingZeros_32(Kills);
+ unsigned KReg = countTrailingZeros(Kills);
DEBUG(dbgs() << "Killing %FP" << KReg << "\n");
freeStackSlotBefore(I, KReg);
Kills &= ~(1 << KReg);
// Load zeros for all the imp-defs.
while(Defs) {
- unsigned DReg = CountTrailingZeros_32(Defs);
+ unsigned DReg = countTrailingZeros(Defs);
DEBUG(dbgs() << "Defining %FP" << DReg << " as 0\n");
BuildMI(*MBB, I, DebugLoc(), TII->get(X86::LD_F0));
pushReg(DReg);
// Now we should have the correct registers live.
DEBUG(dumpStack());
- assert(StackTop == CountPopulation_32(Mask) && "Live count mismatch");
+ assert(StackTop == countPopulation(Mask) && "Live count mismatch");
}
/// shuffleStackTop - emit fxch instructions before I to shuffle the top
// Instruction transformation implementation
//===----------------------------------------------------------------------===//
+void FPS::handleCall(MachineBasicBlock::iterator &I) {
+ unsigned STReturns = 0;
+
+ for (const auto &MO : I->operands()) {
+ if (!MO.isReg())
+ continue;
+
+ unsigned R = MO.getReg() - X86::FP0;
+
+ if (R < 8) {
+ assert(MO.isDef() && MO.isImplicit());
+ STReturns |= 1 << R;
+ }
+ }
+
+ unsigned N = countTrailingOnes(STReturns);
+
+ // FP registers used for function return must be consecutive starting at
+ // FP0.
+ assert(STReturns == 0 || (isMask_32(STReturns) && N <= 2));
+
+ for (unsigned I = 0; I < N; ++I)
+ pushReg(N - I - 1);
+}
+
/// handleZeroArgFP - ST(0) = fld0 ST(0) = flds <mem>
///
void FPS::handleZeroArgFP(MachineBasicBlock::iterator &I) {
// Change from the pseudo instruction to the concrete instruction.
MI->RemoveOperand(0); // Remove the explicit ST(0) operand
MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
-
+
// Result gets pushed on the stack.
pushReg(DestReg);
}
MI->getOpcode() == X86::ISTT_Fp32m80 ||
MI->getOpcode() == X86::ISTT_Fp64m80 ||
MI->getOpcode() == X86::ST_FpP80m)) {
- duplicateToTop(Reg, getScratchReg(), I);
+ duplicateToTop(Reg, ScratchFPReg, I);
} else {
moveToTop(Reg, I); // Move to the top of the stack...
}
-
+
// Convert from the pseudo instruction to the concrete instruction.
MI->RemoveOperand(NumOps-1); // Remove explicit ST(0) operand
MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
// We decide which form to use based on what is on the top of the stack, and
// which operand is killed by this instruction.
- const TableEntry *InstTable;
+ ArrayRef<TableEntry> InstTable;
bool isForward = TOS == Op0;
bool updateST0 = (TOS == Op0 && !KillsOp1) || (TOS == Op1 && !KillsOp0);
if (updateST0) {
InstTable = ReverseSTiTable;
}
- int Opcode = Lookup(InstTable, array_lengthof(ForwardST0Table),
- MI->getOpcode());
+ int Opcode = Lookup(InstTable, MI->getOpcode());
assert(Opcode != -1 && "Unknown TwoArgFP pseudo instruction!");
// NotTOS - The register which is not on the top of stack...
MI->RemoveOperand(1);
MI->getOperand(0).setReg(getSTReg(Op1));
MI->setDesc(TII->get(getConcreteOpcode(MI->getOpcode())));
-
+
// If we kill the second operand, make sure to pop it from the stack.
if (Op0 != Op1 && KillsOp1) {
// Get this value off of the register stack.
/// floating point instructions. This is primarily intended for use by pseudo
/// instructions.
///
-void FPS::handleSpecialFP(MachineBasicBlock::iterator &I) {
- MachineInstr *MI = I;
+void FPS::handleSpecialFP(MachineBasicBlock::iterator &Inst) {
+ MachineInstr *MI = Inst;
+
+ if (MI->isCall()) {
+ handleCall(Inst);
+ return;
+ }
+
switch (MI->getOpcode()) {
default: llvm_unreachable("Unknown SpecialFP instruction!");
case TargetOpcode::COPY: {
// We handle three kinds of copies: FP <- FP, FP <- ST, and ST <- FP.
const MachineOperand &MO1 = MI->getOperand(1);
const MachineOperand &MO0 = MI->getOperand(0);
- unsigned DstST = MO0.getReg() - X86::ST0;
- unsigned SrcST = MO1.getReg() - X86::ST0;
bool KillsSrc = MI->killsRegister(MO1.getReg());
- // ST = COPY FP. Set up a pending ST register.
- if (DstST < 8) {
- unsigned SrcFP = getFPReg(MO1);
- assert(isLive(SrcFP) && "Cannot copy dead register");
- assert(!MO0.isDead() && "Cannot copy to dead ST register");
-
- // Unallocated STs are marked as the nonexistent FP255.
- while (NumPendingSTs <= DstST)
- PendingST[NumPendingSTs++] = NumFPRegs;
-
- // STi could still be live from a previous inline asm.
- if (isScratchReg(PendingST[DstST])) {
- DEBUG(dbgs() << "Clobbering old ST in FP" << unsigned(PendingST[DstST])
- << '\n');
- freeStackSlotBefore(MI, PendingST[DstST]);
- }
-
- // When the source is killed, allocate a scratch FP register.
- if (KillsSrc) {
- unsigned Slot = getSlot(SrcFP);
- unsigned SR = getScratchReg();
- PendingST[DstST] = SR;
- Stack[Slot] = SR;
- RegMap[SR] = Slot;
- } else
- PendingST[DstST] = SrcFP;
- break;
- }
-
- // FP = COPY ST. Extract fixed stack value.
- // Any instruction defining ST registers must have assigned them to a
- // scratch register.
- if (SrcST < 8) {
- unsigned DstFP = getFPReg(MO0);
- assert(!isLive(DstFP) && "Cannot copy ST to live FP register");
- assert(NumPendingSTs > SrcST && "Cannot copy from dead ST register");
- unsigned SrcFP = PendingST[SrcST];
- assert(isScratchReg(SrcFP) && "Expected ST in a scratch register");
- assert(isLive(SrcFP) && "Scratch holding ST is dead");
-
- // DstFP steals the stack slot from SrcFP.
- unsigned Slot = getSlot(SrcFP);
- Stack[Slot] = DstFP;
- RegMap[DstFP] = Slot;
-
- // Always treat the ST as killed.
- PendingST[SrcST] = NumFPRegs;
- while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs)
- --NumPendingSTs;
- break;
- }
-
// FP <- FP copy.
unsigned DstFP = getFPReg(MO0);
unsigned SrcFP = getFPReg(MO1);
} else {
// For COPY we just duplicate the specified value to a new stack slot.
// This could be made better, but would require substantial changes.
- duplicateToTop(SrcFP, DstFP, I);
+ duplicateToTop(SrcFP, DstFP, Inst);
}
break;
}
// All FP registers must be explicitly defined, so load a 0 instead.
unsigned Reg = MI->getOperand(0).getReg() - X86::FP0;
DEBUG(dbgs() << "Emitting LD_F0 for implicit FP" << Reg << '\n');
- BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0));
+ BuildMI(*MBB, Inst, MI->getDebugLoc(), TII->get(X86::LD_F0));
pushReg(Reg);
break;
}
- case X86::FpPOP_RETVAL: {
- // The FpPOP_RETVAL instruction is used after calls that return a value on
- // the floating point stack. We cannot model this with ST defs since CALL
- // instructions have fixed clobber lists. This instruction is interpreted
- // to mean that there is one more live register on the stack than we
- // thought.
- //
- // This means that StackTop does not match the hardware stack between a
- // call and the FpPOP_RETVAL instructions. We do tolerate FP instructions
- // between CALL and FpPOP_RETVAL as long as they don't overflow the
- // hardware stack.
- unsigned DstFP = getFPReg(MI->getOperand(0));
-
- // Move existing stack elements up to reflect reality.
- assert(StackTop < 8 && "Stack overflowed before FpPOP_RETVAL");
- if (StackTop) {
- std::copy_backward(Stack, Stack + StackTop, Stack + StackTop + 1);
- for (unsigned i = 0; i != NumFPRegs; ++i)
- ++RegMap[i];
- }
- ++StackTop;
-
- // DstFP is the new bottom of the stack.
- Stack[0] = DstFP;
- RegMap[DstFP] = 0;
-
- // DstFP will be killed by processBasicBlock if this was a dead def.
- break;
- }
-
case TargetOpcode::INLINEASM: {
// The inline asm MachineInstr currently only *uses* FP registers for the
// 'f' constraint. These should be turned into the current ST(x) register
// only tell clobbers from defs by looking at the asm descriptor.
unsigned STUses = 0, STDefs = 0, STClobbers = 0, STDeadDefs = 0;
unsigned NumOps = 0;
+ SmallSet<unsigned, 1> FRegIdx;
+ unsigned RCID;
+
for (unsigned i = InlineAsm::MIOp_FirstOperand, e = MI->getNumOperands();
i != e && MI->getOperand(i).isImm(); i += 1 + NumOps) {
unsigned Flags = MI->getOperand(i).getImm();
+
NumOps = InlineAsm::getNumOperandRegisters(Flags);
if (NumOps != 1)
continue;
const MachineOperand &MO = MI->getOperand(i + 1);
if (!MO.isReg())
continue;
- unsigned STReg = MO.getReg() - X86::ST0;
+ unsigned STReg = MO.getReg() - X86::FP0;
if (STReg >= 8)
continue;
+ // If the flag has a register class constraint, this must be an operand
+ // with constraint "f". Record its index and continue.
+ if (InlineAsm::hasRegClassConstraint(Flags, RCID)) {
+ FRegIdx.insert(i + 1);
+ continue;
+ }
+
switch (InlineAsm::getKind(Flags)) {
case InlineAsm::Kind_RegUse:
STUses |= (1u << STReg);
if (STUses && !isMask_32(STUses))
MI->emitError("fixed input regs must be last on the x87 stack");
- unsigned NumSTUses = CountTrailingOnes_32(STUses);
+ unsigned NumSTUses = countTrailingOnes(STUses);
// Defs must be contiguous from the stack top. ST0-STn.
if (STDefs && !isMask_32(STDefs)) {
MI->emitError("output regs must be last on the x87 stack");
STDefs = NextPowerOf2(STDefs) - 1;
}
- unsigned NumSTDefs = CountTrailingOnes_32(STDefs);
+ unsigned NumSTDefs = countTrailingOnes(STDefs);
// So must the clobbered stack slots. ST0-STm, m >= n.
if (STClobbers && !isMask_32(STDefs | STClobbers))
unsigned STPopped = STUses & (STDefs | STClobbers);
if (STPopped && !isMask_32(STPopped))
MI->emitError("implicitly popped regs must be last on the x87 stack");
- unsigned NumSTPopped = CountTrailingOnes_32(STPopped);
+ unsigned NumSTPopped = countTrailingOnes(STPopped);
DEBUG(dbgs() << "Asm uses " << NumSTUses << " fixed regs, pops "
<< NumSTPopped << ", and defines " << NumSTDefs << " regs.\n");
- // Scan the instruction for FP uses corresponding to "f" constraints.
- // Collect FP registers to kill afer the instruction.
- // Always kill all the scratch regs.
+#ifndef NDEBUG
+ // If any input operand uses constraint "f", all output register
+ // constraints must be early-clobber defs.
+ for (unsigned I = 0, E = MI->getNumOperands(); I < E; ++I)
+ if (FRegIdx.count(I)) {
+ assert((1 << getFPReg(MI->getOperand(I)) & STDefs) == 0 &&
+ "Operands with constraint \"f\" cannot overlap with defs");
+ }
+#endif
+
+ // Collect all FP registers (register operands with constraints "t", "u",
+ // and "f") to kill afer the instruction.
unsigned FPKills = ((1u << NumFPRegs) - 1) & ~0xff;
- unsigned FPUsed = 0;
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &Op = MI->getOperand(i);
if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
continue;
- if (!Op.isUse())
- MI->emitError("illegal \"f\" output constraint");
unsigned FPReg = getFPReg(Op);
- FPUsed |= 1U << FPReg;
// If we kill this operand, make sure to pop it from the stack after the
// asm. We just remember it for now, and pop them all off at the end in
// a batch.
- if (Op.isKill())
+ if (Op.isUse() && Op.isKill())
FPKills |= 1U << FPReg;
}
- // The popped inputs will be killed by the instruction, so duplicate them
- // if the FP register needs to be live after the instruction, or if it is
- // used in the instruction itself. We effectively treat the popped inputs
- // as early clobbers.
- for (unsigned i = 0; i < NumSTPopped; ++i) {
- if ((FPKills & ~FPUsed) & (1u << PendingST[i]))
- continue;
- unsigned SR = getScratchReg();
- duplicateToTop(PendingST[i], SR, I);
- DEBUG(dbgs() << "Duplicating ST" << i << " in FP"
- << unsigned(PendingST[i]) << " to avoid clobbering it.\n");
- PendingST[i] = SR;
- }
-
- // Make sure we have a unique live register for every fixed use. Some of
- // them could be undef uses, and we need to emit LD_F0 instructions.
- for (unsigned i = 0; i < NumSTUses; ++i) {
- if (i < NumPendingSTs && PendingST[i] < NumFPRegs) {
- // Check for shared assignments.
- for (unsigned j = 0; j < i; ++j) {
- if (PendingST[j] != PendingST[i])
- continue;
- // STi and STj are inn the same register, create a copy.
- unsigned SR = getScratchReg();
- duplicateToTop(PendingST[i], SR, I);
- DEBUG(dbgs() << "Duplicating ST" << i << " in FP"
- << unsigned(PendingST[i])
- << " to avoid collision with ST" << j << '\n');
- PendingST[i] = SR;
- }
- continue;
- }
- unsigned SR = getScratchReg();
- DEBUG(dbgs() << "Emitting LD_F0 for ST" << i << " in FP" << SR << '\n');
- BuildMI(*MBB, I, MI->getDebugLoc(), TII->get(X86::LD_F0));
- pushReg(SR);
- PendingST[i] = SR;
- if (NumPendingSTs == i)
- ++NumPendingSTs;
- }
- assert(NumPendingSTs >= NumSTUses && "Fixed registers should be assigned");
+ // Do not include registers that are implicitly popped by defs/clobbers.
+ FPKills &= ~(STDefs | STClobbers);
// Now we can rearrange the live registers to match what was requested.
- shuffleStackTop(PendingST, NumPendingSTs, I);
+ unsigned char STUsesArray[8];
+
+ for (unsigned I = 0; I < NumSTUses; ++I)
+ STUsesArray[I] = I;
+
+ shuffleStackTop(STUsesArray, NumSTUses, Inst);
DEBUG({dbgs() << "Before asm: "; dumpStack();});
// With the stack layout fixed, rewrite the FP registers.
MachineOperand &Op = MI->getOperand(i);
if (!Op.isReg() || Op.getReg() < X86::FP0 || Op.getReg() > X86::FP6)
continue;
+
unsigned FPReg = getFPReg(Op);
- Op.setReg(getSTReg(FPReg));
+
+ if (FRegIdx.count(i))
+ // Operand with constraint "f".
+ Op.setReg(getSTReg(FPReg));
+ else
+ // Operand with a single register class constraint ("t" or "u").
+ Op.setReg(X86::ST0 + FPReg);
}
// Simulate the inline asm popping its inputs and pushing its outputs.
StackTop -= NumSTPopped;
- // Hold the fixed output registers in scratch FP registers. They will be
- // transferred to real FP registers by copies.
- NumPendingSTs = 0;
- for (unsigned i = 0; i < NumSTDefs; ++i) {
- unsigned SR = getScratchReg();
- pushReg(SR);
- FPKills &= ~(1u << SR);
- }
for (unsigned i = 0; i < NumSTDefs; ++i)
- PendingST[NumPendingSTs++] = getStackEntry(i);
- DEBUG({dbgs() << "After asm: "; dumpStack();});
-
- // If any of the ST defs were dead, pop them immediately. Our caller only
- // handles dead FP defs.
- MachineBasicBlock::iterator InsertPt = MI;
- for (unsigned i = 0; STDefs & (1u << i); ++i) {
- if (!(STDeadDefs & (1u << i)))
- continue;
- freeStackSlotAfter(InsertPt, PendingST[i]);
- PendingST[i] = NumFPRegs;
- }
- while (NumPendingSTs && PendingST[NumPendingSTs - 1] == NumFPRegs)
- --NumPendingSTs;
+ pushReg(NumSTDefs - i - 1);
// If this asm kills any FP registers (is the last use of them) we must
// explicitly emit pop instructions for them. Do this now after the asm has
// Note: this might be a non-optimal pop sequence. We might be able to do
// better by trying to pop in stack order or something.
while (FPKills) {
- unsigned FPReg = CountTrailingZeros_32(FPKills);
+ unsigned FPReg = countTrailingZeros(FPKills);
if (isLive(FPReg))
- freeStackSlotAfter(InsertPt, FPReg);
+ freeStackSlotAfter(Inst, FPReg);
FPKills &= ~(1U << FPReg);
}
+
// Don't delete the inline asm!
return;
}
- case X86::RET:
- case X86::RETI:
+ case X86::RETQ:
+ case X86::RETL:
+ case X86::RETIL:
+ case X86::RETIQ:
// If RET has an FP register use operand, pass the first one in ST(0) and
// the second one in ST(1).
// Assert that the top of stack contains the right FP register.
assert(StackTop == 1 && FirstFPRegOp == getStackEntry(0) &&
"Top of stack not the right register for RET!");
-
+
// Ok, everything is good, mark the value as not being on the stack
// anymore so that our assertion about the stack being empty at end of
// block doesn't fire.
StackTop = 0;
return;
}
-
+
// Otherwise, we are returning two values:
// 2) If returning the same value for both, we only have one thing in the FP
// stack. Consider: RET FP1, FP1
if (StackTop == 1) {
assert(FirstFPRegOp == SecondFPRegOp && FirstFPRegOp == getStackEntry(0)&&
"Stack misconfiguration for RET!");
-
+
// Duplicate the TOS so that we return it twice. Just pick some other FPx
// register to hold it.
- unsigned NewReg = getScratchReg();
+ unsigned NewReg = ScratchFPReg;
duplicateToTop(FirstFPRegOp, NewReg, MI);
FirstFPRegOp = NewReg;
}
-
+
/// Okay we know we have two different FPx operands now:
assert(StackTop == 2 && "Must have two values live!");
-
+
/// 3) If SecondFPRegOp is currently in ST(0) and FirstFPRegOp is currently
/// in ST(1). In this case, emit an fxch.
if (getStackEntry(0) == SecondFPRegOp) {
assert(getStackEntry(1) == FirstFPRegOp && "Unknown regs live");
moveToTop(FirstFPRegOp, MI);
}
-
+
/// 4) Finally, FirstFPRegOp must be in ST(0) and SecondFPRegOp must be in
/// ST(1). Just remove both from our understanding of the stack and return.
assert(getStackEntry(0) == FirstFPRegOp && "Unknown regs live");
return;
}
- I = MBB->erase(I); // Remove the pseudo instruction
+ Inst = MBB->erase(Inst); // Remove the pseudo instruction
// We want to leave I pointing to the previous instruction, but what if we
// just erased the first instruction?
- if (I == MBB->begin()) {
+ if (Inst == MBB->begin()) {
DEBUG(dbgs() << "Inserting dummy KILL\n");
- I = BuildMI(*MBB, I, DebugLoc(), TII->get(TargetOpcode::KILL));
+ Inst = BuildMI(*MBB, Inst, DebugLoc(), TII->get(TargetOpcode::KILL));
} else
- --I;
+ --Inst;
+}
+
+void FPS::setKillFlags(MachineBasicBlock &MBB) const {
+ const TargetRegisterInfo *TRI =
+ MBB.getParent()->getSubtarget().getRegisterInfo();
+ LivePhysRegs LPR(TRI);
+
+ LPR.addLiveOuts(&MBB);
+
+ for (MachineBasicBlock::reverse_iterator I = MBB.rbegin(), E = MBB.rend();
+ I != E; ++I) {
+ if (I->isDebugValue())
+ continue;
+
+ std::bitset<8> Defs;
+ SmallVector<MachineOperand *, 2> Uses;
+ MachineInstr &MI = *I;
+
+ for (auto &MO : I->operands()) {
+ if (!MO.isReg())
+ continue;
+
+ unsigned Reg = MO.getReg() - X86::FP0;
+
+ if (Reg >= 8)
+ continue;
+
+ if (MO.isDef()) {
+ Defs.set(Reg);
+ if (!LPR.contains(MO.getReg()))
+ MO.setIsDead();
+ } else
+ Uses.push_back(&MO);
+ }
+
+ for (auto *MO : Uses)
+ if (Defs.test(getFPReg(*MO)) || !LPR.contains(MO->getReg()))
+ MO->setIsKill();
+
+ LPR.stepBackward(MI);
+ }
}