// same flag that the "cmp" instruction sets and that "bz" uses, then we can
// eliminate the "cmp" instruction.
//
-// - Optimize Bitcast pairs:
+// Another instance, in this code:
//
-// v1 = bitcast v0
-// v2 = bitcast v1
-// = v2
+// sub r1, r3 | sub r1, imm
+// cmp r3, r1 or cmp r1, r3 | cmp r1, imm
+// bge L1
+//
+// If the branch instruction can use flag from "sub", then we can replace
+// "sub" with "subs" and eliminate the "cmp" instruction.
+//
+// - Optimize Loads:
+//
+// Loads that can be folded into a later instruction. A load is foldable
+// if it loads to virtual registers and the virtual register defined has
+// a single use.
+//
+// - Optimize Copies and Bitcast (more generally, target specific copies):
+//
+// Rewrite copies and bitcasts to avoid cross register bank copies
+// when possible.
+// E.g., Consider the following example, where capital and lower
+// letters denote different register file:
+// b = copy A <-- cross-bank copy
+// C = copy b <-- cross-bank copy
// =>
-// v1 = bitcast v0
-// = v0
-//
+// b = copy A <-- cross-bank copy
+// C = copy A <-- same-bank copy
+//
+// E.g., for bitcast:
+// b = bitcast A <-- cross-bank copy
+// C = bitcast b <-- cross-bank copy
+// =>
+// b = bitcast A <-- cross-bank copy
+// C = copy A <-- same-bank copy
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "peephole-opt"
#include "llvm/CodeGen/Passes.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/Support/CommandLine.h"
-#include "llvm/ADT/DenseMap.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/Statistic.h"
+#include "llvm/Target/TargetSubtargetInfo.h"
+#include <utility>
using namespace llvm;
+#define DEBUG_TYPE "peephole-opt"
+
// Optimize Extensions
static cl::opt<bool>
Aggressive("aggressive-ext-opt", cl::Hidden,
DisablePeephole("disable-peephole", cl::Hidden, cl::init(false),
cl::desc("Disable the peephole optimizer"));
+static cl::opt<bool>
+DisableAdvCopyOpt("disable-adv-copy-opt", cl::Hidden, cl::init(false),
+ cl::desc("Disable advanced copy optimization"));
+
+// Limit the number of PHI instructions to process
+// in PeepholeOptimizer::getNextSource.
+static cl::opt<unsigned> RewritePHILimit(
+ "rewrite-phi-limit", cl::Hidden, cl::init(10),
+ cl::desc("Limit the length of PHI chains to lookup"));
+
STATISTIC(NumReuse, "Number of extension results reused");
-STATISTIC(NumBitcasts, "Number of bitcasts eliminated");
STATISTIC(NumCmps, "Number of compares eliminated");
-STATISTIC(NumImmFold, "Number of move immediate foled");
+STATISTIC(NumImmFold, "Number of move immediate folded");
+STATISTIC(NumLoadFold, "Number of loads folded");
+STATISTIC(NumSelects, "Number of selects optimized");
+STATISTIC(NumUncoalescableCopies, "Number of uncoalescable copies optimized");
+STATISTIC(NumRewrittenCopies, "Number of copies rewritten");
namespace {
+ class ValueTrackerResult;
+
class PeepholeOptimizer : public MachineFunctionPass {
- const TargetMachine *TM;
const TargetInstrInfo *TII;
+ const TargetRegisterInfo *TRI;
MachineRegisterInfo *MRI;
MachineDominatorTree *DT; // Machine dominator tree
initializePeepholeOptimizerPass(*PassRegistry::getPassRegistry());
}
- virtual bool runOnMachineFunction(MachineFunction &MF);
+ bool runOnMachineFunction(MachineFunction &MF) override;
- virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ void getAnalysisUsage(AnalysisUsage &AU) const override {
AU.setPreservesCFG();
MachineFunctionPass::getAnalysisUsage(AU);
if (Aggressive) {
}
}
+ /// \brief Track Def -> Use info used for rewriting copies.
+ typedef SmallDenseMap<TargetInstrInfo::RegSubRegPair, ValueTrackerResult>
+ RewriteMapTy;
+
private:
- bool OptimizeBitcastInstr(MachineInstr *MI, MachineBasicBlock *MBB);
- bool OptimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
- bool OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
- SmallPtrSet<MachineInstr*, 8> &LocalMIs);
+ bool optimizeCmpInstr(MachineInstr *MI, MachineBasicBlock *MBB);
+ bool optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
+ SmallPtrSetImpl<MachineInstr*> &LocalMIs);
+ bool optimizeSelect(MachineInstr *MI,
+ SmallPtrSetImpl<MachineInstr *> &LocalMIs);
+ bool optimizeCondBranch(MachineInstr *MI);
+ bool optimizeCopyOrBitcast(MachineInstr *MI);
+ bool optimizeCoalescableCopy(MachineInstr *MI);
+ bool optimizeUncoalescableCopy(MachineInstr *MI,
+ SmallPtrSetImpl<MachineInstr *> &LocalMIs);
+ bool findNextSource(unsigned Reg, unsigned SubReg,
+ RewriteMapTy &RewriteMap);
bool isMoveImmediate(MachineInstr *MI,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
- bool FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
+ bool foldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs);
+ bool isLoadFoldable(MachineInstr *MI,
+ SmallSet<unsigned, 16> &FoldAsLoadDefCandidates);
+
+ /// \brief Check whether \p MI is understood by the register coalescer
+ /// but may require some rewriting.
+ bool isCoalescableCopy(const MachineInstr &MI) {
+ // SubregToRegs are not interesting, because they are already register
+ // coalescer friendly.
+ return MI.isCopy() || (!DisableAdvCopyOpt &&
+ (MI.isRegSequence() || MI.isInsertSubreg() ||
+ MI.isExtractSubreg()));
+ }
+
+ /// \brief Check whether \p MI is a copy like instruction that is
+ /// not recognized by the register coalescer.
+ bool isUncoalescableCopy(const MachineInstr &MI) {
+ return MI.isBitcast() ||
+ (!DisableAdvCopyOpt &&
+ (MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
+ MI.isExtractSubregLike()));
+ }
+ };
+
+ /// \brief Helper class to hold a reply for ValueTracker queries. Contains the
+ /// returned sources for a given search and the instructions where the sources
+ /// were tracked from.
+ class ValueTrackerResult {
+ private:
+ /// Track all sources found by one ValueTracker query.
+ SmallVector<TargetInstrInfo::RegSubRegPair, 2> RegSrcs;
+
+ /// Instruction using the sources in 'RegSrcs'.
+ const MachineInstr *Inst;
+
+ public:
+ ValueTrackerResult() : Inst(nullptr) {}
+ ValueTrackerResult(unsigned Reg, unsigned SubReg) : Inst(nullptr) {
+ addSource(Reg, SubReg);
+ }
+
+ bool isValid() const { return getNumSources() > 0; }
+
+ void setInst(const MachineInstr *I) { Inst = I; }
+ const MachineInstr *getInst() const { return Inst; }
+
+ void clear() {
+ RegSrcs.clear();
+ Inst = nullptr;
+ }
+
+ void addSource(unsigned SrcReg, unsigned SrcSubReg) {
+ RegSrcs.push_back(TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg));
+ }
+
+ void setSource(int Idx, unsigned SrcReg, unsigned SrcSubReg) {
+ assert(Idx < getNumSources() && "Reg pair source out of index");
+ RegSrcs[Idx] = TargetInstrInfo::RegSubRegPair(SrcReg, SrcSubReg);
+ }
+
+ int getNumSources() const { return RegSrcs.size(); }
+
+ unsigned getSrcReg(int Idx) const {
+ assert(Idx < getNumSources() && "Reg source out of index");
+ return RegSrcs[Idx].Reg;
+ }
+
+ unsigned getSrcSubReg(int Idx) const {
+ assert(Idx < getNumSources() && "SubReg source out of index");
+ return RegSrcs[Idx].SubReg;
+ }
+
+ bool operator==(const ValueTrackerResult &Other) {
+ if (Other.getInst() != getInst())
+ return false;
+
+ if (Other.getNumSources() != getNumSources())
+ return false;
+
+ for (int i = 0, e = Other.getNumSources(); i != e; ++i)
+ if (Other.getSrcReg(i) != getSrcReg(i) ||
+ Other.getSrcSubReg(i) != getSrcSubReg(i))
+ return false;
+ return true;
+ }
+ };
+
+ /// \brief Helper class to track the possible sources of a value defined by
+ /// a (chain of) copy related instructions.
+ /// Given a definition (instruction and definition index), this class
+ /// follows the use-def chain to find successive suitable sources.
+ /// The given source can be used to rewrite the definition into
+ /// def = COPY src.
+ ///
+ /// For instance, let us consider the following snippet:
+ /// v0 =
+ /// v2 = INSERT_SUBREG v1, v0, sub0
+ /// def = COPY v2.sub0
+ ///
+ /// Using a ValueTracker for def = COPY v2.sub0 will give the following
+ /// suitable sources:
+ /// v2.sub0 and v0.
+ /// Then, def can be rewritten into def = COPY v0.
+ class ValueTracker {
+ private:
+ /// The current point into the use-def chain.
+ const MachineInstr *Def;
+ /// The index of the definition in Def.
+ unsigned DefIdx;
+ /// The sub register index of the definition.
+ unsigned DefSubReg;
+ /// The register where the value can be found.
+ unsigned Reg;
+ /// Specifiy whether or not the value tracking looks through
+ /// complex instructions. When this is false, the value tracker
+ /// bails on everything that is not a copy or a bitcast.
+ ///
+ /// Note: This could have been implemented as a specialized version of
+ /// the ValueTracker class but that would have complicated the code of
+ /// the users of this class.
+ bool UseAdvancedTracking;
+ /// MachineRegisterInfo used to perform tracking.
+ const MachineRegisterInfo &MRI;
+ /// Optional TargetInstrInfo used to perform some complex
+ /// tracking.
+ const TargetInstrInfo *TII;
+
+ /// \brief Dispatcher to the right underlying implementation of
+ /// getNextSource.
+ ValueTrackerResult getNextSourceImpl();
+ /// \brief Specialized version of getNextSource for Copy instructions.
+ ValueTrackerResult getNextSourceFromCopy();
+ /// \brief Specialized version of getNextSource for Bitcast instructions.
+ ValueTrackerResult getNextSourceFromBitcast();
+ /// \brief Specialized version of getNextSource for RegSequence
+ /// instructions.
+ ValueTrackerResult getNextSourceFromRegSequence();
+ /// \brief Specialized version of getNextSource for InsertSubreg
+ /// instructions.
+ ValueTrackerResult getNextSourceFromInsertSubreg();
+ /// \brief Specialized version of getNextSource for ExtractSubreg
+ /// instructions.
+ ValueTrackerResult getNextSourceFromExtractSubreg();
+ /// \brief Specialized version of getNextSource for SubregToReg
+ /// instructions.
+ ValueTrackerResult getNextSourceFromSubregToReg();
+ /// \brief Specialized version of getNextSource for PHI instructions.
+ ValueTrackerResult getNextSourceFromPHI();
+
+ public:
+ /// \brief Create a ValueTracker instance for the value defined by \p Reg.
+ /// \p DefSubReg represents the sub register index the value tracker will
+ /// track. It does not need to match the sub register index used in the
+ /// definition of \p Reg.
+ /// \p UseAdvancedTracking specifies whether or not the value tracker looks
+ /// through complex instructions. By default (false), it handles only copy
+ /// and bitcast instructions.
+ /// If \p Reg is a physical register, a value tracker constructed with
+ /// this constructor will not find any alternative source.
+ /// Indeed, when \p Reg is a physical register that constructor does not
+ /// know which definition of \p Reg it should track.
+ /// Use the next constructor to track a physical register.
+ ValueTracker(unsigned Reg, unsigned DefSubReg,
+ const MachineRegisterInfo &MRI,
+ bool UseAdvancedTracking = false,
+ const TargetInstrInfo *TII = nullptr)
+ : Def(nullptr), DefIdx(0), DefSubReg(DefSubReg), Reg(Reg),
+ UseAdvancedTracking(UseAdvancedTracking), MRI(MRI), TII(TII) {
+ if (!TargetRegisterInfo::isPhysicalRegister(Reg)) {
+ Def = MRI.getVRegDef(Reg);
+ DefIdx = MRI.def_begin(Reg).getOperandNo();
+ }
+ }
+
+ /// \brief Create a ValueTracker instance for the value defined by
+ /// the pair \p MI, \p DefIdx.
+ /// Unlike the other constructor, the value tracker produced by this one
+ /// may be able to find a new source when the definition is a physical
+ /// register.
+ /// This could be useful to rewrite target specific instructions into
+ /// generic copy instructions.
+ ValueTracker(const MachineInstr &MI, unsigned DefIdx, unsigned DefSubReg,
+ const MachineRegisterInfo &MRI,
+ bool UseAdvancedTracking = false,
+ const TargetInstrInfo *TII = nullptr)
+ : Def(&MI), DefIdx(DefIdx), DefSubReg(DefSubReg),
+ UseAdvancedTracking(UseAdvancedTracking), MRI(MRI), TII(TII) {
+ assert(DefIdx < Def->getDesc().getNumDefs() &&
+ Def->getOperand(DefIdx).isReg() && "Invalid definition");
+ Reg = Def->getOperand(DefIdx).getReg();
+ }
+
+ /// \brief Following the use-def chain, get the next available source
+ /// for the tracked value.
+ /// \return A ValueTrackerResult containing a set of registers
+ /// and sub registers with tracked values. A ValueTrackerResult with
+ /// an empty set of registers means no source was found.
+ ValueTrackerResult getNextSource();
+
+ /// \brief Get the last register where the initial value can be found.
+ /// Initially this is the register of the definition.
+ /// Then, after each successful call to getNextSource, this is the
+ /// register of the last source.
+ unsigned getReg() const { return Reg; }
};
}
char PeepholeOptimizer::ID = 0;
+char &llvm::PeepholeOptimizerID = PeepholeOptimizer::ID;
INITIALIZE_PASS_BEGIN(PeepholeOptimizer, "peephole-opts",
"Peephole Optimizations", false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_END(PeepholeOptimizer, "peephole-opts",
"Peephole Optimizations", false, false)
-FunctionPass *llvm::createPeepholeOptimizerPass() {
- return new PeepholeOptimizer();
-}
-
-/// OptimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
+/// optimizeExtInstr - If instruction is a copy-like instruction, i.e. it reads
/// a single register and writes a single register and it does not modify the
/// source, and if the source value is preserved as a sub-register of the
/// result, then replace all reachable uses of the source with the subreg of the
/// result.
-///
+///
/// Do not generate an EXTRACT that is used only in a debug use, as this changes
/// the code. Since this code does not currently share EXTRACTs, just ignore all
/// debug uses.
bool PeepholeOptimizer::
-OptimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
- SmallPtrSet<MachineInstr*, 8> &LocalMIs) {
+optimizeExtInstr(MachineInstr *MI, MachineBasicBlock *MBB,
+ SmallPtrSetImpl<MachineInstr*> &LocalMIs) {
unsigned SrcReg, DstReg, SubIdx;
if (!TII->isCoalescableExtInstr(*MI, SrcReg, DstReg, SubIdx))
return false;
-
+
if (TargetRegisterInfo::isPhysicalRegister(DstReg) ||
TargetRegisterInfo::isPhysicalRegister(SrcReg))
return false;
- MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(SrcReg);
- if (++UI == MRI->use_nodbg_end())
+ if (MRI->hasOneNonDBGUse(SrcReg))
// No other uses.
return false;
+ // Ensure DstReg can get a register class that actually supports
+ // sub-registers. Don't change the class until we commit.
+ const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
+ DstRC = TRI->getSubClassWithSubReg(DstRC, SubIdx);
+ if (!DstRC)
+ return false;
+
+ // The ext instr may be operating on a sub-register of SrcReg as well.
+ // PPC::EXTSW is a 32 -> 64-bit sign extension, but it reads a 64-bit
+ // register.
+ // If UseSrcSubIdx is Set, SubIdx also applies to SrcReg, and only uses of
+ // SrcReg:SubIdx should be replaced.
+ bool UseSrcSubIdx =
+ TRI->getSubClassWithSubReg(MRI->getRegClass(SrcReg), SubIdx) != nullptr;
+
// The source has other uses. See if we can replace the other uses with use of
// the result of the extension.
SmallPtrSet<MachineBasicBlock*, 4> ReachedBBs;
- UI = MRI->use_nodbg_begin(DstReg);
- for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
- UI != UE; ++UI)
- ReachedBBs.insert(UI->getParent());
+ for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
+ ReachedBBs.insert(UI.getParent());
// Uses that are in the same BB of uses of the result of the instruction.
SmallVector<MachineOperand*, 8> Uses;
SmallVector<MachineOperand*, 8> ExtendedUses;
bool ExtendLife = true;
- UI = MRI->use_nodbg_begin(SrcReg);
- for (MachineRegisterInfo::use_nodbg_iterator UE = MRI->use_nodbg_end();
- UI != UE; ++UI) {
- MachineOperand &UseMO = UI.getOperand();
- MachineInstr *UseMI = &*UI;
+ for (MachineOperand &UseMO : MRI->use_nodbg_operands(SrcReg)) {
+ MachineInstr *UseMI = UseMO.getParent();
if (UseMI == MI)
continue;
continue;
}
+ // Only accept uses of SrcReg:SubIdx.
+ if (UseSrcSubIdx && UseMO.getSubReg() != SubIdx)
+ continue;
+
// It's an error to translate this:
//
// %reg1025 = <sext> %reg1024
if (ExtendLife && !ExtendedUses.empty())
// Extend the liveness of the extension result.
- std::copy(ExtendedUses.begin(), ExtendedUses.end(),
- std::back_inserter(Uses));
+ Uses.append(ExtendedUses.begin(), ExtendedUses.end());
// Now replace all uses.
bool Changed = false;
// Look for PHI uses of the extended result, we don't want to extend the
// liveness of a PHI input. It breaks all kinds of assumptions down
// stream. A PHI use is expected to be the kill of its source values.
- UI = MRI->use_nodbg_begin(DstReg);
- for (MachineRegisterInfo::use_nodbg_iterator
- UE = MRI->use_nodbg_end(); UI != UE; ++UI)
- if (UI->isPHI())
- PHIBBs.insert(UI->getParent());
+ for (MachineInstr &UI : MRI->use_nodbg_instructions(DstReg))
+ if (UI.isPHI())
+ PHIBBs.insert(UI.getParent());
const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
for (unsigned i = 0, e = Uses.size(); i != e; ++i) {
if (PHIBBs.count(UseMBB))
continue;
+ // About to add uses of DstReg, clear DstReg's kill flags.
+ if (!Changed) {
+ MRI->clearKillFlags(DstReg);
+ MRI->constrainRegClass(DstReg, DstRC);
+ }
+
unsigned NewVR = MRI->createVirtualRegister(RC);
- BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
- TII->get(TargetOpcode::COPY), NewVR)
+ MachineInstr *Copy = BuildMI(*UseMBB, UseMI, UseMI->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), NewVR)
.addReg(DstReg, 0, SubIdx);
-
+ // SubIdx applies to both SrcReg and DstReg when UseSrcSubIdx is set.
+ if (UseSrcSubIdx) {
+ Copy->getOperand(0).setSubReg(SubIdx);
+ Copy->getOperand(0).setIsUndef();
+ }
UseMO->setReg(NewVR);
++NumReuse;
Changed = true;
return Changed;
}
-/// OptimizeBitcastInstr - If the instruction is a bitcast instruction A that
-/// cannot be optimized away during isel (e.g. ARM::VMOVSR, which bitcast
-/// a value cross register classes), and the source is defined by another
-/// bitcast instruction B. And if the register class of source of B matches
-/// the register class of instruction A, then it is legal to replace all uses
-/// of the def of A with source of B. e.g.
-/// %vreg0<def> = VMOVSR %vreg1
-/// %vreg3<def> = VMOVRS %vreg0
-/// Replace all uses of vreg3 with vreg1.
-
-bool PeepholeOptimizer::OptimizeBitcastInstr(MachineInstr *MI,
- MachineBasicBlock *MBB) {
- unsigned NumDefs = MI->getDesc().getNumDefs();
- unsigned NumSrcs = MI->getDesc().getNumOperands() - NumDefs;
- if (NumDefs != 1)
+/// optimizeCmpInstr - If the instruction is a compare and the previous
+/// instruction it's comparing against all ready sets (or could be modified to
+/// set) the same flag as the compare, then we can remove the comparison and use
+/// the flag from the previous instruction.
+bool PeepholeOptimizer::optimizeCmpInstr(MachineInstr *MI,
+ MachineBasicBlock *MBB) {
+ // If this instruction is a comparison against zero and isn't comparing a
+ // physical register, we can try to optimize it.
+ unsigned SrcReg, SrcReg2;
+ int CmpMask, CmpValue;
+ if (!TII->analyzeCompare(MI, SrcReg, SrcReg2, CmpMask, CmpValue) ||
+ TargetRegisterInfo::isPhysicalRegister(SrcReg) ||
+ (SrcReg2 != 0 && TargetRegisterInfo::isPhysicalRegister(SrcReg2)))
return false;
- unsigned Def = 0;
- unsigned Src = 0;
- for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
- const MachineOperand &MO = MI->getOperand(i);
- if (!MO.isReg())
- continue;
- unsigned Reg = MO.getReg();
- if (!Reg)
+ // Attempt to optimize the comparison instruction.
+ if (TII->optimizeCompareInstr(MI, SrcReg, SrcReg2, CmpMask, CmpValue, MRI)) {
+ ++NumCmps;
+ return true;
+ }
+
+ return false;
+}
+
+/// Optimize a select instruction.
+bool PeepholeOptimizer::optimizeSelect(MachineInstr *MI,
+ SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
+ unsigned TrueOp = 0;
+ unsigned FalseOp = 0;
+ bool Optimizable = false;
+ SmallVector<MachineOperand, 4> Cond;
+ if (TII->analyzeSelect(MI, Cond, TrueOp, FalseOp, Optimizable))
+ return false;
+ if (!Optimizable)
+ return false;
+ if (!TII->optimizeSelect(MI, LocalMIs))
+ return false;
+ MI->eraseFromParent();
+ ++NumSelects;
+ return true;
+}
+
+/// \brief Check if a simpler conditional branch can be
+// generated
+bool PeepholeOptimizer::optimizeCondBranch(MachineInstr *MI) {
+ return TII->optimizeCondBranch(MI);
+}
+
+/// \brief Check if the registers defined by the pair (RegisterClass, SubReg)
+/// share the same register file.
+static bool shareSameRegisterFile(const TargetRegisterInfo &TRI,
+ const TargetRegisterClass *DefRC,
+ unsigned DefSubReg,
+ const TargetRegisterClass *SrcRC,
+ unsigned SrcSubReg) {
+ // Same register class.
+ if (DefRC == SrcRC)
+ return true;
+
+ // Both operands are sub registers. Check if they share a register class.
+ unsigned SrcIdx, DefIdx;
+ if (SrcSubReg && DefSubReg)
+ return TRI.getCommonSuperRegClass(SrcRC, SrcSubReg, DefRC, DefSubReg,
+ SrcIdx, DefIdx) != nullptr;
+ // At most one of the register is a sub register, make it Src to avoid
+ // duplicating the test.
+ if (!SrcSubReg) {
+ std::swap(DefSubReg, SrcSubReg);
+ std::swap(DefRC, SrcRC);
+ }
+
+ // One of the register is a sub register, check if we can get a superclass.
+ if (SrcSubReg)
+ return TRI.getMatchingSuperRegClass(SrcRC, DefRC, SrcSubReg) != nullptr;
+ // Plain copy.
+ return TRI.getCommonSubClass(DefRC, SrcRC) != nullptr;
+}
+
+/// \brief Try to find the next source that share the same register file
+/// for the value defined by \p Reg and \p SubReg.
+/// When true is returned, the \p RewriteMap can be used by the client to
+/// retrieve all Def -> Use along the way up to the next source. Any found
+/// Use that is not itself a key for another entry, is the next source to
+/// use. During the search for the next source, multiple sources can be found
+/// given multiple incoming sources of a PHI instruction. In this case, we
+/// look in each PHI source for the next source; all found next sources must
+/// share the same register file as \p Reg and \p SubReg. The client should
+/// then be capable to rewrite all intermediate PHIs to get the next source.
+/// \return False if no alternative sources are available. True otherwise.
+bool PeepholeOptimizer::findNextSource(unsigned Reg, unsigned SubReg,
+ RewriteMapTy &RewriteMap) {
+ // Do not try to find a new source for a physical register.
+ // So far we do not have any motivating example for doing that.
+ // Thus, instead of maintaining untested code, we will revisit that if
+ // that changes at some point.
+ if (TargetRegisterInfo::isPhysicalRegister(Reg))
+ return false;
+ const TargetRegisterClass *DefRC = MRI->getRegClass(Reg);
+
+ SmallVector<TargetInstrInfo::RegSubRegPair, 4> SrcToLook;
+ TargetInstrInfo::RegSubRegPair CurSrcPair(Reg, SubReg);
+ SrcToLook.push_back(CurSrcPair);
+
+ unsigned PHICount = 0;
+ while (!SrcToLook.empty() && PHICount < RewritePHILimit) {
+ TargetInstrInfo::RegSubRegPair Pair = SrcToLook.pop_back_val();
+ // As explained above, do not handle physical registers
+ if (TargetRegisterInfo::isPhysicalRegister(Pair.Reg))
+ return false;
+
+ CurSrcPair = Pair;
+ ValueTracker ValTracker(CurSrcPair.Reg, CurSrcPair.SubReg, *MRI,
+ !DisableAdvCopyOpt, TII);
+ ValueTrackerResult Res;
+ bool ShouldRewrite = false;
+
+ do {
+ // Follow the chain of copies until we reach the top of the use-def chain
+ // or find a more suitable source.
+ Res = ValTracker.getNextSource();
+ if (!Res.isValid())
+ break;
+
+ // Insert the Def -> Use entry for the recently found source.
+ ValueTrackerResult CurSrcRes = RewriteMap.lookup(CurSrcPair);
+ if (CurSrcRes.isValid()) {
+ assert(CurSrcRes == Res && "ValueTrackerResult found must match");
+ // An existent entry with multiple sources is a PHI cycle we must avoid.
+ // Otherwise it's an entry with a valid next source we already found.
+ if (CurSrcRes.getNumSources() > 1) {
+ DEBUG(dbgs() << "findNextSource: found PHI cycle, aborting...\n");
+ return false;
+ }
+ break;
+ }
+ RewriteMap.insert(std::make_pair(CurSrcPair, Res));
+
+ // ValueTrackerResult usually have one source unless it's the result from
+ // a PHI instruction. Add the found PHI edges to be looked up further.
+ unsigned NumSrcs = Res.getNumSources();
+ if (NumSrcs > 1) {
+ PHICount++;
+ for (unsigned i = 0; i < NumSrcs; ++i)
+ SrcToLook.push_back(TargetInstrInfo::RegSubRegPair(
+ Res.getSrcReg(i), Res.getSrcSubReg(i)));
+ break;
+ }
+
+ CurSrcPair.Reg = Res.getSrcReg(0);
+ CurSrcPair.SubReg = Res.getSrcSubReg(0);
+ // Do not extend the live-ranges of physical registers as they add
+ // constraints to the register allocator. Moreover, if we want to extend
+ // the live-range of a physical register, unlike SSA virtual register,
+ // we will have to check that they aren't redefine before the related use.
+ if (TargetRegisterInfo::isPhysicalRegister(CurSrcPair.Reg))
+ return false;
+
+ const TargetRegisterClass *SrcRC = MRI->getRegClass(CurSrcPair.Reg);
+
+ // If this source does not incur a cross register bank copy, use it.
+ ShouldRewrite = shareSameRegisterFile(*TRI, DefRC, SubReg, SrcRC,
+ CurSrcPair.SubReg);
+ } while (!ShouldRewrite);
+
+ // Continue looking for new sources...
+ if (Res.isValid())
continue;
- if (MO.isDef())
- Def = Reg;
- else if (Src)
- // Multiple sources?
+
+ // Do not continue searching for a new source if the there's at least
+ // one use-def which cannot be rewritten.
+ if (!ShouldRewrite)
return false;
- else
- Src = Reg;
}
- assert(Def && Src && "Malformed bitcast instruction!");
+ if (PHICount >= RewritePHILimit) {
+ DEBUG(dbgs() << "findNextSource: PHI limit reached\n");
+ return false;
+ }
- MachineInstr *DefMI = MRI->getVRegDef(Src);
- if (!DefMI || !DefMI->getDesc().isBitcast())
+ // If we did not find a more suitable source, there is nothing to optimize.
+ if (CurSrcPair.Reg == Reg)
return false;
- unsigned SrcDef = 0;
- unsigned SrcSrc = 0;
- NumDefs = DefMI->getDesc().getNumDefs();
- NumSrcs = DefMI->getDesc().getNumOperands() - NumDefs;
- if (NumDefs != 1)
+ return true;
+}
+
+/// \brief Insert a PHI instruction with incoming edges \p SrcRegs that are
+/// guaranteed to have the same register class. This is necessary whenever we
+/// successfully traverse a PHI instruction and find suitable sources coming
+/// from its edges. By inserting a new PHI, we provide a rewritten PHI def
+/// suitable to be used in a new COPY instruction.
+MachineInstr *
+insertPHI(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
+ const SmallVectorImpl<TargetInstrInfo::RegSubRegPair> &SrcRegs,
+ MachineInstr *OrigPHI) {
+ assert(!SrcRegs.empty() && "No sources to create a PHI instruction?");
+
+ const TargetRegisterClass *NewRC = MRI->getRegClass(SrcRegs[0].Reg);
+ unsigned NewVR = MRI->createVirtualRegister(NewRC);
+ MachineBasicBlock *MBB = OrigPHI->getParent();
+ MachineInstrBuilder MIB = BuildMI(*MBB, OrigPHI, OrigPHI->getDebugLoc(),
+ TII->get(TargetOpcode::PHI), NewVR);
+
+ unsigned MBBOpIdx = 2;
+ for (auto RegPair : SrcRegs) {
+ MIB.addReg(RegPair.Reg, 0, RegPair.SubReg);
+ MIB.addMBB(OrigPHI->getOperand(MBBOpIdx).getMBB());
+ // Since we're extended the lifetime of RegPair.Reg, clear the
+ // kill flags to account for that and make RegPair.Reg reaches
+ // the new PHI.
+ MRI->clearKillFlags(RegPair.Reg);
+ MBBOpIdx += 2;
+ }
+
+ return MIB;
+}
+
+namespace {
+/// \brief Helper class to rewrite the arguments of a copy-like instruction.
+class CopyRewriter {
+protected:
+ /// The copy-like instruction.
+ MachineInstr &CopyLike;
+ /// The index of the source being rewritten.
+ unsigned CurrentSrcIdx;
+
+public:
+ CopyRewriter(MachineInstr &MI) : CopyLike(MI), CurrentSrcIdx(0) {}
+
+ virtual ~CopyRewriter() {}
+
+ /// \brief Get the next rewritable source (SrcReg, SrcSubReg) and
+ /// the related value that it affects (TrackReg, TrackSubReg).
+ /// A source is considered rewritable if its register class and the
+ /// register class of the related TrackReg may not be register
+ /// coalescer friendly. In other words, given a copy-like instruction
+ /// not all the arguments may be returned at rewritable source, since
+ /// some arguments are none to be register coalescer friendly.
+ ///
+ /// Each call of this method moves the current source to the next
+ /// rewritable source.
+ /// For instance, let CopyLike be the instruction to rewrite.
+ /// CopyLike has one definition and one source:
+ /// dst.dstSubIdx = CopyLike src.srcSubIdx.
+ ///
+ /// The first call will give the first rewritable source, i.e.,
+ /// the only source this instruction has:
+ /// (SrcReg, SrcSubReg) = (src, srcSubIdx).
+ /// This source defines the whole definition, i.e.,
+ /// (TrackReg, TrackSubReg) = (dst, dstSubIdx).
+ ///
+ /// The second and subsequent calls will return false, has there is only one
+ /// rewritable source.
+ ///
+ /// \return True if a rewritable source has been found, false otherwise.
+ /// The output arguments are valid if and only if true is returned.
+ virtual bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
+ unsigned &TrackReg,
+ unsigned &TrackSubReg) {
+ // If CurrentSrcIdx == 1, this means this function has already been
+ // called once. CopyLike has one defintiion and one argument, thus,
+ // there is nothing else to rewrite.
+ if (!CopyLike.isCopy() || CurrentSrcIdx == 1)
+ return false;
+ // This is the first call to getNextRewritableSource.
+ // Move the CurrentSrcIdx to remember that we made that call.
+ CurrentSrcIdx = 1;
+ // The rewritable source is the argument.
+ const MachineOperand &MOSrc = CopyLike.getOperand(1);
+ SrcReg = MOSrc.getReg();
+ SrcSubReg = MOSrc.getSubReg();
+ // What we track are the alternative sources of the definition.
+ const MachineOperand &MODef = CopyLike.getOperand(0);
+ TrackReg = MODef.getReg();
+ TrackSubReg = MODef.getSubReg();
+ return true;
+ }
+
+ /// \brief Rewrite the current source with \p NewReg and \p NewSubReg
+ /// if possible.
+ /// \return True if the rewriting was possible, false otherwise.
+ virtual bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) {
+ if (!CopyLike.isCopy() || CurrentSrcIdx != 1)
+ return false;
+ MachineOperand &MOSrc = CopyLike.getOperand(CurrentSrcIdx);
+ MOSrc.setReg(NewReg);
+ MOSrc.setSubReg(NewSubReg);
+ return true;
+ }
+
+ /// \brief Given a \p Def.Reg and Def.SubReg pair, use \p RewriteMap to find
+ /// the new source to use for rewrite. If \p HandleMultipleSources is true and
+ /// multiple sources for a given \p Def are found along the way, we found a
+ /// PHI instructions that needs to be rewritten.
+ /// TODO: HandleMultipleSources should be removed once we test PHI handling
+ /// with coalescable copies.
+ TargetInstrInfo::RegSubRegPair
+ getNewSource(MachineRegisterInfo *MRI, const TargetInstrInfo *TII,
+ TargetInstrInfo::RegSubRegPair Def,
+ PeepholeOptimizer::RewriteMapTy &RewriteMap,
+ bool HandleMultipleSources = true) {
+
+ TargetInstrInfo::RegSubRegPair LookupSrc(Def.Reg, Def.SubReg);
+ do {
+ ValueTrackerResult Res = RewriteMap.lookup(LookupSrc);
+ // If there are no entries on the map, LookupSrc is the new source.
+ if (!Res.isValid())
+ return LookupSrc;
+
+ // There's only one source for this definition, keep searching...
+ unsigned NumSrcs = Res.getNumSources();
+ if (NumSrcs == 1) {
+ LookupSrc.Reg = Res.getSrcReg(0);
+ LookupSrc.SubReg = Res.getSrcSubReg(0);
+ continue;
+ }
+
+ // TODO: remove once multiple srcs w/ coaslescable copies are supported.
+ if (!HandleMultipleSources)
+ break;
+
+ // Multiple sources, recurse into each source to find a new source
+ // for it. Then, rewrite the PHI accordingly to its new edges.
+ SmallVector<TargetInstrInfo::RegSubRegPair, 4> NewPHISrcs;
+ for (unsigned i = 0; i < NumSrcs; ++i) {
+ TargetInstrInfo::RegSubRegPair PHISrc(Res.getSrcReg(i),
+ Res.getSrcSubReg(i));
+ NewPHISrcs.push_back(
+ getNewSource(MRI, TII, PHISrc, RewriteMap, HandleMultipleSources));
+ }
+
+ // Build the new PHI node and return its def register as the new source.
+ MachineInstr *OrigPHI = const_cast<MachineInstr *>(Res.getInst());
+ MachineInstr *NewPHI = insertPHI(MRI, TII, NewPHISrcs, OrigPHI);
+ DEBUG(dbgs() << "-- getNewSource\n");
+ DEBUG(dbgs() << " Replacing: " << *OrigPHI);
+ DEBUG(dbgs() << " With: " << *NewPHI);
+ const MachineOperand &MODef = NewPHI->getOperand(0);
+ return TargetInstrInfo::RegSubRegPair(MODef.getReg(), MODef.getSubReg());
+
+ } while (1);
+
+ return TargetInstrInfo::RegSubRegPair(0, 0);
+ }
+
+ /// \brief Rewrite the source found through \p Def, by using the \p RewriteMap
+ /// and create a new COPY instruction. More info about RewriteMap in
+ /// PeepholeOptimizer::findNextSource. Right now this is only used to handle
+ /// Uncoalescable copies, since they are copy like instructions that aren't
+ /// recognized by the register allocator.
+ virtual MachineInstr *
+ RewriteSource(TargetInstrInfo::RegSubRegPair Def,
+ PeepholeOptimizer::RewriteMapTy &RewriteMap) {
+ return nullptr;
+ }
+};
+
+/// \brief Helper class to rewrite uncoalescable copy like instructions
+/// into new COPY (coalescable friendly) instructions.
+class UncoalescableRewriter : public CopyRewriter {
+protected:
+ const TargetInstrInfo &TII;
+ MachineRegisterInfo &MRI;
+ /// The number of defs in the bitcast
+ unsigned NumDefs;
+
+public:
+ UncoalescableRewriter(MachineInstr &MI, const TargetInstrInfo &TII,
+ MachineRegisterInfo &MRI)
+ : CopyRewriter(MI), TII(TII), MRI(MRI) {
+ NumDefs = MI.getDesc().getNumDefs();
+ }
+
+ /// \brief Get the next rewritable def source (TrackReg, TrackSubReg)
+ /// All such sources need to be considered rewritable in order to
+ /// rewrite a uncoalescable copy-like instruction. This method return
+ /// each definition that must be checked if rewritable.
+ ///
+ bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
+ unsigned &TrackReg,
+ unsigned &TrackSubReg) override {
+ // Find the next non-dead definition and continue from there.
+ if (CurrentSrcIdx == NumDefs)
+ return false;
+
+ while (CopyLike.getOperand(CurrentSrcIdx).isDead()) {
+ ++CurrentSrcIdx;
+ if (CurrentSrcIdx == NumDefs)
+ return false;
+ }
+
+ // What we track are the alternative sources of the definition.
+ const MachineOperand &MODef = CopyLike.getOperand(CurrentSrcIdx);
+ TrackReg = MODef.getReg();
+ TrackSubReg = MODef.getSubReg();
+
+ CurrentSrcIdx++;
+ return true;
+ }
+
+ /// \brief Rewrite the source found through \p Def, by using the \p RewriteMap
+ /// and create a new COPY instruction. More info about RewriteMap in
+ /// PeepholeOptimizer::findNextSource. Right now this is only used to handle
+ /// Uncoalescable copies, since they are copy like instructions that aren't
+ /// recognized by the register allocator.
+ MachineInstr *
+ RewriteSource(TargetInstrInfo::RegSubRegPair Def,
+ PeepholeOptimizer::RewriteMapTy &RewriteMap) override {
+ assert(!TargetRegisterInfo::isPhysicalRegister(Def.Reg) &&
+ "We do not rewrite physical registers");
+
+ // Find the new source to use in the COPY rewrite.
+ TargetInstrInfo::RegSubRegPair NewSrc =
+ getNewSource(&MRI, &TII, Def, RewriteMap);
+
+ // Insert the COPY.
+ const TargetRegisterClass *DefRC = MRI.getRegClass(Def.Reg);
+ unsigned NewVR = MRI.createVirtualRegister(DefRC);
+
+ MachineInstr *NewCopy =
+ BuildMI(*CopyLike.getParent(), &CopyLike, CopyLike.getDebugLoc(),
+ TII.get(TargetOpcode::COPY), NewVR)
+ .addReg(NewSrc.Reg, 0, NewSrc.SubReg);
+
+ NewCopy->getOperand(0).setSubReg(Def.SubReg);
+ if (Def.SubReg)
+ NewCopy->getOperand(0).setIsUndef();
+
+ DEBUG(dbgs() << "-- RewriteSource\n");
+ DEBUG(dbgs() << " Replacing: " << CopyLike);
+ DEBUG(dbgs() << " With: " << *NewCopy);
+ MRI.replaceRegWith(Def.Reg, NewVR);
+ MRI.clearKillFlags(NewVR);
+
+ // We extended the lifetime of NewSrc.Reg, clear the kill flags to
+ // account for that.
+ MRI.clearKillFlags(NewSrc.Reg);
+
+ return NewCopy;
+ }
+};
+
+/// \brief Specialized rewriter for INSERT_SUBREG instruction.
+class InsertSubregRewriter : public CopyRewriter {
+public:
+ InsertSubregRewriter(MachineInstr &MI) : CopyRewriter(MI) {
+ assert(MI.isInsertSubreg() && "Invalid instruction");
+ }
+
+ /// \brief See CopyRewriter::getNextRewritableSource.
+ /// Here CopyLike has the following form:
+ /// dst = INSERT_SUBREG Src1, Src2.src2SubIdx, subIdx.
+ /// Src1 has the same register class has dst, hence, there is
+ /// nothing to rewrite.
+ /// Src2.src2SubIdx, may not be register coalescer friendly.
+ /// Therefore, the first call to this method returns:
+ /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
+ /// (TrackReg, TrackSubReg) = (dst, subIdx).
+ ///
+ /// Subsequence calls will return false.
+ bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
+ unsigned &TrackReg,
+ unsigned &TrackSubReg) override {
+ // If we already get the only source we can rewrite, return false.
+ if (CurrentSrcIdx == 2)
+ return false;
+ // We are looking at v2 = INSERT_SUBREG v0, v1, sub0.
+ CurrentSrcIdx = 2;
+ const MachineOperand &MOInsertedReg = CopyLike.getOperand(2);
+ SrcReg = MOInsertedReg.getReg();
+ SrcSubReg = MOInsertedReg.getSubReg();
+ const MachineOperand &MODef = CopyLike.getOperand(0);
+
+ // We want to track something that is compatible with the
+ // partial definition.
+ TrackReg = MODef.getReg();
+ if (MODef.getSubReg())
+ // Bails if we have to compose sub-register indices.
+ return false;
+ TrackSubReg = (unsigned)CopyLike.getOperand(3).getImm();
+ return true;
+ }
+ bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
+ if (CurrentSrcIdx != 2)
+ return false;
+ // We are rewriting the inserted reg.
+ MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
+ MO.setReg(NewReg);
+ MO.setSubReg(NewSubReg);
+ return true;
+ }
+};
+
+/// \brief Specialized rewriter for EXTRACT_SUBREG instruction.
+class ExtractSubregRewriter : public CopyRewriter {
+ const TargetInstrInfo &TII;
+
+public:
+ ExtractSubregRewriter(MachineInstr &MI, const TargetInstrInfo &TII)
+ : CopyRewriter(MI), TII(TII) {
+ assert(MI.isExtractSubreg() && "Invalid instruction");
+ }
+
+ /// \brief See CopyRewriter::getNextRewritableSource.
+ /// Here CopyLike has the following form:
+ /// dst.dstSubIdx = EXTRACT_SUBREG Src, subIdx.
+ /// There is only one rewritable source: Src.subIdx,
+ /// which defines dst.dstSubIdx.
+ bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
+ unsigned &TrackReg,
+ unsigned &TrackSubReg) override {
+ // If we already get the only source we can rewrite, return false.
+ if (CurrentSrcIdx == 1)
+ return false;
+ // We are looking at v1 = EXTRACT_SUBREG v0, sub0.
+ CurrentSrcIdx = 1;
+ const MachineOperand &MOExtractedReg = CopyLike.getOperand(1);
+ SrcReg = MOExtractedReg.getReg();
+ // If we have to compose sub-register indices, bails out.
+ if (MOExtractedReg.getSubReg())
+ return false;
+
+ SrcSubReg = CopyLike.getOperand(2).getImm();
+
+ // We want to track something that is compatible with the definition.
+ const MachineOperand &MODef = CopyLike.getOperand(0);
+ TrackReg = MODef.getReg();
+ TrackSubReg = MODef.getSubReg();
+ return true;
+ }
+
+ bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
+ // The only source we can rewrite is the input register.
+ if (CurrentSrcIdx != 1)
+ return false;
+
+ CopyLike.getOperand(CurrentSrcIdx).setReg(NewReg);
+
+ // If we find a source that does not require to extract something,
+ // rewrite the operation with a copy.
+ if (!NewSubReg) {
+ // Move the current index to an invalid position.
+ // We do not want another call to this method to be able
+ // to do any change.
+ CurrentSrcIdx = -1;
+ // Rewrite the operation as a COPY.
+ // Get rid of the sub-register index.
+ CopyLike.RemoveOperand(2);
+ // Morph the operation into a COPY.
+ CopyLike.setDesc(TII.get(TargetOpcode::COPY));
+ return true;
+ }
+ CopyLike.getOperand(CurrentSrcIdx + 1).setImm(NewSubReg);
+ return true;
+ }
+};
+
+/// \brief Specialized rewriter for REG_SEQUENCE instruction.
+class RegSequenceRewriter : public CopyRewriter {
+public:
+ RegSequenceRewriter(MachineInstr &MI) : CopyRewriter(MI) {
+ assert(MI.isRegSequence() && "Invalid instruction");
+ }
+
+ /// \brief See CopyRewriter::getNextRewritableSource.
+ /// Here CopyLike has the following form:
+ /// dst = REG_SEQUENCE Src1.src1SubIdx, subIdx1, Src2.src2SubIdx, subIdx2.
+ /// Each call will return a different source, walking all the available
+ /// source.
+ ///
+ /// The first call returns:
+ /// (SrcReg, SrcSubReg) = (Src1, src1SubIdx).
+ /// (TrackReg, TrackSubReg) = (dst, subIdx1).
+ ///
+ /// The second call returns:
+ /// (SrcReg, SrcSubReg) = (Src2, src2SubIdx).
+ /// (TrackReg, TrackSubReg) = (dst, subIdx2).
+ ///
+ /// And so on, until all the sources have been traversed, then
+ /// it returns false.
+ bool getNextRewritableSource(unsigned &SrcReg, unsigned &SrcSubReg,
+ unsigned &TrackReg,
+ unsigned &TrackSubReg) override {
+ // We are looking at v0 = REG_SEQUENCE v1, sub1, v2, sub2, etc.
+
+ // If this is the first call, move to the first argument.
+ if (CurrentSrcIdx == 0) {
+ CurrentSrcIdx = 1;
+ } else {
+ // Otherwise, move to the next argument and check that it is valid.
+ CurrentSrcIdx += 2;
+ if (CurrentSrcIdx >= CopyLike.getNumOperands())
+ return false;
+ }
+ const MachineOperand &MOInsertedReg = CopyLike.getOperand(CurrentSrcIdx);
+ SrcReg = MOInsertedReg.getReg();
+ // If we have to compose sub-register indices, bails out.
+ if ((SrcSubReg = MOInsertedReg.getSubReg()))
+ return false;
+
+ // We want to track something that is compatible with the related
+ // partial definition.
+ TrackSubReg = CopyLike.getOperand(CurrentSrcIdx + 1).getImm();
+
+ const MachineOperand &MODef = CopyLike.getOperand(0);
+ TrackReg = MODef.getReg();
+ // If we have to compose sub-registers, bails.
+ return MODef.getSubReg() == 0;
+ }
+
+ bool RewriteCurrentSource(unsigned NewReg, unsigned NewSubReg) override {
+ // We cannot rewrite out of bound operands.
+ // Moreover, rewritable sources are at odd positions.
+ if ((CurrentSrcIdx & 1) != 1 || CurrentSrcIdx > CopyLike.getNumOperands())
+ return false;
+
+ MachineOperand &MO = CopyLike.getOperand(CurrentSrcIdx);
+ MO.setReg(NewReg);
+ MO.setSubReg(NewSubReg);
+ return true;
+ }
+};
+} // End namespace.
+
+/// \brief Get the appropriated CopyRewriter for \p MI.
+/// \return A pointer to a dynamically allocated CopyRewriter or nullptr
+/// if no rewriter works for \p MI.
+static CopyRewriter *getCopyRewriter(MachineInstr &MI,
+ const TargetInstrInfo &TII,
+ MachineRegisterInfo &MRI) {
+ // Handle uncoalescable copy-like instructions.
+ if (MI.isBitcast() || (MI.isRegSequenceLike() || MI.isInsertSubregLike() ||
+ MI.isExtractSubregLike()))
+ return new UncoalescableRewriter(MI, TII, MRI);
+
+ switch (MI.getOpcode()) {
+ default:
+ return nullptr;
+ case TargetOpcode::COPY:
+ return new CopyRewriter(MI);
+ case TargetOpcode::INSERT_SUBREG:
+ return new InsertSubregRewriter(MI);
+ case TargetOpcode::EXTRACT_SUBREG:
+ return new ExtractSubregRewriter(MI, TII);
+ case TargetOpcode::REG_SEQUENCE:
+ return new RegSequenceRewriter(MI);
+ }
+ llvm_unreachable(nullptr);
+}
+
+/// \brief Optimize generic copy instructions to avoid cross
+/// register bank copy. The optimization looks through a chain of
+/// copies and tries to find a source that has a compatible register
+/// class.
+/// Two register classes are considered to be compatible if they share
+/// the same register bank.
+/// New copies issued by this optimization are register allocator
+/// friendly. This optimization does not remove any copy as it may
+/// overconstraint the register allocator, but replaces some operands
+/// when possible.
+/// \pre isCoalescableCopy(*MI) is true.
+/// \return True, when \p MI has been rewritten. False otherwise.
+bool PeepholeOptimizer::optimizeCoalescableCopy(MachineInstr *MI) {
+ assert(MI && isCoalescableCopy(*MI) && "Invalid argument");
+ assert(MI->getDesc().getNumDefs() == 1 &&
+ "Coalescer can understand multiple defs?!");
+ const MachineOperand &MODef = MI->getOperand(0);
+ // Do not rewrite physical definitions.
+ if (TargetRegisterInfo::isPhysicalRegister(MODef.getReg()))
return false;
- for (unsigned i = 0, e = NumDefs + NumSrcs; i != e; ++i) {
- const MachineOperand &MO = DefMI->getOperand(i);
- if (!MO.isReg() || MO.isDef())
+
+ bool Changed = false;
+ // Get the right rewriter for the current copy.
+ std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
+ // If none exists, bails out.
+ if (!CpyRewriter)
+ return false;
+ // Rewrite each rewritable source.
+ unsigned SrcReg, SrcSubReg, TrackReg, TrackSubReg;
+ while (CpyRewriter->getNextRewritableSource(SrcReg, SrcSubReg, TrackReg,
+ TrackSubReg)) {
+ // Keep track of PHI nodes and its incoming edges when looking for sources.
+ RewriteMapTy RewriteMap;
+ // Try to find a more suitable source. If we failed to do so, or get the
+ // actual source, move to the next source.
+ if (!findNextSource(TrackReg, TrackSubReg, RewriteMap))
continue;
- unsigned Reg = MO.getReg();
- if (!Reg)
+
+ // Get the new source to rewrite. TODO: Only enable handling of multiple
+ // sources (PHIs) once we have a motivating example and testcases for it.
+ TargetInstrInfo::RegSubRegPair TrackPair(TrackReg, TrackSubReg);
+ TargetInstrInfo::RegSubRegPair NewSrc = CpyRewriter->getNewSource(
+ MRI, TII, TrackPair, RewriteMap, false /* multiple sources */);
+ if (SrcReg == NewSrc.Reg || NewSrc.Reg == 0)
continue;
- if (MO.isDef())
- SrcDef = Reg;
- else if (SrcSrc)
- // Multiple sources?
- return false;
- else
- SrcSrc = Reg;
+
+ // Rewrite source.
+ if (CpyRewriter->RewriteCurrentSource(NewSrc.Reg, NewSrc.SubReg)) {
+ // We may have extended the live-range of NewSrc, account for that.
+ MRI->clearKillFlags(NewSrc.Reg);
+ Changed = true;
+ }
}
+ // TODO: We could have a clean-up method to tidy the instruction.
+ // E.g., v0 = INSERT_SUBREG v1, v1.sub0, sub0
+ // => v0 = COPY v1
+ // Currently we haven't seen motivating example for that and we
+ // want to avoid untested code.
+ NumRewrittenCopies += Changed;
+ return Changed;
+}
- if (MRI->getRegClass(SrcSrc) != MRI->getRegClass(Def))
+/// \brief Optimize copy-like instructions to create
+/// register coalescer friendly instruction.
+/// The optimization tries to kill-off the \p MI by looking
+/// through a chain of copies to find a source that has a compatible
+/// register class.
+/// If such a source is found, it replace \p MI by a generic COPY
+/// operation.
+/// \pre isUncoalescableCopy(*MI) is true.
+/// \return True, when \p MI has been optimized. In that case, \p MI has
+/// been removed from its parent.
+/// All COPY instructions created, are inserted in \p LocalMIs.
+bool PeepholeOptimizer::optimizeUncoalescableCopy(
+ MachineInstr *MI, SmallPtrSetImpl<MachineInstr *> &LocalMIs) {
+ assert(MI && isUncoalescableCopy(*MI) && "Invalid argument");
+
+ // Check if we can rewrite all the values defined by this instruction.
+ SmallVector<TargetInstrInfo::RegSubRegPair, 4> RewritePairs;
+ // Get the right rewriter for the current copy.
+ std::unique_ptr<CopyRewriter> CpyRewriter(getCopyRewriter(*MI, *TII, *MRI));
+ // If none exists, bails out.
+ if (!CpyRewriter)
return false;
- MRI->replaceRegWith(Def, SrcSrc);
- MRI->clearKillFlags(SrcSrc);
+ // Rewrite each rewritable source by generating new COPYs. This works
+ // differently from optimizeCoalescableCopy since it first makes sure that all
+ // definitions can be rewritten.
+ RewriteMapTy RewriteMap;
+ unsigned Reg, SubReg, CopyDefReg, CopyDefSubReg;
+ while (CpyRewriter->getNextRewritableSource(Reg, SubReg, CopyDefReg,
+ CopyDefSubReg)) {
+ // If a physical register is here, this is probably for a good reason.
+ // Do not rewrite that.
+ if (TargetRegisterInfo::isPhysicalRegister(CopyDefReg))
+ return false;
+
+ // If we do not know how to rewrite this definition, there is no point
+ // in trying to kill this instruction.
+ TargetInstrInfo::RegSubRegPair Def(CopyDefReg, CopyDefSubReg);
+ if (!findNextSource(Def.Reg, Def.SubReg, RewriteMap))
+ return false;
+
+ RewritePairs.push_back(Def);
+ }
+
+ // The change is possible for all defs, do it.
+ for (const auto &Def : RewritePairs) {
+ // Rewrite the "copy" in a way the register coalescer understands.
+ MachineInstr *NewCopy = CpyRewriter->RewriteSource(Def, RewriteMap);
+ assert(NewCopy && "Should be able to always generate a new copy");
+ LocalMIs.insert(NewCopy);
+ }
+
+ // MI is now dead.
MI->eraseFromParent();
- ++NumBitcasts;
+ ++NumUncoalescableCopies;
return true;
}
-/// OptimizeCmpInstr - If the instruction is a compare and the previous
-/// instruction it's comparing against all ready sets (or could be modified to
-/// set) the same flag as the compare, then we can remove the comparison and use
-/// the flag from the previous instruction.
-bool PeepholeOptimizer::OptimizeCmpInstr(MachineInstr *MI,
- MachineBasicBlock *MBB) {
- // If this instruction is a comparison against zero and isn't comparing a
- // physical register, we can try to optimize it.
- unsigned SrcReg;
- int CmpMask, CmpValue;
- if (!TII->AnalyzeCompare(MI, SrcReg, CmpMask, CmpValue) ||
- TargetRegisterInfo::isPhysicalRegister(SrcReg))
+/// isLoadFoldable - Check whether MI is a candidate for folding into a later
+/// instruction. We only fold loads to virtual registers and the virtual
+/// register defined has a single use.
+bool PeepholeOptimizer::isLoadFoldable(
+ MachineInstr *MI,
+ SmallSet<unsigned, 16> &FoldAsLoadDefCandidates) {
+ if (!MI->canFoldAsLoad() || !MI->mayLoad())
+ return false;
+ const MCInstrDesc &MCID = MI->getDesc();
+ if (MCID.getNumDefs() != 1)
return false;
- // Attempt to optimize the comparison instruction.
- if (TII->OptimizeCompareInstr(MI, SrcReg, CmpMask, CmpValue, MRI)) {
- ++NumCmps;
+ unsigned Reg = MI->getOperand(0).getReg();
+ // To reduce compilation time, we check MRI->hasOneNonDBGUse when inserting
+ // loads. It should be checked when processing uses of the load, since
+ // uses can be removed during peephole.
+ if (!MI->getOperand(0).getSubReg() &&
+ TargetRegisterInfo::isVirtualRegister(Reg) &&
+ MRI->hasOneNonDBGUse(Reg)) {
+ FoldAsLoadDefCandidates.insert(Reg);
return true;
}
-
return false;
}
bool PeepholeOptimizer::isMoveImmediate(MachineInstr *MI,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
- const TargetInstrDesc &TID = MI->getDesc();
- if (!TID.isMoveImmediate())
+ const MCInstrDesc &MCID = MI->getDesc();
+ if (!MI->isMoveImmediate())
return false;
- if (TID.getNumDefs() != 1)
+ if (MCID.getNumDefs() != 1)
return false;
unsigned Reg = MI->getOperand(0).getReg();
if (TargetRegisterInfo::isVirtualRegister(Reg)) {
ImmDefRegs.insert(Reg);
return true;
}
-
+
return false;
}
-/// FoldImmediate - Try folding register operands that are defined by move
+/// foldImmediate - Try folding register operands that are defined by move
/// immediate instructions, i.e. a trivial constant folding optimization, if
/// and only if the def and use are in the same BB.
-bool PeepholeOptimizer::FoldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
+bool PeepholeOptimizer::foldImmediate(MachineInstr *MI, MachineBasicBlock *MBB,
SmallSet<unsigned, 4> &ImmDefRegs,
DenseMap<unsigned, MachineInstr*> &ImmDefMIs) {
for (unsigned i = 0, e = MI->getDesc().getNumOperands(); i != e; ++i) {
}
bool PeepholeOptimizer::runOnMachineFunction(MachineFunction &MF) {
+ if (skipOptnoneFunction(*MF.getFunction()))
+ return false;
+
+ DEBUG(dbgs() << "********** PEEPHOLE OPTIMIZER **********\n");
+ DEBUG(dbgs() << "********** Function: " << MF.getName() << '\n');
+
if (DisablePeephole)
return false;
-
- TM = &MF.getTarget();
- TII = TM->getInstrInfo();
+
+ TII = MF.getSubtarget().getInstrInfo();
+ TRI = MF.getSubtarget().getRegisterInfo();
MRI = &MF.getRegInfo();
- DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : 0;
+ DT = Aggressive ? &getAnalysis<MachineDominatorTree>() : nullptr;
bool Changed = false;
- SmallPtrSet<MachineInstr*, 8> LocalMIs;
- SmallSet<unsigned, 4> ImmDefRegs;
- DenseMap<unsigned, MachineInstr*> ImmDefMIs;
for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) {
MachineBasicBlock *MBB = &*I;
-
+
bool SeenMoveImm = false;
- LocalMIs.clear();
- ImmDefRegs.clear();
- ImmDefMIs.clear();
- bool First = true;
- MachineBasicBlock::iterator PMII;
+ // During this forward scan, at some point it needs to answer the question
+ // "given a pointer to an MI in the current BB, is it located before or
+ // after the current instruction".
+ // To perform this, the following set keeps track of the MIs already seen
+ // during the scan, if a MI is not in the set, it is assumed to be located
+ // after. Newly created MIs have to be inserted in the set as well.
+ SmallPtrSet<MachineInstr*, 16> LocalMIs;
+ SmallSet<unsigned, 4> ImmDefRegs;
+ DenseMap<unsigned, MachineInstr*> ImmDefMIs;
+ SmallSet<unsigned, 16> FoldAsLoadDefCandidates;
+
for (MachineBasicBlock::iterator
MII = I->begin(), MIE = I->end(); MII != MIE; ) {
MachineInstr *MI = &*MII;
+ // We may be erasing MI below, increment MII now.
+ ++MII;
LocalMIs.insert(MI);
- if (MI->isLabel() || MI->isPHI() || MI->isImplicitDef() ||
- MI->isKill() || MI->isInlineAsm() || MI->isDebugValue() ||
- MI->hasUnmodeledSideEffects()) {
- ++MII;
+ // Skip debug values. They should not affect this peephole optimization.
+ if (MI->isDebugValue())
+ continue;
+
+ // If we run into an instruction we can't fold across, discard
+ // the load candidates.
+ if (MI->isLoadFoldBarrier())
+ FoldAsLoadDefCandidates.clear();
+
+ if (MI->isPosition() || MI->isPHI() || MI->isImplicitDef() ||
+ MI->isKill() || MI->isInlineAsm() ||
+ MI->hasUnmodeledSideEffects())
+ continue;
+
+ if ((isUncoalescableCopy(*MI) &&
+ optimizeUncoalescableCopy(MI, LocalMIs)) ||
+ (MI->isCompare() && optimizeCmpInstr(MI, MBB)) ||
+ (MI->isSelect() && optimizeSelect(MI, LocalMIs))) {
+ // MI is deleted.
+ LocalMIs.erase(MI);
+ Changed = true;
continue;
}
- const TargetInstrDesc &TID = MI->getDesc();
+ if (MI->isConditionalBranch() && optimizeCondBranch(MI)) {
+ Changed = true;
+ continue;
+ }
- if (TID.isBitcast()) {
- if (OptimizeBitcastInstr(MI, MBB)) {
- // MI is deleted.
- Changed = true;
- MII = First ? I->begin() : llvm::next(PMII);
- continue;
- }
- } else if (TID.isCompare()) {
- if (OptimizeCmpInstr(MI, MBB)) {
- // MI is deleted.
- Changed = true;
- MII = First ? I->begin() : llvm::next(PMII);
- continue;
- }
+ if (isCoalescableCopy(*MI) && optimizeCoalescableCopy(MI)) {
+ // MI is just rewritten.
+ Changed = true;
+ continue;
}
if (isMoveImmediate(MI, ImmDefRegs, ImmDefMIs)) {
SeenMoveImm = true;
} else {
- Changed |= OptimizeExtInstr(MI, MBB, LocalMIs);
+ Changed |= optimizeExtInstr(MI, MBB, LocalMIs);
+ // optimizeExtInstr might have created new instructions after MI
+ // and before the already incremented MII. Adjust MII so that the
+ // next iteration sees the new instructions.
+ MII = MI;
+ ++MII;
if (SeenMoveImm)
- Changed |= FoldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
+ Changed |= foldImmediate(MI, MBB, ImmDefRegs, ImmDefMIs);
}
- First = false;
- PMII = MII;
- ++MII;
+ // Check whether MI is a load candidate for folding into a later
+ // instruction. If MI is not a candidate, check whether we can fold an
+ // earlier load into MI.
+ if (!isLoadFoldable(MI, FoldAsLoadDefCandidates) &&
+ !FoldAsLoadDefCandidates.empty()) {
+ const MCInstrDesc &MIDesc = MI->getDesc();
+ for (unsigned i = MIDesc.getNumDefs(); i != MIDesc.getNumOperands();
+ ++i) {
+ const MachineOperand &MOp = MI->getOperand(i);
+ if (!MOp.isReg())
+ continue;
+ unsigned FoldAsLoadDefReg = MOp.getReg();
+ if (FoldAsLoadDefCandidates.count(FoldAsLoadDefReg)) {
+ // We need to fold load after optimizeCmpInstr, since
+ // optimizeCmpInstr can enable folding by converting SUB to CMP.
+ // Save FoldAsLoadDefReg because optimizeLoadInstr() resets it and
+ // we need it for markUsesInDebugValueAsUndef().
+ unsigned FoldedReg = FoldAsLoadDefReg;
+ MachineInstr *DefMI = nullptr;
+ MachineInstr *FoldMI = TII->optimizeLoadInstr(MI, MRI,
+ FoldAsLoadDefReg,
+ DefMI);
+ if (FoldMI) {
+ // Update LocalMIs since we replaced MI with FoldMI and deleted
+ // DefMI.
+ DEBUG(dbgs() << "Replacing: " << *MI);
+ DEBUG(dbgs() << " With: " << *FoldMI);
+ LocalMIs.erase(MI);
+ LocalMIs.erase(DefMI);
+ LocalMIs.insert(FoldMI);
+ MI->eraseFromParent();
+ DefMI->eraseFromParent();
+ MRI->markUsesInDebugValueAsUndef(FoldedReg);
+ FoldAsLoadDefCandidates.erase(FoldedReg);
+ ++NumLoadFold;
+ // MI is replaced with FoldMI.
+ Changed = true;
+ break;
+ }
+ }
+ }
+ }
}
}
return Changed;
}
+
+ValueTrackerResult ValueTracker::getNextSourceFromCopy() {
+ assert(Def->isCopy() && "Invalid definition");
+ // Copy instruction are supposed to be: Def = Src.
+ // If someone breaks this assumption, bad things will happen everywhere.
+ assert(Def->getNumOperands() == 2 && "Invalid number of operands");
+
+ if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
+ // If we look for a different subreg, it means we want a subreg of src.
+ // Bails as we do not support composing subreg yet.
+ return ValueTrackerResult();
+ // Otherwise, we want the whole source.
+ const MachineOperand &Src = Def->getOperand(1);
+ return ValueTrackerResult(Src.getReg(), Src.getSubReg());
+}
+
+ValueTrackerResult ValueTracker::getNextSourceFromBitcast() {
+ assert(Def->isBitcast() && "Invalid definition");
+
+ // Bail if there are effects that a plain copy will not expose.
+ if (Def->hasUnmodeledSideEffects())
+ return ValueTrackerResult();
+
+ // Bitcasts with more than one def are not supported.
+ if (Def->getDesc().getNumDefs() != 1)
+ return ValueTrackerResult();
+ if (Def->getOperand(DefIdx).getSubReg() != DefSubReg)
+ // If we look for a different subreg, it means we want a subreg of the src.
+ // Bails as we do not support composing subreg yet.
+ return ValueTrackerResult();
+
+ unsigned SrcIdx = Def->getNumOperands();
+ for (unsigned OpIdx = DefIdx + 1, EndOpIdx = SrcIdx; OpIdx != EndOpIdx;
+ ++OpIdx) {
+ const MachineOperand &MO = Def->getOperand(OpIdx);
+ if (!MO.isReg() || !MO.getReg())
+ continue;
+ assert(!MO.isDef() && "We should have skipped all the definitions by now");
+ if (SrcIdx != EndOpIdx)
+ // Multiple sources?
+ return ValueTrackerResult();
+ SrcIdx = OpIdx;
+ }
+ const MachineOperand &Src = Def->getOperand(SrcIdx);
+ return ValueTrackerResult(Src.getReg(), Src.getSubReg());
+}
+
+ValueTrackerResult ValueTracker::getNextSourceFromRegSequence() {
+ assert((Def->isRegSequence() || Def->isRegSequenceLike()) &&
+ "Invalid definition");
+
+ if (Def->getOperand(DefIdx).getSubReg())
+ // If we are composing subreg, bails out.
+ // The case we are checking is Def.<subreg> = REG_SEQUENCE.
+ // This should almost never happen as the SSA property is tracked at
+ // the register level (as opposed to the subreg level).
+ // I.e.,
+ // Def.sub0 =
+ // Def.sub1 =
+ // is a valid SSA representation for Def.sub0 and Def.sub1, but not for
+ // Def. Thus, it must not be generated.
+ // However, some code could theoretically generates a single
+ // Def.sub0 (i.e, not defining the other subregs) and we would
+ // have this case.
+ // If we can ascertain (or force) that this never happens, we could
+ // turn that into an assertion.
+ return ValueTrackerResult();
+
+ if (!TII)
+ // We could handle the REG_SEQUENCE here, but we do not want to
+ // duplicate the code from the generic TII.
+ return ValueTrackerResult();
+
+ SmallVector<TargetInstrInfo::RegSubRegPairAndIdx, 8> RegSeqInputRegs;
+ if (!TII->getRegSequenceInputs(*Def, DefIdx, RegSeqInputRegs))
+ return ValueTrackerResult();
+
+ // We are looking at:
+ // Def = REG_SEQUENCE v0, sub0, v1, sub1, ...
+ // Check if one of the operand defines the subreg we are interested in.
+ for (auto &RegSeqInput : RegSeqInputRegs) {
+ if (RegSeqInput.SubIdx == DefSubReg) {
+ if (RegSeqInput.SubReg)
+ // Bails if we have to compose sub registers.
+ return ValueTrackerResult();
+
+ return ValueTrackerResult(RegSeqInput.Reg, RegSeqInput.SubReg);
+ }
+ }
+
+ // If the subreg we are tracking is super-defined by another subreg,
+ // we could follow this value. However, this would require to compose
+ // the subreg and we do not do that for now.
+ return ValueTrackerResult();
+}
+
+ValueTrackerResult ValueTracker::getNextSourceFromInsertSubreg() {
+ assert((Def->isInsertSubreg() || Def->isInsertSubregLike()) &&
+ "Invalid definition");
+
+ if (Def->getOperand(DefIdx).getSubReg())
+ // If we are composing subreg, bails out.
+ // Same remark as getNextSourceFromRegSequence.
+ // I.e., this may be turned into an assert.
+ return ValueTrackerResult();
+
+ if (!TII)
+ // We could handle the REG_SEQUENCE here, but we do not want to
+ // duplicate the code from the generic TII.
+ return ValueTrackerResult();
+
+ TargetInstrInfo::RegSubRegPair BaseReg;
+ TargetInstrInfo::RegSubRegPairAndIdx InsertedReg;
+ if (!TII->getInsertSubregInputs(*Def, DefIdx, BaseReg, InsertedReg))
+ return ValueTrackerResult();
+
+ // We are looking at:
+ // Def = INSERT_SUBREG v0, v1, sub1
+ // There are two cases:
+ // 1. DefSubReg == sub1, get v1.
+ // 2. DefSubReg != sub1, the value may be available through v0.
+
+ // #1 Check if the inserted register matches the required sub index.
+ if (InsertedReg.SubIdx == DefSubReg) {
+ return ValueTrackerResult(InsertedReg.Reg, InsertedReg.SubReg);
+ }
+ // #2 Otherwise, if the sub register we are looking for is not partial
+ // defined by the inserted element, we can look through the main
+ // register (v0).
+ const MachineOperand &MODef = Def->getOperand(DefIdx);
+ // If the result register (Def) and the base register (v0) do not
+ // have the same register class or if we have to compose
+ // subregisters, bails out.
+ if (MRI.getRegClass(MODef.getReg()) != MRI.getRegClass(BaseReg.Reg) ||
+ BaseReg.SubReg)
+ return ValueTrackerResult();
+
+ // Get the TRI and check if the inserted sub-register overlaps with the
+ // sub-register we are tracking.
+ const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo();
+ if (!TRI ||
+ (TRI->getSubRegIndexLaneMask(DefSubReg) &
+ TRI->getSubRegIndexLaneMask(InsertedReg.SubIdx)) != 0)
+ return ValueTrackerResult();
+ // At this point, the value is available in v0 via the same subreg
+ // we used for Def.
+ return ValueTrackerResult(BaseReg.Reg, DefSubReg);
+}
+
+ValueTrackerResult ValueTracker::getNextSourceFromExtractSubreg() {
+ assert((Def->isExtractSubreg() ||
+ Def->isExtractSubregLike()) && "Invalid definition");
+ // We are looking at:
+ // Def = EXTRACT_SUBREG v0, sub0
+
+ // Bails if we have to compose sub registers.
+ // Indeed, if DefSubReg != 0, we would have to compose it with sub0.
+ if (DefSubReg)
+ return ValueTrackerResult();
+
+ if (!TII)
+ // We could handle the EXTRACT_SUBREG here, but we do not want to
+ // duplicate the code from the generic TII.
+ return ValueTrackerResult();
+
+ TargetInstrInfo::RegSubRegPairAndIdx ExtractSubregInputReg;
+ if (!TII->getExtractSubregInputs(*Def, DefIdx, ExtractSubregInputReg))
+ return ValueTrackerResult();
+
+ // Bails if we have to compose sub registers.
+ // Likewise, if v0.subreg != 0, we would have to compose v0.subreg with sub0.
+ if (ExtractSubregInputReg.SubReg)
+ return ValueTrackerResult();
+ // Otherwise, the value is available in the v0.sub0.
+ return ValueTrackerResult(ExtractSubregInputReg.Reg, ExtractSubregInputReg.SubIdx);
+}
+
+ValueTrackerResult ValueTracker::getNextSourceFromSubregToReg() {
+ assert(Def->isSubregToReg() && "Invalid definition");
+ // We are looking at:
+ // Def = SUBREG_TO_REG Imm, v0, sub0
+
+ // Bails if we have to compose sub registers.
+ // If DefSubReg != sub0, we would have to check that all the bits
+ // we track are included in sub0 and if yes, we would have to
+ // determine the right subreg in v0.
+ if (DefSubReg != Def->getOperand(3).getImm())
+ return ValueTrackerResult();
+ // Bails if we have to compose sub registers.
+ // Likewise, if v0.subreg != 0, we would have to compose it with sub0.
+ if (Def->getOperand(2).getSubReg())
+ return ValueTrackerResult();
+
+ return ValueTrackerResult(Def->getOperand(2).getReg(),
+ Def->getOperand(3).getImm());
+}
+
+/// \brief Explore each PHI incoming operand and return its sources
+ValueTrackerResult ValueTracker::getNextSourceFromPHI() {
+ assert(Def->isPHI() && "Invalid definition");
+ ValueTrackerResult Res;
+
+ // If we look for a different subreg, bails as we do not
+ // support composing subreg yet.
+ if (Def->getOperand(0).getSubReg() != DefSubReg)
+ return ValueTrackerResult();
+
+ // Return all register sources for PHI instructions.
+ for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2) {
+ auto &MO = Def->getOperand(i);
+ assert(MO.isReg() && "Invalid PHI instruction");
+ Res.addSource(MO.getReg(), MO.getSubReg());
+ }
+
+ return Res;
+}
+
+ValueTrackerResult ValueTracker::getNextSourceImpl() {
+ assert(Def && "This method needs a valid definition");
+
+ assert(
+ (DefIdx < Def->getDesc().getNumDefs() || Def->getDesc().isVariadic()) &&
+ Def->getOperand(DefIdx).isDef() && "Invalid DefIdx");
+ if (Def->isCopy())
+ return getNextSourceFromCopy();
+ if (Def->isBitcast())
+ return getNextSourceFromBitcast();
+ // All the remaining cases involve "complex" instructions.
+ // Bails if we did not ask for the advanced tracking.
+ if (!UseAdvancedTracking)
+ return ValueTrackerResult();
+ if (Def->isRegSequence() || Def->isRegSequenceLike())
+ return getNextSourceFromRegSequence();
+ if (Def->isInsertSubreg() || Def->isInsertSubregLike())
+ return getNextSourceFromInsertSubreg();
+ if (Def->isExtractSubreg() || Def->isExtractSubregLike())
+ return getNextSourceFromExtractSubreg();
+ if (Def->isSubregToReg())
+ return getNextSourceFromSubregToReg();
+ if (Def->isPHI())
+ return getNextSourceFromPHI();
+ return ValueTrackerResult();
+}
+
+ValueTrackerResult ValueTracker::getNextSource() {
+ // If we reach a point where we cannot move up in the use-def chain,
+ // there is nothing we can get.
+ if (!Def)
+ return ValueTrackerResult();
+
+ ValueTrackerResult Res = getNextSourceImpl();
+ if (Res.isValid()) {
+ // Update definition, definition index, and subregister for the
+ // next call of getNextSource.
+ // Update the current register.
+ bool OneRegSrc = Res.getNumSources() == 1;
+ if (OneRegSrc)
+ Reg = Res.getSrcReg(0);
+ // Update the result before moving up in the use-def chain
+ // with the instruction containing the last found sources.
+ Res.setInst(Def);
+
+ // If we can still move up in the use-def chain, move to the next
+ // definition.
+ if (!TargetRegisterInfo::isPhysicalRegister(Reg) && OneRegSrc) {
+ Def = MRI.getVRegDef(Reg);
+ DefIdx = MRI.def_begin(Reg).getOperandNo();
+ DefSubReg = Res.getSrcSubReg(0);
+ return Res;
+ }
+ }
+ // If we end up here, this means we will not be able to find another source
+ // for the next iteration. Make sure any new call to getNextSource bails out
+ // early by cutting the use-def chain.
+ Def = nullptr;
+ return Res;
+}
projects / oota-llvm.git / blobdiff