diff options
| author | dim <dim@FreeBSD.org> | 2014-03-21 17:53:59 +0000 |
|---|---|---|
| committer | dim <dim@FreeBSD.org> | 2014-03-21 17:53:59 +0000 |
| commit | 9cedb8bb69b89b0f0c529937247a6a80cabdbaec (patch) | |
| tree | c978f0e9ec1ab92dc8123783f30b08a7fd1e2a39 /contrib/llvm/lib/Target/Hexagon | |
| parent | 03fdc2934eb61c44c049a02b02aa974cfdd8a0eb (diff) | |
| download | FreeBSD-src-9cedb8bb69b89b0f0c529937247a6a80cabdbaec.zip FreeBSD-src-9cedb8bb69b89b0f0c529937247a6a80cabdbaec.tar.gz | |
MFC 261991:
Upgrade our copy of llvm/clang to 3.4 release. This version supports
all of the features in the current working draft of the upcoming C++
standard, provisionally named C++1y.
The code generator's performance is greatly increased, and the loop
auto-vectorizer is now enabled at -Os and -O2 in addition to -O3. The
PowerPC backend has made several major improvements to code generation
quality and compile time, and the X86, SPARC, ARM32, Aarch64 and SystemZ
backends have all seen major feature work.
Release notes for llvm and clang can be found here:
<http://llvm.org/releases/3.4/docs/ReleaseNotes.html>
<http://llvm.org/releases/3.4/tools/clang/docs/ReleaseNotes.html>
MFC 262121 (by emaste):
Update lldb for clang/llvm 3.4 import
This commit largely restores the lldb source to the upstream r196259
snapshot with the addition of threaded inferior support and a few bug
fixes.
Specific upstream lldb revisions restored include:
SVN git
181387 779e6ac
181703 7bef4e2
182099 b31044e
182650 f2dcf35
182683 0d91b80
183862 15c1774
183929 99447a6
184177 0b2934b
184948 4dc3761
184954 007e7bc
186990 eebd175
Sponsored by: DARPA, AFRL
MFC 262186 (by emaste):
Fix mismerge in r262121
A break statement was lost in the merge. The error had no functional
impact, but restore it to reduce the diff against upstream.
MFC 262303:
Pull in r197521 from upstream clang trunk (by rdivacky):
Use the integrated assembler by default on FreeBSD/ppc and ppc64.
Requested by: jhibbits
MFC 262611:
Pull in r196874 from upstream llvm trunk:
Fix a crash that occurs when PWD is invalid.
MCJIT needs to be able to run in hostile environments, even when PWD
is invalid. There's no need to crash MCJIT in this case.
The obvious fix is to simply leave MCContext's CompilationDir empty
when PWD can't be determined. This way, MCJIT clients,
and other clients that link with LLVM don't need a valid working directory.
If we do want to guarantee valid CompilationDir, that should be done
only for clients of getCompilationDir(). This is as simple as checking
for an empty string.
The only current use of getCompilationDir is EmitGenDwarfInfo, which
won't conceivably run with an invalid working dir. However, in the
purely hypothetically and untestable case that this happens, the
AT_comp_dir will be omitted from the compilation_unit DIE.
This should help fix assertions occurring with ports-mgmt/tinderbox,
when it is using jails, and sometimes invalidates clang's current
working directory.
Reported by: decke
MFC 262809:
Pull in r203007 from upstream clang trunk:
Don't produce an alias between destructors with different calling conventions.
Fixes pr19007.
(Please note that is an LLVM PR identifier, not a FreeBSD one.)
This should fix Firefox and/or libxul crashes (due to problems with
regparm/stdcall calling conventions) on i386.
Reported by: multiple users on freebsd-current
PR: bin/187103
MFC 263048:
Repair recognition of "CC" as an alias for the C++ compiler, since it
was silently broken by upstream for a Windows-specific use-case.
Apparently some versions of CMake still rely on this archaic feature...
Reported by: rakuco
MFC 263049:
Garbage collect the old way of adding the libstdc++ include directories
in clang's InitHeaderSearch.cpp. This has been superseded by David
Chisnall's commit in r255321.
Moreover, if libc++ is used, the libstdc++ include directories should
not be in the search path at all. These directories are now only used
if you pass -stdlib=libstdc++.
Diffstat (limited to 'contrib/llvm/lib/Target/Hexagon')
41 files changed, 1639 insertions, 2836 deletions
diff --git a/contrib/llvm/lib/Target/Hexagon/Hexagon.h b/contrib/llvm/lib/Target/Hexagon/Hexagon.h index a9b00a2..5467ee3 100644 --- a/contrib/llvm/lib/Target/Hexagon/Hexagon.h +++ b/contrib/llvm/lib/Target/Hexagon/Hexagon.h @@ -29,7 +29,7 @@ namespace llvm { class HexagonTargetMachine; class raw_ostream; - FunctionPass *createHexagonISelDag(const HexagonTargetMachine &TM, + FunctionPass *createHexagonISelDag(HexagonTargetMachine &TM, CodeGenOpt::Level OptLevel); FunctionPass *createHexagonDelaySlotFillerPass(const TargetMachine &TM); FunctionPass *createHexagonFPMoverPass(const TargetMachine &TM); @@ -37,11 +37,15 @@ namespace llvm { FunctionPass *createHexagonCFGOptimizer(const HexagonTargetMachine &TM); FunctionPass *createHexagonSplitTFRCondSets(const HexagonTargetMachine &TM); + FunctionPass *createHexagonSplitConst32AndConst64( + const HexagonTargetMachine &TM); FunctionPass *createHexagonExpandPredSpillCode( const HexagonTargetMachine &TM); FunctionPass *createHexagonHardwareLoops(); FunctionPass *createHexagonPeephole(); FunctionPass *createHexagonFixupHwLoops(); + FunctionPass *createHexagonNewValueJump(); + FunctionPass *createHexagonCopyToCombine(); FunctionPass *createHexagonPacketizer(); FunctionPass *createHexagonNewValueJump(); diff --git a/contrib/llvm/lib/Target/Hexagon/Hexagon.td b/contrib/llvm/lib/Target/Hexagon/Hexagon.td index 9b3a643..568798c 100644 --- a/contrib/llvm/lib/Target/Hexagon/Hexagon.td +++ b/contrib/llvm/lib/Target/Hexagon/Hexagon.td @@ -120,15 +120,39 @@ def getPredNewOpcode : InstrMapping { } //===----------------------------------------------------------------------===// +// Generate mapping table to relate .new predicated instructions with their old +// format. +// +def getPredOldOpcode : InstrMapping { + let FilterClass = "PredNewRel"; + let RowFields = ["BaseOpcode", "PredSense", "isNVStore"]; + let ColFields = ["PNewValue"]; + let KeyCol = ["new"]; + let ValueCols = [[""]]; +} + +//===----------------------------------------------------------------------===// // Generate mapping table to relate store instructions with their new-value // format. // def getNewValueOpcode : InstrMapping { let FilterClass = "NewValueRel"; let RowFields = ["BaseOpcode", "PredSense", "PNewValue"]; - let ColFields = ["isNVStore"]; - let KeyCol = ["0"]; - let ValueCols = [["1"]]; + let ColFields = ["NValueST"]; + let KeyCol = ["false"]; + let ValueCols = [["true"]]; +} + +//===----------------------------------------------------------------------===// +// Generate mapping table to relate new-value store instructions with their old +// format. +// +def getNonNVStore : InstrMapping { + let FilterClass = "NewValueRel"; + let RowFields = ["BaseOpcode", "PredSense", "PNewValue"]; + let ColFields = ["NValueST"]; + let KeyCol = ["true"]; + let ValueCols = [["false"]]; } def getBasedWithImmOffset : InstrMapping { diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp index 88cd3fb..a2e04ba 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonAsmPrinter.cpp @@ -99,7 +99,7 @@ void HexagonAsmPrinter::printOperand(const MachineInstr *MI, unsigned OpNo, return; case MachineOperand::MO_GlobalAddress: // Computing the address of a global symbol, not calling it. - O << *Mang->getSymbol(MO.getGlobal()); + O << *getSymbol(MO.getGlobal()); printOffset(MO.getOffset(), O); return; } @@ -267,7 +267,7 @@ void HexagonAsmPrinter::printGlobalOperand(const MachineInstr *MI, int OpNo, assert( (MO.getType() == MachineOperand::MO_GlobalAddress) && "Expecting global address"); - O << *Mang->getSymbol(MO.getGlobal()); + O << *getSymbol(MO.getGlobal()); if (MO.getOffset() != 0) { O << " + "; O << MO.getOffset(); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.cpp index 2c93d04..f5f958c 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.cpp @@ -25,14 +25,13 @@ using namespace llvm; Hexagon_CCState::Hexagon_CCState(CallingConv::ID CC, bool isVarArg, const TargetMachine &tm, - SmallVector<CCValAssign, 16> &locs, + SmallVectorImpl<CCValAssign> &locs, LLVMContext &c) - : CallingConv(CC), IsVarArg(isVarArg), TM(tm), - TRI(*TM.getRegisterInfo()), Locs(locs), Context(c) { + : CallingConv(CC), IsVarArg(isVarArg), TM(tm), Locs(locs), Context(c) { // No stack is used. StackOffset = 0; - UsedRegs.resize((TRI.getNumRegs()+31)/32); + UsedRegs.resize((TM.getRegisterInfo()->getNumRegs()+31)/32); } // HandleByVal - Allocate a stack slot large enough to pass an argument by @@ -56,6 +55,7 @@ void Hexagon_CCState::HandleByVal(unsigned ValNo, EVT ValVT, /// MarkAllocated - Mark a register and all of its aliases as allocated. void Hexagon_CCState::MarkAllocated(unsigned Reg) { + const TargetRegisterInfo &TRI = *TM.getRegisterInfo(); for (MCRegAliasIterator AI(Reg, &TRI, true); AI.isValid(); ++AI) UsedRegs[*AI/32] |= 1 << (*AI&31); } diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.h b/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.h index 489b3a3..33c8306 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonCallingConvLower.h @@ -48,15 +48,14 @@ class Hexagon_CCState { CallingConv::ID CallingConv; bool IsVarArg; const TargetMachine &TM; - const TargetRegisterInfo &TRI; - SmallVector<CCValAssign, 16> &Locs; + SmallVectorImpl<CCValAssign> &Locs; LLVMContext &Context; unsigned StackOffset; SmallVector<uint32_t, 16> UsedRegs; public: Hexagon_CCState(CallingConv::ID CC, bool isVarArg, const TargetMachine &TM, - SmallVector<CCValAssign, 16> &locs, LLVMContext &c); + SmallVectorImpl<CCValAssign> &locs, LLVMContext &c); void addLoc(const CCValAssign &V) { Locs.push_back(V); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp new file mode 100644 index 0000000..dc440cb --- /dev/null +++ b/contrib/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp @@ -0,0 +1,677 @@ +//===------- HexagonCopyToCombine.cpp - Hexagon Copy-To-Combine Pass ------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// This pass replaces transfer instructions by combine instructions. +// We walk along a basic block and look for two combinable instructions and try +// to move them together. If we can move them next to each other we do so and +// replace them with a combine instruction. +//===----------------------------------------------------------------------===// +#define DEBUG_TYPE "hexagon-copy-combine" + +#include "llvm/PassSupport.h" +#include "llvm/ADT/DenseSet.h" +#include "llvm/ADT/DenseMap.h" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/MachineBasicBlock.h" +#include "llvm/CodeGen/MachineFunction.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineInstr.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Support/CodeGen.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/raw_ostream.h" + +#include "Hexagon.h" +#include "HexagonInstrInfo.h" +#include "HexagonRegisterInfo.h" +#include "HexagonSubtarget.h" +#include "HexagonTargetMachine.h" +#include "HexagonMachineFunctionInfo.h" + +using namespace llvm; + +static +cl::opt<bool> IsCombinesDisabled("disable-merge-into-combines", + cl::Hidden, cl::ZeroOrMore, + cl::init(false), + cl::desc("Disable merging into combines")); +static +cl::opt<unsigned> +MaxNumOfInstsBetweenNewValueStoreAndTFR("max-num-inst-between-tfr-and-nv-store", + cl::Hidden, cl::init(4), + cl::desc("Maximum distance between a tfr feeding a store we " + "consider the store still to be newifiable")); + +namespace llvm { + void initializeHexagonCopyToCombinePass(PassRegistry&); +} + + +namespace { + +class HexagonCopyToCombine : public MachineFunctionPass { + const HexagonInstrInfo *TII; + const TargetRegisterInfo *TRI; + bool ShouldCombineAggressively; + + DenseSet<MachineInstr *> PotentiallyNewifiableTFR; +public: + static char ID; + + HexagonCopyToCombine() : MachineFunctionPass(ID) { + initializeHexagonCopyToCombinePass(*PassRegistry::getPassRegistry()); + } + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + MachineFunctionPass::getAnalysisUsage(AU); + } + + const char *getPassName() const { + return "Hexagon Copy-To-Combine Pass"; + } + + virtual bool runOnMachineFunction(MachineFunction &Fn); + +private: + MachineInstr *findPairable(MachineInstr *I1, bool &DoInsertAtI1); + + void findPotentialNewifiableTFRs(MachineBasicBlock &); + + void combine(MachineInstr *I1, MachineInstr *I2, + MachineBasicBlock::iterator &MI, bool DoInsertAtI1); + + bool isSafeToMoveTogether(MachineInstr *I1, MachineInstr *I2, + unsigned I1DestReg, unsigned I2DestReg, + bool &DoInsertAtI1); + + void emitCombineRR(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineRI(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineIR(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); + + void emitCombineII(MachineBasicBlock::iterator &Before, unsigned DestReg, + MachineOperand &HiOperand, MachineOperand &LoOperand); +}; + +} // End anonymous namespace. + +char HexagonCopyToCombine::ID = 0; + +INITIALIZE_PASS(HexagonCopyToCombine, "hexagon-copy-combine", + "Hexagon Copy-To-Combine Pass", false, false) + +static bool isCombinableInstType(MachineInstr *MI, + const HexagonInstrInfo *TII, + bool ShouldCombineAggressively) { + switch(MI->getOpcode()) { + case Hexagon::TFR: { + // A COPY instruction can be combined if its arguments are IntRegs (32bit). + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isReg()); + + unsigned DestReg = MI->getOperand(0).getReg(); + unsigned SrcReg = MI->getOperand(1).getReg(); + return Hexagon::IntRegsRegClass.contains(DestReg) && + Hexagon::IntRegsRegClass.contains(SrcReg); + } + + case Hexagon::TFRI: { + // A transfer-immediate can be combined if its argument is a signed 8bit + // value. + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isImm()); + unsigned DestReg = MI->getOperand(0).getReg(); + + // Only combine constant extended TFRI if we are in aggressive mode. + return Hexagon::IntRegsRegClass.contains(DestReg) && + (ShouldCombineAggressively || isInt<8>(MI->getOperand(1).getImm())); + } + + case Hexagon::TFRI_V4: { + if (!ShouldCombineAggressively) + return false; + assert(MI->getOperand(0).isReg() && MI->getOperand(1).isGlobal()); + + // Ensure that TargetFlags are MO_NO_FLAG for a global. This is a + // workaround for an ABI bug that prevents GOT relocations on combine + // instructions + if (MI->getOperand(1).getTargetFlags() != HexagonII::MO_NO_FLAG) + return false; + + unsigned DestReg = MI->getOperand(0).getReg(); + return Hexagon::IntRegsRegClass.contains(DestReg); + } + + default: + break; + } + + return false; +} + +static bool isGreaterThan8BitTFRI(MachineInstr *I) { + return I->getOpcode() == Hexagon::TFRI && + !isInt<8>(I->getOperand(1).getImm()); +} +static bool isGreaterThan6BitTFRI(MachineInstr *I) { + return I->getOpcode() == Hexagon::TFRI && + !isUInt<6>(I->getOperand(1).getImm()); +} + +/// areCombinableOperations - Returns true if the two instruction can be merge +/// into a combine (ignoring register constraints). +static bool areCombinableOperations(const TargetRegisterInfo *TRI, + MachineInstr *HighRegInst, + MachineInstr *LowRegInst) { + assert((HighRegInst->getOpcode() == Hexagon::TFR || + HighRegInst->getOpcode() == Hexagon::TFRI || + HighRegInst->getOpcode() == Hexagon::TFRI_V4) && + (LowRegInst->getOpcode() == Hexagon::TFR || + LowRegInst->getOpcode() == Hexagon::TFRI || + LowRegInst->getOpcode() == Hexagon::TFRI_V4) && + "Assume individual instructions are of a combinable type"); + + const HexagonRegisterInfo *QRI = + static_cast<const HexagonRegisterInfo *>(TRI); + + // V4 added some combine variations (mixed immediate and register source + // operands), if we are on < V4 we can only combine 2 register-to-register + // moves and 2 immediate-to-register moves. We also don't have + // constant-extenders. + if (!QRI->Subtarget.hasV4TOps()) + return HighRegInst->getOpcode() == LowRegInst->getOpcode() && + !isGreaterThan8BitTFRI(HighRegInst) && + !isGreaterThan6BitTFRI(LowRegInst); + + // There is no combine of two constant extended values. + if ((HighRegInst->getOpcode() == Hexagon::TFRI_V4 || + isGreaterThan8BitTFRI(HighRegInst)) && + (LowRegInst->getOpcode() == Hexagon::TFRI_V4 || + isGreaterThan6BitTFRI(LowRegInst))) + return false; + + return true; +} + +static bool isEvenReg(unsigned Reg) { + assert(TargetRegisterInfo::isPhysicalRegister(Reg) && + Hexagon::IntRegsRegClass.contains(Reg)); + return (Reg - Hexagon::R0) % 2 == 0; +} + +static void removeKillInfo(MachineInstr *MI, unsigned RegNotKilled) { + for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { + MachineOperand &Op = MI->getOperand(I); + if (!Op.isReg() || Op.getReg() != RegNotKilled || !Op.isKill()) + continue; + Op.setIsKill(false); + } +} + +/// isUnsafeToMoveAcross - Returns true if it is unsafe to move a copy +/// instruction from \p UseReg to \p DestReg over the instruction \p I. +static bool isUnsafeToMoveAcross(MachineInstr *I, unsigned UseReg, + unsigned DestReg, + const TargetRegisterInfo *TRI) { + return (UseReg && (I->modifiesRegister(UseReg, TRI))) || + I->modifiesRegister(DestReg, TRI) || + I->readsRegister(DestReg, TRI) || + I->hasUnmodeledSideEffects() || + I->isInlineAsm() || I->isDebugValue(); +} + +/// isSafeToMoveTogether - Returns true if it is safe to move I1 next to I2 such +/// that the two instructions can be paired in a combine. +bool HexagonCopyToCombine::isSafeToMoveTogether(MachineInstr *I1, + MachineInstr *I2, + unsigned I1DestReg, + unsigned I2DestReg, + bool &DoInsertAtI1) { + + bool IsImmUseReg = I2->getOperand(1).isImm() || I2->getOperand(1).isGlobal(); + unsigned I2UseReg = IsImmUseReg ? 0 : I2->getOperand(1).getReg(); + + // It is not safe to move I1 and I2 into one combine if I2 has a true + // dependence on I1. + if (I2UseReg && I1->modifiesRegister(I2UseReg, TRI)) + return false; + + bool isSafe = true; + + // First try to move I2 towards I1. + { + // A reverse_iterator instantiated like below starts before I2, and I1 + // respectively. + // Look at instructions I in between I2 and (excluding) I1. + MachineBasicBlock::reverse_iterator I(I2), + End = --(MachineBasicBlock::reverse_iterator(I1)); + // At 03 we got better results (dhrystone!) by being more conservative. + if (!ShouldCombineAggressively) + End = MachineBasicBlock::reverse_iterator(I1); + // If I2 kills its operand and we move I2 over an instruction that also + // uses I2's use reg we need to modify that (first) instruction to now kill + // this reg. + unsigned KilledOperand = 0; + if (I2->killsRegister(I2UseReg)) + KilledOperand = I2UseReg; + MachineInstr *KillingInstr = 0; + + for (; I != End; ++I) { + // If the intervening instruction I: + // * modifies I2's use reg + // * modifies I2's def reg + // * reads I2's def reg + // * or has unmodelled side effects + // we can't move I2 across it. + if (isUnsafeToMoveAcross(&*I, I2UseReg, I2DestReg, TRI)) { + isSafe = false; + break; + } + + // Update first use of the killed operand. + if (!KillingInstr && KilledOperand && + I->readsRegister(KilledOperand, TRI)) + KillingInstr = &*I; + } + if (isSafe) { + // Update the intermediate instruction to with the kill flag. + if (KillingInstr) { + bool Added = KillingInstr->addRegisterKilled(KilledOperand, TRI, true); + (void)Added; // supress compiler warning + assert(Added && "Must successfully update kill flag"); + removeKillInfo(I2, KilledOperand); + } + DoInsertAtI1 = true; + return true; + } + } + + // Try to move I1 towards I2. + { + // Look at instructions I in between I1 and (excluding) I2. + MachineBasicBlock::iterator I(I1), End(I2); + // At O3 we got better results (dhrystone) by being more conservative here. + if (!ShouldCombineAggressively) + End = llvm::next(MachineBasicBlock::iterator(I2)); + IsImmUseReg = I1->getOperand(1).isImm() || I1->getOperand(1).isGlobal(); + unsigned I1UseReg = IsImmUseReg ? 0 : I1->getOperand(1).getReg(); + // Track killed operands. If we move across an instruction that kills our + // operand, we need to update the kill information on the moved I1. It kills + // the operand now. + MachineInstr *KillingInstr = 0; + unsigned KilledOperand = 0; + + while(++I != End) { + // If the intervening instruction I: + // * modifies I1's use reg + // * modifies I1's def reg + // * reads I1's def reg + // * or has unmodelled side effects + // We introduce this special case because llvm has no api to remove a + // kill flag for a register (a removeRegisterKilled() analogous to + // addRegisterKilled) that handles aliased register correctly. + // * or has a killed aliased register use of I1's use reg + // %D4<def> = TFRI64 16 + // %R6<def> = TFR %R9 + // %R8<def> = KILL %R8, %D4<imp-use,kill> + // If we want to move R6 = across the KILL instruction we would have + // to remove the %D4<imp-use,kill> operand. For now, we are + // conservative and disallow the move. + // we can't move I1 across it. + if (isUnsafeToMoveAcross(I, I1UseReg, I1DestReg, TRI) || + // Check for an aliased register kill. Bail out if we see one. + (!I->killsRegister(I1UseReg) && I->killsRegister(I1UseReg, TRI))) + return false; + + // Check for an exact kill (registers match). + if (I1UseReg && I->killsRegister(I1UseReg)) { + assert(KillingInstr == 0 && "Should only see one killing instruction"); + KilledOperand = I1UseReg; + KillingInstr = &*I; + } + } + if (KillingInstr) { + removeKillInfo(KillingInstr, KilledOperand); + // Update I1 to set the kill flag. This flag will later be picked up by + // the new COMBINE instruction. + bool Added = I1->addRegisterKilled(KilledOperand, TRI); + (void)Added; // supress compiler warning + assert(Added && "Must successfully update kill flag"); + } + DoInsertAtI1 = false; + } + + return true; +} + +/// findPotentialNewifiableTFRs - Finds tranfers that feed stores that could be +/// newified. (A use of a 64 bit register define can not be newified) +void +HexagonCopyToCombine::findPotentialNewifiableTFRs(MachineBasicBlock &BB) { + DenseMap<unsigned, MachineInstr *> LastDef; + for (MachineBasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ++I) { + MachineInstr *MI = I; + // Mark TFRs that feed a potential new value store as such. + if(TII->mayBeNewStore(MI)) { + // Look for uses of TFR instructions. + for (unsigned OpdIdx = 0, OpdE = MI->getNumOperands(); OpdIdx != OpdE; + ++OpdIdx) { + MachineOperand &Op = MI->getOperand(OpdIdx); + + // Skip over anything except register uses. + if (!Op.isReg() || !Op.isUse() || !Op.getReg()) + continue; + + // Look for the defining instruction. + unsigned Reg = Op.getReg(); + MachineInstr *DefInst = LastDef[Reg]; + if (!DefInst) + continue; + if (!isCombinableInstType(DefInst, TII, ShouldCombineAggressively)) + continue; + + // Only close newifiable stores should influence the decision. + MachineBasicBlock::iterator It(DefInst); + unsigned NumInstsToDef = 0; + while (&*It++ != MI) + ++NumInstsToDef; + + if (NumInstsToDef > MaxNumOfInstsBetweenNewValueStoreAndTFR) + continue; + + PotentiallyNewifiableTFR.insert(DefInst); + } + // Skip to next instruction. + continue; + } + + // Put instructions that last defined integer or double registers into the + // map. + for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) { + MachineOperand &Op = MI->getOperand(I); + if (!Op.isReg() || !Op.isDef() || !Op.getReg()) + continue; + unsigned Reg = Op.getReg(); + if (Hexagon::DoubleRegsRegClass.contains(Reg)) { + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) { + LastDef[*SubRegs] = MI; + } + } else if (Hexagon::IntRegsRegClass.contains(Reg)) + LastDef[Reg] = MI; + } + } +} + +bool HexagonCopyToCombine::runOnMachineFunction(MachineFunction &MF) { + + if (IsCombinesDisabled) return false; + + bool HasChanged = false; + + // Get target info. + TRI = MF.getTarget().getRegisterInfo(); + TII = static_cast<const HexagonInstrInfo *>(MF.getTarget().getInstrInfo()); + + // Combine aggressively (for code size) + ShouldCombineAggressively = + MF.getTarget().getOptLevel() <= CodeGenOpt::Default; + + // Traverse basic blocks. + for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE; + ++BI) { + PotentiallyNewifiableTFR.clear(); + findPotentialNewifiableTFRs(*BI); + + // Traverse instructions in basic block. + for(MachineBasicBlock::iterator MI = BI->begin(), End = BI->end(); + MI != End;) { + MachineInstr *I1 = MI++; + // Don't combine a TFR whose user could be newified (instructions that + // define double registers can not be newified - Programmer's Ref Manual + // 5.4.2 New-value stores). + if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I1)) + continue; + + // Ignore instructions that are not combinable. + if (!isCombinableInstType(I1, TII, ShouldCombineAggressively)) + continue; + + // Find a second instruction that can be merged into a combine + // instruction. + bool DoInsertAtI1 = false; + MachineInstr *I2 = findPairable(I1, DoInsertAtI1); + if (I2) { + HasChanged = true; + combine(I1, I2, MI, DoInsertAtI1); + } + } + } + + return HasChanged; +} + +/// findPairable - Returns an instruction that can be merged with \p I1 into a +/// COMBINE instruction or 0 if no such instruction can be found. Returns true +/// in \p DoInsertAtI1 if the combine must be inserted at instruction \p I1 +/// false if the combine must be inserted at the returned instruction. +MachineInstr *HexagonCopyToCombine::findPairable(MachineInstr *I1, + bool &DoInsertAtI1) { + MachineBasicBlock::iterator I2 = llvm::next(MachineBasicBlock::iterator(I1)); + unsigned I1DestReg = I1->getOperand(0).getReg(); + + for (MachineBasicBlock::iterator End = I1->getParent()->end(); I2 != End; + ++I2) { + // Bail out early if we see a second definition of I1DestReg. + if (I2->modifiesRegister(I1DestReg, TRI)) + break; + + // Ignore non-combinable instructions. + if (!isCombinableInstType(I2, TII, ShouldCombineAggressively)) + continue; + + // Don't combine a TFR whose user could be newified. + if (ShouldCombineAggressively && PotentiallyNewifiableTFR.count(I2)) + continue; + + unsigned I2DestReg = I2->getOperand(0).getReg(); + + // Check that registers are adjacent and that the first destination register + // is even. + bool IsI1LowReg = (I2DestReg - I1DestReg) == 1; + bool IsI2LowReg = (I1DestReg - I2DestReg) == 1; + unsigned FirstRegIndex = IsI1LowReg ? I1DestReg : I2DestReg; + if ((!IsI1LowReg && !IsI2LowReg) || !isEvenReg(FirstRegIndex)) + continue; + + // Check that the two instructions are combinable. V4 allows more + // instructions to be merged into a combine. + // The order matters because in a TFRI we might can encode a int8 as the + // hi reg operand but only a uint6 as the low reg operand. + if ((IsI2LowReg && !areCombinableOperations(TRI, I1, I2)) || + (IsI1LowReg && !areCombinableOperations(TRI, I2, I1))) + break; + + if (isSafeToMoveTogether(I1, I2, I1DestReg, I2DestReg, + DoInsertAtI1)) + return I2; + + // Not safe. Stop searching. + break; + } + return 0; +} + +void HexagonCopyToCombine::combine(MachineInstr *I1, MachineInstr *I2, + MachineBasicBlock::iterator &MI, + bool DoInsertAtI1) { + // We are going to delete I2. If MI points to I2 advance it to the next + // instruction. + if ((MachineInstr *)MI == I2) ++MI; + + // Figure out whether I1 or I2 goes into the lowreg part. + unsigned I1DestReg = I1->getOperand(0).getReg(); + unsigned I2DestReg = I2->getOperand(0).getReg(); + bool IsI1Loreg = (I2DestReg - I1DestReg) == 1; + unsigned LoRegDef = IsI1Loreg ? I1DestReg : I2DestReg; + + // Get the double word register. + unsigned DoubleRegDest = + TRI->getMatchingSuperReg(LoRegDef, Hexagon::subreg_loreg, + &Hexagon::DoubleRegsRegClass); + assert(DoubleRegDest != 0 && "Expect a valid register"); + + + // Setup source operands. + MachineOperand &LoOperand = IsI1Loreg ? I1->getOperand(1) : + I2->getOperand(1); + MachineOperand &HiOperand = IsI1Loreg ? I2->getOperand(1) : + I1->getOperand(1); + + // Figure out which source is a register and which a constant. + bool IsHiReg = HiOperand.isReg(); + bool IsLoReg = LoOperand.isReg(); + + MachineBasicBlock::iterator InsertPt(DoInsertAtI1 ? I1 : I2); + // Emit combine. + if (IsHiReg && IsLoReg) + emitCombineRR(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else if (IsHiReg) + emitCombineRI(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else if (IsLoReg) + emitCombineIR(InsertPt, DoubleRegDest, HiOperand, LoOperand); + else + emitCombineII(InsertPt, DoubleRegDest, HiOperand, LoOperand); + + I1->eraseFromParent(); + I2->eraseFromParent(); +} + +void HexagonCopyToCombine::emitCombineII(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle globals. + if (HiOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(), + HiOperand.getTargetFlags()) + .addImm(LoOperand.getImm()); + return; + } + if (LoOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(), + LoOperand.getTargetFlags()); + return; + } + + // Handle constant extended immediates. + if (!isInt<8>(HiOperand.getImm())) { + assert(isInt<8>(LoOperand.getImm())); + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); + return; + } + + if (!isUInt<6>(LoOperand.getImm())) { + assert(isInt<8>(HiOperand.getImm())); + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_iI_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); + return; + } + + // Insert new combine instruction. + // DoubleRegDest = combine #HiImm, #LoImm + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ii), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addImm(LoOperand.getImm()); +} + +void HexagonCopyToCombine::emitCombineIR(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned LoReg = LoOperand.getReg(); + unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill()); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle global. + if (HiOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg) + .addGlobalAddress(HiOperand.getGlobal(), HiOperand.getOffset(), + HiOperand.getTargetFlags()) + .addReg(LoReg, LoRegKillFlag); + return; + } + // Insert new combine instruction. + // DoubleRegDest = combine #HiImm, LoReg + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_Ir_V4), DoubleDestReg) + .addImm(HiOperand.getImm()) + .addReg(LoReg, LoRegKillFlag); +} + +void HexagonCopyToCombine::emitCombineRI(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill()); + unsigned HiReg = HiOperand.getReg(); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Handle global. + if (LoOperand.isGlobal()) { + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addGlobalAddress(LoOperand.getGlobal(), LoOperand.getOffset(), + LoOperand.getTargetFlags()); + return; + } + + // Insert new combine instruction. + // DoubleRegDest = combine HiReg, #LoImm + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rI_V4), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addImm(LoOperand.getImm()); +} + +void HexagonCopyToCombine::emitCombineRR(MachineBasicBlock::iterator &InsertPt, + unsigned DoubleDestReg, + MachineOperand &HiOperand, + MachineOperand &LoOperand) { + unsigned LoRegKillFlag = getKillRegState(LoOperand.isKill()); + unsigned HiRegKillFlag = getKillRegState(HiOperand.isKill()); + unsigned LoReg = LoOperand.getReg(); + unsigned HiReg = HiOperand.getReg(); + + DebugLoc DL = InsertPt->getDebugLoc(); + MachineBasicBlock *BB = InsertPt->getParent(); + + // Insert new combine instruction. + // DoubleRegDest = combine HiReg, LoReg + BuildMI(*BB, InsertPt, DL, TII->get(Hexagon::COMBINE_rr), DoubleDestReg) + .addReg(HiReg, HiRegKillFlag) + .addReg(LoReg, LoRegKillFlag); +} + +FunctionPass *llvm::createHexagonCopyToCombine() { + return new HexagonCopyToCombine(); +} diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp index de993ee..2b04f25 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonFrameLowering.cpp @@ -76,17 +76,12 @@ void HexagonFrameLowering::determineFrameLayout(MachineFunction &MF) const { void HexagonFrameLowering::emitPrologue(MachineFunction &MF) const { MachineBasicBlock &MBB = MF.front(); MachineFrameInfo *MFI = MF.getFrameInfo(); - MachineModuleInfo &MMI = MF.getMMI(); MachineBasicBlock::iterator MBBI = MBB.begin(); const HexagonRegisterInfo *QRI = static_cast<const HexagonRegisterInfo *>(MF.getTarget().getRegisterInfo()); DebugLoc dl = MBBI != MBB.end() ? MBBI->getDebugLoc() : DebugLoc(); determineFrameLayout(MF); - // Check if frame moves are needed for EH. - bool needsFrameMoves = MMI.hasDebugInfo() || - !MF.getFunction()->needsUnwindTableEntry(); - // Get the number of bytes to allocate from the FrameInfo. int NumBytes = (int) MFI->getStackSize(); @@ -113,28 +108,6 @@ void HexagonFrameLowering::emitPrologue(MachineFunction &MF) const { MO.setImm(MFI->getMaxCallFrameSize()); } - std::vector<MachineMove> &Moves = MMI.getFrameMoves(); - - if (needsFrameMoves) { - // Advance CFA. DW_CFA_def_cfa - unsigned FPReg = QRI->getFrameRegister(); - unsigned RAReg = QRI->getRARegister(); - - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(FPReg, -8); - Moves.push_back(MachineMove(0, Dst, Src)); - - // R31 = (R31 - #4) - MachineLocation LRDst(RAReg, -4); - MachineLocation LRSrc(RAReg); - Moves.push_back(MachineMove(0, LRDst, LRSrc)); - - // R30 = (R30 - #8) - MachineLocation SPDst(FPReg, -8); - MachineLocation SPSrc(FPReg); - Moves.push_back(MachineMove(0, SPDst, SPSrc)); - } - // // Only insert ALLOCFRAME if we need to. // @@ -174,30 +147,55 @@ void HexagonFrameLowering::emitEpilogue(MachineFunction &MF, MachineBasicBlock::iterator MBBI = prior(MBB.end()); DebugLoc dl = MBBI->getDebugLoc(); // - // Only insert deallocframe if we need to. + // Only insert deallocframe if we need to. Also at -O0. See comment + // in emitPrologue above. // - if (hasFP(MF)) { + if (hasFP(MF) || MF.getTarget().getOptLevel() == CodeGenOpt::None) { MachineBasicBlock::iterator MBBI = prior(MBB.end()); MachineBasicBlock::iterator MBBI_end = MBB.end(); - // - // For Hexagon, we don't need the frame size. - // - MachineFrameInfo *MFI = MF.getFrameInfo(); - int NumBytes = (int) MFI->getStackSize(); const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo(); - + // Handle EH_RETURN. + if (MBBI->getOpcode() == Hexagon::EH_RETURN_JMPR) { + assert(MBBI->getOperand(0).isReg() && "Offset should be in register!"); + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)); + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::ADD_rr), + Hexagon::R29).addReg(Hexagon::R29).addReg(Hexagon::R28); + return; + } // Replace 'jumpr r31' instruction with dealloc_return for V4 and higher // versions. if (STI.hasV4TOps() && MBBI->getOpcode() == Hexagon::JMPret && !DisableDeallocRet) { - // Remove jumpr node. - MBB.erase(MBBI); + // Check for RESTORE_DEALLOC_RET_JMP_V4 call. Don't emit an extra DEALLOC + // instruction if we encounter it. + MachineBasicBlock::iterator BeforeJMPR = + MBB.begin() == MBBI ? MBBI : prior(MBBI); + if (BeforeJMPR != MBBI && + BeforeJMPR->getOpcode() == Hexagon::RESTORE_DEALLOC_RET_JMP_V4) { + // Remove the JMPR node. + MBB.erase(MBBI); + return; + } + // Add dealloc_return. - BuildMI(MBB, MBBI_end, dl, TII.get(Hexagon::DEALLOC_RET_V4)) - .addImm(NumBytes); - } else { // Add deallocframe for V2 and V3. - BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)).addImm(NumBytes); + MachineInstrBuilder MIB = + BuildMI(MBB, MBBI_end, dl, TII.get(Hexagon::DEALLOC_RET_V4)); + // Transfer the function live-out registers. + MIB->copyImplicitOps(*MBB.getParent(), &*MBBI); + // Remove the JUMPR node. + MBB.erase(MBBI); + } else { // Add deallocframe for V2 and V3, and V4 tail calls. + // Check for RESTORE_DEALLOC_BEFORE_TAILCALL_V4. We don't need an extra + // DEALLOCFRAME instruction after it. + MachineBasicBlock::iterator Term = MBB.getFirstTerminator(); + MachineBasicBlock::iterator I = + Term == MBB.begin() ? MBB.end() : prior(Term); + if (I != MBB.end() && + I->getOpcode() == Hexagon::RESTORE_DEALLOC_BEFORE_TAILCALL_V4) + return; + + BuildMI(MBB, MBBI, dl, TII.get(Hexagon::DEALLOCFRAME)); } } } diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp index d002788..52d5ab2 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp @@ -134,7 +134,7 @@ namespace { /// has a computable trip count and, if so, return a value that represents /// the trip count expression. CountValue *getLoopTripCount(MachineLoop *L, - SmallVector<MachineInstr*, 2> &OldInsts); + SmallVectorImpl<MachineInstr *> &OldInsts); /// \brief Return the expression that represents the number of times /// a loop iterates. The function takes the operands that represent the @@ -164,7 +164,7 @@ namespace { /// \brief Return true if the instruction is now dead. bool isDead(const MachineInstr *MI, - SmallVector<MachineInstr*, 1> &DeadPhis) const; + SmallVectorImpl<MachineInstr *> &DeadPhis) const; /// \brief Remove the instruction if it is now dead. void removeIfDead(MachineInstr *MI); @@ -428,7 +428,7 @@ bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L, /// induction variable patterns that are used in the calculation for /// the number of time the loop is executed. CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, - SmallVector<MachineInstr*, 2> &OldInsts) { + SmallVectorImpl<MachineInstr *> &OldInsts) { MachineBasicBlock *TopMBB = L->getTopBlock(); MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(); assert(PI != TopMBB->pred_end() && @@ -871,7 +871,7 @@ bool HexagonHardwareLoops::isInvalidLoopOperation( /// \brief - Return true if the loop contains an instruction that inhibits /// the use of the hardware loop function. bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const { - const std::vector<MachineBasicBlock*> Blocks = L->getBlocks(); + const std::vector<MachineBasicBlock *> &Blocks = L->getBlocks(); for (unsigned i = 0, e = Blocks.size(); i != e; ++i) { MachineBasicBlock *MBB = Blocks[i]; for (MachineBasicBlock::iterator @@ -890,7 +890,7 @@ bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const { /// for inline asm, physical registers and instructions with side effects /// removed. bool HexagonHardwareLoops::isDead(const MachineInstr *MI, - SmallVector<MachineInstr*, 1> &DeadPhis) const { + SmallVectorImpl<MachineInstr *> &DeadPhis) const { // Examine each operand. for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI->getOperand(i); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp index 491ccc4..5ae93284 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonISelDAGToDAG.cpp @@ -50,15 +50,13 @@ class HexagonDAGToDAGISel : public SelectionDAGISel { // Keep a reference to HexagonTargetMachine. const HexagonTargetMachine& TM; - const HexagonInstrInfo *TII; DenseMap<const GlobalValue *, unsigned> GlobalAddressUseCountMap; public: - explicit HexagonDAGToDAGISel(const HexagonTargetMachine &targetmachine, + explicit HexagonDAGToDAGISel(HexagonTargetMachine &targetmachine, CodeGenOpt::Level OptLevel) : SelectionDAGISel(targetmachine, OptLevel), Subtarget(targetmachine.getSubtarget<HexagonSubtarget>()), - TM(targetmachine), - TII(static_cast<const HexagonInstrInfo*>(TM.getInstrInfo())) { + TM(targetmachine) { initializeHexagonDAGToDAGISelPass(*PassRegistry::getPassRegistry()); } bool hasNumUsesBelowThresGA(SDNode *N) const; @@ -92,14 +90,14 @@ public: bool SelectAddr(SDNode *Op, SDValue Addr, SDValue &Base, SDValue &Offset); SDNode *SelectLoad(SDNode *N); - SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl); - SDNode *SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl); + SDNode *SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl); + SDNode *SelectIndexedLoad(LoadSDNode *LD, SDLoc dl); SDNode *SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl); + SDLoc dl); SDNode *SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl); - SDNode *SelectBaseOffsetStore(StoreSDNode *ST, DebugLoc dl); - SDNode *SelectIndexedStore(StoreSDNode *ST, DebugLoc dl); + SDLoc dl); + SDNode *SelectBaseOffsetStore(StoreSDNode *ST, SDLoc dl); + SDNode *SelectIndexedStore(StoreSDNode *ST, SDLoc dl); SDNode *SelectStore(SDNode *N); SDNode *SelectSHL(SDNode *N); SDNode *SelectSelect(SDNode *N); @@ -180,7 +178,7 @@ inline SDValue XformUToUM1Imm(unsigned Imm) { /// createHexagonISelDag - This pass converts a legalized DAG into a /// Hexagon-specific DAG, ready for instruction scheduling. /// -FunctionPass *llvm::createHexagonISelDag(const HexagonTargetMachine &TM, +FunctionPass *llvm::createHexagonISelDag(HexagonTargetMachine &TM, CodeGenOpt::Level OptLevel) { return new HexagonDAGToDAGISel(TM, OptLevel); } @@ -385,7 +383,7 @@ static bool OffsetFitsS11(EVT MemType, int64_t Offset) { // lowering for GlobalAddress nodes has already turned it into a // CONST32. // -SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, SDLoc dl) { SDValue Chain = LD->getChain(); SDNode* Const32 = LD->getBasePtr().getNode(); unsigned Opcode = 0; @@ -396,7 +394,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { EVT LoadedVT = LD->getMemoryVT(); int64_t Offset = cast<GlobalAddressSDNode>(Base)->getOffset(); if (Offset != 0 && OffsetFitsS11(LoadedVT, Offset)) { - MVT PointerTy = TLI.getPointerTy(); + MVT PointerTy = getTargetLowering()->getPointerTy(); const GlobalValue* GV = cast<GlobalAddressSDNode>(Base)->getGlobal(); SDValue TargAddr = @@ -433,7 +431,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetLoad(LoadSDNode *LD, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl) + SDLoc dl) { SDValue Chain = LD->getChain(); EVT LoadedVT = LD->getMemoryVT(); @@ -444,8 +442,11 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, SDValue N1 = LD->getOperand(1); SDValue CPTmpN1_0; SDValue CPTmpN1_1; + if (SelectADDRriS11_2(N1, CPTmpN1_0, CPTmpN1_1) && N1.getNode()->getValueType(0) == MVT::i32) { + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(LoadedVT, Val)) { SDValue TargetConst = CurDAG->getTargetConstant(Val, MVT::i32); SDNode *Result_1 = CurDAG->getMachineNode(Opcode, dl, MVT::i32, MVT::i32, @@ -497,7 +498,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadSignExtend64(LoadSDNode *LD, SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, unsigned Opcode, - DebugLoc dl) + SDLoc dl) { SDValue Chain = LD->getChain(); EVT LoadedVT = LD->getMemoryVT(); @@ -508,8 +509,11 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, SDValue N1 = LD->getOperand(1); SDValue CPTmpN1_0; SDValue CPTmpN1_1; + if (SelectADDRriS11_2(N1, CPTmpN1_0, CPTmpN1_1) && N1.getNode()->getValueType(0) == MVT::i32) { + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(LoadedVT, Val)) { SDValue TargetConstVal = CurDAG->getTargetConstant(Val, MVT::i32); SDValue TargetConst0 = CurDAG->getTargetConstant(0, MVT::i32); @@ -572,7 +576,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoadZeroExtend64(LoadSDNode *LD, } -SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, SDLoc dl) { SDValue Chain = LD->getChain(); SDValue Base = LD->getBasePtr(); SDValue Offset = LD->getOffset(); @@ -586,6 +590,8 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { bool zextval = (LD->getExtensionType() == ISD::ZEXTLOAD); // Figure out the opcode. + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (LoadedVT == MVT::i64) { if (TII->isValidAutoIncImm(LoadedVT, Val)) Opcode = Hexagon::POST_LDrid; @@ -667,7 +673,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) { SDNode *result; - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); LoadSDNode *LD = cast<LoadSDNode>(N); ISD::MemIndexedMode AM = LD->getAddressingMode(); @@ -682,7 +688,7 @@ SDNode *HexagonDAGToDAGISel::SelectLoad(SDNode *N) { } -SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { +SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, SDLoc dl) { SDValue Chain = ST->getChain(); SDValue Base = ST->getBasePtr(); SDValue Offset = ST->getOffset(); @@ -694,6 +700,8 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { // Offset value must be within representable range // and must have correct alignment properties. + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); if (TII->isValidAutoIncImm(StoredVT, Val)) { SDValue Ops[] = {Base, CurDAG->getTargetConstant(Val, MVT::i32), Value, Chain}; @@ -751,7 +759,7 @@ SDNode *HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, DebugLoc dl) { SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, - DebugLoc dl) { + SDLoc dl) { SDValue Chain = ST->getChain(); SDNode* Const32 = ST->getBasePtr().getNode(); SDValue Value = ST->getValue(); @@ -769,7 +777,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, EVT StoredVT = ST->getMemoryVT(); int64_t Offset = cast<GlobalAddressSDNode>(Base)->getOffset(); if (Offset != 0 && OffsetFitsS11(StoredVT, Offset)) { - MVT PointerTy = TLI.getPointerTy(); + MVT PointerTy = getTargetLowering()->getPointerTy(); const GlobalValue* GV = cast<GlobalAddressSDNode>(Base)->getGlobal(); SDValue TargAddr = @@ -805,7 +813,7 @@ SDNode *HexagonDAGToDAGISel::SelectBaseOffsetStore(StoreSDNode *ST, SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); StoreSDNode *ST = cast<StoreSDNode>(N); ISD::MemIndexedMode AM = ST->getAddressingMode(); @@ -818,7 +826,7 @@ SDNode *HexagonDAGToDAGISel::SelectStore(SDNode *N) { } SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); // // %conv.i = sext i32 %tmp1 to i64 @@ -902,7 +910,7 @@ SDNode *HexagonDAGToDAGISel::SelectMul(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectSelect(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDValue N0 = N->getOperand(0); if (N0.getOpcode() == ISD::SETCC) { SDValue N00 = N0.getOperand(0); @@ -969,7 +977,7 @@ SDNode *HexagonDAGToDAGISel::SelectSelect(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectTruncate(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDValue Shift = N->getOperand(0); // @@ -1082,7 +1090,7 @@ SDNode *HexagonDAGToDAGISel::SelectTruncate(SDNode *N) { SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) == MVT::i32) { SDValue Shl_0 = N->getOperand(0); SDValue Shl_1 = N->getOperand(1); @@ -1158,7 +1166,7 @@ SDNode *HexagonDAGToDAGISel::SelectSHL(SDNode *N) { // We want to preserve all the lower 8-bits and, not just 1 LSB bit. // SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); SDNode *IsIntrinsic = N->getOperand(0).getNode(); if ((IsIntrinsic->getOpcode() == ISD::INTRINSIC_WO_CHAIN)) { unsigned ID = @@ -1201,7 +1209,7 @@ SDNode *HexagonDAGToDAGISel::SelectZeroExtend(SDNode *N) { // and lowering to the actual intrinsic. // SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); unsigned ID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); unsigned IntrinsicWithPred = doesIntrinsicContainPredicate(ID); @@ -1209,6 +1217,8 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { // as at least one of the operands. if (IntrinsicWithPred) { SmallVector<SDValue, 8> Ops; + const HexagonInstrInfo *TII = + static_cast<const HexagonInstrInfo*>(TM.getInstrInfo()); const MCInstrDesc &MCID = TII->get(IntrinsicWithPred); const TargetRegisterInfo *TRI = TM.getRegisterInfo(); @@ -1251,7 +1261,7 @@ SDNode *HexagonDAGToDAGISel::SelectIntrinsicWOChain(SDNode *N) { // Map floating point constant values. // SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N); APFloat APF = CN->getValueAPF(); if (N->getValueType(0) == MVT::f32) { @@ -1271,7 +1281,7 @@ SDNode *HexagonDAGToDAGISel::SelectConstantFP(SDNode *N) { // Map predicate true (encoded as -1 in LLVM) to a XOR. // SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) == MVT::i1) { SDNode* Result; int32_t Val = cast<ConstantSDNode>(N)->getSExtValue(); @@ -1310,7 +1320,7 @@ SDNode *HexagonDAGToDAGISel::SelectConstant(SDNode *N) { // Map add followed by a asr -> asr +=. // SDNode *HexagonDAGToDAGISel::SelectAdd(SDNode *N) { - DebugLoc dl = N->getDebugLoc(); + SDLoc dl(N); if (N->getValueType(0) != MVT::i32) { return SelectCode(N); } @@ -1662,7 +1672,7 @@ bool HexagonDAGToDAGISel::foldGlobalAddressImpl(SDValue &N, SDValue &R, !hasNumUsesBelowThresGA(GA)) return false; R = CurDAG->getTargetGlobalAddress(GA->getGlobal(), - Const->getDebugLoc(), + SDLoc(Const), N.getValueType(), GA->getOffset() + (uint64_t)Const->getSExtValue()); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp index 0e5b8dc..1374179 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp @@ -39,13 +39,24 @@ using namespace llvm; -const unsigned Hexagon_MAX_RET_SIZE = 64; - static cl::opt<bool> EmitJumpTables("hexagon-emit-jump-tables", cl::init(true), cl::Hidden, cl::desc("Control jump table emission on Hexagon target")); -int NumNamedVarArgParams = -1; +namespace { +class HexagonCCState : public CCState { + int NumNamedVarArgParams; + +public: + HexagonCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF, + const TargetMachine &TM, SmallVectorImpl<CCValAssign> &locs, + LLVMContext &C, int NumNamedVarArgParams) + : CCState(CC, isVarArg, MF, TM, locs, C), + NumNamedVarArgParams(NumNamedVarArgParams) {} + + int getNumNamedVarArgParams() const { return NumNamedVarArgParams; } +}; +} // Implement calling convention for Hexagon. static bool @@ -82,12 +93,13 @@ static bool CC_Hexagon_VarArg (unsigned ValNo, MVT ValVT, MVT LocVT, CCValAssign::LocInfo LocInfo, ISD::ArgFlagsTy ArgFlags, CCState &State) { + HexagonCCState &HState = static_cast<HexagonCCState &>(State); // NumNamedVarArgParams can not be zero for a VarArg function. - assert ( (NumNamedVarArgParams > 0) && - "NumNamedVarArgParams is not bigger than zero."); + assert((HState.getNumNamedVarArgParams() > 0) && + "NumNamedVarArgParams is not bigger than zero."); - if ( (int)ValNo < NumNamedVarArgParams ) { + if ((int)ValNo < HState.getNumNamedVarArgParams()) { // Deal with named arguments. return CC_Hexagon(ValNo, ValVT, LocVT, LocInfo, ArgFlags, State); } @@ -285,7 +297,7 @@ const { static SDValue CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, ISD::ArgFlagsTy Flags, SelectionDAG &DAG, - DebugLoc dl) { + SDLoc dl) { SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), @@ -302,7 +314,7 @@ HexagonTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, - DebugLoc dl, SelectionDAG &DAG) const { + SDLoc dl, SelectionDAG &DAG) const { // CCValAssign - represent the assignment of the return value to locations. SmallVector<CCValAssign, 16> RVLocs; @@ -351,7 +363,7 @@ HexagonTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, const SmallVectorImpl<SDValue> &OutVals, SDValue Callee) const { @@ -382,10 +394,10 @@ SDValue HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, SmallVectorImpl<SDValue> &InVals) const { SelectionDAG &DAG = CLI.DAG; - DebugLoc &dl = CLI.DL; - SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; - SmallVector<SDValue, 32> &OutVals = CLI.OutVals; - SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins; + SDLoc &dl = CLI.DL; + SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; + SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; + SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; SDValue Chain = CLI.Chain; SDValue Callee = CLI.Callee; bool &isTailCall = CLI.IsTailCall; @@ -394,13 +406,8 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet(); - // Analyze operands of the call, assigning locations to each operand. - SmallVector<CCValAssign, 16> ArgLocs; - CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), - getTargetMachine(), ArgLocs, *DAG.getContext()); - // Check for varargs. - NumNamedVarArgParams = -1; + int NumNamedVarArgParams = -1; if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Callee)) { const Function* CalleeFn = NULL; @@ -417,6 +424,12 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, } } + // Analyze operands of the call, assigning locations to each operand. + SmallVector<CCValAssign, 16> ArgLocs; + HexagonCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), + getTargetMachine(), ArgLocs, *DAG.getContext(), + NumNamedVarArgParams); + if (NumNamedVarArgParams > 0) CCInfo.AnalyzeCallOperands(Outs, CC_Hexagon_VarArg); else @@ -513,7 +526,8 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, if (!isTailCall) Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(NumBytes, - getPointerTy(), true)); + getPointerTy(), true), + dl); // Build a sequence of copy-to-reg nodes chained together with token // chain and flag operands which copy the outgoing args into registers. @@ -588,7 +602,7 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, // Create the CALLSEQ_END node. Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), - DAG.getIntPtrConstant(0, true), InFlag); + DAG.getIntPtrConstant(0, true), InFlag, dl); InFlag = Chain.getValue(1); // Handle result values, copying them out of physregs into vregs that we @@ -730,7 +744,7 @@ LowerBR_JT(SDValue Op, SelectionDAG &DAG) const SDValue Chain = Op.getOperand(0); SDValue Table = Op.getOperand(1); SDValue Index = Op.getOperand(2); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); unsigned JTI = JT->getIndex(); MachineFunction &MF = DAG.getMachineFunction(); @@ -766,7 +780,7 @@ HexagonTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const { SDValue Chain = Op.getOperand(0); SDValue Size = Op.getOperand(1); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned SPReg = getStackPointerRegisterToSaveRestore(); @@ -812,7 +826,7 @@ HexagonTargetLowering::LowerFormalArguments(SDValue Chain, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { @@ -925,7 +939,7 @@ HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { HexagonMachineFunctionInfo *QFI = MF.getInfo<HexagonMachineFunctionInfo>(); SDValue Addr = DAG.getFrameIndex(QFI->getVarArgsFrameIndex(), MVT::i32); const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); - return DAG.getStore(Op.getOperand(0), Op.getDebugLoc(), Addr, + return DAG.getStore(Op.getOperand(0), SDLoc(Op), Addr, Op.getOperand(1), MachinePointerInfo(SV), false, false, 0); } @@ -937,7 +951,7 @@ HexagonTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { SDValue CC = Op.getOperand(4); SDValue TrueVal = Op.getOperand(2); SDValue FalseVal = Op.getOperand(3); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); SDNode* OpNode = Op.getNode(); EVT SVT = OpNode->getValueType(0); @@ -948,8 +962,7 @@ HexagonTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const { SDValue HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const { EVT ValTy = Op.getValueType(); - - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); SDValue Res; if (CP->isMachineConstantPoolEntry()) @@ -969,7 +982,7 @@ HexagonTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const { MFI->setReturnAddressIsTaken(true); EVT VT = Op.getValueType(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); if (Depth) { SDValue FrameAddr = LowerFRAMEADDR(Op, DAG); @@ -991,7 +1004,7 @@ HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { MFI->setFrameAddressIsTaken(true); EVT VT = Op.getValueType(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, TRI->getFrameRegister(), VT); @@ -1004,7 +1017,7 @@ HexagonTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const { SDValue HexagonTargetLowering::LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const { - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); return DAG.getNode(HexagonISD::BARRIER, dl, MVT::Other, Op.getOperand(0)); } @@ -1014,7 +1027,7 @@ SDValue HexagonTargetLowering::LowerGLOBALADDRESS(SDValue Op, SDValue Result; const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset(); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset); const HexagonTargetObjectFile &TLOF = @@ -1030,7 +1043,7 @@ SDValue HexagonTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const { const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress(); SDValue BA_SD = DAG.getTargetBlockAddress(BA, MVT::i32); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); return DAG.getNode(HexagonISD::CONST32_GP, dl, getPointerTy(), BA_SD); } @@ -1361,7 +1374,6 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine // Increase jump tables cutover to 5, was 4. setMinimumJumpTableEntries(5); - setOperationAction(ISD::BR_CC, MVT::Other, Expand); setOperationAction(ISD::BR_CC, MVT::f32, Expand); setOperationAction(ISD::BR_CC, MVT::f64, Expand); setOperationAction(ISD::BR_CC, MVT::i1, Expand); @@ -1429,11 +1441,6 @@ HexagonTargetLowering::HexagonTargetLowering(HexagonTargetMachine setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand); setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand); - setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand); - setOperationAction(ISD::EHSELECTION, MVT::i64, Expand); - setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand); - setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); - setOperationAction(ISD::EH_RETURN, MVT::Other, Custom); if (TM.getSubtargetImpl()->isSubtargetV2()) { @@ -1510,12 +1517,25 @@ bool HexagonTargetLowering::isTruncateFree(EVT VT1, EVT VT2) const { return ((VT1.getSimpleVT() == MVT::i64) && (VT2.getSimpleVT() == MVT::i32)); } +bool +HexagonTargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const { + // Assuming the caller does not have either a signext or zeroext modifier, and + // only one value is accepted, any reasonable truncation is allowed. + if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy()) + return false; + + // FIXME: in principle up to 64-bit could be made safe, but it would be very + // fragile at the moment: any support for multiple value returns would be + // liable to disallow tail calls involving i64 -> iN truncation in many cases. + return Ty1->getPrimitiveSizeInBits() <= 32; +} + SDValue HexagonTargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const { SDValue Chain = Op.getOperand(0); SDValue Offset = Op.getOperand(1); SDValue Handler = Op.getOperand(2); - DebugLoc dl = Op.getDebugLoc(); + SDLoc dl(Op); // Mark function as containing a call to EH_RETURN. HexagonMachineFunctionInfo *FuncInfo = @@ -1591,11 +1611,11 @@ const { std::pair<unsigned, const TargetRegisterClass*> HexagonTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, - EVT VT) const { + MVT VT) const { if (Constraint.size() == 1) { switch (Constraint[0]) { case 'r': // R0-R31 - switch (VT.getSimpleVT().SimpleTy) { + switch (VT.SimpleTy) { default: llvm_unreachable("getRegForInlineAsmConstraint Unhandled data type"); case MVT::i32: diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.h b/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.h index bb1acc1..73da226 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonISelLowering.h @@ -95,6 +95,8 @@ namespace llvm { virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const; virtual bool isTruncateFree(EVT VT1, EVT VT2) const; + virtual bool allowTruncateForTailCall(Type *Ty1, Type *Ty2) const; + virtual SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const; virtual const char *getTargetNodeName(unsigned Opcode) const; @@ -106,7 +108,7 @@ namespace llvm { SDValue LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const; SDValue LowerGLOBALADDRESS(SDValue Op, SelectionDAG &DAG) const; SDValue LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const; @@ -117,7 +119,7 @@ namespace llvm { SDValue LowerCallResult(SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, - DebugLoc dl, SelectionDAG &DAG, + SDLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, const SmallVectorImpl<SDValue> &OutVals, SDValue Callee) const; @@ -131,7 +133,7 @@ namespace llvm { CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, - DebugLoc dl, SelectionDAG &DAG) const; + SDLoc dl, SelectionDAG &DAG) const; virtual MachineBasicBlock *EmitInstrWithCustomInserter(MachineInstr *MI, @@ -139,8 +141,11 @@ namespace llvm { SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) const; SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) const; - virtual EVT getSetCCResultType(EVT VT) const { - return MVT::i1; + virtual EVT getSetCCResultType(LLVMContext &C, EVT VT) const { + if (!VT.isVector()) + return MVT::i1; + else + return EVT::getVectorVT(C, MVT::i1, VT.getVectorNumElements()); } virtual bool getPostIndexedAddressParts(SDNode *N, SDNode *Op, @@ -150,7 +155,7 @@ namespace llvm { std::pair<unsigned, const TargetRegisterClass*> getRegForInlineAsmConstraint(const std::string &Constraint, - EVT VT) const; + MVT VT) const; // Intrinsics virtual SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrFormats.td b/contrib/llvm/lib/Target/Hexagon/HexagonInstrFormats.td index 587fa7d..d25bfa8 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrFormats.td +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrFormats.td @@ -54,6 +54,7 @@ def AbsoluteSet : AddrModeType<2>; // Absolute set addressing mode def BaseImmOffset : AddrModeType<3>; // Indirect with offset def BaseLongOffset : AddrModeType<4>; // Indirect with long offset def BaseRegOffset : AddrModeType<5>; // Indirect with register offset +def PostInc : AddrModeType<6>; // Post increment addressing mode class MemAccessSize<bits<3> value> { bits<3> Value = value; @@ -62,7 +63,7 @@ class MemAccessSize<bits<3> value> { def NoMemAccess : MemAccessSize<0>;// Not a memory acces instruction. def ByteAccess : MemAccessSize<1>;// Byte access instruction (memb). def HalfWordAccess : MemAccessSize<2>;// Half word access instruction (memh). -def WordAccess : MemAccessSize<3>;// Word access instrution (memw). +def WordAccess : MemAccessSize<3>;// Word access instruction (memw). def DoubleWordAccess : MemAccessSize<4>;// Double word access instruction (memd) @@ -157,6 +158,7 @@ class InstHexagon<dag outs, dag ins, string asmstr, list<dag> pattern, string CextOpcode = ""; string PredSense = ""; string PNewValue = ""; + string NValueST = ""; // Set to "true" for new-value stores. string InputType = ""; // Input is "imm" or "reg" type. string isMEMri = "false"; // Set to "true" for load/store with MEMri operand. string isFloat = "false"; // Set to "true" for the floating-point load/store. @@ -165,6 +167,7 @@ class InstHexagon<dag outs, dag ins, string asmstr, list<dag> pattern, let PredSense = !if(isPredicated, !if(isPredicatedFalse, "false", "true"), ""); let PNewValue = !if(isPredicatedNew, "new", ""); + let NValueST = !if(isNVStore, "true", "false"); // *** Must match MCTargetDesc/HexagonBaseInfo.h *** } diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp index f114170..6b97609 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp @@ -26,7 +26,7 @@ #include "llvm/Support/Debug.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/raw_ostream.h" -#define GET_INSTRINFO_CTOR +#define GET_INSTRINFO_CTOR_DTOR #define GET_INSTRMAP_INFO #include "HexagonGenInstrInfo.inc" #include "HexagonGenDFAPacketizer.inc" @@ -55,10 +55,12 @@ const int Hexagon_MEMH_AUTOINC_MIN = -16; const int Hexagon_MEMB_AUTOINC_MAX = 7; const int Hexagon_MEMB_AUTOINC_MIN = -8; +// Pin the vtable to this file. +void HexagonInstrInfo::anchor() {} HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST) : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP), - RI(ST, *this), Subtarget(ST) { + RI(ST), Subtarget(ST) { } @@ -558,16 +560,6 @@ MachineInstr *HexagonInstrInfo::foldMemoryOperandImpl(MachineFunction &MF, return(0); } -MachineInstr* -HexagonInstrInfo::emitFrameIndexDebugValue(MachineFunction &MF, - int FrameIx, uint64_t Offset, - const MDNode *MDPtr, - DebugLoc DL) const { - MachineInstrBuilder MIB = BuildMI(MF, DL, get(Hexagon::DBG_VALUE)) - .addImm(0).addImm(Offset).addMetadata(MDPtr); - return &*MIB; -} - unsigned HexagonInstrInfo::createVR(MachineFunction* MF, MVT VT) const { MachineRegisterInfo &RegInfo = MF->getRegInfo(); @@ -630,188 +622,6 @@ bool HexagonInstrInfo::isBranch (const MachineInstr *MI) const { return MI->getDesc().isBranch(); } -bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { - switch (MI->getOpcode()) { - default: return false; - // Store Byte - case Hexagon::STrib_nv_V4: - case Hexagon::STrib_indexed_nv_V4: - case Hexagon::STrib_indexed_shl_nv_V4: - case Hexagon::STrib_shl_nv_V4: - case Hexagon::STb_GP_nv_V4: - case Hexagon::POST_STbri_nv_V4: - case Hexagon::STrib_cPt_nv_V4: - case Hexagon::STrib_cdnPt_nv_V4: - case Hexagon::STrib_cNotPt_nv_V4: - case Hexagon::STrib_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_cPt_nv_V4: - case Hexagon::STrib_indexed_cdnPt_nv_V4: - case Hexagon::STrib_indexed_cNotPt_nv_V4: - case Hexagon::STrib_indexed_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cPt_nv_V4: - case Hexagon::STrib_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_STbri_cPt_nv_V4: - case Hexagon::POST_STbri_cdnPt_nv_V4: - case Hexagon::POST_STbri_cNotPt_nv_V4: - case Hexagon::POST_STbri_cdnNotPt_nv_V4: - case Hexagon::STb_GP_cPt_nv_V4: - case Hexagon::STb_GP_cNotPt_nv_V4: - case Hexagon::STb_GP_cdnPt_nv_V4: - case Hexagon::STb_GP_cdnNotPt_nv_V4: - case Hexagon::STrib_abs_nv_V4: - case Hexagon::STrib_abs_cPt_nv_V4: - case Hexagon::STrib_abs_cdnPt_nv_V4: - case Hexagon::STrib_abs_cNotPt_nv_V4: - case Hexagon::STrib_abs_cdnNotPt_nv_V4: - - // Store Halfword - case Hexagon::STrih_nv_V4: - case Hexagon::STrih_indexed_nv_V4: - case Hexagon::STrih_indexed_shl_nv_V4: - case Hexagon::STrih_shl_nv_V4: - case Hexagon::STh_GP_nv_V4: - case Hexagon::POST_SThri_nv_V4: - case Hexagon::STrih_cPt_nv_V4: - case Hexagon::STrih_cdnPt_nv_V4: - case Hexagon::STrih_cNotPt_nv_V4: - case Hexagon::STrih_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_cPt_nv_V4: - case Hexagon::STrih_indexed_cdnPt_nv_V4: - case Hexagon::STrih_indexed_cNotPt_nv_V4: - case Hexagon::STrih_indexed_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_shl_cPt_nv_V4: - case Hexagon::STrih_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4: - case Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_SThri_cPt_nv_V4: - case Hexagon::POST_SThri_cdnPt_nv_V4: - case Hexagon::POST_SThri_cNotPt_nv_V4: - case Hexagon::POST_SThri_cdnNotPt_nv_V4: - case Hexagon::STh_GP_cPt_nv_V4: - case Hexagon::STh_GP_cNotPt_nv_V4: - case Hexagon::STh_GP_cdnPt_nv_V4: - case Hexagon::STh_GP_cdnNotPt_nv_V4: - case Hexagon::STrih_abs_nv_V4: - case Hexagon::STrih_abs_cPt_nv_V4: - case Hexagon::STrih_abs_cdnPt_nv_V4: - case Hexagon::STrih_abs_cNotPt_nv_V4: - case Hexagon::STrih_abs_cdnNotPt_nv_V4: - - // Store Word - case Hexagon::STriw_nv_V4: - case Hexagon::STriw_indexed_nv_V4: - case Hexagon::STriw_indexed_shl_nv_V4: - case Hexagon::STriw_shl_nv_V4: - case Hexagon::STw_GP_nv_V4: - case Hexagon::POST_STwri_nv_V4: - case Hexagon::STriw_cPt_nv_V4: - case Hexagon::STriw_cdnPt_nv_V4: - case Hexagon::STriw_cNotPt_nv_V4: - case Hexagon::STriw_cdnNotPt_nv_V4: - case Hexagon::STriw_indexed_cPt_nv_V4: - case Hexagon::STriw_indexed_cdnPt_nv_V4: - case Hexagon::STriw_indexed_cNotPt_nv_V4: - case Hexagon::STriw_indexed_cdnNotPt_nv_V4: - case Hexagon::STriw_indexed_shl_cPt_nv_V4: - case Hexagon::STriw_indexed_shl_cdnPt_nv_V4: - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4: - case Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::POST_STwri_cPt_nv_V4: - case Hexagon::POST_STwri_cdnPt_nv_V4: - case Hexagon::POST_STwri_cNotPt_nv_V4: - case Hexagon::POST_STwri_cdnNotPt_nv_V4: - case Hexagon::STw_GP_cPt_nv_V4: - case Hexagon::STw_GP_cNotPt_nv_V4: - case Hexagon::STw_GP_cdnPt_nv_V4: - case Hexagon::STw_GP_cdnNotPt_nv_V4: - case Hexagon::STriw_abs_nv_V4: - case Hexagon::STriw_abs_cPt_nv_V4: - case Hexagon::STriw_abs_cdnPt_nv_V4: - case Hexagon::STriw_abs_cNotPt_nv_V4: - case Hexagon::STriw_abs_cdnNotPt_nv_V4: - return true; - } -} - -bool HexagonInstrInfo::isPostIncrement (const MachineInstr* MI) const { - switch (MI->getOpcode()) - { - default: return false; - // Load Byte - case Hexagon::POST_LDrib: - case Hexagon::POST_LDrib_cPt: - case Hexagon::POST_LDrib_cNotPt: - case Hexagon::POST_LDrib_cdnPt_V4: - case Hexagon::POST_LDrib_cdnNotPt_V4: - - // Load unsigned byte - case Hexagon::POST_LDriub: - case Hexagon::POST_LDriub_cPt: - case Hexagon::POST_LDriub_cNotPt: - case Hexagon::POST_LDriub_cdnPt_V4: - case Hexagon::POST_LDriub_cdnNotPt_V4: - - // Load halfword - case Hexagon::POST_LDrih: - case Hexagon::POST_LDrih_cPt: - case Hexagon::POST_LDrih_cNotPt: - case Hexagon::POST_LDrih_cdnPt_V4: - case Hexagon::POST_LDrih_cdnNotPt_V4: - - // Load unsigned halfword - case Hexagon::POST_LDriuh: - case Hexagon::POST_LDriuh_cPt: - case Hexagon::POST_LDriuh_cNotPt: - case Hexagon::POST_LDriuh_cdnPt_V4: - case Hexagon::POST_LDriuh_cdnNotPt_V4: - - // Load word - case Hexagon::POST_LDriw: - case Hexagon::POST_LDriw_cPt: - case Hexagon::POST_LDriw_cNotPt: - case Hexagon::POST_LDriw_cdnPt_V4: - case Hexagon::POST_LDriw_cdnNotPt_V4: - - // Load double word - case Hexagon::POST_LDrid: - case Hexagon::POST_LDrid_cPt: - case Hexagon::POST_LDrid_cNotPt: - case Hexagon::POST_LDrid_cdnPt_V4: - case Hexagon::POST_LDrid_cdnNotPt_V4: - - // Store byte - case Hexagon::POST_STbri: - case Hexagon::POST_STbri_cPt: - case Hexagon::POST_STbri_cNotPt: - case Hexagon::POST_STbri_cdnPt_V4: - case Hexagon::POST_STbri_cdnNotPt_V4: - - // Store halfword - case Hexagon::POST_SThri: - case Hexagon::POST_SThri_cPt: - case Hexagon::POST_SThri_cNotPt: - case Hexagon::POST_SThri_cdnPt_V4: - case Hexagon::POST_SThri_cdnNotPt_V4: - - // Store word - case Hexagon::POST_STwri: - case Hexagon::POST_STwri_cPt: - case Hexagon::POST_STwri_cNotPt: - case Hexagon::POST_STwri_cdnPt_V4: - case Hexagon::POST_STwri_cdnNotPt_V4: - - // Store double word - case Hexagon::POST_STdri: - case Hexagon::POST_STdri_cPt: - case Hexagon::POST_STdri_cNotPt: - case Hexagon::POST_STdri_cdnPt_V4: - case Hexagon::POST_STdri_cdnNotPt_V4: - return true; - } -} - bool HexagonInstrInfo::isNewValueInst(const MachineInstr *MI) const { if (isNewValueJump(MI)) return true; @@ -917,19 +727,6 @@ bool HexagonInstrInfo::isPredicable(MachineInstr *MI) const { // cPt ---> cNotPt // cNotPt ---> cPt // -// however, these inversiones are NOT included: -// -// cdnPt -X-> cdnNotPt -// cdnNotPt -X-> cdnPt -// cPt_nv -X-> cNotPt_nv (new value stores) -// cNotPt_nv -X-> cPt_nv (new value stores) -// -// because only the following transformations are allowed: -// -// cNotPt ---> cdnNotPt -// cPt ---> cdnPt -// cNotPt ---> cNotPt_nv -// cPt ---> cPt_nv unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { int InvPredOpcode; InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc) @@ -939,332 +736,12 @@ unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { switch(Opc) { default: llvm_unreachable("Unexpected predicated instruction"); - case Hexagon::TFR_cPt: - return Hexagon::TFR_cNotPt; - case Hexagon::TFR_cNotPt: - return Hexagon::TFR_cPt; - - case Hexagon::TFRI_cPt: - return Hexagon::TFRI_cNotPt; - case Hexagon::TFRI_cNotPt: - return Hexagon::TFRI_cPt; - - case Hexagon::JMP_t: - return Hexagon::JMP_f; - case Hexagon::JMP_f: - return Hexagon::JMP_t; - - case Hexagon::ADD_ri_cPt: - return Hexagon::ADD_ri_cNotPt; - case Hexagon::ADD_ri_cNotPt: - return Hexagon::ADD_ri_cPt; - - case Hexagon::ADD_rr_cPt: - return Hexagon::ADD_rr_cNotPt; - case Hexagon::ADD_rr_cNotPt: - return Hexagon::ADD_rr_cPt; - - case Hexagon::XOR_rr_cPt: - return Hexagon::XOR_rr_cNotPt; - case Hexagon::XOR_rr_cNotPt: - return Hexagon::XOR_rr_cPt; - - case Hexagon::AND_rr_cPt: - return Hexagon::AND_rr_cNotPt; - case Hexagon::AND_rr_cNotPt: - return Hexagon::AND_rr_cPt; - - case Hexagon::OR_rr_cPt: - return Hexagon::OR_rr_cNotPt; - case Hexagon::OR_rr_cNotPt: - return Hexagon::OR_rr_cPt; - - case Hexagon::SUB_rr_cPt: - return Hexagon::SUB_rr_cNotPt; - case Hexagon::SUB_rr_cNotPt: - return Hexagon::SUB_rr_cPt; - case Hexagon::COMBINE_rr_cPt: return Hexagon::COMBINE_rr_cNotPt; case Hexagon::COMBINE_rr_cNotPt: return Hexagon::COMBINE_rr_cPt; - case Hexagon::ASLH_cPt_V4: - return Hexagon::ASLH_cNotPt_V4; - case Hexagon::ASLH_cNotPt_V4: - return Hexagon::ASLH_cPt_V4; - - case Hexagon::ASRH_cPt_V4: - return Hexagon::ASRH_cNotPt_V4; - case Hexagon::ASRH_cNotPt_V4: - return Hexagon::ASRH_cPt_V4; - - case Hexagon::SXTB_cPt_V4: - return Hexagon::SXTB_cNotPt_V4; - case Hexagon::SXTB_cNotPt_V4: - return Hexagon::SXTB_cPt_V4; - - case Hexagon::SXTH_cPt_V4: - return Hexagon::SXTH_cNotPt_V4; - case Hexagon::SXTH_cNotPt_V4: - return Hexagon::SXTH_cPt_V4; - - case Hexagon::ZXTB_cPt_V4: - return Hexagon::ZXTB_cNotPt_V4; - case Hexagon::ZXTB_cNotPt_V4: - return Hexagon::ZXTB_cPt_V4; - - case Hexagon::ZXTH_cPt_V4: - return Hexagon::ZXTH_cNotPt_V4; - case Hexagon::ZXTH_cNotPt_V4: - return Hexagon::ZXTH_cPt_V4; - - - case Hexagon::JMPR_t: - return Hexagon::JMPR_f; - case Hexagon::JMPR_f: - return Hexagon::JMPR_t; - - // V4 indexed+scaled load. - case Hexagon::LDrid_indexed_shl_cPt_V4: - return Hexagon::LDrid_indexed_shl_cNotPt_V4; - case Hexagon::LDrid_indexed_shl_cNotPt_V4: - return Hexagon::LDrid_indexed_shl_cPt_V4; - - case Hexagon::LDrib_indexed_shl_cPt_V4: - return Hexagon::LDrib_indexed_shl_cNotPt_V4; - case Hexagon::LDrib_indexed_shl_cNotPt_V4: - return Hexagon::LDrib_indexed_shl_cPt_V4; - - case Hexagon::LDriub_indexed_shl_cPt_V4: - return Hexagon::LDriub_indexed_shl_cNotPt_V4; - case Hexagon::LDriub_indexed_shl_cNotPt_V4: - return Hexagon::LDriub_indexed_shl_cPt_V4; - - case Hexagon::LDrih_indexed_shl_cPt_V4: - return Hexagon::LDrih_indexed_shl_cNotPt_V4; - case Hexagon::LDrih_indexed_shl_cNotPt_V4: - return Hexagon::LDrih_indexed_shl_cPt_V4; - - case Hexagon::LDriuh_indexed_shl_cPt_V4: - return Hexagon::LDriuh_indexed_shl_cNotPt_V4; - case Hexagon::LDriuh_indexed_shl_cNotPt_V4: - return Hexagon::LDriuh_indexed_shl_cPt_V4; - - case Hexagon::LDriw_indexed_shl_cPt_V4: - return Hexagon::LDriw_indexed_shl_cNotPt_V4; - case Hexagon::LDriw_indexed_shl_cNotPt_V4: - return Hexagon::LDriw_indexed_shl_cPt_V4; - - // Byte. - case Hexagon::POST_STbri_cPt: - return Hexagon::POST_STbri_cNotPt; - case Hexagon::POST_STbri_cNotPt: - return Hexagon::POST_STbri_cPt; - - case Hexagon::STrib_cPt: - return Hexagon::STrib_cNotPt; - case Hexagon::STrib_cNotPt: - return Hexagon::STrib_cPt; - - case Hexagon::STrib_indexed_cPt: - return Hexagon::STrib_indexed_cNotPt; - case Hexagon::STrib_indexed_cNotPt: - return Hexagon::STrib_indexed_cPt; - - case Hexagon::STrib_imm_cPt_V4: - return Hexagon::STrib_imm_cNotPt_V4; - case Hexagon::STrib_imm_cNotPt_V4: - return Hexagon::STrib_imm_cPt_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4: - return Hexagon::STrib_indexed_shl_cNotPt_V4; - case Hexagon::STrib_indexed_shl_cNotPt_V4: - return Hexagon::STrib_indexed_shl_cPt_V4; - - // Halfword. - case Hexagon::POST_SThri_cPt: - return Hexagon::POST_SThri_cNotPt; - case Hexagon::POST_SThri_cNotPt: - return Hexagon::POST_SThri_cPt; - - case Hexagon::STrih_cPt: - return Hexagon::STrih_cNotPt; - case Hexagon::STrih_cNotPt: - return Hexagon::STrih_cPt; - - case Hexagon::STrih_indexed_cPt: - return Hexagon::STrih_indexed_cNotPt; - case Hexagon::STrih_indexed_cNotPt: - return Hexagon::STrih_indexed_cPt; - - case Hexagon::STrih_imm_cPt_V4: - return Hexagon::STrih_imm_cNotPt_V4; - case Hexagon::STrih_imm_cNotPt_V4: - return Hexagon::STrih_imm_cPt_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4: - return Hexagon::STrih_indexed_shl_cNotPt_V4; - case Hexagon::STrih_indexed_shl_cNotPt_V4: - return Hexagon::STrih_indexed_shl_cPt_V4; - - // Word. - case Hexagon::POST_STwri_cPt: - return Hexagon::POST_STwri_cNotPt; - case Hexagon::POST_STwri_cNotPt: - return Hexagon::POST_STwri_cPt; - - case Hexagon::STriw_cPt: - return Hexagon::STriw_cNotPt; - case Hexagon::STriw_cNotPt: - return Hexagon::STriw_cPt; - - case Hexagon::STriw_indexed_cPt: - return Hexagon::STriw_indexed_cNotPt; - case Hexagon::STriw_indexed_cNotPt: - return Hexagon::STriw_indexed_cPt; - - case Hexagon::STriw_indexed_shl_cPt_V4: - return Hexagon::STriw_indexed_shl_cNotPt_V4; - case Hexagon::STriw_indexed_shl_cNotPt_V4: - return Hexagon::STriw_indexed_shl_cPt_V4; - - case Hexagon::STriw_imm_cPt_V4: - return Hexagon::STriw_imm_cNotPt_V4; - case Hexagon::STriw_imm_cNotPt_V4: - return Hexagon::STriw_imm_cPt_V4; - - // Double word. - case Hexagon::POST_STdri_cPt: - return Hexagon::POST_STdri_cNotPt; - case Hexagon::POST_STdri_cNotPt: - return Hexagon::POST_STdri_cPt; - - case Hexagon::STrid_cPt: - return Hexagon::STrid_cNotPt; - case Hexagon::STrid_cNotPt: - return Hexagon::STrid_cPt; - - case Hexagon::STrid_indexed_cPt: - return Hexagon::STrid_indexed_cNotPt; - case Hexagon::STrid_indexed_cNotPt: - return Hexagon::STrid_indexed_cPt; - - case Hexagon::STrid_indexed_shl_cPt_V4: - return Hexagon::STrid_indexed_shl_cNotPt_V4; - case Hexagon::STrid_indexed_shl_cNotPt_V4: - return Hexagon::STrid_indexed_shl_cPt_V4; - - // V4 Store to global address. - case Hexagon::STd_GP_cPt_V4: - return Hexagon::STd_GP_cNotPt_V4; - case Hexagon::STd_GP_cNotPt_V4: - return Hexagon::STd_GP_cPt_V4; - - case Hexagon::STb_GP_cPt_V4: - return Hexagon::STb_GP_cNotPt_V4; - case Hexagon::STb_GP_cNotPt_V4: - return Hexagon::STb_GP_cPt_V4; - - case Hexagon::STh_GP_cPt_V4: - return Hexagon::STh_GP_cNotPt_V4; - case Hexagon::STh_GP_cNotPt_V4: - return Hexagon::STh_GP_cPt_V4; - - case Hexagon::STw_GP_cPt_V4: - return Hexagon::STw_GP_cNotPt_V4; - case Hexagon::STw_GP_cNotPt_V4: - return Hexagon::STw_GP_cPt_V4; - - // Load. - case Hexagon::LDrid_cPt: - return Hexagon::LDrid_cNotPt; - case Hexagon::LDrid_cNotPt: - return Hexagon::LDrid_cPt; - - case Hexagon::LDriw_cPt: - return Hexagon::LDriw_cNotPt; - case Hexagon::LDriw_cNotPt: - return Hexagon::LDriw_cPt; - - case Hexagon::LDrih_cPt: - return Hexagon::LDrih_cNotPt; - case Hexagon::LDrih_cNotPt: - return Hexagon::LDrih_cPt; - - case Hexagon::LDriuh_cPt: - return Hexagon::LDriuh_cNotPt; - case Hexagon::LDriuh_cNotPt: - return Hexagon::LDriuh_cPt; - - case Hexagon::LDrib_cPt: - return Hexagon::LDrib_cNotPt; - case Hexagon::LDrib_cNotPt: - return Hexagon::LDrib_cPt; - - case Hexagon::LDriub_cPt: - return Hexagon::LDriub_cNotPt; - case Hexagon::LDriub_cNotPt: - return Hexagon::LDriub_cPt; - - // Load Indexed. - case Hexagon::LDrid_indexed_cPt: - return Hexagon::LDrid_indexed_cNotPt; - case Hexagon::LDrid_indexed_cNotPt: - return Hexagon::LDrid_indexed_cPt; - - case Hexagon::LDriw_indexed_cPt: - return Hexagon::LDriw_indexed_cNotPt; - case Hexagon::LDriw_indexed_cNotPt: - return Hexagon::LDriw_indexed_cPt; - - case Hexagon::LDrih_indexed_cPt: - return Hexagon::LDrih_indexed_cNotPt; - case Hexagon::LDrih_indexed_cNotPt: - return Hexagon::LDrih_indexed_cPt; - - case Hexagon::LDriuh_indexed_cPt: - return Hexagon::LDriuh_indexed_cNotPt; - case Hexagon::LDriuh_indexed_cNotPt: - return Hexagon::LDriuh_indexed_cPt; - - case Hexagon::LDrib_indexed_cPt: - return Hexagon::LDrib_indexed_cNotPt; - case Hexagon::LDrib_indexed_cNotPt: - return Hexagon::LDrib_indexed_cPt; - - case Hexagon::LDriub_indexed_cPt: - return Hexagon::LDriub_indexed_cNotPt; - case Hexagon::LDriub_indexed_cNotPt: - return Hexagon::LDriub_indexed_cPt; - - // Post Inc Load. - case Hexagon::POST_LDrid_cPt: - return Hexagon::POST_LDrid_cNotPt; - case Hexagon::POST_LDriw_cNotPt: - return Hexagon::POST_LDriw_cPt; - - case Hexagon::POST_LDrih_cPt: - return Hexagon::POST_LDrih_cNotPt; - case Hexagon::POST_LDrih_cNotPt: - return Hexagon::POST_LDrih_cPt; - - case Hexagon::POST_LDriuh_cPt: - return Hexagon::POST_LDriuh_cNotPt; - case Hexagon::POST_LDriuh_cNotPt: - return Hexagon::POST_LDriuh_cPt; - - case Hexagon::POST_LDrib_cPt: - return Hexagon::POST_LDrib_cNotPt; - case Hexagon::POST_LDrib_cNotPt: - return Hexagon::POST_LDrib_cPt; - - case Hexagon::POST_LDriub_cPt: - return Hexagon::POST_LDriub_cNotPt; - case Hexagon::POST_LDriub_cNotPt: - return Hexagon::POST_LDriub_cPt; - - // Dealloc_return. + // Dealloc_return. case Hexagon::DEALLOC_RET_cPt_V4: return Hexagon::DEALLOC_RET_cNotPt_V4; case Hexagon::DEALLOC_RET_cNotPt_V4: @@ -1272,6 +749,18 @@ unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const { } } +// New Value Store instructions. +bool HexagonInstrInfo::isNewValueStore(const MachineInstr *MI) const { + const uint64_t F = MI->getDesc().TSFlags; + + return ((F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask); +} + +bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const { + const uint64_t F = get(Opcode).TSFlags; + + return ((F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask); +} int HexagonInstrInfo:: getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { @@ -1285,218 +774,21 @@ getMatchingCondBranchOpcode(int Opc, bool invertPredicate) const { // This switch case will be removed once all the instructions have been // modified to use relation maps. switch(Opc) { - case Hexagon::TFR: - return !invertPredicate ? Hexagon::TFR_cPt : - Hexagon::TFR_cNotPt; case Hexagon::TFRI_f: return !invertPredicate ? Hexagon::TFRI_cPt_f : Hexagon::TFRI_cNotPt_f; - case Hexagon::TFRI: - return !invertPredicate ? Hexagon::TFRI_cPt : - Hexagon::TFRI_cNotPt; - case Hexagon::JMP: - return !invertPredicate ? Hexagon::JMP_t : - Hexagon::JMP_f; - case Hexagon::COMBINE_rr: return !invertPredicate ? Hexagon::COMBINE_rr_cPt : Hexagon::COMBINE_rr_cNotPt; - case Hexagon::ASLH: - return !invertPredicate ? Hexagon::ASLH_cPt_V4 : - Hexagon::ASLH_cNotPt_V4; - case Hexagon::ASRH: - return !invertPredicate ? Hexagon::ASRH_cPt_V4 : - Hexagon::ASRH_cNotPt_V4; - case Hexagon::SXTB: - return !invertPredicate ? Hexagon::SXTB_cPt_V4 : - Hexagon::SXTB_cNotPt_V4; - case Hexagon::SXTH: - return !invertPredicate ? Hexagon::SXTH_cPt_V4 : - Hexagon::SXTH_cNotPt_V4; - case Hexagon::ZXTB: - return !invertPredicate ? Hexagon::ZXTB_cPt_V4 : - Hexagon::ZXTB_cNotPt_V4; - case Hexagon::ZXTH: - return !invertPredicate ? Hexagon::ZXTH_cPt_V4 : - Hexagon::ZXTH_cNotPt_V4; - - case Hexagon::JMPR: - return !invertPredicate ? Hexagon::JMPR_t : - Hexagon::JMPR_f; - - // V4 indexed+scaled load. - case Hexagon::LDrid_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrid_indexed_shl_cPt_V4 : - Hexagon::LDrid_indexed_shl_cNotPt_V4; - case Hexagon::LDrib_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrib_indexed_shl_cPt_V4 : - Hexagon::LDrib_indexed_shl_cNotPt_V4; - case Hexagon::LDriub_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriub_indexed_shl_cPt_V4 : - Hexagon::LDriub_indexed_shl_cNotPt_V4; - case Hexagon::LDrih_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDrih_indexed_shl_cPt_V4 : - Hexagon::LDrih_indexed_shl_cNotPt_V4; - case Hexagon::LDriuh_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriuh_indexed_shl_cPt_V4 : - Hexagon::LDriuh_indexed_shl_cNotPt_V4; - case Hexagon::LDriw_indexed_shl_V4: - return !invertPredicate ? Hexagon::LDriw_indexed_shl_cPt_V4 : - Hexagon::LDriw_indexed_shl_cNotPt_V4; - - // V4 Load from global address - case Hexagon::LDd_GP_V4: - return !invertPredicate ? Hexagon::LDd_GP_cPt_V4 : - Hexagon::LDd_GP_cNotPt_V4; - case Hexagon::LDb_GP_V4: - return !invertPredicate ? Hexagon::LDb_GP_cPt_V4 : - Hexagon::LDb_GP_cNotPt_V4; - case Hexagon::LDub_GP_V4: - return !invertPredicate ? Hexagon::LDub_GP_cPt_V4 : - Hexagon::LDub_GP_cNotPt_V4; - case Hexagon::LDh_GP_V4: - return !invertPredicate ? Hexagon::LDh_GP_cPt_V4 : - Hexagon::LDh_GP_cNotPt_V4; - case Hexagon::LDuh_GP_V4: - return !invertPredicate ? Hexagon::LDuh_GP_cPt_V4 : - Hexagon::LDuh_GP_cNotPt_V4; - case Hexagon::LDw_GP_V4: - return !invertPredicate ? Hexagon::LDw_GP_cPt_V4 : - Hexagon::LDw_GP_cNotPt_V4; - - // Byte. - case Hexagon::POST_STbri: - return !invertPredicate ? Hexagon::POST_STbri_cPt : - Hexagon::POST_STbri_cNotPt; - case Hexagon::STrib: - return !invertPredicate ? Hexagon::STrib_cPt : - Hexagon::STrib_cNotPt; - case Hexagon::STrib_indexed: - return !invertPredicate ? Hexagon::STrib_indexed_cPt : - Hexagon::STrib_indexed_cNotPt; - case Hexagon::STrib_imm_V4: - return !invertPredicate ? Hexagon::STrib_imm_cPt_V4 : - Hexagon::STrib_imm_cNotPt_V4; - case Hexagon::STrib_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrib_indexed_shl_cPt_V4 : - Hexagon::STrib_indexed_shl_cNotPt_V4; - // Halfword. - case Hexagon::POST_SThri: - return !invertPredicate ? Hexagon::POST_SThri_cPt : - Hexagon::POST_SThri_cNotPt; - case Hexagon::STrih: - return !invertPredicate ? Hexagon::STrih_cPt : - Hexagon::STrih_cNotPt; - case Hexagon::STrih_indexed: - return !invertPredicate ? Hexagon::STrih_indexed_cPt : - Hexagon::STrih_indexed_cNotPt; - case Hexagon::STrih_imm_V4: - return !invertPredicate ? Hexagon::STrih_imm_cPt_V4 : - Hexagon::STrih_imm_cNotPt_V4; - case Hexagon::STrih_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrih_indexed_shl_cPt_V4 : - Hexagon::STrih_indexed_shl_cNotPt_V4; + // Word. - case Hexagon::POST_STwri: - return !invertPredicate ? Hexagon::POST_STwri_cPt : - Hexagon::POST_STwri_cNotPt; - case Hexagon::STriw: + case Hexagon::STriw_f: return !invertPredicate ? Hexagon::STriw_cPt : Hexagon::STriw_cNotPt; - case Hexagon::STriw_indexed: + case Hexagon::STriw_indexed_f: return !invertPredicate ? Hexagon::STriw_indexed_cPt : Hexagon::STriw_indexed_cNotPt; - case Hexagon::STriw_indexed_shl_V4: - return !invertPredicate ? Hexagon::STriw_indexed_shl_cPt_V4 : - Hexagon::STriw_indexed_shl_cNotPt_V4; - case Hexagon::STriw_imm_V4: - return !invertPredicate ? Hexagon::STriw_imm_cPt_V4 : - Hexagon::STriw_imm_cNotPt_V4; - // Double word. - case Hexagon::POST_STdri: - return !invertPredicate ? Hexagon::POST_STdri_cPt : - Hexagon::POST_STdri_cNotPt; - case Hexagon::STrid: - return !invertPredicate ? Hexagon::STrid_cPt : - Hexagon::STrid_cNotPt; - case Hexagon::STrid_indexed: - return !invertPredicate ? Hexagon::STrid_indexed_cPt : - Hexagon::STrid_indexed_cNotPt; - case Hexagon::STrid_indexed_shl_V4: - return !invertPredicate ? Hexagon::STrid_indexed_shl_cPt_V4 : - Hexagon::STrid_indexed_shl_cNotPt_V4; - - // V4 Store to global address - case Hexagon::STd_GP_V4: - return !invertPredicate ? Hexagon::STd_GP_cPt_V4 : - Hexagon::STd_GP_cNotPt_V4; - case Hexagon::STb_GP_V4: - return !invertPredicate ? Hexagon::STb_GP_cPt_V4 : - Hexagon::STb_GP_cNotPt_V4; - case Hexagon::STh_GP_V4: - return !invertPredicate ? Hexagon::STh_GP_cPt_V4 : - Hexagon::STh_GP_cNotPt_V4; - case Hexagon::STw_GP_V4: - return !invertPredicate ? Hexagon::STw_GP_cPt_V4 : - Hexagon::STw_GP_cNotPt_V4; - - // Load. - case Hexagon::LDrid: - return !invertPredicate ? Hexagon::LDrid_cPt : - Hexagon::LDrid_cNotPt; - case Hexagon::LDriw: - return !invertPredicate ? Hexagon::LDriw_cPt : - Hexagon::LDriw_cNotPt; - case Hexagon::LDrih: - return !invertPredicate ? Hexagon::LDrih_cPt : - Hexagon::LDrih_cNotPt; - case Hexagon::LDriuh: - return !invertPredicate ? Hexagon::LDriuh_cPt : - Hexagon::LDriuh_cNotPt; - case Hexagon::LDrib: - return !invertPredicate ? Hexagon::LDrib_cPt : - Hexagon::LDrib_cNotPt; - case Hexagon::LDriub: - return !invertPredicate ? Hexagon::LDriub_cPt : - Hexagon::LDriub_cNotPt; - // Load Indexed. - case Hexagon::LDrid_indexed: - return !invertPredicate ? Hexagon::LDrid_indexed_cPt : - Hexagon::LDrid_indexed_cNotPt; - case Hexagon::LDriw_indexed: - return !invertPredicate ? Hexagon::LDriw_indexed_cPt : - Hexagon::LDriw_indexed_cNotPt; - case Hexagon::LDrih_indexed: - return !invertPredicate ? Hexagon::LDrih_indexed_cPt : - Hexagon::LDrih_indexed_cNotPt; - case Hexagon::LDriuh_indexed: - return !invertPredicate ? Hexagon::LDriuh_indexed_cPt : - Hexagon::LDriuh_indexed_cNotPt; - case Hexagon::LDrib_indexed: - return !invertPredicate ? Hexagon::LDrib_indexed_cPt : - Hexagon::LDrib_indexed_cNotPt; - case Hexagon::LDriub_indexed: - return !invertPredicate ? Hexagon::LDriub_indexed_cPt : - Hexagon::LDriub_indexed_cNotPt; - // Post Increment Load. - case Hexagon::POST_LDrid: - return !invertPredicate ? Hexagon::POST_LDrid_cPt : - Hexagon::POST_LDrid_cNotPt; - case Hexagon::POST_LDriw: - return !invertPredicate ? Hexagon::POST_LDriw_cPt : - Hexagon::POST_LDriw_cNotPt; - case Hexagon::POST_LDrih: - return !invertPredicate ? Hexagon::POST_LDrih_cPt : - Hexagon::POST_LDrih_cNotPt; - case Hexagon::POST_LDriuh: - return !invertPredicate ? Hexagon::POST_LDriuh_cPt : - Hexagon::POST_LDriuh_cNotPt; - case Hexagon::POST_LDrib: - return !invertPredicate ? Hexagon::POST_LDrib_cPt : - Hexagon::POST_LDrib_cNotPt; - case Hexagon::POST_LDriub: - return !invertPredicate ? Hexagon::POST_LDriub_cPt : - Hexagon::POST_LDriub_cNotPt; + // DEALLOC_RETURN. case Hexagon::DEALLOC_RET_V4: return !invertPredicate ? Hexagon::DEALLOC_RET_cPt_V4 : @@ -1727,6 +1019,16 @@ bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const { return ((F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask); } +// Returns true, if a ST insn can be promoted to a new-value store. +bool HexagonInstrInfo::mayBeNewStore(const MachineInstr *MI) const { + const HexagonRegisterInfo& QRI = getRegisterInfo(); + const uint64_t F = MI->getDesc().TSFlags; + + return ((F >> HexagonII::mayNVStorePos) & + HexagonII::mayNVStoreMask & + QRI.Subtarget.hasV4TOps()); +} + bool HexagonInstrInfo::DefinesPredicate(MachineInstr *MI, std::vector<MachineOperand> &Pred) const { @@ -1911,6 +1213,8 @@ isValidAutoIncImm(const EVT VT, const int Offset) const { bool HexagonInstrInfo:: isMemOp(const MachineInstr *MI) const { +// return MI->getDesc().mayLoad() && MI->getDesc().mayStore(); + switch (MI->getOpcode()) { default: return false; @@ -2202,6 +1506,10 @@ bool HexagonInstrInfo::isNewValueJump(const MachineInstr *MI) const { return false; } +bool HexagonInstrInfo::isPostIncrement (const MachineInstr* MI) const { + return (getAddrMode(MI) == HexagonII::PostInc); +} + bool HexagonInstrInfo::isNewValue(const MachineInstr* MI) const { const uint64_t F = MI->getDesc().TSFlags; return ((F >> HexagonII::NewValuePos) & HexagonII::NewValueMask); @@ -2214,6 +1522,97 @@ bool HexagonInstrInfo::isDotNewInst (const MachineInstr* MI) const { (isPredicated(MI) && isPredicatedNew(MI))); } +// Returns the most basic instruction for the .new predicated instructions and +// new-value stores. +// For example, all of the following instructions will be converted back to the +// same instruction: +// 1) if (p0.new) memw(R0+#0) = R1.new ---> +// 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 +// 3) if (p0.new) memw(R0+#0) = R1 ---> +// + +int HexagonInstrInfo::GetDotOldOp(const int opc) const { + int NewOp = opc; + if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form + NewOp = Hexagon::getPredOldOpcode(NewOp); + if (NewOp < 0) + assert(0 && "Couldn't change predicate new instruction to its old form."); + } + + if (isNewValueStore(NewOp)) { // Convert into non new-value format + NewOp = Hexagon::getNonNVStore(NewOp); + if (NewOp < 0) + assert(0 && "Couldn't change new-value store to its old form."); + } + return NewOp; +} + +// Return the new value instruction for a given store. +int HexagonInstrInfo::GetDotNewOp(const MachineInstr* MI) const { + int NVOpcode = Hexagon::getNewValueOpcode(MI->getOpcode()); + if (NVOpcode >= 0) // Valid new-value store instruction. + return NVOpcode; + + switch (MI->getOpcode()) { + default: llvm_unreachable("Unknown .new type"); + // store new value byte + case Hexagon::STrib_shl_V4: + return Hexagon::STrib_shl_nv_V4; + + case Hexagon::STrih_shl_V4: + return Hexagon::STrih_shl_nv_V4; + + case Hexagon::STriw_f: + return Hexagon::STriw_nv_V4; + + case Hexagon::STriw_indexed_f: + return Hexagon::STriw_indexed_nv_V4; + + case Hexagon::STriw_shl_V4: + return Hexagon::STriw_shl_nv_V4; + + } + return 0; +} + +// Return .new predicate version for an instruction. +int HexagonInstrInfo::GetDotNewPredOp(MachineInstr *MI, + const MachineBranchProbabilityInfo + *MBPI) const { + + int NewOpcode = Hexagon::getPredNewOpcode(MI->getOpcode()); + if (NewOpcode >= 0) // Valid predicate new instruction + return NewOpcode; + + switch (MI->getOpcode()) { + default: llvm_unreachable("Unknown .new type"); + // Condtional Jumps + case Hexagon::JMP_t: + case Hexagon::JMP_f: + return getDotNewPredJumpOp(MI, MBPI); + + case Hexagon::JMPR_t: + return Hexagon::JMPR_tnew_tV3; + + case Hexagon::JMPR_f: + return Hexagon::JMPR_fnew_tV3; + + case Hexagon::JMPret_t: + return Hexagon::JMPret_tnew_tV3; + + case Hexagon::JMPret_f: + return Hexagon::JMPret_fnew_tV3; + + + // Conditional combine + case Hexagon::COMBINE_rr_cPt : + return Hexagon::COMBINE_rr_cdnPt; + case Hexagon::COMBINE_rr_cNotPt : + return Hexagon::COMBINE_rr_cdnNotPt; + } +} + + unsigned HexagonInstrInfo::getAddrMode(const MachineInstr* MI) const { const uint64_t F = MI->getDesc().TSFlags; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.h b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.h index b721da4..3f45b8b 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.h @@ -26,8 +26,9 @@ namespace llvm { class HexagonInstrInfo : public HexagonGenInstrInfo { + virtual void anchor(); const HexagonRegisterInfo RI; - const HexagonSubtarget& Subtarget; + const HexagonSubtarget &Subtarget; typedef unsigned Opcode_t; public: @@ -148,11 +149,6 @@ public: isProfitableToDupForIfCvt(MachineBasicBlock &MBB,unsigned NumCycles, const BranchProbability &Probability) const; - virtual MachineInstr *emitFrameIndexDebugValue(MachineFunction &MF, - int FrameIx, - uint64_t Offset, - const MDNode *MDPtr, - DebugLoc DL) const; virtual DFAPacketizer* CreateTargetScheduleState(const TargetMachine *TM, const ScheduleDAG *DAG) const; @@ -185,12 +181,19 @@ public: bool isNewValueInst(const MachineInstr* MI) const; bool isNewValue(const MachineInstr* MI) const; bool isDotNewInst(const MachineInstr* MI) const; + int GetDotOldOp(const int opc) const; + int GetDotNewOp(const MachineInstr* MI) const; + int GetDotNewPredOp(MachineInstr *MI, + const MachineBranchProbabilityInfo + *MBPI) const; + bool mayBeNewStore(const MachineInstr* MI) const; bool isDeallocRet(const MachineInstr *MI) const; unsigned getInvertedPredicatedOpcode(const int Opc) const; bool isExtendable(const MachineInstr* MI) const; bool isExtended(const MachineInstr* MI) const; bool isPostIncrement(const MachineInstr* MI) const; bool isNewValueStore(const MachineInstr* MI) const; + bool isNewValueStore(unsigned Opcode) const; bool isNewValueJump(const MachineInstr* MI) const; bool isNewValueJumpCandidate(const MachineInstr *MI) const; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.td b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.td index 2a4b17b..c96aaca 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.td +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfo.td @@ -384,6 +384,12 @@ def TFCR : CRInst<(outs CRRegs:$dst), (ins IntRegs:$src1), // ALU32/PERM + //===----------------------------------------------------------------------===// +let neverHasSideEffects = 1 in +def COMBINE_ii : ALU32_ii<(outs DoubleRegs:$dst), + (ins s8Imm:$src1, s8Imm:$src2), + "$dst = combine(#$src1, #$src2)", + []>; + // Mux. def VMUX_prr64 : ALU64_rr<(outs DoubleRegs:$dst), (ins PredRegs:$src1, DoubleRegs:$src2, @@ -932,12 +938,21 @@ multiclass LD_MEMri<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, isMEMri = "true" in { - defm LDrib: LD_MEMri < "memb", "LDrib", IntRegs, 11, 6>, AddrModeRel; - defm LDriub: LD_MEMri < "memub" , "LDriub", IntRegs, 11, 6>, AddrModeRel; - defm LDrih: LD_MEMri < "memh", "LDrih", IntRegs, 12, 7>, AddrModeRel; - defm LDriuh: LD_MEMri < "memuh", "LDriuh", IntRegs, 12, 7>, AddrModeRel; - defm LDriw: LD_MEMri < "memw", "LDriw", IntRegs, 13, 8>, AddrModeRel; - defm LDrid: LD_MEMri < "memd", "LDrid", DoubleRegs, 14, 9>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib: LD_MEMri < "memb", "LDrib", IntRegs, 11, 6>, AddrModeRel; + defm LDriub: LD_MEMri < "memub" , "LDriub", IntRegs, 11, 6>, AddrModeRel; + } + + let accessSize = HalfWordAccess in { + defm LDrih: LD_MEMri < "memh", "LDrih", IntRegs, 12, 7>, AddrModeRel; + defm LDriuh: LD_MEMri < "memuh", "LDriuh", IntRegs, 12, 7>, AddrModeRel; + } + + let accessSize = WordAccess in + defm LDriw: LD_MEMri < "memw", "LDriw", IntRegs, 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid: LD_MEMri < "memd", "LDrid", DoubleRegs, 14, 9>, AddrModeRel; } def : Pat < (i32 (sextloadi8 ADDRriS11_0:$addr)), @@ -1000,18 +1015,25 @@ multiclass LD_Idxd<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset in { - defm LDrib_indexed: LD_Idxd <"memb", "LDrib", IntRegs, s11_0Ext, u6_0Ext, - 11, 6>, AddrModeRel; - defm LDriub_indexed: LD_Idxd <"memub" , "LDriub", IntRegs, s11_0Ext, u6_0Ext, - 11, 6>, AddrModeRel; - defm LDrih_indexed: LD_Idxd <"memh", "LDrih", IntRegs, s11_1Ext, u6_1Ext, - 12, 7>, AddrModeRel; - defm LDriuh_indexed: LD_Idxd <"memuh", "LDriuh", IntRegs, s11_1Ext, u6_1Ext, - 12, 7>, AddrModeRel; - defm LDriw_indexed: LD_Idxd <"memw", "LDriw", IntRegs, s11_2Ext, u6_2Ext, - 13, 8>, AddrModeRel; - defm LDrid_indexed: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext, - 14, 9>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib_indexed: LD_Idxd <"memb", "LDrib", IntRegs, s11_0Ext, u6_0Ext, + 11, 6>, AddrModeRel; + defm LDriub_indexed: LD_Idxd <"memub" , "LDriub", IntRegs, s11_0Ext, u6_0Ext, + 11, 6>, AddrModeRel; + } + let accessSize = HalfWordAccess in { + defm LDrih_indexed: LD_Idxd <"memh", "LDrih", IntRegs, s11_1Ext, u6_1Ext, + 12, 7>, AddrModeRel; + defm LDriuh_indexed: LD_Idxd <"memuh", "LDriuh", IntRegs, s11_1Ext, u6_1Ext, + 12, 7>, AddrModeRel; + } + let accessSize = WordAccess in + defm LDriw_indexed: LD_Idxd <"memw", "LDriw", IntRegs, s11_2Ext, u6_2Ext, + 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid_indexed: LD_Idxd <"memd", "LDrid", DoubleRegs, s11_3Ext, u6_3Ext, + 14, 9>, AddrModeRel; } let AddedComplexity = 20 in { @@ -1036,8 +1058,6 @@ def : Pat < (i64 (load (add IntRegs:$src1, s11_3ExtPred:$offset))), //===----------------------------------------------------------------------===// // Post increment load -// Make sure that in post increment load, the first operand is always the post -// increment operand. //===----------------------------------------------------------------------===// multiclass LD_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, @@ -1079,7 +1099,7 @@ multiclass LD_PostInc<string mnemonic, string BaseOp, RegisterClass RC, } } -let hasCtrlDep = 1, neverHasSideEffects = 1 in { +let hasCtrlDep = 1, neverHasSideEffects = 1, addrMode = PostInc in { defm POST_LDrib : LD_PostInc<"memb", "LDrib", IntRegs, s4_0Imm>, PredNewRel; defm POST_LDriub : LD_PostInc<"memub", "LDriub", IntRegs, s4_0Imm>, @@ -1382,7 +1402,7 @@ multiclass ST_PostInc_Pbase<string mnemonic, RegisterClass RC, Operand ImmOp, multiclass ST_PostInc_Pred<string mnemonic, RegisterClass RC, Operand ImmOp, bit PredNot> { let isPredicatedFalse = PredNot in { - defm _c#NAME# : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; + defm _c#NAME : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 0>; // Predicate new let Predicates = [HasV4T], validSubTargets = HasV4SubT in defm _cdn#NAME#_V4 : ST_PostInc_Pbase<mnemonic, RC, ImmOp, PredNot, 1>; @@ -1397,7 +1417,7 @@ multiclass ST_PostInc<string mnemonic, string BaseOp, RegisterClass RC, let isPredicable = 1 in def NAME : STInst2PI<(outs IntRegs:$dst), (ins IntRegs:$src1, ImmOp:$offset, RC:$src2), - #mnemonic#"($src1++#$offset) = $src2", + mnemonic#"($src1++#$offset) = $src2", [], "$src1 = $dst">; @@ -1474,12 +1494,17 @@ multiclass ST_MEMri<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, isMEMri = "true" in { - defm STrib: ST_MEMri < "memb", "STrib", IntRegs, 11, 6>, AddrModeRel; - defm STrih: ST_MEMri < "memh", "STrih", IntRegs, 12, 7>, AddrModeRel; - defm STriw: ST_MEMri < "memw", "STriw", IntRegs, 13, 8>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib: ST_MEMri < "memb", "STrib", IntRegs, 11, 6>, AddrModeRel; + + let accessSize = HalfWordAccess in + defm STrih: ST_MEMri < "memh", "STrih", IntRegs, 12, 7>, AddrModeRel; - let isNVStorable = 0 in - defm STrid: ST_MEMri < "memd", "STrid", DoubleRegs, 14, 9>, AddrModeRel; + let accessSize = WordAccess in + defm STriw: ST_MEMri < "memw", "STriw", IntRegs, 13, 8>, AddrModeRel; + + let accessSize = DoubleWordAccess, isNVStorable = 0 in + defm STrid: ST_MEMri < "memd", "STrid", DoubleRegs, 14, 9>, AddrModeRel; } def : Pat<(truncstorei8 (i32 IntRegs:$src1), ADDRriS11_0:$addr), @@ -1541,15 +1566,21 @@ multiclass ST_Idxd<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, InputType = "reg" in { - defm STrib_indexed: ST_Idxd < "memb", "STrib", IntRegs, s11_0Ext, - u6_0Ext, 11, 6>, AddrModeRel, ImmRegRel; - defm STrih_indexed: ST_Idxd < "memh", "STrih", IntRegs, s11_1Ext, - u6_1Ext, 12, 7>, AddrModeRel, ImmRegRel; - defm STriw_indexed: ST_Idxd < "memw", "STriw", IntRegs, s11_2Ext, - u6_2Ext, 13, 8>, AddrModeRel, ImmRegRel; - let isNVStorable = 0 in - defm STrid_indexed: ST_Idxd < "memd", "STrid", DoubleRegs, s11_3Ext, - u6_3Ext, 14, 9>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed: ST_Idxd < "memb", "STrib", IntRegs, s11_0Ext, + u6_0Ext, 11, 6>, AddrModeRel, ImmRegRel; + + let accessSize = HalfWordAccess in + defm STrih_indexed: ST_Idxd < "memh", "STrih", IntRegs, s11_1Ext, + u6_1Ext, 12, 7>, AddrModeRel, ImmRegRel; + + let accessSize = WordAccess in + defm STriw_indexed: ST_Idxd < "memw", "STriw", IntRegs, s11_2Ext, + u6_2Ext, 13, 8>, AddrModeRel, ImmRegRel; + + let accessSize = DoubleWordAccess, isNVStorable = 0 in + defm STrid_indexed: ST_Idxd < "memd", "STrid", DoubleRegs, s11_3Ext, + u6_3Ext, 14, 9>, AddrModeRel; } let AddedComplexity = 10 in { diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td index 933239d..475c23d 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV4.td @@ -213,7 +213,7 @@ def COMBINE_iI_V4 : ALU32_ii<(outs DoubleRegs:$dst), // Template class for load instructions with Absolute set addressing mode. //===----------------------------------------------------------------------===// let isExtended = 1, opExtendable = 2, neverHasSideEffects = 1, -validSubTargets = HasV4SubT in +validSubTargets = HasV4SubT, addrMode = AbsoluteSet in class T_LD_abs_set<string mnemonic, RegisterClass RC>: LDInst2<(outs RC:$dst1, IntRegs:$dst2), (ins u0AlwaysExt:$addr), @@ -266,12 +266,23 @@ multiclass ld_idxd_shl<string mnemonic, string CextOp, RegisterClass RC> { } let addrMode = BaseRegOffset in { - defm LDrib_indexed_shl: ld_idxd_shl<"memb", "LDrib", IntRegs>, AddrModeRel; - defm LDriub_indexed_shl: ld_idxd_shl<"memub", "LDriub", IntRegs>, AddrModeRel; - defm LDrih_indexed_shl: ld_idxd_shl<"memh", "LDrih", IntRegs>, AddrModeRel; - defm LDriuh_indexed_shl: ld_idxd_shl<"memuh", "LDriuh", IntRegs>, AddrModeRel; - defm LDriw_indexed_shl: ld_idxd_shl<"memw", "LDriw", IntRegs>, AddrModeRel; - defm LDrid_indexed_shl: ld_idxd_shl<"memd", "LDrid", DoubleRegs>, AddrModeRel; + let accessSize = ByteAccess in { + defm LDrib_indexed_shl: ld_idxd_shl<"memb", "LDrib", IntRegs>, + AddrModeRel; + defm LDriub_indexed_shl: ld_idxd_shl<"memub", "LDriub", IntRegs>, + AddrModeRel; + } + let accessSize = HalfWordAccess in { + defm LDrih_indexed_shl: ld_idxd_shl<"memh", "LDrih", IntRegs>, AddrModeRel; + defm LDriuh_indexed_shl: ld_idxd_shl<"memuh", "LDriuh", IntRegs>, + AddrModeRel; + } + let accessSize = WordAccess in + defm LDriw_indexed_shl: ld_idxd_shl<"memw", "LDriw", IntRegs>, AddrModeRel; + + let accessSize = DoubleWordAccess in + defm LDrid_indexed_shl: ld_idxd_shl<"memd", "LDrid", DoubleRegs>, + AddrModeRel; } // 'def pats' for load instructions with base + register offset and non-zero @@ -456,7 +467,8 @@ def: Pat <(i64 (extloadi32 (i32 (add IntRegs:$src1, s11_2ExtPred:$offset)))), //===----------------------------------------------------------------------===// // Template class for store instructions with Absolute set addressing mode. //===----------------------------------------------------------------------===// -let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT in +let isExtended = 1, opExtendable = 2, validSubTargets = HasV4SubT, +addrMode = AbsoluteSet in class T_ST_abs_set<string mnemonic, RegisterClass RC>: STInst2<(outs IntRegs:$dst1), (ins RC:$src1, u0AlwaysExt:$src2), @@ -551,17 +563,20 @@ multiclass ST_Idxd_shl_nv<string mnemonic, string CextOp, RegisterClass RC> { let addrMode = BaseRegOffset, neverHasSideEffects = 1, validSubTargets = HasV4SubT in { - defm STrib_indexed_shl: ST_Idxd_shl<"memb", "STrib", IntRegs>, - ST_Idxd_shl_nv<"memb", "STrib", IntRegs>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed_shl: ST_Idxd_shl<"memb", "STrib", IntRegs>, + ST_Idxd_shl_nv<"memb", "STrib", IntRegs>, AddrModeRel; - defm STrih_indexed_shl: ST_Idxd_shl<"memh", "STrih", IntRegs>, - ST_Idxd_shl_nv<"memh", "STrih", IntRegs>, AddrModeRel; + let accessSize = HalfWordAccess in + defm STrih_indexed_shl: ST_Idxd_shl<"memh", "STrih", IntRegs>, + ST_Idxd_shl_nv<"memh", "STrih", IntRegs>, AddrModeRel; - defm STriw_indexed_shl: ST_Idxd_shl<"memw", "STriw", IntRegs>, - ST_Idxd_shl_nv<"memw", "STriw", IntRegs>, AddrModeRel; + let accessSize = WordAccess in + defm STriw_indexed_shl: ST_Idxd_shl<"memw", "STriw", IntRegs>, + ST_Idxd_shl_nv<"memw", "STriw", IntRegs>, AddrModeRel; - let isNVStorable = 0 in - defm STrid_indexed_shl: ST_Idxd_shl<"memd", "STrid", DoubleRegs>, AddrModeRel; + let isNVStorable = 0, accessSize = DoubleWordAccess in + defm STrid_indexed_shl: ST_Idxd_shl<"memd", "STrid", DoubleRegs>, AddrModeRel; } let Predicates = [HasV4T], AddedComplexity = 10 in { @@ -695,10 +710,15 @@ multiclass ST_Imm<string mnemonic, string CextOp, Operand OffsetOp> { } let addrMode = BaseImmOffset, InputType = "imm", - validSubTargets = HasV4SubT in { - defm STrib_imm : ST_Imm<"memb", "STrib", u6_0Imm>, ImmRegRel, PredNewRel; - defm STrih_imm : ST_Imm<"memh", "STrih", u6_1Imm>, ImmRegRel, PredNewRel; - defm STriw_imm : ST_Imm<"memw", "STriw", u6_2Imm>, ImmRegRel, PredNewRel; +validSubTargets = HasV4SubT in { + let accessSize = ByteAccess in + defm STrib_imm : ST_Imm<"memb", "STrib", u6_0Imm>, ImmRegRel, PredNewRel; + + let accessSize = HalfWordAccess in + defm STrih_imm : ST_Imm<"memh", "STrih", u6_1Imm>, ImmRegRel, PredNewRel; + + let accessSize = WordAccess in + defm STriw_imm : ST_Imm<"memw", "STriw", u6_2Imm>, ImmRegRel, PredNewRel; } let Predicates = [HasV4T], AddedComplexity = 10 in { @@ -834,12 +854,17 @@ multiclass ST_Idxd_nv<string mnemonic, string CextOp, RegisterClass RC, } let addrMode = BaseImmOffset, validSubTargets = HasV4SubT in { - defm STrib_indexed: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext, - u6_0Ext, 11, 6>, AddrModeRel; - defm STrih_indexed: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext, - u6_1Ext, 12, 7>, AddrModeRel; - defm STriw_indexed: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext, - u6_2Ext, 13, 8>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib_indexed: ST_Idxd_nv<"memb", "STrib", IntRegs, s11_0Ext, + u6_0Ext, 11, 6>, AddrModeRel; + + let accessSize = HalfWordAccess in + defm STrih_indexed: ST_Idxd_nv<"memh", "STrih", IntRegs, s11_1Ext, + u6_1Ext, 12, 7>, AddrModeRel; + + let accessSize = WordAccess in + defm STriw_indexed: ST_Idxd_nv<"memw", "STriw", IntRegs, s11_2Ext, + u6_2Ext, 13, 8>, AddrModeRel; } // multiclass for new-value store instructions with base + immediate offset. @@ -887,9 +912,14 @@ multiclass ST_MEMri_nv<string mnemonic, string CextOp, RegisterClass RC, let addrMode = BaseImmOffset, isMEMri = "true", validSubTargets = HasV4SubT, mayStore = 1 in { - defm STrib: ST_MEMri_nv<"memb", "STrib", IntRegs, 11, 6>, AddrModeRel; - defm STrih: ST_MEMri_nv<"memh", "STrih", IntRegs, 12, 7>, AddrModeRel; - defm STriw: ST_MEMri_nv<"memw", "STriw", IntRegs, 13, 8>, AddrModeRel; + let accessSize = ByteAccess in + defm STrib: ST_MEMri_nv<"memb", "STrib", IntRegs, 11, 6>, AddrModeRel; + + let accessSize = HalfWordAccess in + defm STrih: ST_MEMri_nv<"memh", "STrih", IntRegs, 12, 7>, AddrModeRel; + + let accessSize = WordAccess in + defm STriw: ST_MEMri_nv<"memw", "STriw", IntRegs, 13, 8>, AddrModeRel; } //===----------------------------------------------------------------------===// @@ -939,7 +969,7 @@ multiclass ST_PostInc_nv<string mnemonic, string BaseOp, RegisterClass RC, } } -let validSubTargets = HasV4SubT in { +let addrMode = PostInc, validSubTargets = HasV4SubT in { defm POST_STbri: ST_PostInc_nv <"memb", "STrib", IntRegs, s4_0Imm>, AddrModeRel; defm POST_SThri: ST_PostInc_nv <"memh", "STrih", IntRegs, s4_1Imm>, AddrModeRel; defm POST_STwri: ST_PostInc_nv <"memw", "STriw", IntRegs, s4_2Imm>, AddrModeRel; @@ -2534,8 +2564,9 @@ def NTSTBIT_ri : SInst<(outs PredRegs:$dst), (ins IntRegs:$src1, u5Imm:$src2), //Deallocate frame and return. // dealloc_return let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicable = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1 in { - def DEALLOC_RET_V4 : NVInst_V4<(outs), (ins i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_V4 : LD0Inst<(outs), (ins), "dealloc_return", []>, Requires<[HasV4T]>; @@ -2544,9 +2575,10 @@ let isReturn = 1, isTerminator = 1, isBarrier = 1, isPredicable = 1, // Restore registers and dealloc return function call. let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1, Defs = [R29, R30, R31, PC] in { +let validSubTargets = HasV4SubT in def RESTORE_DEALLOC_RET_JMP_V4 : JInst<(outs), (ins calltarget:$dst), - "jump $dst // Restore_and_dealloc_return", + "jump $dst", []>, Requires<[HasV4T]>; } @@ -2554,9 +2586,10 @@ let isCall = 1, isBarrier = 1, isReturn = 1, isTerminator = 1, // Restore registers and dealloc frame before a tail call. let isCall = 1, isBarrier = 1, Defs = [R29, R30, R31, PC] in { +let validSubTargets = HasV4SubT in def RESTORE_DEALLOC_BEFORE_TAILCALL_V4 : JInst<(outs), (ins calltarget:$dst), - "call $dst // Restore_and_dealloc_before_tailcall", + "call $dst", []>, Requires<[HasV4T]>; } @@ -2573,10 +2606,11 @@ let isCall = 1, isBarrier = 1, // if (Ps) dealloc_return let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cPt_V4 : NVInst_V4<(outs), - (ins PredRegs:$src1, i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cPt_V4 : LD0Inst<(outs), + (ins PredRegs:$src1), "if ($src1) dealloc_return", []>, Requires<[HasV4T]>; @@ -2584,10 +2618,10 @@ let isReturn = 1, isTerminator = 1, // if (!Ps) dealloc_return let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1) dealloc_return", []>, Requires<[HasV4T]>; @@ -2595,10 +2629,10 @@ let isReturn = 1, isTerminator = 1, // if (Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cdnPnt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cdnPnt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if ($src1.new) dealloc_return:nt", []>, Requires<[HasV4T]>; @@ -2606,10 +2640,10 @@ let isReturn = 1, isTerminator = 1, // if (!Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotdnPnt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotdnPnt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1.new) dealloc_return:nt", []>, Requires<[HasV4T]>; @@ -2617,21 +2651,21 @@ let isReturn = 1, isTerminator = 1, // if (Ps.new) dealloc_return:t let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, isPredicated = 1 in { - def DEALLOC_RET_cdnPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cdnPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if ($src1.new) dealloc_return:t", []>, Requires<[HasV4T]>; } -// if (!Ps.new) dealloc_return:nt +// if (!Ps.new) dealloc_return:nt let isReturn = 1, isTerminator = 1, - Defs = [R29, R30, R31, PC], Uses = [R29, R31], neverHasSideEffects = 1, - isPredicated = 1 in { - def DEALLOC_RET_cNotdnPt_V4 : NVInst_V4<(outs), (ins PredRegs:$src1, - i32imm:$amt1), + Defs = [R29, R30, R31, PC], Uses = [R30], neverHasSideEffects = 1, + isPredicated = 1, isPredicatedFalse = 1 in { +let validSubTargets = HasV4SubT in + def DEALLOC_RET_cNotdnPt_V4 : LD0Inst<(outs), (ins PredRegs:$src1), "if (!$src1.new) dealloc_return:t", []>, Requires<[HasV4T]>; @@ -3033,37 +3067,42 @@ def : Pat<(HexagonCONST32_GP tblockaddress:$src1), (TFRI_V4 tblockaddress:$src1)>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if($src1) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if($src1) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if(!$src1) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if(!$src1) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cdnPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if($src1.new) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if($src1.new) $dst = #$src2", []>, Requires<[HasV4T]>; -let AddedComplexity=50, neverHasSideEffects = 1, isPredicated = 1 in +let isExtended = 1, opExtendable = 2, AddedComplexity=50, isPredicatedFalse = 1, +neverHasSideEffects = 1, isPredicated = 1, validSubTargets = HasV4SubT in def TFRI_cdnNotPt_V4 : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, globaladdress:$src2), - "if(!$src1.new) $dst = ##$src2", + (ins PredRegs:$src1, s16Ext:$src2), + "if(!$src1.new) $dst = #$src2", []>, Requires<[HasV4T]>; let AddedComplexity = 50, Predicates = [HasV4T] in def : Pat<(HexagonCONST32_GP tglobaladdr:$src1), - (TFRI_V4 tglobaladdr:$src1)>; + (TFRI_V4 tglobaladdr:$src1)>, + Requires<[HasV4T]>; // Load - Indirect with long offset: These instructions take global address @@ -3149,6 +3188,93 @@ def STriw_offset_ext_V4 : STInst<(outs), (add IntRegs:$src1, u6_2ImmPred:$src2))]>, Requires<[HasV4T]>; +def : Pat<(i64 (ctlz (i64 DoubleRegs:$src1))), + (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTLZ64_rr DoubleRegs:$src1))))>, + Requires<[HasV4T]>; + +def : Pat<(i64 (cttz (i64 DoubleRegs:$src1))), + (i64 (COMBINE_Ir_V4 (i32 0), (i32 (CTTZ64_rr DoubleRegs:$src1))))>, + Requires<[HasV4T]>; + + +// i8 -> i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi8. +let Predicates = [HasV4T], AddedComplexity = 120 in { +def: Pat <(i64 (extloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (zextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (sextloadi8 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDrib_abs_V4 tglobaladdr:$addr)))>; + +def: Pat <(i64 (extloadi8 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDrib_abs_V4 FoldGlobalAddr:$addr)))>; + +def: Pat <(i64 (zextloadi8 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriub_abs_V4 FoldGlobalAddr:$addr)))>; + +def: Pat <(i64 (sextloadi8 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDrib_abs_V4 FoldGlobalAddr:$addr)))>; +} +// i16 -> i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi16. +let AddedComplexity = 120 in { +def: Pat <(i64 (extloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi16 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDrih_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (extloadi16 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDrih_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi16 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriuh_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi16 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDrih_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; +} +// i32->i64 loads +// We need a complexity of 120 here to overide preceeding handling of +// zextloadi32. +let AddedComplexity = 120 in { +def: Pat <(i64 (extloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi32 (NumUsesBelowThresCONST32 tglobaladdr:$addr))), + (i64 (SXTW (LDriw_abs_V4 tglobaladdr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (extloadi32 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (zextloadi32 FoldGlobalAddr:$addr)), + (i64 (COMBINE_Ir_V4 0, (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; + +def: Pat <(i64 (sextloadi32 FoldGlobalAddr:$addr)), + (i64 (SXTW (LDriw_abs_V4 FoldGlobalAddr:$addr)))>, + Requires<[HasV4T]>; +} // Indexed store double word - global address. // memw(Rs+#u6:2)=#S8 @@ -3264,4 +3390,3 @@ def : Pat<(i32 (load FoldGlobalAddrGP:$addr)), def : Pat<(atomic_load_32 FoldGlobalAddrGP:$addr), (i32 (LDriw_abs_V4 FoldGlobalAddrGP:$addr))>, Requires<[HasV4T]>; - diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td index 92d098c..9da6074 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td +++ b/contrib/llvm/lib/Target/Hexagon/HexagonInstrInfoV5.td @@ -26,22 +26,29 @@ def CONST32_Float_Real : LDInst<(outs IntRegs:$dst), (ins f32imm:$src1), // Only works with single precision fp value. // For double precision, use CONST64_float_real, as 64bit transfer // can only hold 40-bit values - 32 from const ext + 8 bit immediate. -let isMoveImm = 1, isReMaterializable = 1, isPredicable = 1 in -def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32imm:$src1), - "$dst = ##$src1", +// Make sure that complexity is more than the CONST32 pattern in +// HexagonInstrInfo.td patterns. +let isExtended = 1, opExtendable = 1, isMoveImm = 1, isReMaterializable = 1, +isPredicable = 1, AddedComplexity = 30, validSubTargets = HasV5SubT, +isCodeGenOnly = 1 in +def TFRI_f : ALU32_ri<(outs IntRegs:$dst), (ins f32Ext:$src1), + "$dst = #$src1", [(set IntRegs:$dst, fpimm:$src1)]>, Requires<[HasV5T]>; +let isExtended = 1, opExtendable = 2, isPredicated = 1, +neverHasSideEffects = 1, validSubTargets = HasV5SubT in def TFRI_cPt_f : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, f32imm:$src2), - "if ($src1) $dst = ##$src2", + (ins PredRegs:$src1, f32Ext:$src2), + "if ($src1) $dst = #$src2", []>, Requires<[HasV5T]>; -let isPredicated = 1 in +let isExtended = 1, opExtendable = 2, isPredicated = 1, isPredicatedFalse = 1, +neverHasSideEffects = 1, validSubTargets = HasV5SubT in def TFRI_cNotPt_f : ALU32_ri<(outs IntRegs:$dst), - (ins PredRegs:$src1, f32imm:$src2), - "if (!$src1) $dst = ##$src2", + (ins PredRegs:$src1, f32Ext:$src2), + "if (!$src1) $dst =#$src2", []>, Requires<[HasV5T]>; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonMCInstLower.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonMCInstLower.cpp index f011d51..bbb2fa4 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonMCInstLower.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonMCInstLower.cpp @@ -73,7 +73,7 @@ void llvm::HexagonLowerToMC(const MachineInstr* MI, HexagonMCInst& MCI, AP.OutContext)); break; case MachineOperand::MO_GlobalAddress: - MCO = GetSymbolRef(MO, AP.Mang->getSymbol(MO.getGlobal()), AP); + MCO = GetSymbolRef(MO, AP.getSymbol(MO.getGlobal()), AP); break; case MachineOperand::MO_ExternalSymbol: MCO = GetSymbolRef(MO, AP.GetExternalSymbolSymbol(MO.getSymbolName()), diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.cpp new file mode 100644 index 0000000..9579c8b --- /dev/null +++ b/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.cpp @@ -0,0 +1,16 @@ +//= HexagonMachineFunctionInfo.cpp - Hexagon machine function info *- C++ -*-=// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// + +#include "HexagonMachineFunctionInfo.h" + +using namespace llvm; + +// pin vtable to this file +void HexagonMachineFunctionInfo::anchor() {} + diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.h b/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.h index bd7b26a..a59c8c9 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonMachineFunctionInfo.h @@ -1,4 +1,4 @@ -//=- HexagonMachineFuctionInfo.h - Hexagon machine function info --*- C++ -*-=// +//=- HexagonMachineFunctionInfo.h - Hexagon machine function info -*- C++ -*-=// // // The LLVM Compiler Infrastructure // @@ -10,6 +10,7 @@ #ifndef HexagonMACHINEFUNCTIONINFO_H #define HexagonMACHINEFUNCTIONINFO_H +#include <map> #include "llvm/CodeGen/MachineFunction.h" namespace llvm { @@ -30,9 +31,8 @@ class HexagonMachineFunctionInfo : public MachineFunctionInfo { int VarArgsFrameIndex; bool HasClobberLR; bool HasEHReturn; - std::map<const MachineInstr*, unsigned> PacketInfo; - + virtual void anchor(); public: HexagonMachineFunctionInfo() : SRetReturnReg(0), HasClobberLR(0), diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.cpp index 1388ad4..c94f081 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.cpp @@ -195,7 +195,6 @@ void VLIWMachineScheduler::schedule() { void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) { DAG = static_cast<VLIWMachineScheduler*>(dag); SchedModel = DAG->getSchedModel(); - TRI = DAG->TRI; Top.init(DAG, SchedModel); Bot.init(DAG, SchedModel); @@ -209,6 +208,8 @@ void ConvergingVLIWScheduler::initialize(ScheduleDAGMI *dag) { Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); + delete Top.ResourceModel; + delete Bot.ResourceModel; Top.ResourceModel = new VLIWResourceModel(TM, DAG->getSchedModel()); Bot.ResourceModel = new VLIWResourceModel(TM, DAG->getSchedModel()); @@ -223,7 +224,7 @@ void ConvergingVLIWScheduler::releaseTopNode(SUnit *SU) { for (SUnit::succ_iterator I = SU->Preds.begin(), E = SU->Preds.end(); I != E; ++I) { unsigned PredReadyCycle = I->getSUnit()->TopReadyCycle; - unsigned MinLatency = I->getMinLatency(); + unsigned MinLatency = I->getLatency(); #ifndef NDEBUG Top.MaxMinLatency = std::max(MinLatency, Top.MaxMinLatency); #endif @@ -242,7 +243,7 @@ void ConvergingVLIWScheduler::releaseBottomNode(SUnit *SU) { for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); I != E; ++I) { unsigned SuccReadyCycle = I->getSUnit()->BotReadyCycle; - unsigned MinLatency = I->getMinLatency(); + unsigned MinLatency = I->getLatency(); #ifndef NDEBUG Bot.MaxMinLatency = std::max(MinLatency, Bot.MaxMinLatency); #endif @@ -406,11 +407,11 @@ SUnit *ConvergingVLIWScheduler::SchedBoundary::pickOnlyChoice() { #ifndef NDEBUG void ConvergingVLIWScheduler::traceCandidate(const char *Label, const ReadyQueue &Q, - SUnit *SU, PressureElement P) { + SUnit *SU, PressureChange P) { dbgs() << Label << " " << Q.getName() << " "; if (P.isValid()) - dbgs() << TRI->getRegPressureSetName(P.PSetID) << ":" << P.UnitIncrease - << " "; + dbgs() << DAG->TRI->getRegPressureSetName(P.getPSet()) << ":" + << P.getUnitInc() << " "; else dbgs() << " "; SU->dump(DAG); @@ -456,9 +457,7 @@ static SUnit *getSingleUnscheduledSucc(SUnit *SU) { // Constants used to denote relative importance of // heuristic components for cost computation. static const unsigned PriorityOne = 200; -static const unsigned PriorityTwo = 100; -static const unsigned PriorityThree = 50; -static const unsigned PriorityFour = 20; +static const unsigned PriorityTwo = 50; static const unsigned ScaleTwo = 10; static const unsigned FactorOne = 2; @@ -516,8 +515,8 @@ int ConvergingVLIWScheduler::SchedulingCost(ReadyQueue &Q, SUnit *SU, ResCount += (NumNodesBlocking * ScaleTwo); // Factor in reg pressure as a heuristic. - ResCount -= (Delta.Excess.UnitIncrease*PriorityThree); - ResCount -= (Delta.CriticalMax.UnitIncrease*PriorityThree); + ResCount -= (Delta.Excess.getUnitInc()*PriorityTwo); + ResCount -= (Delta.CriticalMax.getUnitInc()*PriorityTwo); DEBUG(if (verbose) dbgs() << " Total(" << ResCount << ")"); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.h b/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.h index f68dadf..8ac333f 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonMachineScheduler.h @@ -190,7 +190,6 @@ class ConvergingVLIWScheduler : public MachineSchedStrategy { VLIWMachineScheduler *DAG; const TargetSchedModel *SchedModel; - const TargetRegisterInfo *TRI; // State of the top and bottom scheduled instruction boundaries. SchedBoundary Top; @@ -205,7 +204,7 @@ public: }; ConvergingVLIWScheduler(): - DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} + DAG(0), SchedModel(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} virtual void initialize(ScheduleDAGMI *dag); @@ -234,7 +233,7 @@ protected: SchedCandidate &Candidate); #ifndef NDEBUG void traceCandidate(const char *Label, const ReadyQueue &Q, SUnit *SU, - PressureElement P = PressureElement()); + PressureChange P = PressureChange()); #endif }; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonNewValueJump.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonNewValueJump.cpp index 05e6968..f7c4513 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonNewValueJump.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonNewValueJump.cpp @@ -631,6 +631,7 @@ bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { .addMBB(jmpTarget); assert(NewMI && "New Value Jump Instruction Not created!"); + (void)NewMI; if (cmpInstr->getOperand(0).isReg() && cmpInstr->getOperand(0).isKill()) cmpInstr->getOperand(0).setIsKill(false); diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonPeephole.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonPeephole.cpp index 89e3406..5490ecd 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonPeephole.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonPeephole.cpp @@ -29,7 +29,7 @@ // // Note: The peephole pass makes the instrucstions like // %vreg170<def> = SXTW %vreg166 or %vreg16<def> = NOT_p %vreg15<kill> -// redundant and relies on some form of dead removal instrucions, like +// redundant and relies on some form of dead removal instructions, like // DCE or DIE to actually eliminate them. diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.cpp index d8b4e2f..1786e9d 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.cpp @@ -38,11 +38,9 @@ using namespace llvm; -HexagonRegisterInfo::HexagonRegisterInfo(HexagonSubtarget &st, - const HexagonInstrInfo &tii) +HexagonRegisterInfo::HexagonRegisterInfo(HexagonSubtarget &st) : HexagonGenRegisterInfo(Hexagon::R31), - Subtarget(st), - TII(tii) { + Subtarget(st) { } const uint16_t* HexagonRegisterInfo::getCalleeSavedRegs(const MachineFunction @@ -130,6 +128,8 @@ void HexagonRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, // Addressable stack objects are accessed using neg. offsets from %fp. MachineFunction &MF = *MI.getParent()->getParent(); + const HexagonInstrInfo &TII = + *static_cast<const HexagonInstrInfo*>(MF.getTarget().getInstrInfo()); int Offset = MF.getFrameInfo()->getObjectOffset(FrameIndex); MachineFrameInfo &MFI = *MF.getFrameInfo(); @@ -295,23 +295,5 @@ unsigned HexagonRegisterInfo::getStackRegister() const { return Hexagon::R29; } -void HexagonRegisterInfo::getInitialFrameState(std::vector<MachineMove> - &Moves) const -{ - // VirtualFP = (R30 + #0). - unsigned FPReg = getFrameRegister(); - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(FPReg, 0); - Moves.push_back(MachineMove(0, Dst, Src)); -} - -unsigned HexagonRegisterInfo::getEHExceptionRegister() const { - llvm_unreachable("What is the exception register"); -} - -unsigned HexagonRegisterInfo::getEHHandlerRegister() const { - llvm_unreachable("What is the exception handler register"); -} - #define GET_REGINFO_TARGET_DESC #include "HexagonGenRegisterInfo.inc" diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.h b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.h index 8a3f94a..89af7c3 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.h @@ -44,9 +44,8 @@ class Type; struct HexagonRegisterInfo : public HexagonGenRegisterInfo { HexagonSubtarget &Subtarget; - const HexagonInstrInfo &TII; - HexagonRegisterInfo(HexagonSubtarget &st, const HexagonInstrInfo &tii); + HexagonRegisterInfo(HexagonSubtarget &st); /// Code Generation virtual methods... const uint16_t *getCalleeSavedRegs(const MachineFunction *MF = 0) const; @@ -78,12 +77,7 @@ struct HexagonRegisterInfo : public HexagonGenRegisterInfo { unsigned getRARegister() const; unsigned getFrameRegister(const MachineFunction &MF) const; unsigned getFrameRegister() const; - void getInitialFrameState(std::vector<MachineMove> &Moves) const; unsigned getStackRegister() const; - - // Exception handling queries. - unsigned getEHExceptionRegister() const; - unsigned getEHHandlerRegister() const; }; } // end namespace llvm diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.td b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.td index fe41fc3..8ea1b7e 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.td +++ b/contrib/llvm/lib/Target/Hexagon/HexagonRegisterInfo.td @@ -57,8 +57,8 @@ let Namespace = "Hexagon" in { let Aliases = [R]; } - def subreg_loreg : SubRegIndex; - def subreg_hireg : SubRegIndex; + def subreg_loreg : SubRegIndex<32>; + def subreg_hireg : SubRegIndex<32, 32>; // Integer registers. def R0 : Ri< 0, "r0">, DwarfRegNum<[0]>; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp index a52c604..c37bf9f 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp @@ -27,7 +27,7 @@ HexagonSelectionDAGInfo::~HexagonSelectionDAGInfo() { SDValue HexagonSelectionDAGInfo:: -EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, SDValue Chain, +EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src, SDValue Size, unsigned Align, bool isVolatile, bool AlwaysInline, MachinePointerInfo DstPtrInfo, diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.h b/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.h index 0673e4d..31f278a 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonSelectionDAGInfo.h @@ -26,7 +26,7 @@ public: ~HexagonSelectionDAGInfo(); virtual - SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, + SDValue EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src, SDValue Size, unsigned Align, diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp new file mode 100644 index 0000000..5166f8e --- /dev/null +++ b/contrib/llvm/lib/Target/Hexagon/HexagonSplitConst32AndConst64.cpp @@ -0,0 +1,174 @@ +//=== HexagonSplitConst32AndConst64.cpp - split CONST32/Const64 into HI/LO ===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// When the compiler is invoked with no small data, for instance, with the -G0 +// command line option, then all CONST32_* opcodes should be broken down into +// appropriate LO and HI instructions. This splitting is done by this pass. +// The only reason this is not done in the DAG lowering itself is that there +// is no simple way of getting the register allocator to allot the same hard +// register to the result of LO and HI instructions. This pass is always +// scheduled after register allocation. +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "xfer" + +#include "HexagonTargetMachine.h" +#include "HexagonSubtarget.h" +#include "HexagonMachineFunctionInfo.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/CodeGen/LatencyPriorityQueue.h" +#include "llvm/CodeGen/MachineDominators.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" +#include "llvm/CodeGen/MachineLoopInfo.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/ScheduleDAGInstrs.h" +#include "llvm/CodeGen/ScheduleHazardRecognizer.h" +#include "llvm/CodeGen/SchedulerRegistry.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Support/Compiler.h" +#include "llvm/Support/Debug.h" +#include "llvm/Support/CommandLine.h" +#include "llvm/Support/MathExtras.h" +#include <map> + +using namespace llvm; + +namespace { + +class HexagonSplitConst32AndConst64 : public MachineFunctionPass { + const HexagonTargetMachine& QTM; + const HexagonSubtarget &QST; + + public: + static char ID; + HexagonSplitConst32AndConst64(const HexagonTargetMachine& TM) + : MachineFunctionPass(ID), QTM(TM), QST(*TM.getSubtargetImpl()) {} + + const char *getPassName() const { + return "Hexagon Split Const32s and Const64s"; + } + bool runOnMachineFunction(MachineFunction &Fn); +}; + + +char HexagonSplitConst32AndConst64::ID = 0; + + +bool HexagonSplitConst32AndConst64::runOnMachineFunction(MachineFunction &Fn) { + + const TargetInstrInfo *TII = QTM.getInstrInfo(); + + // Loop over all of the basic blocks + for (MachineFunction::iterator MBBb = Fn.begin(), MBBe = Fn.end(); + MBBb != MBBe; ++MBBb) { + MachineBasicBlock* MBB = MBBb; + // Traverse the basic block + MachineBasicBlock::iterator MII = MBB->begin(); + MachineBasicBlock::iterator MIE = MBB->end (); + while (MII != MIE) { + MachineInstr *MI = MII; + int Opc = MI->getOpcode(); + if (Opc == Hexagon::CONST32_set) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_set_jt) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO_jt), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI_jt), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_Label) { + int DestReg = MI->getOperand(0).getReg(); + MachineOperand &Symbol = MI->getOperand (1); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LO_label), DestReg).addOperand(Symbol); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HI_label), DestReg).addOperand(Symbol); + // MBB->erase returns the iterator to the next instruction, which is the + // one we want to process next + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST32_Int_Real) { + int DestReg = MI->getOperand(0).getReg(); + int64_t ImmValue = MI->getOperand(1).getImm (); + + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestReg).addImm(ImmValue); + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestReg).addImm(ImmValue); + MII = MBB->erase (MI); + continue; + } + else if (Opc == Hexagon::CONST64_Int_Real) { + int DestReg = MI->getOperand(0).getReg(); + int64_t ImmValue = MI->getOperand(1).getImm (); + unsigned DestLo = + QTM.getRegisterInfo()->getSubReg (DestReg, Hexagon::subreg_loreg); + unsigned DestHi = + QTM.getRegisterInfo()->getSubReg (DestReg, Hexagon::subreg_hireg); + + int32_t LowWord = (ImmValue & 0xFFFFFFFF); + int32_t HighWord = (ImmValue >> 32) & 0xFFFFFFFF; + + // Lower Registers Lower Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestLo).addImm(LowWord); + // Lower Registers Higher Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestLo).addImm(LowWord); + // Higher Registers Lower Half + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::LOi), DestHi).addImm(HighWord); + // Higher Registers Higher Half. + BuildMI (*MBB, MII, MI->getDebugLoc(), + TII->get(Hexagon::HIi), DestHi).addImm(HighWord); + MII = MBB->erase (MI); + continue; + } + ++MII; + } + } + + return true; +} + +} + +//===----------------------------------------------------------------------===// +// Public Constructor Functions +//===----------------------------------------------------------------------===// + +FunctionPass * +llvm::createHexagonSplitConst32AndConst64(const HexagonTargetMachine &TM) { + return new HexagonSplitConst32AndConst64(TM); +} diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp index 07d5ce1..fca6707 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp @@ -86,3 +86,5 @@ HexagonSubtarget::HexagonSubtarget(StringRef TT, StringRef CPU, StringRef FS): ModeIEEERndNear = false; } +// Pin the vtable to this file. +void HexagonSubtarget::anchor() {} diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.h b/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.h index 76a8fba..690bef0 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonSubtarget.h @@ -27,7 +27,7 @@ namespace llvm { class HexagonSubtarget : public HexagonGenSubtargetInfo { - + virtual void anchor(); bool UseMemOps; bool ModeIEEERndNear; diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonTargetMachine.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonTargetMachine.cpp index caa1ba4..bb950a0 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonTargetMachine.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonTargetMachine.cpp @@ -15,6 +15,7 @@ #include "Hexagon.h" #include "HexagonISelLowering.h" #include "HexagonMachineScheduler.h" +#include "HexagonTargetObjectFile.h" #include "llvm/CodeGen/Passes.h" #include "llvm/IR/Module.h" #include "llvm/PassManager.h" @@ -78,6 +79,7 @@ HexagonTargetMachine::HexagonTargetMachine(const Target &T, StringRef TT, FrameLowering(Subtarget), InstrItins(&Subtarget.getInstrItineraryData()) { setMCUseCFI(false); + initAsmInfo(); } // addPassesForOptimizations - Allow the backend (target) to add Target @@ -100,17 +102,25 @@ class HexagonPassConfig : public TargetPassConfig { public: HexagonPassConfig(HexagonTargetMachine *TM, PassManagerBase &PM) : TargetPassConfig(TM, PM) { - // Enable MI scheduler. - if (!DisableHexagonMISched) { + // FIXME: Rather than calling enablePass(&MachineSchedulerID) below, define + // HexagonSubtarget::enableMachineScheduler() { return true; }. + // That will bypass the SelectionDAG VLIW scheduler, which is probably just + // hurting compile time and will be removed eventually anyway. + if (DisableHexagonMISched) + disablePass(&MachineSchedulerID); + else enablePass(&MachineSchedulerID); - MachineSchedRegistry::setDefault(createVLIWMachineSched); - } } HexagonTargetMachine &getHexagonTargetMachine() const { return getTM<HexagonTargetMachine>(); } + virtual ScheduleDAGInstrs * + createMachineScheduler(MachineSchedContext *C) const { + return createVLIWMachineSched(C); + } + virtual bool addInstSelector(); virtual bool addPreRegAlloc(); virtual bool addPostRegAlloc(); @@ -124,7 +134,7 @@ TargetPassConfig *HexagonTargetMachine::createPassConfig(PassManagerBase &PM) { } bool HexagonPassConfig::addInstSelector() { - const HexagonTargetMachine &TM = getHexagonTargetMachine(); + HexagonTargetMachine &TM = getHexagonTargetMachine(); bool NoOpt = (getOptLevel() == CodeGenOpt::None); if (!NoOpt) @@ -156,9 +166,18 @@ bool HexagonPassConfig::addPostRegAlloc() { } bool HexagonPassConfig::addPreSched2() { + const HexagonTargetMachine &TM = getHexagonTargetMachine(); + const HexagonTargetObjectFile &TLOF = + (const HexagonTargetObjectFile &)getTargetLowering()->getObjFileLowering(); + + addPass(createHexagonCopyToCombine()); if (getOptLevel() != CodeGenOpt::None) addPass(&IfConverterID); - return false; + if (!TLOF.IsSmallDataEnabled()) { + addPass(createHexagonSplitConst32AndConst64(TM)); + printAndVerify("After hexagon split const32/64 pass"); + } + return true; } bool HexagonPassConfig::addPreEmitPass() { diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.cpp index 993fcfa..7773cff 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.cpp @@ -25,7 +25,8 @@ using namespace llvm; static cl::opt<int> SmallDataThreshold("hexagon-small-data-threshold", - cl::init(8), cl::Hidden); + cl::init(8), cl::Hidden, + cl::desc("The maximum size of an object in the sdata section")); void HexagonTargetObjectFile::Initialize(MCContext &Ctx, const TargetMachine &TM) { @@ -46,6 +47,11 @@ void HexagonTargetObjectFile::Initialize(MCContext &Ctx, static bool IsInSmallSection(uint64_t Size) { return Size > 0 && Size <= (uint64_t)SmallDataThreshold; } + +bool HexagonTargetObjectFile::IsSmallDataEnabled () const { + return SmallDataThreshold > 0; +} + /// IsGlobalInSmallSection - Return true if this global value should be /// placed into small data/bss section. bool HexagonTargetObjectFile::IsGlobalInSmallSection(const GlobalValue *GV, diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.h b/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.h index 6933450..41f6792 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.h +++ b/contrib/llvm/lib/Target/Hexagon/HexagonTargetObjectFile.h @@ -29,6 +29,7 @@ namespace llvm { bool IsGlobalInSmallSection(const GlobalValue *GV, const TargetMachine &TM) const; + bool IsSmallDataEnabled () const; const MCSection* SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind, Mangler *Mang, diff --git a/contrib/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp b/contrib/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp index 39995e1..41e382d 100644 --- a/contrib/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp +++ b/contrib/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp @@ -164,7 +164,6 @@ namespace { unsigned, std::map <MachineInstr*, SUnit*>); bool isNewifiable(MachineInstr* MI); bool isCondInst(MachineInstr* MI); - bool IsNewifyStore (MachineInstr* MI); bool tryAllocateResourcesForConstExt(MachineInstr* MI); bool canReserveResourcesForConstExt(MachineInstr *MI); void reserveResourcesForConstExt(MachineInstr* MI); @@ -383,104 +382,6 @@ static bool IsControlFlow(MachineInstr* MI) { return (MI->getDesc().isTerminator() || MI->getDesc().isCall()); } -// Function returns true if an instruction can be promoted to the new-value -// store. It will always return false for v2 and v3. -// It lists all the conditional and unconditional stores that can be promoted -// to the new-value stores. - -bool HexagonPacketizerList::IsNewifyStore (MachineInstr* MI) { - const HexagonRegisterInfo* QRI = - (const HexagonRegisterInfo *) TM.getRegisterInfo(); - switch (MI->getOpcode()) - { - // store byte - case Hexagon::STrib: - case Hexagon::STrib_indexed: - case Hexagon::STrib_indexed_shl_V4: - case Hexagon::STrib_shl_V4: - case Hexagon::STb_GP_V4: - case Hexagon::POST_STbri: - case Hexagon::STrib_cPt: - case Hexagon::STrib_cdnPt_V4: - case Hexagon::STrib_cNotPt: - case Hexagon::STrib_cdnNotPt_V4: - case Hexagon::STrib_indexed_cPt: - case Hexagon::STrib_indexed_cdnPt_V4: - case Hexagon::STrib_indexed_cNotPt: - case Hexagon::STrib_indexed_cdnNotPt_V4: - case Hexagon::STrib_indexed_shl_cPt_V4: - case Hexagon::STrib_indexed_shl_cdnPt_V4: - case Hexagon::STrib_indexed_shl_cNotPt_V4: - case Hexagon::STrib_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_STbri_cPt: - case Hexagon::POST_STbri_cdnPt_V4: - case Hexagon::POST_STbri_cNotPt: - case Hexagon::POST_STbri_cdnNotPt_V4: - case Hexagon::STb_GP_cPt_V4: - case Hexagon::STb_GP_cNotPt_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - - // store halfword - case Hexagon::STrih: - case Hexagon::STrih_indexed: - case Hexagon::STrih_indexed_shl_V4: - case Hexagon::STrih_shl_V4: - case Hexagon::STh_GP_V4: - case Hexagon::POST_SThri: - case Hexagon::STrih_cPt: - case Hexagon::STrih_cdnPt_V4: - case Hexagon::STrih_cNotPt: - case Hexagon::STrih_cdnNotPt_V4: - case Hexagon::STrih_indexed_cPt: - case Hexagon::STrih_indexed_cdnPt_V4: - case Hexagon::STrih_indexed_cNotPt: - case Hexagon::STrih_indexed_cdnNotPt_V4: - case Hexagon::STrih_indexed_shl_cPt_V4: - case Hexagon::STrih_indexed_shl_cdnPt_V4: - case Hexagon::STrih_indexed_shl_cNotPt_V4: - case Hexagon::STrih_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_SThri_cPt: - case Hexagon::POST_SThri_cdnPt_V4: - case Hexagon::POST_SThri_cNotPt: - case Hexagon::POST_SThri_cdnNotPt_V4: - case Hexagon::STh_GP_cPt_V4: - case Hexagon::STh_GP_cNotPt_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - - // store word - case Hexagon::STriw: - case Hexagon::STriw_indexed: - case Hexagon::STriw_indexed_shl_V4: - case Hexagon::STriw_shl_V4: - case Hexagon::STw_GP_V4: - case Hexagon::POST_STwri: - case Hexagon::STriw_cPt: - case Hexagon::STriw_cdnPt_V4: - case Hexagon::STriw_cNotPt: - case Hexagon::STriw_cdnNotPt_V4: - case Hexagon::STriw_indexed_cPt: - case Hexagon::STriw_indexed_cdnPt_V4: - case Hexagon::STriw_indexed_cNotPt: - case Hexagon::STriw_indexed_cdnNotPt_V4: - case Hexagon::STriw_indexed_shl_cPt_V4: - case Hexagon::STriw_indexed_shl_cdnPt_V4: - case Hexagon::STriw_indexed_shl_cNotPt_V4: - case Hexagon::STriw_indexed_shl_cdnNotPt_V4: - case Hexagon::POST_STwri_cPt: - case Hexagon::POST_STwri_cdnPt_V4: - case Hexagon::POST_STwri_cNotPt: - case Hexagon::POST_STwri_cdnNotPt_V4: - case Hexagon::STw_GP_cPt_V4: - case Hexagon::STw_GP_cNotPt_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - return QRI->Subtarget.hasV4TOps(); - } - return false; -} - static bool IsLoopN(MachineInstr *MI) { return (MI->getOpcode() == Hexagon::LOOP0_i || MI->getOpcode() == Hexagon::LOOP0_r); @@ -498,769 +399,11 @@ static bool DoesModifyCalleeSavedReg(MachineInstr *MI, return false; } -// Return the new value instruction for a given store. -static int GetDotNewOp(const int opc) { - switch (opc) { - default: llvm_unreachable("Unknown .new type"); - // store new value byte - case Hexagon::STrib: - return Hexagon::STrib_nv_V4; - - case Hexagon::STrib_indexed: - return Hexagon::STrib_indexed_nv_V4; - - case Hexagon::STrib_indexed_shl_V4: - return Hexagon::STrib_indexed_shl_nv_V4; - - case Hexagon::STrib_shl_V4: - return Hexagon::STrib_shl_nv_V4; - - case Hexagon::STb_GP_V4: - return Hexagon::STb_GP_nv_V4; - - case Hexagon::POST_STbri: - return Hexagon::POST_STbri_nv_V4; - - case Hexagon::STrib_cPt: - return Hexagon::STrib_cPt_nv_V4; - - case Hexagon::STrib_cdnPt_V4: - return Hexagon::STrib_cdnPt_nv_V4; - - case Hexagon::STrib_cNotPt: - return Hexagon::STrib_cNotPt_nv_V4; - - case Hexagon::STrib_cdnNotPt_V4: - return Hexagon::STrib_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_cPt: - return Hexagon::STrib_indexed_cPt_nv_V4; - - case Hexagon::STrib_indexed_cdnPt_V4: - return Hexagon::STrib_indexed_cdnPt_nv_V4; - - case Hexagon::STrib_indexed_cNotPt: - return Hexagon::STrib_indexed_cNotPt_nv_V4; - - case Hexagon::STrib_indexed_cdnNotPt_V4: - return Hexagon::STrib_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4: - return Hexagon::STrib_indexed_shl_cPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cdnPt_V4: - return Hexagon::STrib_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cNotPt_V4: - return Hexagon::STrib_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cdnNotPt_V4: - return Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STbri_cPt: - return Hexagon::POST_STbri_cPt_nv_V4; - - case Hexagon::POST_STbri_cdnPt_V4: - return Hexagon::POST_STbri_cdnPt_nv_V4; - - case Hexagon::POST_STbri_cNotPt: - return Hexagon::POST_STbri_cNotPt_nv_V4; - - case Hexagon::POST_STbri_cdnNotPt_V4: - return Hexagon::POST_STbri_cdnNotPt_nv_V4; - - case Hexagon::STb_GP_cPt_V4: - return Hexagon::STb_GP_cPt_nv_V4; - - case Hexagon::STb_GP_cNotPt_V4: - return Hexagon::STb_GP_cNotPt_nv_V4; - - case Hexagon::STb_GP_cdnPt_V4: - return Hexagon::STb_GP_cdnPt_nv_V4; - - case Hexagon::STb_GP_cdnNotPt_V4: - return Hexagon::STb_GP_cdnNotPt_nv_V4; - - // store new value halfword - case Hexagon::STrih: - return Hexagon::STrih_nv_V4; - - case Hexagon::STrih_indexed: - return Hexagon::STrih_indexed_nv_V4; - - case Hexagon::STrih_indexed_shl_V4: - return Hexagon::STrih_indexed_shl_nv_V4; - - case Hexagon::STrih_shl_V4: - return Hexagon::STrih_shl_nv_V4; - - case Hexagon::STh_GP_V4: - return Hexagon::STh_GP_nv_V4; - - case Hexagon::POST_SThri: - return Hexagon::POST_SThri_nv_V4; - - case Hexagon::STrih_cPt: - return Hexagon::STrih_cPt_nv_V4; - - case Hexagon::STrih_cdnPt_V4: - return Hexagon::STrih_cdnPt_nv_V4; - - case Hexagon::STrih_cNotPt: - return Hexagon::STrih_cNotPt_nv_V4; - - case Hexagon::STrih_cdnNotPt_V4: - return Hexagon::STrih_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_cPt: - return Hexagon::STrih_indexed_cPt_nv_V4; - - case Hexagon::STrih_indexed_cdnPt_V4: - return Hexagon::STrih_indexed_cdnPt_nv_V4; - - case Hexagon::STrih_indexed_cNotPt: - return Hexagon::STrih_indexed_cNotPt_nv_V4; - - case Hexagon::STrih_indexed_cdnNotPt_V4: - return Hexagon::STrih_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4: - return Hexagon::STrih_indexed_shl_cPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cdnPt_V4: - return Hexagon::STrih_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cNotPt_V4: - return Hexagon::STrih_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cdnNotPt_V4: - return Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_SThri_cPt: - return Hexagon::POST_SThri_cPt_nv_V4; - - case Hexagon::POST_SThri_cdnPt_V4: - return Hexagon::POST_SThri_cdnPt_nv_V4; - - case Hexagon::POST_SThri_cNotPt: - return Hexagon::POST_SThri_cNotPt_nv_V4; - - case Hexagon::POST_SThri_cdnNotPt_V4: - return Hexagon::POST_SThri_cdnNotPt_nv_V4; - - case Hexagon::STh_GP_cPt_V4: - return Hexagon::STh_GP_cPt_nv_V4; - - case Hexagon::STh_GP_cNotPt_V4: - return Hexagon::STh_GP_cNotPt_nv_V4; - - case Hexagon::STh_GP_cdnPt_V4: - return Hexagon::STh_GP_cdnPt_nv_V4; - - case Hexagon::STh_GP_cdnNotPt_V4: - return Hexagon::STh_GP_cdnNotPt_nv_V4; - - // store new value word - case Hexagon::STriw: - return Hexagon::STriw_nv_V4; - - case Hexagon::STriw_indexed: - return Hexagon::STriw_indexed_nv_V4; - - case Hexagon::STriw_indexed_shl_V4: - return Hexagon::STriw_indexed_shl_nv_V4; - - case Hexagon::STriw_shl_V4: - return Hexagon::STriw_shl_nv_V4; - - case Hexagon::STw_GP_V4: - return Hexagon::STw_GP_nv_V4; - - case Hexagon::POST_STwri: - return Hexagon::POST_STwri_nv_V4; - - case Hexagon::STriw_cPt: - return Hexagon::STriw_cPt_nv_V4; - - case Hexagon::STriw_cdnPt_V4: - return Hexagon::STriw_cdnPt_nv_V4; - - case Hexagon::STriw_cNotPt: - return Hexagon::STriw_cNotPt_nv_V4; - - case Hexagon::STriw_cdnNotPt_V4: - return Hexagon::STriw_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_cPt: - return Hexagon::STriw_indexed_cPt_nv_V4; - - case Hexagon::STriw_indexed_cdnPt_V4: - return Hexagon::STriw_indexed_cdnPt_nv_V4; - - case Hexagon::STriw_indexed_cNotPt: - return Hexagon::STriw_indexed_cNotPt_nv_V4; - - case Hexagon::STriw_indexed_cdnNotPt_V4: - return Hexagon::STriw_indexed_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cPt_V4: - return Hexagon::STriw_indexed_shl_cPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cdnPt_V4: - return Hexagon::STriw_indexed_shl_cdnPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cNotPt_V4: - return Hexagon::STriw_indexed_shl_cNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cdnNotPt_V4: - return Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STwri_cPt: - return Hexagon::POST_STwri_cPt_nv_V4; - - case Hexagon::POST_STwri_cdnPt_V4: - return Hexagon::POST_STwri_cdnPt_nv_V4; - - case Hexagon::POST_STwri_cNotPt: - return Hexagon::POST_STwri_cNotPt_nv_V4; - - case Hexagon::POST_STwri_cdnNotPt_V4: - return Hexagon::POST_STwri_cdnNotPt_nv_V4; - - case Hexagon::STw_GP_cPt_V4: - return Hexagon::STw_GP_cPt_nv_V4; - - case Hexagon::STw_GP_cNotPt_V4: - return Hexagon::STw_GP_cNotPt_nv_V4; - - case Hexagon::STw_GP_cdnPt_V4: - return Hexagon::STw_GP_cdnPt_nv_V4; - - case Hexagon::STw_GP_cdnNotPt_V4: - return Hexagon::STw_GP_cdnNotPt_nv_V4; - - } -} - -// Return .new predicate version for an instruction -static int GetDotNewPredOp(MachineInstr *MI, - const MachineBranchProbabilityInfo *MBPI, - const HexagonInstrInfo *QII) { - switch (MI->getOpcode()) { - default: llvm_unreachable("Unknown .new type"); - // Conditional stores - // Store byte conditionally - case Hexagon::STrib_cPt : - return Hexagon::STrib_cdnPt_V4; - - case Hexagon::STrib_cNotPt : - return Hexagon::STrib_cdnNotPt_V4; - - case Hexagon::STrib_indexed_cPt : - return Hexagon::STrib_indexed_cdnPt_V4; - - case Hexagon::STrib_indexed_cNotPt : - return Hexagon::STrib_indexed_cdnNotPt_V4; - - case Hexagon::STrib_imm_cPt_V4 : - return Hexagon::STrib_imm_cdnPt_V4; - - case Hexagon::STrib_imm_cNotPt_V4 : - return Hexagon::STrib_imm_cdnNotPt_V4; - - case Hexagon::POST_STbri_cPt : - return Hexagon::POST_STbri_cdnPt_V4; - - case Hexagon::POST_STbri_cNotPt : - return Hexagon::POST_STbri_cdnNotPt_V4; - - case Hexagon::STrib_indexed_shl_cPt_V4 : - return Hexagon::STrib_indexed_shl_cdnPt_V4; - - case Hexagon::STrib_indexed_shl_cNotPt_V4 : - return Hexagon::STrib_indexed_shl_cdnNotPt_V4; - - case Hexagon::STb_GP_cPt_V4 : - return Hexagon::STb_GP_cdnPt_V4; - - case Hexagon::STb_GP_cNotPt_V4 : - return Hexagon::STb_GP_cdnNotPt_V4; - - // Store doubleword conditionally - case Hexagon::STrid_cPt : - return Hexagon::STrid_cdnPt_V4; - - case Hexagon::STrid_cNotPt : - return Hexagon::STrid_cdnNotPt_V4; - - case Hexagon::STrid_indexed_cPt : - return Hexagon::STrid_indexed_cdnPt_V4; - - case Hexagon::STrid_indexed_cNotPt : - return Hexagon::STrid_indexed_cdnNotPt_V4; - - case Hexagon::STrid_indexed_shl_cPt_V4 : - return Hexagon::STrid_indexed_shl_cdnPt_V4; - - case Hexagon::STrid_indexed_shl_cNotPt_V4 : - return Hexagon::STrid_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_STdri_cPt : - return Hexagon::POST_STdri_cdnPt_V4; - - case Hexagon::POST_STdri_cNotPt : - return Hexagon::POST_STdri_cdnNotPt_V4; - - case Hexagon::STd_GP_cPt_V4 : - return Hexagon::STd_GP_cdnPt_V4; - - case Hexagon::STd_GP_cNotPt_V4 : - return Hexagon::STd_GP_cdnNotPt_V4; - - // Store halfword conditionally - case Hexagon::STrih_cPt : - return Hexagon::STrih_cdnPt_V4; - - case Hexagon::STrih_cNotPt : - return Hexagon::STrih_cdnNotPt_V4; - - case Hexagon::STrih_indexed_cPt : - return Hexagon::STrih_indexed_cdnPt_V4; - - case Hexagon::STrih_indexed_cNotPt : - return Hexagon::STrih_indexed_cdnNotPt_V4; - - case Hexagon::STrih_imm_cPt_V4 : - return Hexagon::STrih_imm_cdnPt_V4; - - case Hexagon::STrih_imm_cNotPt_V4 : - return Hexagon::STrih_imm_cdnNotPt_V4; - - case Hexagon::STrih_indexed_shl_cPt_V4 : - return Hexagon::STrih_indexed_shl_cdnPt_V4; - - case Hexagon::STrih_indexed_shl_cNotPt_V4 : - return Hexagon::STrih_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_SThri_cPt : - return Hexagon::POST_SThri_cdnPt_V4; - - case Hexagon::POST_SThri_cNotPt : - return Hexagon::POST_SThri_cdnNotPt_V4; - - case Hexagon::STh_GP_cPt_V4 : - return Hexagon::STh_GP_cdnPt_V4; - - case Hexagon::STh_GP_cNotPt_V4 : - return Hexagon::STh_GP_cdnNotPt_V4; - - // Store word conditionally - case Hexagon::STriw_cPt : - return Hexagon::STriw_cdnPt_V4; - - case Hexagon::STriw_cNotPt : - return Hexagon::STriw_cdnNotPt_V4; - - case Hexagon::STriw_indexed_cPt : - return Hexagon::STriw_indexed_cdnPt_V4; - - case Hexagon::STriw_indexed_cNotPt : - return Hexagon::STriw_indexed_cdnNotPt_V4; - - case Hexagon::STriw_imm_cPt_V4 : - return Hexagon::STriw_imm_cdnPt_V4; - - case Hexagon::STriw_imm_cNotPt_V4 : - return Hexagon::STriw_imm_cdnNotPt_V4; - - case Hexagon::STriw_indexed_shl_cPt_V4 : - return Hexagon::STriw_indexed_shl_cdnPt_V4; - - case Hexagon::STriw_indexed_shl_cNotPt_V4 : - return Hexagon::STriw_indexed_shl_cdnNotPt_V4; - - case Hexagon::POST_STwri_cPt : - return Hexagon::POST_STwri_cdnPt_V4; - - case Hexagon::POST_STwri_cNotPt : - return Hexagon::POST_STwri_cdnNotPt_V4; - - case Hexagon::STw_GP_cPt_V4 : - return Hexagon::STw_GP_cdnPt_V4; - - case Hexagon::STw_GP_cNotPt_V4 : - return Hexagon::STw_GP_cdnNotPt_V4; - - // Condtional Jumps - case Hexagon::JMP_t: - case Hexagon::JMP_f: - return QII->getDotNewPredJumpOp(MI, MBPI); - - case Hexagon::JMPR_t: - return Hexagon::JMPR_tnew_tV3; - - case Hexagon::JMPR_f: - return Hexagon::JMPR_fnew_tV3; - - // Conditional Transfers - case Hexagon::TFR_cPt: - return Hexagon::TFR_cdnPt; - - case Hexagon::TFR_cNotPt: - return Hexagon::TFR_cdnNotPt; - - case Hexagon::TFRI_cPt: - return Hexagon::TFRI_cdnPt; - - case Hexagon::TFRI_cNotPt: - return Hexagon::TFRI_cdnNotPt; - - // Load double word - case Hexagon::LDrid_cPt : - return Hexagon::LDrid_cdnPt; - - case Hexagon::LDrid_cNotPt : - return Hexagon::LDrid_cdnNotPt; - - case Hexagon::LDrid_indexed_cPt : - return Hexagon::LDrid_indexed_cdnPt; - - case Hexagon::LDrid_indexed_cNotPt : - return Hexagon::LDrid_indexed_cdnNotPt; - - case Hexagon::POST_LDrid_cPt : - return Hexagon::POST_LDrid_cdnPt_V4; - - case Hexagon::POST_LDrid_cNotPt : - return Hexagon::POST_LDrid_cdnNotPt_V4; - - // Load word - case Hexagon::LDriw_cPt : - return Hexagon::LDriw_cdnPt; - - case Hexagon::LDriw_cNotPt : - return Hexagon::LDriw_cdnNotPt; - - case Hexagon::LDriw_indexed_cPt : - return Hexagon::LDriw_indexed_cdnPt; - - case Hexagon::LDriw_indexed_cNotPt : - return Hexagon::LDriw_indexed_cdnNotPt; - - case Hexagon::POST_LDriw_cPt : - return Hexagon::POST_LDriw_cdnPt_V4; - - case Hexagon::POST_LDriw_cNotPt : - return Hexagon::POST_LDriw_cdnNotPt_V4; - - // Load halfword - case Hexagon::LDrih_cPt : - return Hexagon::LDrih_cdnPt; - - case Hexagon::LDrih_cNotPt : - return Hexagon::LDrih_cdnNotPt; - - case Hexagon::LDrih_indexed_cPt : - return Hexagon::LDrih_indexed_cdnPt; - - case Hexagon::LDrih_indexed_cNotPt : - return Hexagon::LDrih_indexed_cdnNotPt; - - case Hexagon::POST_LDrih_cPt : - return Hexagon::POST_LDrih_cdnPt_V4; - - case Hexagon::POST_LDrih_cNotPt : - return Hexagon::POST_LDrih_cdnNotPt_V4; - - // Load byte - case Hexagon::LDrib_cPt : - return Hexagon::LDrib_cdnPt; - - case Hexagon::LDrib_cNotPt : - return Hexagon::LDrib_cdnNotPt; - - case Hexagon::LDrib_indexed_cPt : - return Hexagon::LDrib_indexed_cdnPt; - - case Hexagon::LDrib_indexed_cNotPt : - return Hexagon::LDrib_indexed_cdnNotPt; - - case Hexagon::POST_LDrib_cPt : - return Hexagon::POST_LDrib_cdnPt_V4; - - case Hexagon::POST_LDrib_cNotPt : - return Hexagon::POST_LDrib_cdnNotPt_V4; - - // Load unsigned halfword - case Hexagon::LDriuh_cPt : - return Hexagon::LDriuh_cdnPt; - - case Hexagon::LDriuh_cNotPt : - return Hexagon::LDriuh_cdnNotPt; - - case Hexagon::LDriuh_indexed_cPt : - return Hexagon::LDriuh_indexed_cdnPt; - - case Hexagon::LDriuh_indexed_cNotPt : - return Hexagon::LDriuh_indexed_cdnNotPt; - - case Hexagon::POST_LDriuh_cPt : - return Hexagon::POST_LDriuh_cdnPt_V4; - - case Hexagon::POST_LDriuh_cNotPt : - return Hexagon::POST_LDriuh_cdnNotPt_V4; - - // Load unsigned byte - case Hexagon::LDriub_cPt : - return Hexagon::LDriub_cdnPt; - - case Hexagon::LDriub_cNotPt : - return Hexagon::LDriub_cdnNotPt; - - case Hexagon::LDriub_indexed_cPt : - return Hexagon::LDriub_indexed_cdnPt; - - case Hexagon::LDriub_indexed_cNotPt : - return Hexagon::LDriub_indexed_cdnNotPt; - - case Hexagon::POST_LDriub_cPt : - return Hexagon::POST_LDriub_cdnPt_V4; - - case Hexagon::POST_LDriub_cNotPt : - return Hexagon::POST_LDriub_cdnNotPt_V4; - - // V4 indexed+scaled load - - case Hexagon::LDrid_indexed_shl_cPt_V4 : - return Hexagon::LDrid_indexed_shl_cdnPt_V4; - - case Hexagon::LDrid_indexed_shl_cNotPt_V4 : - return Hexagon::LDrid_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDrib_indexed_shl_cPt_V4 : - return Hexagon::LDrib_indexed_shl_cdnPt_V4; - - case Hexagon::LDrib_indexed_shl_cNotPt_V4 : - return Hexagon::LDrib_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriub_indexed_shl_cPt_V4 : - return Hexagon::LDriub_indexed_shl_cdnPt_V4; - - case Hexagon::LDriub_indexed_shl_cNotPt_V4 : - return Hexagon::LDriub_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDrih_indexed_shl_cPt_V4 : - return Hexagon::LDrih_indexed_shl_cdnPt_V4; - - case Hexagon::LDrih_indexed_shl_cNotPt_V4 : - return Hexagon::LDrih_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriuh_indexed_shl_cPt_V4 : - return Hexagon::LDriuh_indexed_shl_cdnPt_V4; - - case Hexagon::LDriuh_indexed_shl_cNotPt_V4 : - return Hexagon::LDriuh_indexed_shl_cdnNotPt_V4; - - case Hexagon::LDriw_indexed_shl_cPt_V4 : - return Hexagon::LDriw_indexed_shl_cdnPt_V4; - - case Hexagon::LDriw_indexed_shl_cNotPt_V4 : - return Hexagon::LDriw_indexed_shl_cdnNotPt_V4; - - // V4 global address load - - case Hexagon::LDd_GP_cPt_V4: - return Hexagon::LDd_GP_cdnPt_V4; - - case Hexagon::LDd_GP_cNotPt_V4: - return Hexagon::LDd_GP_cdnNotPt_V4; - - case Hexagon::LDb_GP_cPt_V4: - return Hexagon::LDb_GP_cdnPt_V4; - - case Hexagon::LDb_GP_cNotPt_V4: - return Hexagon::LDb_GP_cdnNotPt_V4; - - case Hexagon::LDub_GP_cPt_V4: - return Hexagon::LDub_GP_cdnPt_V4; - - case Hexagon::LDub_GP_cNotPt_V4: - return Hexagon::LDub_GP_cdnNotPt_V4; - - case Hexagon::LDh_GP_cPt_V4: - return Hexagon::LDh_GP_cdnPt_V4; - - case Hexagon::LDh_GP_cNotPt_V4: - return Hexagon::LDh_GP_cdnNotPt_V4; - - case Hexagon::LDuh_GP_cPt_V4: - return Hexagon::LDuh_GP_cdnPt_V4; - - case Hexagon::LDuh_GP_cNotPt_V4: - return Hexagon::LDuh_GP_cdnNotPt_V4; - - case Hexagon::LDw_GP_cPt_V4: - return Hexagon::LDw_GP_cdnPt_V4; - - case Hexagon::LDw_GP_cNotPt_V4: - return Hexagon::LDw_GP_cdnNotPt_V4; - - // Conditional store new-value byte - case Hexagon::STrib_cPt_nv_V4 : - return Hexagon::STrib_cdnPt_nv_V4; - case Hexagon::STrib_cNotPt_nv_V4 : - return Hexagon::STrib_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_cPt_nv_V4 : - return Hexagon::STrib_indexed_cdnPt_nv_V4; - case Hexagon::STrib_indexed_cNotPt_nv_V4 : - return Hexagon::STrib_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrib_indexed_shl_cPt_nv_V4 : - return Hexagon::STrib_indexed_shl_cdnPt_nv_V4; - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STbri_cPt_nv_V4 : - return Hexagon::POST_STbri_cdnPt_nv_V4; - case Hexagon::POST_STbri_cNotPt_nv_V4 : - return Hexagon::POST_STbri_cdnNotPt_nv_V4; - - case Hexagon::STb_GP_cPt_nv_V4 : - return Hexagon::STb_GP_cdnPt_nv_V4; - - case Hexagon::STb_GP_cNotPt_nv_V4 : - return Hexagon::STb_GP_cdnNotPt_nv_V4; - - // Conditional store new-value halfword - case Hexagon::STrih_cPt_nv_V4 : - return Hexagon::STrih_cdnPt_nv_V4; - case Hexagon::STrih_cNotPt_nv_V4 : - return Hexagon::STrih_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_cPt_nv_V4 : - return Hexagon::STrih_indexed_cdnPt_nv_V4; - case Hexagon::STrih_indexed_cNotPt_nv_V4 : - return Hexagon::STrih_indexed_cdnNotPt_nv_V4; - - case Hexagon::STrih_indexed_shl_cPt_nv_V4 : - return Hexagon::STrih_indexed_shl_cdnPt_nv_V4; - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_SThri_cPt_nv_V4 : - return Hexagon::POST_SThri_cdnPt_nv_V4; - case Hexagon::POST_SThri_cNotPt_nv_V4 : - return Hexagon::POST_SThri_cdnNotPt_nv_V4; - - case Hexagon::STh_GP_cPt_nv_V4 : - return Hexagon::STh_GP_cdnPt_nv_V4; - - case Hexagon::STh_GP_cNotPt_nv_V4 : - return Hexagon::STh_GP_cdnNotPt_nv_V4; - - // Conditional store new-value word - case Hexagon::STriw_cPt_nv_V4 : - return Hexagon::STriw_cdnPt_nv_V4; - case Hexagon::STriw_cNotPt_nv_V4 : - return Hexagon::STriw_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_cPt_nv_V4 : - return Hexagon::STriw_indexed_cdnPt_nv_V4; - case Hexagon::STriw_indexed_cNotPt_nv_V4 : - return Hexagon::STriw_indexed_cdnNotPt_nv_V4; - - case Hexagon::STriw_indexed_shl_cPt_nv_V4 : - return Hexagon::STriw_indexed_shl_cdnPt_nv_V4; - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4 : - return Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4; - - case Hexagon::POST_STwri_cPt_nv_V4 : - return Hexagon::POST_STwri_cdnPt_nv_V4; - case Hexagon::POST_STwri_cNotPt_nv_V4: - return Hexagon::POST_STwri_cdnNotPt_nv_V4; - - case Hexagon::STw_GP_cPt_nv_V4 : - return Hexagon::STw_GP_cdnPt_nv_V4; - - case Hexagon::STw_GP_cNotPt_nv_V4 : - return Hexagon::STw_GP_cdnNotPt_nv_V4; - - // Conditional add - case Hexagon::ADD_ri_cPt : - return Hexagon::ADD_ri_cdnPt; - case Hexagon::ADD_ri_cNotPt : - return Hexagon::ADD_ri_cdnNotPt; - - case Hexagon::ADD_rr_cPt : - return Hexagon::ADD_rr_cdnPt; - case Hexagon::ADD_rr_cNotPt : - return Hexagon::ADD_rr_cdnNotPt; - - // Conditional logical Operations - case Hexagon::XOR_rr_cPt : - return Hexagon::XOR_rr_cdnPt; - case Hexagon::XOR_rr_cNotPt : - return Hexagon::XOR_rr_cdnNotPt; - - case Hexagon::AND_rr_cPt : - return Hexagon::AND_rr_cdnPt; - case Hexagon::AND_rr_cNotPt : - return Hexagon::AND_rr_cdnNotPt; - - case Hexagon::OR_rr_cPt : - return Hexagon::OR_rr_cdnPt; - case Hexagon::OR_rr_cNotPt : - return Hexagon::OR_rr_cdnNotPt; - - // Conditional Subtract - case Hexagon::SUB_rr_cPt : - return Hexagon::SUB_rr_cdnPt; - case Hexagon::SUB_rr_cNotPt : - return Hexagon::SUB_rr_cdnNotPt; - - // Conditional combine - case Hexagon::COMBINE_rr_cPt : - return Hexagon::COMBINE_rr_cdnPt; - case Hexagon::COMBINE_rr_cNotPt : - return Hexagon::COMBINE_rr_cdnNotPt; - - case Hexagon::ASLH_cPt_V4 : - return Hexagon::ASLH_cdnPt_V4; - case Hexagon::ASLH_cNotPt_V4 : - return Hexagon::ASLH_cdnNotPt_V4; - - case Hexagon::ASRH_cPt_V4 : - return Hexagon::ASRH_cdnPt_V4; - case Hexagon::ASRH_cNotPt_V4 : - return Hexagon::ASRH_cdnNotPt_V4; - - case Hexagon::SXTB_cPt_V4 : - return Hexagon::SXTB_cdnPt_V4; - case Hexagon::SXTB_cNotPt_V4 : - return Hexagon::SXTB_cdnNotPt_V4; - - case Hexagon::SXTH_cPt_V4 : - return Hexagon::SXTH_cdnPt_V4; - case Hexagon::SXTH_cNotPt_V4 : - return Hexagon::SXTH_cdnNotPt_V4; - - case Hexagon::ZXTB_cPt_V4 : - return Hexagon::ZXTB_cdnPt_V4; - case Hexagon::ZXTB_cNotPt_V4 : - return Hexagon::ZXTB_cdnNotPt_V4; - - case Hexagon::ZXTH_cPt_V4 : - return Hexagon::ZXTH_cdnPt_V4; - case Hexagon::ZXTH_cNotPt_V4 : - return Hexagon::ZXTH_cdnNotPt_V4; - } -} - // Returns true if an instruction can be promoted to .new predicate // or new-value store. bool HexagonPacketizerList::isNewifiable(MachineInstr* MI) { - if ( isCondInst(MI) || IsNewifyStore(MI)) + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + if ( isCondInst(MI) || QII->mayBeNewStore(MI)) return true; else return false; @@ -1294,896 +437,38 @@ bool HexagonPacketizerList::PromoteToDotNew(MachineInstr* MI, int NewOpcode; if (RC == &Hexagon::PredRegsRegClass) - NewOpcode = GetDotNewPredOp(MI, MBPI, QII); + NewOpcode = QII->GetDotNewPredOp(MI, MBPI); else - NewOpcode = GetDotNewOp(MI->getOpcode()); + NewOpcode = QII->GetDotNewOp(MI); MI->setDesc(QII->get(NewOpcode)); return true; } -// Returns the most basic instruction for the .new predicated instructions and -// new-value stores. -// For example, all of the following instructions will be converted back to the -// same instruction: -// 1) if (p0.new) memw(R0+#0) = R1.new ---> -// 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1 -// 3) if (p0.new) memw(R0+#0) = R1 ---> -// -// To understand the translation of instruction 1 to its original form, consider -// a packet with 3 instructions. -// { p0 = cmp.eq(R0,R1) -// if (p0.new) R2 = add(R3, R4) -// R5 = add (R3, R1) -// } -// if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet -// -// This instruction can be part of the previous packet only if both p0 and R2 -// are promoted to .new values. This promotion happens in steps, first -// predicate register is promoted to .new and in the next iteration R2 is -// promoted. Therefore, in case of dependence check failure (due to R5) during -// next iteration, it should be converted back to its most basic form. - -static int GetDotOldOp(const int opc) { - switch (opc) { - default: llvm_unreachable("Unknown .old type"); - case Hexagon::TFR_cdnPt: - return Hexagon::TFR_cPt; - - case Hexagon::TFR_cdnNotPt: - return Hexagon::TFR_cNotPt; - - case Hexagon::TFRI_cdnPt: - return Hexagon::TFRI_cPt; - - case Hexagon::TFRI_cdnNotPt: - return Hexagon::TFRI_cNotPt; - - case Hexagon::JMP_tnew_t: - return Hexagon::JMP_t; - - case Hexagon::JMP_fnew_t: - return Hexagon::JMP_f; - - case Hexagon::JMPR_tnew_tV3: - return Hexagon::JMPR_t; - - case Hexagon::JMPR_fnew_tV3: - return Hexagon::JMPR_f; - - // Load double word - - case Hexagon::LDrid_cdnPt : - return Hexagon::LDrid_cPt; - - case Hexagon::LDrid_cdnNotPt : - return Hexagon::LDrid_cNotPt; - - case Hexagon::LDrid_indexed_cdnPt : - return Hexagon::LDrid_indexed_cPt; - - case Hexagon::LDrid_indexed_cdnNotPt : - return Hexagon::LDrid_indexed_cNotPt; - - case Hexagon::POST_LDrid_cdnPt_V4 : - return Hexagon::POST_LDrid_cPt; - - case Hexagon::POST_LDrid_cdnNotPt_V4 : - return Hexagon::POST_LDrid_cNotPt; - - // Load word - - case Hexagon::LDriw_cdnPt : - return Hexagon::LDriw_cPt; - - case Hexagon::LDriw_cdnNotPt : - return Hexagon::LDriw_cNotPt; - - case Hexagon::LDriw_indexed_cdnPt : - return Hexagon::LDriw_indexed_cPt; - - case Hexagon::LDriw_indexed_cdnNotPt : - return Hexagon::LDriw_indexed_cNotPt; - - case Hexagon::POST_LDriw_cdnPt_V4 : - return Hexagon::POST_LDriw_cPt; - - case Hexagon::POST_LDriw_cdnNotPt_V4 : - return Hexagon::POST_LDriw_cNotPt; - - // Load half - - case Hexagon::LDrih_cdnPt : - return Hexagon::LDrih_cPt; - - case Hexagon::LDrih_cdnNotPt : - return Hexagon::LDrih_cNotPt; - - case Hexagon::LDrih_indexed_cdnPt : - return Hexagon::LDrih_indexed_cPt; - - case Hexagon::LDrih_indexed_cdnNotPt : - return Hexagon::LDrih_indexed_cNotPt; - - case Hexagon::POST_LDrih_cdnPt_V4 : - return Hexagon::POST_LDrih_cPt; - - case Hexagon::POST_LDrih_cdnNotPt_V4 : - return Hexagon::POST_LDrih_cNotPt; - - // Load byte - - case Hexagon::LDrib_cdnPt : - return Hexagon::LDrib_cPt; - - case Hexagon::LDrib_cdnNotPt : - return Hexagon::LDrib_cNotPt; - - case Hexagon::LDrib_indexed_cdnPt : - return Hexagon::LDrib_indexed_cPt; - - case Hexagon::LDrib_indexed_cdnNotPt : - return Hexagon::LDrib_indexed_cNotPt; - - case Hexagon::POST_LDrib_cdnPt_V4 : - return Hexagon::POST_LDrib_cPt; - - case Hexagon::POST_LDrib_cdnNotPt_V4 : - return Hexagon::POST_LDrib_cNotPt; - - // Load unsigned half - - case Hexagon::LDriuh_cdnPt : - return Hexagon::LDriuh_cPt; - - case Hexagon::LDriuh_cdnNotPt : - return Hexagon::LDriuh_cNotPt; - - case Hexagon::LDriuh_indexed_cdnPt : - return Hexagon::LDriuh_indexed_cPt; - - case Hexagon::LDriuh_indexed_cdnNotPt : - return Hexagon::LDriuh_indexed_cNotPt; - - case Hexagon::POST_LDriuh_cdnPt_V4 : - return Hexagon::POST_LDriuh_cPt; - - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - return Hexagon::POST_LDriuh_cNotPt; - - // Load unsigned byte - case Hexagon::LDriub_cdnPt : - return Hexagon::LDriub_cPt; - - case Hexagon::LDriub_cdnNotPt : - return Hexagon::LDriub_cNotPt; - - case Hexagon::LDriub_indexed_cdnPt : - return Hexagon::LDriub_indexed_cPt; - - case Hexagon::LDriub_indexed_cdnNotPt : - return Hexagon::LDriub_indexed_cNotPt; - - case Hexagon::POST_LDriub_cdnPt_V4 : - return Hexagon::POST_LDriub_cPt; - - case Hexagon::POST_LDriub_cdnNotPt_V4 : - return Hexagon::POST_LDriub_cNotPt; - - // V4 indexed+scaled Load - - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - return Hexagon::LDrid_indexed_shl_cPt_V4; - - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrid_indexed_shl_cNotPt_V4; - - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - return Hexagon::LDrib_indexed_shl_cPt_V4; - - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrib_indexed_shl_cNotPt_V4; - - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - return Hexagon::LDriub_indexed_shl_cPt_V4; - - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriub_indexed_shl_cNotPt_V4; - - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - return Hexagon::LDrih_indexed_shl_cPt_V4; - - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDrih_indexed_shl_cNotPt_V4; - - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - return Hexagon::LDriuh_indexed_shl_cPt_V4; - - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriuh_indexed_shl_cNotPt_V4; - - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - return Hexagon::LDriw_indexed_shl_cPt_V4; - - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - return Hexagon::LDriw_indexed_shl_cNotPt_V4; - - // V4 global address load - - case Hexagon::LDd_GP_cdnPt_V4: - return Hexagon::LDd_GP_cPt_V4; - - case Hexagon::LDd_GP_cdnNotPt_V4: - return Hexagon::LDd_GP_cNotPt_V4; - - case Hexagon::LDb_GP_cdnPt_V4: - return Hexagon::LDb_GP_cPt_V4; - - case Hexagon::LDb_GP_cdnNotPt_V4: - return Hexagon::LDb_GP_cNotPt_V4; - - case Hexagon::LDub_GP_cdnPt_V4: - return Hexagon::LDub_GP_cPt_V4; - - case Hexagon::LDub_GP_cdnNotPt_V4: - return Hexagon::LDub_GP_cNotPt_V4; - - case Hexagon::LDh_GP_cdnPt_V4: - return Hexagon::LDh_GP_cPt_V4; - - case Hexagon::LDh_GP_cdnNotPt_V4: - return Hexagon::LDh_GP_cNotPt_V4; - - case Hexagon::LDuh_GP_cdnPt_V4: - return Hexagon::LDuh_GP_cPt_V4; - - case Hexagon::LDuh_GP_cdnNotPt_V4: - return Hexagon::LDuh_GP_cNotPt_V4; - - case Hexagon::LDw_GP_cdnPt_V4: - return Hexagon::LDw_GP_cPt_V4; - - case Hexagon::LDw_GP_cdnNotPt_V4: - return Hexagon::LDw_GP_cNotPt_V4; - - // Conditional add - - case Hexagon::ADD_ri_cdnPt : - return Hexagon::ADD_ri_cPt; - case Hexagon::ADD_ri_cdnNotPt : - return Hexagon::ADD_ri_cNotPt; - - case Hexagon::ADD_rr_cdnPt : - return Hexagon::ADD_rr_cPt; - case Hexagon::ADD_rr_cdnNotPt: - return Hexagon::ADD_rr_cNotPt; - - // Conditional logical Operations - - case Hexagon::XOR_rr_cdnPt : - return Hexagon::XOR_rr_cPt; - case Hexagon::XOR_rr_cdnNotPt : - return Hexagon::XOR_rr_cNotPt; - - case Hexagon::AND_rr_cdnPt : - return Hexagon::AND_rr_cPt; - case Hexagon::AND_rr_cdnNotPt : - return Hexagon::AND_rr_cNotPt; - - case Hexagon::OR_rr_cdnPt : - return Hexagon::OR_rr_cPt; - case Hexagon::OR_rr_cdnNotPt : - return Hexagon::OR_rr_cNotPt; - - // Conditional Subtract - - case Hexagon::SUB_rr_cdnPt : - return Hexagon::SUB_rr_cPt; - case Hexagon::SUB_rr_cdnNotPt : - return Hexagon::SUB_rr_cNotPt; - - // Conditional combine - - case Hexagon::COMBINE_rr_cdnPt : - return Hexagon::COMBINE_rr_cPt; - case Hexagon::COMBINE_rr_cdnNotPt : - return Hexagon::COMBINE_rr_cNotPt; - -// Conditional shift operations - - case Hexagon::ASLH_cdnPt_V4 : - return Hexagon::ASLH_cPt_V4; - case Hexagon::ASLH_cdnNotPt_V4 : - return Hexagon::ASLH_cNotPt_V4; - - case Hexagon::ASRH_cdnPt_V4 : - return Hexagon::ASRH_cPt_V4; - case Hexagon::ASRH_cdnNotPt_V4 : - return Hexagon::ASRH_cNotPt_V4; - - case Hexagon::SXTB_cdnPt_V4 : - return Hexagon::SXTB_cPt_V4; - case Hexagon::SXTB_cdnNotPt_V4 : - return Hexagon::SXTB_cNotPt_V4; - - case Hexagon::SXTH_cdnPt_V4 : - return Hexagon::SXTH_cPt_V4; - case Hexagon::SXTH_cdnNotPt_V4 : - return Hexagon::SXTH_cNotPt_V4; - - case Hexagon::ZXTB_cdnPt_V4 : - return Hexagon::ZXTB_cPt_V4; - case Hexagon::ZXTB_cdnNotPt_V4 : - return Hexagon::ZXTB_cNotPt_V4; - - case Hexagon::ZXTH_cdnPt_V4 : - return Hexagon::ZXTH_cPt_V4; - case Hexagon::ZXTH_cdnNotPt_V4 : - return Hexagon::ZXTH_cNotPt_V4; - - // Store byte - - case Hexagon::STrib_imm_cdnPt_V4 : - return Hexagon::STrib_imm_cPt_V4; - - case Hexagon::STrib_imm_cdnNotPt_V4 : - return Hexagon::STrib_imm_cNotPt_V4; - - case Hexagon::STrib_cdnPt_nv_V4 : - case Hexagon::STrib_cPt_nv_V4 : - case Hexagon::STrib_cdnPt_V4 : - return Hexagon::STrib_cPt; - - case Hexagon::STrib_cdnNotPt_nv_V4 : - case Hexagon::STrib_cNotPt_nv_V4 : - case Hexagon::STrib_cdnNotPt_V4 : - return Hexagon::STrib_cNotPt; - - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_cPt_nv_V4 : - case Hexagon::STrib_indexed_cdnPt_nv_V4 : - return Hexagon::STrib_indexed_cPt; - - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cNotPt_nv_V4 : - case Hexagon::STrib_indexed_cdnNotPt_nv_V4 : - return Hexagon::STrib_indexed_cNotPt; - - case Hexagon::STrib_indexed_shl_cdnPt_nv_V4: - case Hexagon::STrib_indexed_shl_cPt_nv_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - return Hexagon::STrib_indexed_shl_cPt_V4; - - case Hexagon::STrib_indexed_shl_cdnNotPt_nv_V4: - case Hexagon::STrib_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrib_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STbri_cdnPt_nv_V4 : - case Hexagon::POST_STbri_cPt_nv_V4 : - case Hexagon::POST_STbri_cdnPt_V4 : - return Hexagon::POST_STbri_cPt; - - case Hexagon::POST_STbri_cdnNotPt_nv_V4 : - case Hexagon::POST_STbri_cNotPt_nv_V4: - case Hexagon::POST_STbri_cdnNotPt_V4 : - return Hexagon::POST_STbri_cNotPt; - - case Hexagon::STb_GP_cdnPt_nv_V4: - case Hexagon::STb_GP_cdnPt_V4: - case Hexagon::STb_GP_cPt_nv_V4: - return Hexagon::STb_GP_cPt_V4; - - case Hexagon::STb_GP_cdnNotPt_nv_V4: - case Hexagon::STb_GP_cdnNotPt_V4: - case Hexagon::STb_GP_cNotPt_nv_V4: - return Hexagon::STb_GP_cNotPt_V4; - - // Store new-value byte - unconditional - case Hexagon::STrib_nv_V4: - return Hexagon::STrib; - - case Hexagon::STrib_indexed_nv_V4: - return Hexagon::STrib_indexed; - - case Hexagon::STrib_indexed_shl_nv_V4: - return Hexagon::STrib_indexed_shl_V4; - - case Hexagon::STrib_shl_nv_V4: - return Hexagon::STrib_shl_V4; - - case Hexagon::STb_GP_nv_V4: - return Hexagon::STb_GP_V4; - - case Hexagon::POST_STbri_nv_V4: - return Hexagon::POST_STbri; - - // Store halfword - case Hexagon::STrih_imm_cdnPt_V4 : - return Hexagon::STrih_imm_cPt_V4; - - case Hexagon::STrih_imm_cdnNotPt_V4 : - return Hexagon::STrih_imm_cNotPt_V4; - - case Hexagon::STrih_cdnPt_nv_V4 : - case Hexagon::STrih_cPt_nv_V4 : - case Hexagon::STrih_cdnPt_V4 : - return Hexagon::STrih_cPt; - - case Hexagon::STrih_cdnNotPt_nv_V4 : - case Hexagon::STrih_cNotPt_nv_V4 : - case Hexagon::STrih_cdnNotPt_V4 : - return Hexagon::STrih_cNotPt; - - case Hexagon::STrih_indexed_cdnPt_nv_V4: - case Hexagon::STrih_indexed_cPt_nv_V4 : - case Hexagon::STrih_indexed_cdnPt_V4 : - return Hexagon::STrih_indexed_cPt; - - case Hexagon::STrih_indexed_cdnNotPt_nv_V4: - case Hexagon::STrih_indexed_cNotPt_nv_V4 : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - return Hexagon::STrih_indexed_cNotPt; - - case Hexagon::STrih_indexed_shl_cdnPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - return Hexagon::STrih_indexed_shl_cPt_V4; - - case Hexagon::STrih_indexed_shl_cdnNotPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrih_indexed_shl_cNotPt_V4; - - case Hexagon::POST_SThri_cdnPt_nv_V4 : - case Hexagon::POST_SThri_cPt_nv_V4 : - case Hexagon::POST_SThri_cdnPt_V4 : - return Hexagon::POST_SThri_cPt; - - case Hexagon::POST_SThri_cdnNotPt_nv_V4 : - case Hexagon::POST_SThri_cNotPt_nv_V4 : - case Hexagon::POST_SThri_cdnNotPt_V4 : - return Hexagon::POST_SThri_cNotPt; - - case Hexagon::STh_GP_cdnPt_nv_V4: - case Hexagon::STh_GP_cdnPt_V4: - case Hexagon::STh_GP_cPt_nv_V4: - return Hexagon::STh_GP_cPt_V4; - - case Hexagon::STh_GP_cdnNotPt_nv_V4: - case Hexagon::STh_GP_cdnNotPt_V4: - case Hexagon::STh_GP_cNotPt_nv_V4: - return Hexagon::STh_GP_cNotPt_V4; - - // Store new-value halfword - unconditional - - case Hexagon::STrih_nv_V4: - return Hexagon::STrih; - - case Hexagon::STrih_indexed_nv_V4: - return Hexagon::STrih_indexed; - - case Hexagon::STrih_indexed_shl_nv_V4: - return Hexagon::STrih_indexed_shl_V4; - - case Hexagon::STrih_shl_nv_V4: - return Hexagon::STrih_shl_V4; - - case Hexagon::STh_GP_nv_V4: - return Hexagon::STh_GP_V4; - - case Hexagon::POST_SThri_nv_V4: - return Hexagon::POST_SThri; - - // Store word - - case Hexagon::STriw_imm_cdnPt_V4 : - return Hexagon::STriw_imm_cPt_V4; - - case Hexagon::STriw_imm_cdnNotPt_V4 : - return Hexagon::STriw_imm_cNotPt_V4; - - case Hexagon::STriw_cdnPt_nv_V4 : - case Hexagon::STriw_cPt_nv_V4 : - case Hexagon::STriw_cdnPt_V4 : - return Hexagon::STriw_cPt; - - case Hexagon::STriw_cdnNotPt_nv_V4 : - case Hexagon::STriw_cNotPt_nv_V4 : - case Hexagon::STriw_cdnNotPt_V4 : - return Hexagon::STriw_cNotPt; - - case Hexagon::STriw_indexed_cdnPt_nv_V4 : - case Hexagon::STriw_indexed_cPt_nv_V4 : - case Hexagon::STriw_indexed_cdnPt_V4 : - return Hexagon::STriw_indexed_cPt; - - case Hexagon::STriw_indexed_cdnNotPt_nv_V4 : - case Hexagon::STriw_indexed_cNotPt_nv_V4 : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - return Hexagon::STriw_indexed_cNotPt; - - case Hexagon::STriw_indexed_shl_cdnPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - return Hexagon::STriw_indexed_shl_cPt_V4; - - case Hexagon::STriw_indexed_shl_cdnNotPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cNotPt_nv_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - return Hexagon::STriw_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STwri_cdnPt_nv_V4 : - case Hexagon::POST_STwri_cPt_nv_V4 : - case Hexagon::POST_STwri_cdnPt_V4 : - return Hexagon::POST_STwri_cPt; - - case Hexagon::POST_STwri_cdnNotPt_nv_V4 : - case Hexagon::POST_STwri_cNotPt_nv_V4 : - case Hexagon::POST_STwri_cdnNotPt_V4 : - return Hexagon::POST_STwri_cNotPt; - - case Hexagon::STw_GP_cdnPt_nv_V4: - case Hexagon::STw_GP_cdnPt_V4: - case Hexagon::STw_GP_cPt_nv_V4: - return Hexagon::STw_GP_cPt_V4; - - case Hexagon::STw_GP_cdnNotPt_nv_V4: - case Hexagon::STw_GP_cdnNotPt_V4: - case Hexagon::STw_GP_cNotPt_nv_V4: - return Hexagon::STw_GP_cNotPt_V4; - - // Store new-value word - unconditional - - case Hexagon::STriw_nv_V4: - return Hexagon::STriw; - - case Hexagon::STriw_indexed_nv_V4: - return Hexagon::STriw_indexed; - - case Hexagon::STriw_indexed_shl_nv_V4: - return Hexagon::STriw_indexed_shl_V4; - - case Hexagon::STriw_shl_nv_V4: - return Hexagon::STriw_shl_V4; - - case Hexagon::STw_GP_nv_V4: - return Hexagon::STw_GP_V4; - - case Hexagon::POST_STwri_nv_V4: - return Hexagon::POST_STwri; - - // Store doubleword - - case Hexagon::STrid_cdnPt_V4 : - return Hexagon::STrid_cPt; - - case Hexagon::STrid_cdnNotPt_V4 : - return Hexagon::STrid_cNotPt; - - case Hexagon::STrid_indexed_cdnPt_V4 : - return Hexagon::STrid_indexed_cPt; - - case Hexagon::STrid_indexed_cdnNotPt_V4 : - return Hexagon::STrid_indexed_cNotPt; - - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - return Hexagon::STrid_indexed_shl_cPt_V4; - - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - return Hexagon::STrid_indexed_shl_cNotPt_V4; - - case Hexagon::POST_STdri_cdnPt_V4 : - return Hexagon::POST_STdri_cPt; - - case Hexagon::POST_STdri_cdnNotPt_V4 : - return Hexagon::POST_STdri_cNotPt; - - case Hexagon::STd_GP_cdnPt_V4 : - return Hexagon::STd_GP_cPt_V4; - - case Hexagon::STd_GP_cdnNotPt_V4 : - return Hexagon::STd_GP_cNotPt_V4; - - } -} - bool HexagonPacketizerList::DemoteToDotOld(MachineInstr* MI) { const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - int NewOpcode = GetDotOldOp(MI->getOpcode()); + int NewOpcode = QII->GetDotOldOp(MI->getOpcode()); MI->setDesc(QII->get(NewOpcode)); return true; } -// Returns true if an instruction is predicated on p0 and false if it's -// predicated on !p0. +enum PredicateKind { + PK_False, + PK_True, + PK_Unknown +}; -static bool GetPredicateSense(MachineInstr* MI, - const HexagonInstrInfo *QII) { +/// Returns true if an instruction is predicated on p0 and false if it's +/// predicated on !p0. +static PredicateKind getPredicateSense(MachineInstr* MI, + const HexagonInstrInfo *QII) { + if (!QII->isPredicated(MI)) + return PK_Unknown; - switch (MI->getOpcode()) { - default: llvm_unreachable("Unknown predicate sense of the instruction"); - case Hexagon::TFR_cPt: - case Hexagon::TFR_cdnPt: - case Hexagon::TFRI_cPt: - case Hexagon::TFRI_cdnPt: - case Hexagon::STrib_cPt : - case Hexagon::STrib_cdnPt_V4 : - case Hexagon::STrib_indexed_cPt : - case Hexagon::STrib_indexed_cdnPt_V4 : - case Hexagon::STrib_indexed_shl_cPt_V4 : - case Hexagon::STrib_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STbri_cPt : - case Hexagon::POST_STbri_cdnPt_V4 : - case Hexagon::STrih_cPt : - case Hexagon::STrih_cdnPt_V4 : - case Hexagon::STrih_indexed_cPt : - case Hexagon::STrih_indexed_cdnPt_V4 : - case Hexagon::STrih_indexed_shl_cPt_V4 : - case Hexagon::STrih_indexed_shl_cdnPt_V4 : - case Hexagon::POST_SThri_cPt : - case Hexagon::POST_SThri_cdnPt_V4 : - case Hexagon::STriw_cPt : - case Hexagon::STriw_cdnPt_V4 : - case Hexagon::STriw_indexed_cPt : - case Hexagon::STriw_indexed_cdnPt_V4 : - case Hexagon::STriw_indexed_shl_cPt_V4 : - case Hexagon::STriw_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STwri_cPt : - case Hexagon::POST_STwri_cdnPt_V4 : - case Hexagon::STrib_imm_cPt_V4 : - case Hexagon::STrib_imm_cdnPt_V4 : - case Hexagon::STrid_cPt : - case Hexagon::STrid_cdnPt_V4 : - case Hexagon::STrid_indexed_cPt : - case Hexagon::STrid_indexed_cdnPt_V4 : - case Hexagon::STrid_indexed_shl_cPt_V4 : - case Hexagon::STrid_indexed_shl_cdnPt_V4 : - case Hexagon::POST_STdri_cPt : - case Hexagon::POST_STdri_cdnPt_V4 : - case Hexagon::STrih_imm_cPt_V4 : - case Hexagon::STrih_imm_cdnPt_V4 : - case Hexagon::STriw_imm_cPt_V4 : - case Hexagon::STriw_imm_cdnPt_V4 : - case Hexagon::JMP_tnew_t : - case Hexagon::LDrid_cPt : - case Hexagon::LDrid_cdnPt : - case Hexagon::LDrid_indexed_cPt : - case Hexagon::LDrid_indexed_cdnPt : - case Hexagon::POST_LDrid_cPt : - case Hexagon::POST_LDrid_cdnPt_V4 : - case Hexagon::LDriw_cPt : - case Hexagon::LDriw_cdnPt : - case Hexagon::LDriw_indexed_cPt : - case Hexagon::LDriw_indexed_cdnPt : - case Hexagon::POST_LDriw_cPt : - case Hexagon::POST_LDriw_cdnPt_V4 : - case Hexagon::LDrih_cPt : - case Hexagon::LDrih_cdnPt : - case Hexagon::LDrih_indexed_cPt : - case Hexagon::LDrih_indexed_cdnPt : - case Hexagon::POST_LDrih_cPt : - case Hexagon::POST_LDrih_cdnPt_V4 : - case Hexagon::LDrib_cPt : - case Hexagon::LDrib_cdnPt : - case Hexagon::LDrib_indexed_cPt : - case Hexagon::LDrib_indexed_cdnPt : - case Hexagon::POST_LDrib_cPt : - case Hexagon::POST_LDrib_cdnPt_V4 : - case Hexagon::LDriuh_cPt : - case Hexagon::LDriuh_cdnPt : - case Hexagon::LDriuh_indexed_cPt : - case Hexagon::LDriuh_indexed_cdnPt : - case Hexagon::POST_LDriuh_cPt : - case Hexagon::POST_LDriuh_cdnPt_V4 : - case Hexagon::LDriub_cPt : - case Hexagon::LDriub_cdnPt : - case Hexagon::LDriub_indexed_cPt : - case Hexagon::LDriub_indexed_cdnPt : - case Hexagon::POST_LDriub_cPt : - case Hexagon::POST_LDriub_cdnPt_V4 : - case Hexagon::LDrid_indexed_shl_cPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnPt_V4 : - case Hexagon::LDrib_indexed_shl_cPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnPt_V4 : - case Hexagon::LDriub_indexed_shl_cPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnPt_V4 : - case Hexagon::LDrih_indexed_shl_cPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnPt_V4 : - case Hexagon::LDriuh_indexed_shl_cPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnPt_V4 : - case Hexagon::LDriw_indexed_shl_cPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnPt_V4 : - case Hexagon::ADD_ri_cPt : - case Hexagon::ADD_ri_cdnPt : - case Hexagon::ADD_rr_cPt : - case Hexagon::ADD_rr_cdnPt : - case Hexagon::XOR_rr_cPt : - case Hexagon::XOR_rr_cdnPt : - case Hexagon::AND_rr_cPt : - case Hexagon::AND_rr_cdnPt : - case Hexagon::OR_rr_cPt : - case Hexagon::OR_rr_cdnPt : - case Hexagon::SUB_rr_cPt : - case Hexagon::SUB_rr_cdnPt : - case Hexagon::COMBINE_rr_cPt : - case Hexagon::COMBINE_rr_cdnPt : - case Hexagon::ASLH_cPt_V4 : - case Hexagon::ASLH_cdnPt_V4 : - case Hexagon::ASRH_cPt_V4 : - case Hexagon::ASRH_cdnPt_V4 : - case Hexagon::SXTB_cPt_V4 : - case Hexagon::SXTB_cdnPt_V4 : - case Hexagon::SXTH_cPt_V4 : - case Hexagon::SXTH_cdnPt_V4 : - case Hexagon::ZXTB_cPt_V4 : - case Hexagon::ZXTB_cdnPt_V4 : - case Hexagon::ZXTH_cPt_V4 : - case Hexagon::ZXTH_cdnPt_V4 : - case Hexagon::LDd_GP_cPt_V4 : - case Hexagon::LDb_GP_cPt_V4 : - case Hexagon::LDub_GP_cPt_V4 : - case Hexagon::LDh_GP_cPt_V4 : - case Hexagon::LDuh_GP_cPt_V4 : - case Hexagon::LDw_GP_cPt_V4 : - case Hexagon::STd_GP_cPt_V4 : - case Hexagon::STb_GP_cPt_V4 : - case Hexagon::STh_GP_cPt_V4 : - case Hexagon::STw_GP_cPt_V4 : - case Hexagon::LDd_GP_cdnPt_V4 : - case Hexagon::LDb_GP_cdnPt_V4 : - case Hexagon::LDub_GP_cdnPt_V4 : - case Hexagon::LDh_GP_cdnPt_V4 : - case Hexagon::LDuh_GP_cdnPt_V4 : - case Hexagon::LDw_GP_cdnPt_V4 : - case Hexagon::STd_GP_cdnPt_V4 : - case Hexagon::STb_GP_cdnPt_V4 : - case Hexagon::STh_GP_cdnPt_V4 : - case Hexagon::STw_GP_cdnPt_V4 : - return true; + if (QII->isPredicatedTrue(MI)) + return PK_True; - case Hexagon::TFR_cNotPt: - case Hexagon::TFR_cdnNotPt: - case Hexagon::TFRI_cNotPt: - case Hexagon::TFRI_cdnNotPt: - case Hexagon::STrib_cNotPt : - case Hexagon::STrib_cdnNotPt_V4 : - case Hexagon::STrib_indexed_cNotPt : - case Hexagon::STrib_indexed_cdnNotPt_V4 : - case Hexagon::STrib_indexed_shl_cNotPt_V4 : - case Hexagon::STrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STbri_cNotPt : - case Hexagon::POST_STbri_cdnNotPt_V4 : - case Hexagon::STrih_cNotPt : - case Hexagon::STrih_cdnNotPt_V4 : - case Hexagon::STrih_indexed_cNotPt : - case Hexagon::STrih_indexed_cdnNotPt_V4 : - case Hexagon::STrih_indexed_shl_cNotPt_V4 : - case Hexagon::STrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_SThri_cNotPt : - case Hexagon::POST_SThri_cdnNotPt_V4 : - case Hexagon::STriw_cNotPt : - case Hexagon::STriw_cdnNotPt_V4 : - case Hexagon::STriw_indexed_cNotPt : - case Hexagon::STriw_indexed_cdnNotPt_V4 : - case Hexagon::STriw_indexed_shl_cNotPt_V4 : - case Hexagon::STriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STwri_cNotPt : - case Hexagon::POST_STwri_cdnNotPt_V4 : - case Hexagon::STrib_imm_cNotPt_V4 : - case Hexagon::STrib_imm_cdnNotPt_V4 : - case Hexagon::STrid_cNotPt : - case Hexagon::STrid_cdnNotPt_V4 : - case Hexagon::STrid_indexed_cdnNotPt_V4 : - case Hexagon::STrid_indexed_cNotPt : - case Hexagon::STrid_indexed_shl_cNotPt_V4 : - case Hexagon::STrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::POST_STdri_cNotPt : - case Hexagon::POST_STdri_cdnNotPt_V4 : - case Hexagon::STrih_imm_cNotPt_V4 : - case Hexagon::STrih_imm_cdnNotPt_V4 : - case Hexagon::STriw_imm_cNotPt_V4 : - case Hexagon::STriw_imm_cdnNotPt_V4 : - case Hexagon::JMP_fnew_t : - case Hexagon::LDrid_cNotPt : - case Hexagon::LDrid_cdnNotPt : - case Hexagon::LDrid_indexed_cNotPt : - case Hexagon::LDrid_indexed_cdnNotPt : - case Hexagon::POST_LDrid_cNotPt : - case Hexagon::POST_LDrid_cdnNotPt_V4 : - case Hexagon::LDriw_cNotPt : - case Hexagon::LDriw_cdnNotPt : - case Hexagon::LDriw_indexed_cNotPt : - case Hexagon::LDriw_indexed_cdnNotPt : - case Hexagon::POST_LDriw_cNotPt : - case Hexagon::POST_LDriw_cdnNotPt_V4 : - case Hexagon::LDrih_cNotPt : - case Hexagon::LDrih_cdnNotPt : - case Hexagon::LDrih_indexed_cNotPt : - case Hexagon::LDrih_indexed_cdnNotPt : - case Hexagon::POST_LDrih_cNotPt : - case Hexagon::POST_LDrih_cdnNotPt_V4 : - case Hexagon::LDrib_cNotPt : - case Hexagon::LDrib_cdnNotPt : - case Hexagon::LDrib_indexed_cNotPt : - case Hexagon::LDrib_indexed_cdnNotPt : - case Hexagon::POST_LDrib_cNotPt : - case Hexagon::POST_LDrib_cdnNotPt_V4 : - case Hexagon::LDriuh_cNotPt : - case Hexagon::LDriuh_cdnNotPt : - case Hexagon::LDriuh_indexed_cNotPt : - case Hexagon::LDriuh_indexed_cdnNotPt : - case Hexagon::POST_LDriuh_cNotPt : - case Hexagon::POST_LDriuh_cdnNotPt_V4 : - case Hexagon::LDriub_cNotPt : - case Hexagon::LDriub_cdnNotPt : - case Hexagon::LDriub_indexed_cNotPt : - case Hexagon::LDriub_indexed_cdnNotPt : - case Hexagon::POST_LDriub_cNotPt : - case Hexagon::POST_LDriub_cdnNotPt_V4 : - case Hexagon::LDrid_indexed_shl_cNotPt_V4 : - case Hexagon::LDrid_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cNotPt_V4 : - case Hexagon::LDrib_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cNotPt_V4 : - case Hexagon::LDriub_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cNotPt_V4 : - case Hexagon::LDrih_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cNotPt_V4 : - case Hexagon::LDriuh_indexed_shl_cdnNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cNotPt_V4 : - case Hexagon::LDriw_indexed_shl_cdnNotPt_V4 : - case Hexagon::ADD_ri_cNotPt : - case Hexagon::ADD_ri_cdnNotPt : - case Hexagon::ADD_rr_cNotPt : - case Hexagon::ADD_rr_cdnNotPt : - case Hexagon::XOR_rr_cNotPt : - case Hexagon::XOR_rr_cdnNotPt : - case Hexagon::AND_rr_cNotPt : - case Hexagon::AND_rr_cdnNotPt : - case Hexagon::OR_rr_cNotPt : - case Hexagon::OR_rr_cdnNotPt : - case Hexagon::SUB_rr_cNotPt : - case Hexagon::SUB_rr_cdnNotPt : - case Hexagon::COMBINE_rr_cNotPt : - case Hexagon::COMBINE_rr_cdnNotPt : - case Hexagon::ASLH_cNotPt_V4 : - case Hexagon::ASLH_cdnNotPt_V4 : - case Hexagon::ASRH_cNotPt_V4 : - case Hexagon::ASRH_cdnNotPt_V4 : - case Hexagon::SXTB_cNotPt_V4 : - case Hexagon::SXTB_cdnNotPt_V4 : - case Hexagon::SXTH_cNotPt_V4 : - case Hexagon::SXTH_cdnNotPt_V4 : - case Hexagon::ZXTB_cNotPt_V4 : - case Hexagon::ZXTB_cdnNotPt_V4 : - case Hexagon::ZXTH_cNotPt_V4 : - case Hexagon::ZXTH_cdnNotPt_V4 : - - case Hexagon::LDd_GP_cNotPt_V4 : - case Hexagon::LDb_GP_cNotPt_V4 : - case Hexagon::LDub_GP_cNotPt_V4 : - case Hexagon::LDh_GP_cNotPt_V4 : - case Hexagon::LDuh_GP_cNotPt_V4 : - case Hexagon::LDw_GP_cNotPt_V4 : - case Hexagon::STd_GP_cNotPt_V4 : - case Hexagon::STb_GP_cNotPt_V4 : - case Hexagon::STh_GP_cNotPt_V4 : - case Hexagon::STw_GP_cNotPt_V4 : - case Hexagon::LDd_GP_cdnNotPt_V4 : - case Hexagon::LDb_GP_cdnNotPt_V4 : - case Hexagon::LDub_GP_cdnNotPt_V4 : - case Hexagon::LDh_GP_cdnNotPt_V4 : - case Hexagon::LDuh_GP_cdnNotPt_V4 : - case Hexagon::LDw_GP_cdnNotPt_V4 : - case Hexagon::STd_GP_cdnNotPt_V4 : - case Hexagon::STb_GP_cdnNotPt_V4 : - case Hexagon::STh_GP_cdnNotPt_V4 : - case Hexagon::STw_GP_cdnNotPt_V4 : - return false; - } - // return *some value* to avoid compiler warning - return false; + return PK_False; } static MachineOperand& GetPostIncrementOperand(MachineInstr *MI, @@ -2252,10 +537,10 @@ static MachineOperand& GetStoreValueOperand(MachineInstr *MI) { // Arch Spec: 3.4.4.2 bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, MachineInstr *PacketMI, unsigned DepReg, - std::map <MachineInstr*, SUnit*> MIToSUnit) -{ - // Make sure we are looking at the store - if (!IsNewifyStore(MI)) + std::map <MachineInstr*, SUnit*> MIToSUnit) { + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; + // Make sure we are looking at the store, that can be promoted. + if (!QII->mayBeNewStore(MI)) return false; // Make sure there is dependency and can be new value'ed @@ -2263,12 +548,11 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, GetStoreValueOperand(MI).getReg() != DepReg) return false; - const HexagonRegisterInfo* QRI = + const HexagonRegisterInfo* QRI = (const HexagonRegisterInfo *) TM.getRegisterInfo(); const MCInstrDesc& MCID = PacketMI->getDesc(); // first operand is always the result - const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; const TargetRegisterClass* PacketRC = QII->getRegClass(MCID, 0, QRI, MF); // if there is already an store in the packet, no can do new value store @@ -2311,7 +595,7 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, } // If the source that feeds the store is predicated, new value store must - // also be also predicated. + // also be predicated. if (QII->isPredicated(PacketMI)) { if (!QII->isPredicated(MI)) return false; @@ -2357,7 +641,7 @@ bool HexagonPacketizerList::CanPromoteToNewValueStore( MachineInstr *MI, if (( predRegNumDst != predRegNumSrc) || QII->isDotNewInst(PacketMI) != QII->isDotNewInst(MI) || - GetPredicateSense(MI, QII) != GetPredicateSense(PacketMI, QII)) { + getPredicateSense(MI, QII) != getPredicateSense(PacketMI, QII)) { return false; } } @@ -2438,10 +722,11 @@ bool HexagonPacketizerList::CanPromoteToNewValue( MachineInstr *MI, MachineBasicBlock::iterator &MII) { + const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; const HexagonRegisterInfo* QRI = (const HexagonRegisterInfo *) TM.getRegisterInfo(); if (!QRI->Subtarget.hasV4TOps() || - !IsNewifyStore(MI)) + !QII->mayBeNewStore(MI)) return false; MachineInstr *PacketMI = PacketSU->getInstr(); @@ -2468,7 +753,7 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, { const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; // Already a dot new instruction. - if (QII->isDotNewInst(MI) && !IsNewifyStore(MI)) + if (QII->isDotNewInst(MI) && !QII->mayBeNewStore(MI)) return false; if (!isNewifiable(MI)) @@ -2478,12 +763,12 @@ bool HexagonPacketizerList::CanPromoteToDotNew( MachineInstr *MI, if (RC == &Hexagon::PredRegsRegClass && isCondInst(MI)) return true; else if (RC != &Hexagon::PredRegsRegClass && - !IsNewifyStore(MI)) // MI is not a new-value store + !QII->mayBeNewStore(MI)) // MI is not a new-value store return false; else { // Create a dot new machine instruction to see if resources can be // allocated. If not, bail out now. - int NewOpcode = GetDotNewOp(MI->getOpcode()); + int NewOpcode = QII->GetDotNewOp(MI); const MCInstrDesc &desc = QII->get(NewOpcode); DebugLoc dl; MachineInstr *NewMI = @@ -2552,16 +837,39 @@ bool HexagonPacketizerList::RestrictingDepExistInPacket (MachineInstr* MI, } +/// Gets the predicate register of a predicated instruction. +static unsigned getPredicatedRegister(MachineInstr *MI, + const HexagonInstrInfo *QII) { + /// We use the following rule: The first predicate register that is a use is + /// the predicate register of a predicated instruction. + + assert(QII->isPredicated(MI) && "Must be predicated instruction"); + + for (MachineInstr::mop_iterator OI = MI->operands_begin(), + OE = MI->operands_end(); OI != OE; ++OI) { + MachineOperand &Op = *OI; + if (Op.isReg() && Op.getReg() && Op.isUse() && + Hexagon::PredRegsRegClass.contains(Op.getReg())) + return Op.getReg(); + } + + llvm_unreachable("Unknown instruction operand layout"); + + return 0; +} + // Given two predicated instructions, this function detects whether // the predicates are complements bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, MachineInstr* MI2, std::map <MachineInstr*, SUnit*> MIToSUnit) { const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII; - // Currently can only reason about conditional transfers - if (!QII->isConditionalTransfer(MI1) || !QII->isConditionalTransfer(MI2)) { + + // If we don't know the predicate sense of the instructions bail out early, we + // need it later. + if (getPredicateSense(MI1, QII) == PK_Unknown || + getPredicateSense(MI2, QII) == PK_Unknown) return false; - } // Scheduling unit for candidate SUnit* SU = MIToSUnit[MI1]; @@ -2600,9 +908,9 @@ bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, // there already exist anti dep on the same pred in // the packet. if (PacketSU->Succs[i].getSUnit() == SU && + PacketSU->Succs[i].getKind() == SDep::Data && Hexagon::PredRegsRegClass.contains( PacketSU->Succs[i].getReg()) && - PacketSU->Succs[i].getKind() == SDep::Data && // Here I know that *VIN is predicate setting instruction // with true data dep to candidate on the register // we care about - c) in the above example. @@ -2623,8 +931,12 @@ bool HexagonPacketizerList::ArePredicatesComplements (MachineInstr* MI1, // that the predicate sense is different // We also need to differentiate .old vs. .new: // !p0 is not complimentary to p0.new - return ((MI1->getOperand(1).getReg() == MI2->getOperand(1).getReg()) && - (GetPredicateSense(MI1, QII) != GetPredicateSense(MI2, QII)) && + unsigned PReg1 = getPredicatedRegister(MI1, QII); + unsigned PReg2 = getPredicatedRegister(MI2, QII); + return ((PReg1 == PReg2) && + Hexagon::PredRegsRegClass.contains(PReg1) && + Hexagon::PredRegsRegClass.contains(PReg2) && + (getPredicateSense(MI1, QII) != getPredicateSense(MI2, QII)) && (QII->isDotNewInst(MI1) == QII->isDotNewInst(MI2))); } @@ -2722,24 +1034,21 @@ bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) { } // A LoopN instruction cannot appear in the same packet as a jump or call. - if (IsLoopN(I) && ( IsDirectJump(J) - || MCIDJ.isCall() - || QII->isDeallocRet(J))) { + if (IsLoopN(I) && + (IsDirectJump(J) || MCIDJ.isCall() || QII->isDeallocRet(J))) { Dependence = true; return false; } - if (IsLoopN(J) && ( IsDirectJump(I) - || MCIDI.isCall() - || QII->isDeallocRet(I))) { + if (IsLoopN(J) && + (IsDirectJump(I) || MCIDI.isCall() || QII->isDeallocRet(I))) { Dependence = true; return false; } // dealloc_return cannot appear in the same packet as a conditional or // unconditional jump. - if (QII->isDeallocRet(I) && ( MCIDJ.isBranch() - || MCIDJ.isCall() - || MCIDJ.isBarrier())) { + if (QII->isDeallocRet(I) && + (MCIDJ.isBranch() || MCIDJ.isCall() || MCIDJ.isBarrier())) { Dependence = true; return false; } @@ -2764,7 +1073,7 @@ bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) { } //if dealloc_return - if (MCIDJ.mayStore() && QII->isDeallocRet(I)){ + if (MCIDJ.mayStore() && QII->isDeallocRet(I)) { Dependence = true; return false; } diff --git a/contrib/llvm/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp b/contrib/llvm/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp index 36da6df..7c41507 100644 --- a/contrib/llvm/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp +++ b/contrib/llvm/lib/Target/Hexagon/InstPrinter/HexagonInstPrinter.cpp @@ -21,7 +21,6 @@ #include "llvm/MC/MCAsmInfo.h" #include "llvm/MC/MCExpr.h" #include "llvm/Support/raw_ostream.h" -#include <cstdio> using namespace llvm; @@ -179,7 +178,7 @@ void HexagonInstPrinter::printBranchOperand(const MCInst *MI, unsigned OpNo, raw_ostream &O) const { // Branches can take an immediate operand. This is used by the branch // selection pass to print $+8, an eight byte displacement from the PC. - assert("Unknown branch operand."); + llvm_unreachable("Unknown branch operand."); } void HexagonInstPrinter::printCallOperand(const MCInst *MI, unsigned OpNo, @@ -196,15 +195,9 @@ void HexagonInstPrinter::printPredicateOperand(const MCInst *MI, unsigned OpNo, void HexagonInstPrinter::printSymbol(const MCInst *MI, unsigned OpNo, raw_ostream &O, bool hi) const { - const MCOperand& MO = MI->getOperand(OpNo); + assert(MI->getOperand(OpNo).isImm() && "Unknown symbol operand"); - O << '#' << (hi? "HI": "LO") << '('; - if (MO.isImm()) { - O << '#'; - printOperand(MI, OpNo, O); - } else { - assert("Unknown symbol operand"); - printOperand(MI, OpNo, O); - } + O << '#' << (hi ? "HI" : "LO") << "(#"; + printOperand(MI, OpNo, O); O << ')'; } diff --git a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h index d4a93b5..8519cf3 100644 --- a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonBaseInfo.h @@ -65,14 +65,15 @@ namespace HexagonII { AbsoluteSet = 2, // Absolute set addressing mode BaseImmOffset = 3, // Indirect with offset BaseLongOffset = 4, // Indirect with long offset - BaseRegOffset = 5 // Indirect with register offset + BaseRegOffset = 5, // Indirect with register offset + PostInc = 6 // Post increment addressing mode }; enum MemAccessSize { NoMemAccess = 0, // Not a memory acces instruction. ByteAccess = 1, // Byte access instruction (memb). HalfWordAccess = 2, // Half word access instruction (memh). - WordAccess = 3, // Word access instrution (memw). + WordAccess = 3, // Word access instruction (memw). DoubleWordAccess = 4 // Double word access instruction (memd) }; diff --git a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp index 3deb8d1..3f9415b 100644 --- a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp +++ b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.cpp @@ -15,7 +15,10 @@ using namespace llvm; -HexagonMCAsmInfo::HexagonMCAsmInfo(const Target &T, StringRef TT) { +// Pin the vtable to this file. +void HexagonMCAsmInfo::anchor() {} + +HexagonMCAsmInfo::HexagonMCAsmInfo(StringRef TT) { Data16bitsDirective = "\t.half\t"; Data32bitsDirective = "\t.word\t"; Data64bitsDirective = 0; // .xword is only supported by V9. @@ -29,7 +32,6 @@ HexagonMCAsmInfo::HexagonMCAsmInfo(const Target &T, StringRef TT) { InlineAsmEnd = "# InlineAsm End"; ZeroDirective = "\t.space\t"; AscizDirective = "\t.string\t"; - WeakRefDirective = "\t.weak\t"; SupportsDebugInformation = true; UsesELFSectionDirectiveForBSS = true; diff --git a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h index d336cd5..bd8cb76 100644 --- a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h +++ b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCAsmInfo.h @@ -15,14 +15,13 @@ #define HexagonMCASMINFO_H #include "llvm/ADT/StringRef.h" -#include "llvm/MC/MCAsmInfo.h" +#include "llvm/MC/MCAsmInfoELF.h" namespace llvm { - class Target; - - class HexagonMCAsmInfo : public MCAsmInfo { + class HexagonMCAsmInfo : public MCAsmInfoELF { + virtual void anchor(); public: - explicit HexagonMCAsmInfo(const Target &T, StringRef TT); + explicit HexagonMCAsmInfo(StringRef TT); }; } // namespace llvm diff --git a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp index 6b1d2d1..2f93a52 100644 --- a/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp +++ b/contrib/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp @@ -54,13 +54,14 @@ static MCSubtargetInfo *createHexagonMCSubtargetInfo(StringRef TT, return X; } -static MCAsmInfo *createHexagonMCAsmInfo(const Target &T, StringRef TT) { - MCAsmInfo *MAI = new HexagonMCAsmInfo(T, TT); +static MCAsmInfo *createHexagonMCAsmInfo(const MCRegisterInfo &MRI, + StringRef TT) { + MCAsmInfo *MAI = new HexagonMCAsmInfo(TT); // VirtualFP = (R30 + #0). - MachineLocation Dst(MachineLocation::VirtualFP); - MachineLocation Src(Hexagon::R30, 0); - MAI->addInitialFrameState(0, Dst, Src); + MCCFIInstruction Inst = MCCFIInstruction::createDefCfa( + 0, Hexagon::R30, 0); + MAI->addInitialFrameState(Inst); return MAI; } |
