diff options
| author | Shih-wei Liao <sliao@google.com> | 2010-04-07 12:21:42 -0700 |
|---|---|---|
| committer | Shih-wei Liao <sliao@google.com> | 2010-04-07 12:21:42 -0700 |
| commit | e4454320b3cfffe926a487c33fbeb454366de2f8 (patch) | |
| tree | 133c05da684edf4a3b2529bcacfa996298c455f6 /lib/CodeGen | |
| parent | 20570085304f0a4ab4f112a01d77958bbd2827a1 (diff) | |
| download | external_llvm-e4454320b3cfffe926a487c33fbeb454366de2f8.zip external_llvm-e4454320b3cfffe926a487c33fbeb454366de2f8.tar.gz external_llvm-e4454320b3cfffe926a487c33fbeb454366de2f8.tar.bz2 | |
libbcc
Change-Id: Ieaa3ebd5a38f370752495549f8870b534eeedfc5
Diffstat (limited to 'lib/CodeGen')
68 files changed, 4607 insertions, 1090 deletions
diff --git a/lib/CodeGen/Android.mk b/lib/CodeGen/Android.mk new file mode 100644 index 0000000..9fa2ecd --- /dev/null +++ b/lib/CodeGen/Android.mk @@ -0,0 +1,99 @@ +LOCAL_PATH:= $(call my-dir) + +codegen_SRC_FILES := \ + AggressiveAntiDepBreaker.cpp \ + BranchFolding.cpp \ + CalcSpillWeights.cpp \ + CodePlacementOpt.cpp \ + CriticalAntiDepBreaker.cpp \ + DeadMachineInstructionElim.cpp \ + DwarfEHPrepare.cpp \ + ELFCodeEmitter.cpp \ + ELFWriter.cpp \ + ExactHazardRecognizer.cpp \ + GCMetadata.cpp \ + GCMetadataPrinter.cpp \ + GCStrategy.cpp \ + IfConversion.cpp \ + IntrinsicLowering.cpp \ + LLVMTargetMachine.cpp \ + LatencyPriorityQueue.cpp \ + LiveInterval.cpp \ + LiveIntervalAnalysis.cpp \ + LiveStackAnalysis.cpp \ + LiveVariables.cpp \ + LowerSubregs.cpp \ + MachineBasicBlock.cpp \ + MachineCSE.cpp \ + MachineDominators.cpp \ + MachineFunction.cpp \ + MachineFunctionAnalysis.cpp \ + MachineFunctionPass.cpp \ + MachineInstr.cpp \ + MachineLICM.cpp \ + MachineLoopInfo.cpp \ + MachineModuleInfo.cpp \ + MachineModuleInfoImpls.cpp \ + MachinePassRegistry.cpp \ + MachineRegisterInfo.cpp \ + MachineSSAUpdater.cpp \ + MachineSink.cpp \ + MachineVerifier.cpp \ + ObjectCodeEmitter.cpp \ + OcamlGC.cpp \ + OptimizeExts.cpp \ + OptimizePHIs.cpp \ + PHIElimination.cpp \ + Passes.cpp \ + PostRASchedulerList.cpp \ + PreAllocSplitting.cpp \ + ProcessImplicitDefs.cpp \ + PrologEpilogInserter.cpp \ + PseudoSourceValue.cpp \ + RegAllocLinearScan.cpp \ + RegAllocLocal.cpp \ + RegAllocPBQP.cpp \ + RegisterCoalescer.cpp \ + RegisterScavenging.cpp \ + ScheduleDAG.cpp \ + ScheduleDAGEmit.cpp \ + ScheduleDAGInstrs.cpp \ + ScheduleDAGPrinter.cpp \ + ShadowStackGC.cpp \ + ShrinkWrapping.cpp \ + SimpleRegisterCoalescing.cpp \ + SjLjEHPrepare.cpp \ + SlotIndexes.cpp \ + Spiller.cpp \ + StackProtector.cpp \ + StackSlotColoring.cpp \ + StrongPHIElimination.cpp \ + TailDuplication.cpp \ + TargetInstrInfoImpl.cpp \ + TargetLoweringObjectFileImpl.cpp \ + TwoAddressInstructionPass.cpp \ + UnreachableBlockElim.cpp \ + VirtRegMap.cpp \ + VirtRegRewriter.cpp + +# For the host +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_SRC_FILES) +LOCAL_MODULE:= libLLVMCodeGen + +include $(LLVM_HOST_BUILD_MK) +include $(LLVM_GEN_INTRINSICS_MK) +include $(BUILD_HOST_STATIC_LIBRARY) + +# For the device +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_SRC_FILES) +LOCAL_MODULE:= libLLVMCodeGen + +include $(LLVM_DEVICE_BUILD_MK) +include $(LLVM_GEN_INTRINSICS_MK) +include $(BUILD_STATIC_LIBRARY) diff --git a/lib/CodeGen/AsmPrinter/Android.mk b/lib/CodeGen/AsmPrinter/Android.mk new file mode 100644 index 0000000..62601f0 --- /dev/null +++ b/lib/CodeGen/AsmPrinter/Android.mk @@ -0,0 +1,31 @@ +LOCAL_PATH := $(call my-dir) + +codegen_asmprinter_SRC_FILES := \ + AsmPrinter.cpp \ + DIE.cpp \ + DwarfDebug.cpp \ + DwarfException.cpp \ + DwarfLabel.cpp \ + DwarfPrinter.cpp \ + DwarfWriter.cpp \ + OcamlGCPrinter.cpp + +# For the host +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_asmprinter_SRC_FILES) +LOCAL_MODULE:= libLLVMAsmPrinter + +include $(LLVM_HOST_BUILD_MK) +include $(BUILD_HOST_STATIC_LIBRARY) + +# For the device +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_asmprinter_SRC_FILES) +LOCAL_MODULE:= libLLVMAsmPrinter + +include $(LLVM_DEVICE_BUILD_MK) +include $(BUILD_STATIC_LIBRARY) diff --git a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp index fc08384..bbeb026 100644 --- a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp +++ b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp @@ -50,6 +50,7 @@ #include "llvm/Support/Format.h" #include "llvm/Support/FormattedStream.h" #include <cerrno> +#include <ctype.h> using namespace llvm; STATISTIC(EmittedInsts, "Number of machine instrs printed"); @@ -917,11 +918,10 @@ void AsmPrinter::EmitAlignment(unsigned NumBits, const GlobalValue *GV, if (NumBits == 0) return; // No need to emit alignment. - unsigned FillValue = 0; if (getCurrentSection()->getKind().isText()) - FillValue = MAI->getTextAlignFillValue(); - - OutStreamer.EmitValueToAlignment(1 << NumBits, FillValue, 1, 0); + OutStreamer.EmitCodeAlignment(1 << NumBits); + else + OutStreamer.EmitValueToAlignment(1 << NumBits, 0, 1, 0); } /// LowerConstant - Lower the specified LLVM Constant to an MCExpr. @@ -1717,7 +1717,7 @@ void AsmPrinter::EmitBasicBlockStart(const MachineBasicBlock *MBB) const { } // Print the main label for the block. - if (MBB->pred_empty() || MBB->isOnlyReachableByFallthrough()) { + if (MBB->pred_empty() || isBlockOnlyReachableByFallthrough(MBB)) { if (VerboseAsm) { // NOTE: Want this comment at start of line. O << MAI->getCommentString() << " BB#" << MBB->getNumber() << ':'; @@ -1764,6 +1764,39 @@ void AsmPrinter::printOffset(int64_t Offset) const { O << Offset; } +/// isBlockOnlyReachableByFallthough - Return true if the basic block has +/// exactly one predecessor and the control transfer mechanism between +/// the predecessor and this block is a fall-through. +bool AsmPrinter::isBlockOnlyReachableByFallthrough(const MachineBasicBlock *MBB) + const { + // If this is a landing pad, it isn't a fall through. If it has no preds, + // then nothing falls through to it. + if (MBB->isLandingPad() || MBB->pred_empty()) + return false; + + // If there isn't exactly one predecessor, it can't be a fall through. + MachineBasicBlock::const_pred_iterator PI = MBB->pred_begin(), PI2 = PI; + ++PI2; + if (PI2 != MBB->pred_end()) + return false; + + // The predecessor has to be immediately before this block. + const MachineBasicBlock *Pred = *PI; + + if (!Pred->isLayoutSuccessor(MBB)) + return false; + + // If the block is completely empty, then it definitely does fall through. + if (Pred->empty()) + return true; + + // Otherwise, check the last instruction. + const MachineInstr &LastInst = Pred->back(); + return !LastInst.getDesc().isBarrier(); +} + + + GCMetadataPrinter *AsmPrinter::GetOrCreateGCPrinter(GCStrategy *S) { if (!S->usesMetadata()) return 0; diff --git a/lib/CodeGen/AsmPrinter/DIE.cpp b/lib/CodeGen/AsmPrinter/DIE.cpp index 349e0ac..63360c0 100644 --- a/lib/CodeGen/AsmPrinter/DIE.cpp +++ b/lib/CodeGen/AsmPrinter/DIE.cpp @@ -313,6 +313,7 @@ void DIESectionOffset::EmitValue(DwarfPrinter *D, unsigned Form) const { D->EmitSectionOffset(Label.getTag(), Section.getTag(), Label.getNumber(), Section.getNumber(), IsSmall, IsEH, UseSet); + D->getAsm()->O << '\n'; // FIXME: Necesssary? } /// SizeOf - Determine size of delta value in bytes. diff --git a/lib/CodeGen/AsmPrinter/DwarfDebug.cpp b/lib/CodeGen/AsmPrinter/DwarfDebug.cpp index 5093dd9..5ad1e5e 100644 --- a/lib/CodeGen/AsmPrinter/DwarfDebug.cpp +++ b/lib/CodeGen/AsmPrinter/DwarfDebug.cpp @@ -238,7 +238,18 @@ public: LIndex = DSI; } } - setLastInsn(LastInsn); + + unsigned CurrentLastInsnIndex = 0; + if (const MachineInstr *CL = getLastInsn()) + CurrentLastInsnIndex = MIIndexMap[CL]; + unsigned FIndex = MIIndexMap[getFirstInsn()]; + + // Set LastInsn as the last instruction for this scope only if + // it follows + // 1) this scope's first instruction and + // 2) current last instruction for this scope, if any. + if (LIndex >= CurrentLastInsnIndex && LIndex >= FIndex) + setLastInsn(LastInsn); } #ifndef NDEBUG @@ -1166,7 +1177,9 @@ DIE *DwarfDebug::createSubprogramDIE(const DISubprogram &SP, bool MakeDecl) { return SPDie; SPDie = new DIE(dwarf::DW_TAG_subprogram); - addString(SPDie, dwarf::DW_AT_name, dwarf::DW_FORM_string, SP.getName()); + // Constructors and operators for anonymous aggregates do not have names. + if (!SP.getName().empty()) + addString(SPDie, dwarf::DW_AT_name, dwarf::DW_FORM_string, SP.getName()); StringRef LinkageName = SP.getLinkageName(); if (!LinkageName.empty()) diff --git a/lib/CodeGen/AsmPrinter/DwarfException.cpp b/lib/CodeGen/AsmPrinter/DwarfException.cpp index b6801dc..2b08ba4 100644 --- a/lib/CodeGen/AsmPrinter/DwarfException.cpp +++ b/lib/CodeGen/AsmPrinter/DwarfException.cpp @@ -50,26 +50,6 @@ DwarfException::~DwarfException() { delete ExceptionTimer; } -/// SizeOfEncodedValue - Return the size of the encoding in bytes. -unsigned DwarfException::SizeOfEncodedValue(unsigned Encoding) { - if (Encoding == dwarf::DW_EH_PE_omit) - return 0; - - switch (Encoding & 0x07) { - case dwarf::DW_EH_PE_absptr: - return TD->getPointerSize(); - case dwarf::DW_EH_PE_udata2: - return 2; - case dwarf::DW_EH_PE_udata4: - return 4; - case dwarf::DW_EH_PE_udata8: - return 8; - } - - assert(0 && "Invalid encoded value."); - return 0; -} - /// CreateLabelDiff - Emit a label and subtract it from the expression we /// already have. This is equivalent to emitting "foo - .", but we have to emit /// the label for "." directly. @@ -100,7 +80,7 @@ void DwarfException::EmitCIE(const Function *PersonalityFn, unsigned Index) { TD->getPointerSize() : -TD->getPointerSize(); const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); - + // Begin eh frame section. Asm->OutStreamer.SwitchSection(TLOF.getEHFrameSection()); @@ -128,30 +108,16 @@ void DwarfException::EmitCIE(const Function *PersonalityFn, unsigned Index) { // The personality presence indicates that language specific information will // show up in the eh frame. Find out how we are supposed to lower the // personality function reference: - const MCExpr *PersonalityRef = 0; - bool IsPersonalityIndirect = false, IsPersonalityPCRel = false; - if (PersonalityFn) { - // FIXME: HANDLE STATIC CODEGEN MODEL HERE. - - // In non-static mode, ask the object file how to represent this reference. - PersonalityRef = - TLOF.getSymbolForDwarfGlobalReference(PersonalityFn, Asm->Mang, - Asm->MMI, - IsPersonalityIndirect, - IsPersonalityPCRel); - } - - unsigned PerEncoding = dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4; - if (IsPersonalityIndirect) - PerEncoding |= dwarf::DW_EH_PE_indirect; - unsigned LSDAEncoding = dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4; - unsigned FDEEncoding = dwarf::DW_EH_PE_pcrel | dwarf::DW_EH_PE_sdata4; + + unsigned LSDAEncoding = TLOF.getLSDAEncoding(); + unsigned FDEEncoding = TLOF.getFDEEncoding(); + unsigned PerEncoding = TLOF.getPersonalityEncoding(); char Augmentation[6] = { 0 }; unsigned AugmentationSize = 0; char *APtr = Augmentation + 1; - if (PersonalityRef) { + if (PersonalityFn) { // There is a personality function. *APtr++ = 'P'; AugmentationSize += 1 + SizeOfEncodedValue(PerEncoding); @@ -181,20 +147,19 @@ void DwarfException::EmitCIE(const Function *PersonalityFn, unsigned Index) { Asm->EmitInt8(RI->getDwarfRegNum(RI->getRARegister(), true)); EOL("CIE Return Address Column"); - EmitULEB128(AugmentationSize, "Augmentation Size"); - EmitEncodingByte(PerEncoding, "Personality"); - - // If there is a personality, we need to indicate the function's location. - if (PersonalityRef) { - if (!IsPersonalityPCRel) - PersonalityRef = CreateLabelDiff(PersonalityRef, "personalityref_addr", - Index); + if (Augmentation[0]) { + EmitULEB128(AugmentationSize, "Augmentation Size"); - O << MAI->getData32bitsDirective() << *PersonalityRef; - EOL("Personality"); - - EmitEncodingByte(LSDAEncoding, "LSDA"); - EmitEncodingByte(FDEEncoding, "FDE"); + // If there is a personality, we need to indicate the function's location. + if (PersonalityFn) { + EmitEncodingByte(PerEncoding, "Personality"); + EmitReference(PersonalityFn, PerEncoding); + EOL("Personality"); + } + if (UsesLSDA[Index]) + EmitEncodingByte(LSDAEncoding, "LSDA"); + if (FDEEncoding != dwarf::DW_EH_PE_absptr) + EmitEncodingByte(FDEEncoding, "FDE"); } // Indicate locations of general callee saved registers in frame. @@ -216,8 +181,12 @@ void DwarfException::EmitFDE(const FunctionEHFrameInfo &EHFrameInfo) { "Should not emit 'available externally' functions at all"); const Function *TheFunc = EHFrameInfo.function; + const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); - Asm->OutStreamer.SwitchSection(Asm->getObjFileLowering().getEHFrameSection()); + unsigned LSDAEncoding = TLOF.getLSDAEncoding(); + unsigned FDEEncoding = TLOF.getFDEEncoding(); + + Asm->OutStreamer.SwitchSection(TLOF.getEHFrameSection()); // Externally visible entry into the functions eh frame info. If the // corresponding function is static, this should not be externally visible. @@ -255,7 +224,8 @@ void DwarfException::EmitFDE(const FunctionEHFrameInfo &EHFrameInfo) { // EH frame header. EmitDifference("eh_frame_end", EHFrameInfo.Number, - "eh_frame_begin", EHFrameInfo.Number, true); + "eh_frame_begin", EHFrameInfo.Number, + true); EOL("Length of Frame Information Entry"); EmitLabel("eh_frame_begin", EHFrameInfo.Number); @@ -266,33 +236,23 @@ void DwarfException::EmitFDE(const FunctionEHFrameInfo &EHFrameInfo) { EOL("FDE CIE offset"); - EmitReference("eh_func_begin", EHFrameInfo.Number, true, true); + EmitReference("eh_func_begin", EHFrameInfo.Number, FDEEncoding); EOL("FDE initial location"); EmitDifference("eh_func_end", EHFrameInfo.Number, - "eh_func_begin", EHFrameInfo.Number, true); + "eh_func_begin", EHFrameInfo.Number, + SizeOfEncodedValue(FDEEncoding) == 4); EOL("FDE address range"); // If there is a personality and landing pads then point to the language // specific data area in the exception table. if (MMI->getPersonalities()[0] != NULL) { + unsigned Size = SizeOfEncodedValue(LSDAEncoding); - if (Asm->TM.getLSDAEncoding() != DwarfLSDAEncoding::EightByte) { - EmitULEB128(4, "Augmentation size"); - - if (EHFrameInfo.hasLandingPads) - EmitReference("exception", EHFrameInfo.Number, true, true); - else - Asm->OutStreamer.EmitIntValue(0, 4/*size*/, 0/*addrspace*/); - } else { - EmitULEB128(TD->getPointerSize(), "Augmentation size"); - - if (EHFrameInfo.hasLandingPads) { - EmitReference("exception", EHFrameInfo.Number, true, false); - } else { - Asm->OutStreamer.EmitIntValue(0, TD->getPointerSize(), - 0/*addrspace*/); - } - } + EmitULEB128(Size, "Augmentation size"); + if (EHFrameInfo.hasLandingPads) + EmitReference("exception", EHFrameInfo.Number, LSDAEncoding); + else + Asm->OutStreamer.EmitIntValue(0, Size/*size*/, 0/*addrspace*/); EOL("Language Specific Data Area"); } else { @@ -407,20 +367,22 @@ ComputeActionsTable(const SmallVectorImpl<const LandingPadInfo*> &LandingPads, if (NumShared < TypeIds.size()) { unsigned SizeAction = 0; - ActionEntry *PrevAction = 0; + unsigned PrevAction = (unsigned)-1; if (NumShared) { const unsigned SizePrevIds = PrevLPI->TypeIds.size(); assert(Actions.size()); - PrevAction = &Actions.back(); - SizeAction = MCAsmInfo::getSLEB128Size(PrevAction->NextAction) + - MCAsmInfo::getSLEB128Size(PrevAction->ValueForTypeID); + PrevAction = Actions.size() - 1; + SizeAction = + MCAsmInfo::getSLEB128Size(Actions[PrevAction].NextAction) + + MCAsmInfo::getSLEB128Size(Actions[PrevAction].ValueForTypeID); for (unsigned j = NumShared; j != SizePrevIds; ++j) { + assert(PrevAction != (unsigned)-1 && "PrevAction is invalid!"); SizeAction -= - MCAsmInfo::getSLEB128Size(PrevAction->ValueForTypeID); - SizeAction += -PrevAction->NextAction; - PrevAction = PrevAction->Previous; + MCAsmInfo::getSLEB128Size(Actions[PrevAction].ValueForTypeID); + SizeAction += -Actions[PrevAction].NextAction; + PrevAction = Actions[PrevAction].Previous; } } @@ -437,7 +399,7 @@ ComputeActionsTable(const SmallVectorImpl<const LandingPadInfo*> &LandingPads, ActionEntry Action = { ValueForTypeID, NextAction, PrevAction }; Actions.push_back(Action); - PrevAction = &Actions.back(); + PrevAction = Actions.size() - 1; } // Record the first action of the landing pad site. @@ -447,7 +409,7 @@ ComputeActionsTable(const SmallVectorImpl<const LandingPadInfo*> &LandingPads, // Information used when created the call-site table. The action record // field of the call site record is the offset of the first associated // action record, relative to the start of the actions table. This value is - // biased by 1 (1 in dicating the start of the actions table), and 0 + // biased by 1 (1 indicating the start of the actions table), and 0 // indicates that there are no actions. FirstActions.push_back(FirstAction); @@ -648,8 +610,7 @@ void DwarfException::EmitExceptionTable() { // landing pad site. SmallVector<ActionEntry, 32> Actions; SmallVector<unsigned, 64> FirstActions; - unsigned SizeActions = ComputeActionsTable(LandingPads, Actions, - FirstActions); + unsigned SizeActions=ComputeActionsTable(LandingPads, Actions, FirstActions); // Invokes and nounwind calls have entries in PadMap (due to being bracketed // by try-range labels when lowered). Ordinary calls do not, so appropriate @@ -677,29 +638,29 @@ void DwarfException::EmitExceptionTable() { const unsigned LandingPadSize = SizeOfEncodedValue(dwarf::DW_EH_PE_udata4); bool IsSJLJ = MAI->getExceptionHandlingType() == ExceptionHandling::SjLj; bool HaveTTData = IsSJLJ ? (!TypeInfos.empty() || !FilterIds.empty()) : true; - unsigned SizeSites; + unsigned CallSiteTableLength; if (IsSJLJ) - SizeSites = 0; + CallSiteTableLength = 0; else - SizeSites = CallSites.size() * + CallSiteTableLength = CallSites.size() * (SiteStartSize + SiteLengthSize + LandingPadSize); for (unsigned i = 0, e = CallSites.size(); i < e; ++i) { - SizeSites += MCAsmInfo::getULEB128Size(CallSites[i].Action); + CallSiteTableLength += MCAsmInfo::getULEB128Size(CallSites[i].Action); if (IsSJLJ) - SizeSites += MCAsmInfo::getULEB128Size(i); + CallSiteTableLength += MCAsmInfo::getULEB128Size(i); } // Type infos. const MCSection *LSDASection = Asm->getObjFileLowering().getLSDASection(); - unsigned TTypeFormat; + unsigned TTypeEncoding; unsigned TypeFormatSize; if (!HaveTTData) { // For SjLj exceptions, if there is no TypeInfo, then we just explicitly say // that we're omitting that bit. - TTypeFormat = dwarf::DW_EH_PE_omit; + TTypeEncoding = dwarf::DW_EH_PE_omit; TypeFormatSize = SizeOfEncodedValue(dwarf::DW_EH_PE_absptr); } else { // Okay, we have actual filters or typeinfos to emit. As such, we need to @@ -729,21 +690,28 @@ void DwarfException::EmitExceptionTable() { // somewhere. This predicate should be moved to a shared location that is // in target-independent code. // - if (LSDASection->getKind().isWriteable() || - Asm->TM.getRelocationModel() == Reloc::Static) - TTypeFormat = dwarf::DW_EH_PE_absptr; - else - TTypeFormat = dwarf::DW_EH_PE_indirect | dwarf::DW_EH_PE_pcrel | - dwarf::DW_EH_PE_sdata4; - - TypeFormatSize = SizeOfEncodedValue(TTypeFormat); + TTypeEncoding = Asm->getObjFileLowering().getTTypeEncoding(); + TypeFormatSize = SizeOfEncodedValue(TTypeEncoding); } // Begin the exception table. Asm->OutStreamer.SwitchSection(LSDASection); Asm->EmitAlignment(2, 0, 0, false); + // Emit the LSDA. O << "GCC_except_table" << SubprogramCount << ":\n"; + EmitLabel("exception", SubprogramCount); + + if (IsSJLJ) { + SmallString<16> LSDAName; + raw_svector_ostream(LSDAName) << MAI->getPrivateGlobalPrefix() << + "_LSDA_" << Asm->getFunctionNumber(); + O << LSDAName.str() << ":\n"; + } + + // Emit the LSDA header. + EmitEncodingByte(dwarf::DW_EH_PE_omit, "@LPStart"); + EmitEncodingByte(TTypeEncoding, "@TType"); // The type infos need to be aligned. GCC does this by inserting padding just // before the type infos. However, this changes the size of the exception @@ -752,7 +720,7 @@ void DwarfException::EmitExceptionTable() { // So by increasing the size by inserting padding, you may increase the number // of bytes used for writing the size. If it increases, say by one byte, then // you now need to output one less byte of padding to get the type infos - // aligned. However this decreases the size of the exception table. This + // aligned. However this decreases the size of the exception table. This // changes the value you have to output for the exception table size. Due to // the variable length encoding, the number of bytes used for writing the // length may decrease. If so, you then have to increase the amount of @@ -761,41 +729,35 @@ void DwarfException::EmitExceptionTable() { // We chose another solution: don't output padding inside the table like GCC // does, instead output it before the table. unsigned SizeTypes = TypeInfos.size() * TypeFormatSize; - unsigned TyOffset = sizeof(int8_t) + // Call site format - MCAsmInfo::getULEB128Size(SizeSites) + // Call site table length - SizeSites + SizeActions + SizeTypes; - unsigned TotalSize = sizeof(int8_t) + // LPStart format - sizeof(int8_t) + // TType format - (HaveTTData ? - MCAsmInfo::getULEB128Size(TyOffset) : 0) + // TType base offset - TyOffset; + unsigned CallSiteTableLengthSize = + MCAsmInfo::getULEB128Size(CallSiteTableLength); + unsigned TTypeBaseOffset = + sizeof(int8_t) + // Call site format + CallSiteTableLengthSize + // Call site table length size + CallSiteTableLength + // Call site table length + SizeActions + // Actions size + SizeTypes; + unsigned TTypeBaseOffsetSize = MCAsmInfo::getULEB128Size(TTypeBaseOffset); + unsigned TotalSize = + sizeof(int8_t) + // LPStart format + sizeof(int8_t) + // TType format + (HaveTTData ? TTypeBaseOffsetSize : 0) + // TType base offset size + TTypeBaseOffset; // TType base offset unsigned SizeAlign = (4 - TotalSize) & 3; - for (unsigned i = 0; i != SizeAlign; ++i) { - Asm->EmitInt8(0); - EOL("Padding"); - } - - EmitLabel("exception", SubprogramCount); - - if (IsSJLJ) { - SmallString<16> LSDAName; - raw_svector_ostream(LSDAName) << MAI->getPrivateGlobalPrefix() << - "_LSDA_" << Asm->getFunctionNumber(); - O << LSDAName.str() << ":\n"; + if (HaveTTData) { + // Account for any extra padding that will be added to the call site table + // length. + EmitULEB128(TTypeBaseOffset, "@TType base offset", SizeAlign); + SizeAlign = 0; } - // Emit the header. - EmitEncodingByte(dwarf::DW_EH_PE_omit, "@LPStart"); - EmitEncodingByte(TTypeFormat, "@TType"); - - if (HaveTTData) - EmitULEB128(TyOffset, "@TType base offset"); - // SjLj Exception handling if (IsSJLJ) { EmitEncodingByte(dwarf::DW_EH_PE_udata4, "Call site"); - EmitULEB128(SizeSites, "Call site table length"); + + // Add extra padding if it wasn't added to the TType base offset. + EmitULEB128(CallSiteTableLength, "Call site table length", SizeAlign); // Emit the landing pad site information. unsigned idx = 0; @@ -836,7 +798,9 @@ void DwarfException::EmitExceptionTable() { // Emit the landing pad call site table. EmitEncodingByte(dwarf::DW_EH_PE_udata4, "Call site"); - EmitULEB128(SizeSites, "Call site table length"); + + // Add extra padding if it wasn't added to the TType base offset. + EmitULEB128(CallSiteTableLength, "Call site table length", SizeAlign); for (SmallVectorImpl<CallSiteEntry>::const_iterator I = CallSites.begin(), E = CallSites.end(); I != E; ++I) { @@ -906,23 +870,23 @@ void DwarfException::EmitExceptionTable() { } // Emit the Catch TypeInfos. - if (TypeInfos.size() != 0) EOL("-- Catch TypeInfos --"); + if (!TypeInfos.empty()) EOL("-- Catch TypeInfos --"); for (std::vector<GlobalVariable *>::const_reverse_iterator I = TypeInfos.rbegin(), E = TypeInfos.rend(); I != E; ++I) { const GlobalVariable *GV = *I; - PrintRelDirective(); if (GV) { - O << *Asm->GetGlobalValueSymbol(GV); + EmitReference(GV, TTypeEncoding); EOL("TypeInfo"); } else { + PrintRelDirective(TTypeEncoding); O << "0x0"; EOL(""); } } // Emit the Exception Specifications. - if (FilterIds.size() != 0) EOL("-- Filter IDs --"); + if (!FilterIds.empty()) EOL("-- Filter IDs --"); for (std::vector<unsigned>::const_iterator I = FilterIds.begin(), E = FilterIds.end(); I < E; ++I) { unsigned TypeID = *I; diff --git a/lib/CodeGen/AsmPrinter/DwarfException.h b/lib/CodeGen/AsmPrinter/DwarfException.h index 06033a1..3db1a00 100644 --- a/lib/CodeGen/AsmPrinter/DwarfException.h +++ b/lib/CodeGen/AsmPrinter/DwarfException.h @@ -76,9 +76,6 @@ class DwarfException : public DwarfPrinter { /// ExceptionTimer - Timer for the Dwarf exception writer. Timer *ExceptionTimer; - /// SizeOfEncodedValue - Return the size of the encoding in bytes. - unsigned SizeOfEncodedValue(unsigned Encoding); - /// EmitCIE - Emit a Common Information Entry (CIE). This holds information /// that is shared among many Frame Description Entries. There is at least /// one CIE in every non-empty .debug_frame section. @@ -135,7 +132,7 @@ class DwarfException : public DwarfPrinter { struct ActionEntry { int ValueForTypeID; // The value to write - may not be equal to the type id. int NextAction; - struct ActionEntry *Previous; + unsigned Previous; }; /// CallSiteEntry - Structure describing an entry in the call-site table. diff --git a/lib/CodeGen/AsmPrinter/DwarfPrinter.cpp b/lib/CodeGen/AsmPrinter/DwarfPrinter.cpp index 415390b..28ff0eb 100644 --- a/lib/CodeGen/AsmPrinter/DwarfPrinter.cpp +++ b/lib/CodeGen/AsmPrinter/DwarfPrinter.cpp @@ -8,7 +8,7 @@ //===----------------------------------------------------------------------===// // // Emit general DWARF directives. -// +// //===----------------------------------------------------------------------===// #include "DwarfPrinter.h" @@ -18,13 +18,17 @@ #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineModuleInfo.h" #include "llvm/MC/MCAsmInfo.h" +#include "llvm/MC/MCContext.h" +#include "llvm/MC/MCExpr.h" #include "llvm/MC/MCStreamer.h" #include "llvm/MC/MCSymbol.h" #include "llvm/Target/TargetData.h" #include "llvm/Target/TargetFrameInfo.h" +#include "llvm/Target/TargetLoweringObjectFile.h" #include "llvm/Target/TargetRegisterInfo.h" #include "llvm/Support/Dwarf.h" #include "llvm/Support/ErrorHandling.h" +#include "llvm/ADT/SmallString.h" using namespace llvm; DwarfPrinter::DwarfPrinter(raw_ostream &OS, AsmPrinter *A, const MCAsmInfo *T, @@ -33,6 +37,26 @@ DwarfPrinter::DwarfPrinter(raw_ostream &OS, AsmPrinter *A, const MCAsmInfo *T, RI(Asm->TM.getRegisterInfo()), M(NULL), MF(NULL), MMI(NULL), SubprogramCount(0), Flavor(flavor), SetCounter(1) {} +/// SizeOfEncodedValue - Return the size of the encoding in bytes. +unsigned DwarfPrinter::SizeOfEncodedValue(unsigned Encoding) const { + if (Encoding == dwarf::DW_EH_PE_omit) + return 0; + + switch (Encoding & 0x07) { + case dwarf::DW_EH_PE_absptr: + return TD->getPointerSize(); + case dwarf::DW_EH_PE_udata2: + return 2; + case dwarf::DW_EH_PE_udata4: + return 4; + case dwarf::DW_EH_PE_udata8: + return 8; + } + + assert(0 && "Invalid encoded value."); + return 0; +} + void DwarfPrinter::PrintRelDirective(bool Force32Bit, bool isInSection) const { if (isInSection && MAI->getDwarfSectionOffsetDirective()) O << MAI->getDwarfSectionOffsetDirective(); @@ -42,6 +66,14 @@ void DwarfPrinter::PrintRelDirective(bool Force32Bit, bool isInSection) const { O << MAI->getData64bitsDirective(); } +void DwarfPrinter::PrintRelDirective(unsigned Encoding) const { + unsigned Size = SizeOfEncodedValue(Encoding); + assert((Size == 4 || Size == 8) && "Do not support other types or rels!"); + + O << (Size == 4 ? + MAI->getData32bitsDirective() : MAI->getData64bitsDirective()); +} + /// EOL - Print a newline character to asm stream. If a comment is present /// then it will be printed first. Comments should not contain '\n'. void DwarfPrinter::EOL(const Twine &Comment) const { @@ -127,29 +159,35 @@ void DwarfPrinter::EmitSLEB128(int Value, const char *Desc) const { Value >>= 7; IsMore = Value != Sign || ((Byte ^ Sign) & 0x40) != 0; if (IsMore) Byte |= 0x80; - Asm->OutStreamer.EmitIntValue(Byte, 1, /*addrspace*/0); } while (IsMore); } /// EmitULEB128 - emit the specified signed leb128 value. -void DwarfPrinter::EmitULEB128(unsigned Value, const char *Desc) const { +void DwarfPrinter::EmitULEB128(unsigned Value, const char *Desc, + unsigned PadTo) const { if (Asm->VerboseAsm && Desc) Asm->OutStreamer.AddComment(Desc); - if (MAI->hasLEB128()) { + if (MAI->hasLEB128() && PadTo == 0) { O << "\t.uleb128\t" << Value; Asm->OutStreamer.AddBlankLine(); return; } - // If we don't have .uleb128, emit as .bytes. + // If we don't have .uleb128 or we want to emit padding, emit as .bytes. do { unsigned char Byte = static_cast<unsigned char>(Value & 0x7f); Value >>= 7; - if (Value) Byte |= 0x80; + if (Value || PadTo != 0) Byte |= 0x80; Asm->OutStreamer.EmitIntValue(Byte, 1, /*addrspace*/0); } while (Value); + + if (PadTo) { + if (PadTo > 1) + Asm->OutStreamer.EmitFill(PadTo - 1, 0x80/*fillval*/, 0/*addrspace*/); + Asm->OutStreamer.EmitFill(1, 0/*fillval*/, 0/*addrspace*/); + } } @@ -195,6 +233,31 @@ void DwarfPrinter::EmitReference(const MCSymbol *Sym, bool IsPCRelative, if (IsPCRelative) O << "-" << MAI->getPCSymbol(); } +void DwarfPrinter::EmitReference(const char *Tag, unsigned Number, + unsigned Encoding) const { + SmallString<64> Name; + raw_svector_ostream(Name) << MAI->getPrivateGlobalPrefix() + << Tag << Number; + + MCSymbol *Sym = Asm->OutContext.GetOrCreateSymbol(Name.str()); + EmitReference(Sym, Encoding); +} + +void DwarfPrinter::EmitReference(const MCSymbol *Sym, unsigned Encoding) const { + const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); + + PrintRelDirective(Encoding); + O << *TLOF.getSymbolForDwarfReference(Sym, Asm->MMI, Encoding);; +} + +void DwarfPrinter::EmitReference(const GlobalValue *GV, unsigned Encoding)const { + const TargetLoweringObjectFile &TLOF = Asm->getObjFileLowering(); + + PrintRelDirective(Encoding); + O << *TLOF.getSymbolForDwarfGlobalReference(GV, Asm->Mang, + Asm->MMI, Encoding);; +} + /// EmitDifference - Emit the difference between two labels. If this assembler /// supports .set, we emit a .set of a temporary and then use it in the .word. void DwarfPrinter::EmitDifference(const char *TagHi, unsigned NumberHi, @@ -248,7 +311,6 @@ void DwarfPrinter::EmitSectionOffset(const char* Label, const char* Section, PrintRelDirective(IsSmall); PrintLabelName("set", SetCounter, Flavor); ++SetCounter; - O << "\n"; } else { PrintRelDirective(IsSmall, true); PrintLabelName(Label, LabelNumber); @@ -257,7 +319,6 @@ void DwarfPrinter::EmitSectionOffset(const char* Label, const char* Section, O << "-"; PrintLabelName(Section, SectionNumber); } - O << "\n"; } } diff --git a/lib/CodeGen/AsmPrinter/DwarfPrinter.h b/lib/CodeGen/AsmPrinter/DwarfPrinter.h index 69d9c27..bd715f2 100644 --- a/lib/CodeGen/AsmPrinter/DwarfPrinter.h +++ b/lib/CodeGen/AsmPrinter/DwarfPrinter.h @@ -28,6 +28,7 @@ class Module; class MCAsmInfo; class TargetData; class TargetRegisterInfo; +class GlobalValue; class MCSymbol; class Twine; @@ -85,6 +86,10 @@ public: const MCAsmInfo *getMCAsmInfo() const { return MAI; } const TargetData *getTargetData() const { return TD; } + /// SizeOfEncodedValue - Return the size of the encoding in bytes. + unsigned SizeOfEncodedValue(unsigned Encoding) const; + + void PrintRelDirective(unsigned Encoding) const; void PrintRelDirective(bool Force32Bit = false, bool isInSection = false) const; @@ -106,7 +111,8 @@ public: void EmitSLEB128(int Value, const char *Desc) const; /// EmitULEB128 - emit the specified unsigned leb128 value. - void EmitULEB128(unsigned Value, const char *Desc = 0) const; + void EmitULEB128(unsigned Value, const char *Desc = 0, + unsigned PadTo = 0) const; /// PrintLabelName - Print label name in form used by Dwarf writer. @@ -140,6 +146,10 @@ public: void EmitReference(const MCSymbol *Sym, bool IsPCRelative = false, bool Force32Bit = false) const; + void EmitReference(const char *Tag, unsigned Number, unsigned Encoding) const; + void EmitReference(const MCSymbol *Sym, unsigned Encoding) const; + void EmitReference(const GlobalValue *GV, unsigned Encoding) const; + /// EmitDifference - Emit the difference between two labels. void EmitDifference(const DWLabel &LabelHi, const DWLabel &LabelLo, bool IsSmall = false) { diff --git a/lib/CodeGen/AsmPrinter/OcamlGCPrinter.cpp b/lib/CodeGen/AsmPrinter/OcamlGCPrinter.cpp index 3531ed6..a9502fd 100644 --- a/lib/CodeGen/AsmPrinter/OcamlGCPrinter.cpp +++ b/lib/CodeGen/AsmPrinter/OcamlGCPrinter.cpp @@ -22,6 +22,7 @@ #include "llvm/Target/TargetMachine.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/FormattedStream.h" +#include <ctype.h> using namespace llvm; namespace { diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp index faf4d95..d94729a 100644 --- a/lib/CodeGen/BranchFolding.cpp +++ b/lib/CodeGen/BranchFolding.cpp @@ -334,7 +334,9 @@ static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1, unsigned TailLen = 0; while (I1 != MBB1->begin() && I2 != MBB2->begin()) { --I1; --I2; - if (!I1->isIdenticalTo(I2) || + // Don't merge debugging pseudos. + if (I1->isDebugValue() || I2->isDebugValue() || + !I1->isIdenticalTo(I2) || // FIXME: This check is dubious. It's used to get around a problem where // people incorrectly expect inline asm directives to remain in the same // relative order. This is untenable because normal compiler @@ -412,6 +414,8 @@ static unsigned EstimateRuntime(MachineBasicBlock::iterator I, MachineBasicBlock::iterator E) { unsigned Time = 0; for (; I != E; ++I) { + if (I->isDebugValue()) + continue; const TargetInstrDesc &TID = I->getDesc(); if (TID.isCall()) Time += 10; diff --git a/lib/CodeGen/CMakeLists.txt b/lib/CodeGen/CMakeLists.txt index 9fcbea9..d385b86 100644 --- a/lib/CodeGen/CMakeLists.txt +++ b/lib/CodeGen/CMakeLists.txt @@ -22,6 +22,7 @@ add_llvm_library(LLVMCodeGen LiveVariables.cpp LowerSubregs.cpp MachineBasicBlock.cpp + MachineCSE.cpp MachineDominators.cpp MachineFunction.cpp MachineFunctionAnalysis.cpp @@ -39,6 +40,7 @@ add_llvm_library(LLVMCodeGen ObjectCodeEmitter.cpp OcamlGC.cpp OptimizeExts.cpp + OptimizePHIs.cpp PHIElimination.cpp Passes.cpp PostRASchedulerList.cpp @@ -66,6 +68,7 @@ add_llvm_library(LLVMCodeGen StrongPHIElimination.cpp TailDuplication.cpp TargetInstrInfoImpl.cpp + TargetLoweringObjectFileImpl.cpp TwoAddressInstructionPass.cpp UnreachableBlockElim.cpp VirtRegMap.cpp diff --git a/lib/CodeGen/CalcSpillWeights.cpp b/lib/CodeGen/CalcSpillWeights.cpp index 2bedd04..a328d0e 100644 --- a/lib/CodeGen/CalcSpillWeights.cpp +++ b/lib/CodeGen/CalcSpillWeights.cpp @@ -131,10 +131,7 @@ bool CalculateSpillWeights::runOnMachineFunction(MachineFunction &fn) { if (Hint.first || Hint.second) li.weight *= 1.01F; - // Divide the weight of the interval by its size. This encourages - // spilling of intervals that are large and have few uses, and - // discourages spilling of small intervals with many uses. - li.weight /= lis->getApproximateInstructionCount(li) * SlotIndex::NUM; + lis->normalizeSpillWeight(li); } } diff --git a/lib/CodeGen/CodePlacementOpt.cpp b/lib/CodeGen/CodePlacementOpt.cpp index 05a57d4..3ff2a04 100644 --- a/lib/CodeGen/CodePlacementOpt.cpp +++ b/lib/CodeGen/CodePlacementOpt.cpp @@ -102,22 +102,23 @@ bool CodePlacementOpt::HasAnalyzableTerminator(MachineBasicBlock *MBB) { // Conservatively ignore EH landing pads. if (MBB->isLandingPad()) return false; - // Ignore blocks which look like they might have EH-related control flow. - // At the time of this writing, there are blocks which AnalyzeBranch - // thinks end in single uncoditional branches, yet which have two CFG - // successors. Code in this file is not prepared to reason about such things. - if (!MBB->empty() && MBB->back().isEHLabel()) - return false; - // Aggressively handle return blocks and similar constructs. if (MBB->succ_empty()) return true; // Ask the target's AnalyzeBranch if it can handle this block. MachineBasicBlock *TBB = 0, *FBB = 0; SmallVector<MachineOperand, 4> Cond; - // Make the terminator is understood. + // Make sure the terminator is understood. if (TII->AnalyzeBranch(*MBB, TBB, FBB, Cond)) return false; + // Ignore blocks which look like they might have EH-related control flow. + // AnalyzeBranch thinks it knows how to analyze such things, but it doesn't + // recognize the possibility of a control transfer through an unwind. + // Such blocks contain EH_LABEL instructions, however they may be in the + // middle of the block. Instead of searching for them, just check to see + // if the CFG disagrees with AnalyzeBranch. + if (1u + !Cond.empty() != MBB->succ_size()) + return false; // Make sure we have the option of reversing the condition. if (!Cond.empty() && TII->ReverseBranchCondition(Cond)) return false; diff --git a/lib/CodeGen/CriticalAntiDepBreaker.cpp b/lib/CodeGen/CriticalAntiDepBreaker.cpp index 056e2d5..7d3de89 100644 --- a/lib/CodeGen/CriticalAntiDepBreaker.cpp +++ b/lib/CodeGen/CriticalAntiDepBreaker.cpp @@ -119,6 +119,8 @@ void CriticalAntiDepBreaker::FinishBlock() { void CriticalAntiDepBreaker::Observe(MachineInstr *MI, unsigned Count, unsigned InsertPosIndex) { + if (MI->isDebugValue()) + return; assert(Count < InsertPosIndex && "Instruction index out of expected range!"); // Any register which was defined within the previous scheduling region @@ -409,6 +411,8 @@ BreakAntiDependencies(std::vector<SUnit>& SUnits, for (MachineBasicBlock::iterator I = End, E = Begin; I != E; --Count) { MachineInstr *MI = --I; + if (MI->isDebugValue()) + continue; // Check if this instruction has a dependence on the critical path that // is an anti-dependence that we may be able to break. If it is, set diff --git a/lib/CodeGen/DeadMachineInstructionElim.cpp b/lib/CodeGen/DeadMachineInstructionElim.cpp index b0cb24d..d69c995 100644 --- a/lib/CodeGen/DeadMachineInstructionElim.cpp +++ b/lib/CodeGen/DeadMachineInstructionElim.cpp @@ -55,7 +55,7 @@ FunctionPass *llvm::createDeadMachineInstructionElimPass() { bool DeadMachineInstructionElim::isDead(const MachineInstr *MI) const { // Don't delete instructions with side effects. bool SawStore = false; - if (!MI->isSafeToMove(TII, SawStore, 0)) + if (!MI->isSafeToMove(TII, 0, SawStore) && !MI->isPHI()) return false; // Examine each operand. @@ -64,8 +64,8 @@ bool DeadMachineInstructionElim::isDead(const MachineInstr *MI) const { if (MO.isReg() && MO.isDef()) { unsigned Reg = MO.getReg(); if (TargetRegisterInfo::isPhysicalRegister(Reg) ? - LivePhysRegs[Reg] : !MRI->use_empty(Reg)) { - // This def has a use. Don't delete the instruction! + LivePhysRegs[Reg] : !MRI->use_nodbg_empty(Reg)) { + // This def has a non-debug use. Don't delete the instruction! return false; } } @@ -111,23 +111,31 @@ bool DeadMachineInstructionElim::runOnMachineFunction(MachineFunction &MF) { MIE = MBB->rend(); MII != MIE; ) { MachineInstr *MI = &*MII; - if (MI->isDebugValue()) { - // Don't delete the DBG_VALUE itself, but if its Value operand is - // a vreg and this is the only use, substitute an undef operand; - // the former operand will then be deleted normally. - if (MI->getNumOperands()==3 && MI->getOperand(0).isReg()) { - unsigned Reg = MI->getOperand(0).getReg(); - MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg); - assert(I != MRI->use_end()); - if (++I == MRI->use_end()) - // only one use, which must be this DBG_VALUE. - MI->getOperand(0).setReg(0U); - } - } - // If the instruction is dead, delete it! if (isDead(MI)) { DEBUG(dbgs() << "DeadMachineInstructionElim: DELETING: " << *MI); + // It is possible that some DBG_VALUE instructions refer to this + // instruction. Examine each def operand for such references; + // if found, mark the DBG_VALUE as undef (but don't delete it). + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || !MO.isDef()) + continue; + unsigned Reg = MO.getReg(); + if (!TargetRegisterInfo::isVirtualRegister(Reg)) + continue; + MachineRegisterInfo::use_iterator nextI; + for (MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg), + E = MRI->use_end(); I!=E; I=nextI) { + nextI = llvm::next(I); // I is invalidated by the setReg + MachineOperand& Use = I.getOperand(); + MachineInstr *UseMI = Use.getParent(); + if (UseMI==MI) + continue; + assert(Use.isDebug()); + UseMI->getOperand(0).setReg(0U); + } + } AnyChanges = true; MI->eraseFromParent(); ++NumDeletes; diff --git a/lib/CodeGen/IntrinsicLowering.cpp b/lib/CodeGen/IntrinsicLowering.cpp index 9997a48..87ab7ef 100644 --- a/lib/CodeGen/IntrinsicLowering.cpp +++ b/lib/CodeGen/IntrinsicLowering.cpp @@ -155,7 +155,7 @@ void IntrinsicLowering::AddPrototypes(Module &M) { /// LowerBSWAP - Emit the code to lower bswap of V before the specified /// instruction IP. static Value *LowerBSWAP(LLVMContext &Context, Value *V, Instruction *IP) { - assert(V->getType()->isInteger() && "Can't bswap a non-integer type!"); + assert(V->getType()->isIntegerTy() && "Can't bswap a non-integer type!"); unsigned BitSize = V->getType()->getPrimitiveSizeInBits(); @@ -251,7 +251,7 @@ static Value *LowerBSWAP(LLVMContext &Context, Value *V, Instruction *IP) { /// LowerCTPOP - Emit the code to lower ctpop of V before the specified /// instruction IP. static Value *LowerCTPOP(LLVMContext &Context, Value *V, Instruction *IP) { - assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!"); + assert(V->getType()->isIntegerTy() && "Can't ctpop a non-integer type!"); static const uint64_t MaskValues[6] = { 0x5555555555555555ULL, 0x3333333333333333ULL, diff --git a/lib/CodeGen/LLVMTargetMachine.cpp b/lib/CodeGen/LLVMTargetMachine.cpp index 40e0150..5e88865 100644 --- a/lib/CodeGen/LLVMTargetMachine.cpp +++ b/lib/CodeGen/LLVMTargetMachine.cpp @@ -14,6 +14,7 @@ #include "llvm/Target/TargetMachine.h" #include "llvm/PassManager.h" #include "llvm/Pass.h" +#include "llvm/Analysis/Verifier.h" #include "llvm/Assembly/PrintModulePass.h" #include "llvm/CodeGen/AsmPrinter.h" #include "llvm/CodeGen/Passes.h" @@ -66,6 +67,9 @@ static cl::opt<bool> VerifyMachineCode("verify-machineinstrs", cl::Hidden, cl::desc("Verify generated machine code"), cl::init(getenv("LLVM_VERIFY_MACHINEINSTRS")!=NULL)); +static cl::opt<bool> EnableMachineCSE("enable-machine-cse", cl::Hidden, + cl::desc("Enable Machine CSE")); + static cl::opt<cl::boolOrDefault> AsmVerbose("asm-verbose", cl::desc("Add comments to directives."), cl::init(cl::BOU_UNSET)); @@ -114,9 +118,10 @@ LLVMTargetMachine::setCodeModelForStatic() { bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM, formatted_raw_ostream &Out, CodeGenFileType FileType, - CodeGenOpt::Level OptLevel) { + CodeGenOpt::Level OptLevel, + bool DisableVerify) { // Add common CodeGen passes. - if (addCommonCodeGenPasses(PM, OptLevel)) + if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify)) return true; OwningPtr<MCContext> Context(new MCContext()); @@ -140,7 +145,7 @@ bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM, case CGFT_ObjectFile: { // Create the code emitter for the target if it exists. If not, .o file // emission fails. - MCCodeEmitter *MCE = getTarget().createCodeEmitter(*this); + MCCodeEmitter *MCE = getTarget().createCodeEmitter(*this, *Context); if (MCE == 0) return true; @@ -192,12 +197,13 @@ bool LLVMTargetMachine::addPassesToEmitFile(PassManagerBase &PM, /// bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM, JITCodeEmitter &JCE, - CodeGenOpt::Level OptLevel) { + CodeGenOpt::Level OptLevel, + bool DisableVerify) { // Make sure the code model is set. setCodeModelForJIT(); // Add common CodeGen passes. - if (addCommonCodeGenPasses(PM, OptLevel)) + if (addCommonCodeGenPasses(PM, OptLevel, DisableVerify)) return true; addCodeEmitter(PM, OptLevel, JCE); @@ -206,6 +212,12 @@ bool LLVMTargetMachine::addPassesToEmitMachineCode(PassManagerBase &PM, return false; // success! } +static void printNoVerify(PassManagerBase &PM, + const char *Banner) { + if (PrintMachineCode) + PM.add(createMachineFunctionPrinterPass(dbgs(), Banner)); +} + static void printAndVerify(PassManagerBase &PM, const char *Banner, bool allowDoubleDefs = false) { @@ -220,13 +232,19 @@ static void printAndVerify(PassManagerBase &PM, /// emitting to assembly files or machine code output. /// bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, - CodeGenOpt::Level OptLevel) { + CodeGenOpt::Level OptLevel, + bool DisableVerify) { // Standard LLVM-Level Passes. + // Before running any passes, run the verifier to determine if the input + // coming from the front-end and/or optimizer is valid. + if (!DisableVerify) + PM.add(createVerifierPass()); + // Optionally, tun split-GEPs and no-load GVN. if (EnableSplitGEPGVN) { PM.add(createGEPSplitterPass()); - PM.add(createGVNPass(/*NoPRE=*/false, /*NoLoads=*/true)); + PM.add(createGVNPass(/*NoLoads=*/true)); } // Run loop strength reduction before anything else. @@ -273,6 +291,11 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, "*** Final LLVM Code input to ISel ***\n", &dbgs())); + // All passes which modify the LLVM IR are now complete; run the verifier + // to ensure that the IR is valid. + if (!DisableVerify) + PM.add(createVerifierPass()); + // Standard Lower-Level Passes. // Set up a MachineFunction for the rest of CodeGen to work on. @@ -291,6 +314,10 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, printAndVerify(PM, "After Instruction Selection", /* allowDoubleDefs= */ true); + // Optimize PHIs before DCE: removing dead PHI cycles may make more + // instructions dead. + if (OptLevel != CodeGenOpt::None) + PM.add(createOptimizePHIsPass()); // Delete dead machine instructions regardless of optimization level. PM.add(createDeadMachineInstructionElimPass()); @@ -301,6 +328,8 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, PM.add(createOptimizeExtsPass()); if (!DisableMachineLICM) PM.add(createMachineLICMPass()); + if (EnableMachineCSE) + PM.add(createMachineCSEPass()); if (!DisableMachineSink) PM.add(createMachineSinkingPass()); printAndVerify(PM, "After MachineLICM and MachineSinking", @@ -355,13 +384,13 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, // Branch folding must be run after regalloc and prolog/epilog insertion. if (OptLevel != CodeGenOpt::None && !DisableBranchFold) { PM.add(createBranchFoldingPass(getEnableTailMergeDefault())); - printAndVerify(PM, "After BranchFolding"); + printNoVerify(PM, "After BranchFolding"); } // Tail duplication. if (OptLevel != CodeGenOpt::None && !DisableTailDuplicate) { PM.add(createTailDuplicatePass(false)); - printAndVerify(PM, "After TailDuplicate"); + printNoVerify(PM, "After TailDuplicate"); } PM.add(createGCMachineCodeAnalysisPass()); @@ -371,11 +400,11 @@ bool LLVMTargetMachine::addCommonCodeGenPasses(PassManagerBase &PM, if (OptLevel != CodeGenOpt::None && !DisableCodePlace) { PM.add(createCodePlacementOptPass()); - printAndVerify(PM, "After CodePlacementOpt"); + printNoVerify(PM, "After CodePlacementOpt"); } if (addPreEmitPass(PM, OptLevel)) - printAndVerify(PM, "After PreEmit passes"); + printNoVerify(PM, "After PreEmit passes"); return false; } diff --git a/lib/CodeGen/LiveIntervalAnalysis.cpp b/lib/CodeGen/LiveIntervalAnalysis.cpp index 432409a..ccda66f 100644 --- a/lib/CodeGen/LiveIntervalAnalysis.cpp +++ b/lib/CodeGen/LiveIntervalAnalysis.cpp @@ -329,24 +329,43 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, DEBUG(dbgs() << " +" << NewLR); interval.addRange(NewLR); - // Iterate over all of the blocks that the variable is completely - // live in, adding [insrtIndex(begin), instrIndex(end)+4) to the - // live interval. - for (SparseBitVector<>::iterator I = vi.AliveBlocks.begin(), - E = vi.AliveBlocks.end(); I != E; ++I) { - MachineBasicBlock *aliveBlock = mf_->getBlockNumbered(*I); - LiveRange LR(getMBBStartIdx(aliveBlock), getMBBEndIdx(aliveBlock), ValNo); - interval.addRange(LR); - DEBUG(dbgs() << " +" << LR); + bool PHIJoin = lv_->isPHIJoin(interval.reg); + + if (PHIJoin) { + // A phi join register is killed at the end of the MBB and revived as a new + // valno in the killing blocks. + assert(vi.AliveBlocks.empty() && "Phi join can't pass through blocks"); + DEBUG(dbgs() << " phi-join"); + ValNo->addKill(indexes_->getTerminatorGap(mbb)); + ValNo->setHasPHIKill(true); + } else { + // Iterate over all of the blocks that the variable is completely + // live in, adding [insrtIndex(begin), instrIndex(end)+4) to the + // live interval. + for (SparseBitVector<>::iterator I = vi.AliveBlocks.begin(), + E = vi.AliveBlocks.end(); I != E; ++I) { + MachineBasicBlock *aliveBlock = mf_->getBlockNumbered(*I); + LiveRange LR(getMBBStartIdx(aliveBlock), getMBBEndIdx(aliveBlock), ValNo); + interval.addRange(LR); + DEBUG(dbgs() << " +" << LR); + } } // Finally, this virtual register is live from the start of any killing // block to the 'use' slot of the killing instruction. for (unsigned i = 0, e = vi.Kills.size(); i != e; ++i) { MachineInstr *Kill = vi.Kills[i]; - SlotIndex killIdx = - getInstructionIndex(Kill).getDefIndex(); - LiveRange LR(getMBBStartIdx(Kill->getParent()), killIdx, ValNo); + SlotIndex Start = getMBBStartIdx(Kill->getParent()); + SlotIndex killIdx = getInstructionIndex(Kill).getDefIndex(); + + // Create interval with one of a NEW value number. Note that this value + // number isn't actually defined by an instruction, weird huh? :) + if (PHIJoin) { + ValNo = interval.getNextValue(SlotIndex(Start, true), 0, false, + VNInfoAllocator); + ValNo->setIsPHIDef(true); + } + LiveRange LR(Start, killIdx, ValNo); interval.addRange(LR); ValNo->addKill(killIdx); DEBUG(dbgs() << " +" << LR); @@ -409,48 +428,11 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, interval.print(dbgs(), tri_); }); } else { - // Otherwise, this must be because of phi elimination. If this is the - // first redefinition of the vreg that we have seen, go back and change - // the live range in the PHI block to be a different value number. - if (interval.containsOneValue()) { - - VNInfo *VNI = interval.getValNumInfo(0); - // Phi elimination may have reused the register for multiple identical - // phi nodes. There will be a kill per phi. Remove the old ranges that - // we now know have an incorrect number. - for (unsigned ki=0, ke=vi.Kills.size(); ki != ke; ++ki) { - MachineInstr *Killer = vi.Kills[ki]; - SlotIndex Start = getMBBStartIdx(Killer->getParent()); - SlotIndex End = getInstructionIndex(Killer).getDefIndex(); - DEBUG({ - dbgs() << "\n\t\trenaming [" << Start << "," << End << "] in: "; - interval.print(dbgs(), tri_); - }); - interval.removeRange(Start, End); - - // Replace the interval with one of a NEW value number. Note that - // this value number isn't actually defined by an instruction, weird - // huh? :) - LiveRange LR(Start, End, - interval.getNextValue(SlotIndex(Start, true), - 0, false, VNInfoAllocator)); - LR.valno->setIsPHIDef(true); - interval.addRange(LR); - LR.valno->addKill(End); - } - - MachineBasicBlock *killMBB = getMBBFromIndex(VNI->def); - VNI->addKill(indexes_->getTerminatorGap(killMBB)); - VNI->setHasPHIKill(true); - DEBUG({ - dbgs() << " RESULT: "; - interval.print(dbgs(), tri_); - }); - } - + assert(lv_->isPHIJoin(interval.reg) && "Multiply defined register"); // In the case of PHI elimination, each variable definition is only // live until the end of the block. We've already taken care of the // rest of the live range. + SlotIndex defIndex = MIIdx.getDefIndex(); if (MO.isEarlyClobber()) defIndex = MIIdx.getUseIndex(); @@ -468,7 +450,7 @@ void LiveIntervals::handleVirtualRegisterDef(MachineBasicBlock *mbb, interval.addRange(LR); ValNo->addKill(indexes_->getTerminatorGap(mbb)); ValNo->setHasPHIKill(true); - DEBUG(dbgs() << " +" << LR); + DEBUG(dbgs() << " phi-join +" << LR); } } @@ -613,6 +595,9 @@ void LiveIntervals::handleLiveInRegister(MachineBasicBlock *MBB, while (mi != E) { if (mi->isDebugValue()) { ++mi; + if (mi != E && !mi->isDebugValue()) { + baseIndex = indexes_->getNextNonNullIndex(baseIndex); + } continue; } if (mi->killsRegister(interval.reg, tri_)) { @@ -1355,11 +1340,9 @@ rewriteInstructionsForSpills(const LiveInterval &li, bool TrySplit, MachineBasicBlock *MBB = MI->getParent(); if (ImpUse && MI != ReMatDefMI) { - // Re-matting an instruction with virtual register use. Update the - // register interval's spill weight to HUGE_VALF to prevent it from - // being spilled. - LiveInterval &ImpLi = getInterval(ImpUse); - ImpLi.weight = HUGE_VALF; + // Re-matting an instruction with virtual register use. Prevent interval + // from being spilled. + getInterval(ImpUse).markNotSpillable(); } unsigned MBBId = MBB->getNumber(); @@ -1411,7 +1394,7 @@ rewriteInstructionsForSpills(const LiveInterval &li, bool TrySplit, LiveInterval &nI = getOrCreateInterval(NewVReg); if (!TrySplit) { // The spill weight is now infinity as it cannot be spilled again. - nI.weight = HUGE_VALF; + nI.markNotSpillable(); continue; } @@ -1559,6 +1542,28 @@ LiveIntervals::handleSpilledImpDefs(const LiveInterval &li, VirtRegMap &vrm, } } +float +LiveIntervals::getSpillWeight(bool isDef, bool isUse, unsigned loopDepth) { + // Limit the loop depth ridiculousness. + if (loopDepth > 200) + loopDepth = 200; + + // The loop depth is used to roughly estimate the number of times the + // instruction is executed. Something like 10^d is simple, but will quickly + // overflow a float. This expression behaves like 10^d for small d, but is + // more tempered for large d. At d=200 we get 6.7e33 which leaves a bit of + // headroom before overflow. + float lc = powf(1 + (100.0f / (loopDepth+10)), (float)loopDepth); + + return (isDef + isUse) * lc; +} + +void +LiveIntervals::normalizeSpillWeights(std::vector<LiveInterval*> &NewLIs) { + for (unsigned i = 0, e = NewLIs.size(); i != e; ++i) + normalizeSpillWeight(*NewLIs[i]); +} + std::vector<LiveInterval*> LiveIntervals:: addIntervalsForSpillsFast(const LiveInterval &li, const MachineLoopInfo *loopInfo, @@ -1567,8 +1572,7 @@ addIntervalsForSpillsFast(const LiveInterval &li, std::vector<LiveInterval*> added; - assert(li.weight != HUGE_VALF && - "attempt to spill already spilled interval!"); + assert(li.isSpillable() && "attempt to spill already spilled interval!"); DEBUG({ dbgs() << "\t\t\t\tadding intervals for spills for interval: "; @@ -1604,10 +1608,7 @@ addIntervalsForSpillsFast(const LiveInterval &li, // create a new register for this spill LiveInterval &nI = getOrCreateInterval(NewVReg); - - // the spill weight is now infinity as it - // cannot be spilled again - nI.weight = HUGE_VALF; + nI.markNotSpillable(); // Rewrite register operands to use the new vreg. for (SmallVectorImpl<unsigned>::iterator I = Indices.begin(), @@ -1661,8 +1662,7 @@ addIntervalsForSpills(const LiveInterval &li, if (EnableFastSpilling) return addIntervalsForSpillsFast(li, loopInfo, vrm); - assert(li.weight != HUGE_VALF && - "attempt to spill already spilled interval!"); + assert(li.isSpillable() && "attempt to spill already spilled interval!"); DEBUG({ dbgs() << "\t\t\t\tadding intervals for spills for interval: "; @@ -1736,6 +1736,7 @@ addIntervalsForSpills(const LiveInterval &li, } handleSpilledImpDefs(li, vrm, rc, NewLIs); + normalizeSpillWeights(NewLIs); return NewLIs; } @@ -1811,6 +1812,7 @@ addIntervalsForSpills(const LiveInterval &li, // Insert spills / restores if we are splitting. if (!TrySplit) { handleSpilledImpDefs(li, vrm, rc, NewLIs); + normalizeSpillWeights(NewLIs); return NewLIs; } @@ -1927,11 +1929,10 @@ addIntervalsForSpills(const LiveInterval &li, unsigned ImpUse = getReMatImplicitUse(li, ReMatDefMI); if (ImpUse) { // Re-matting an instruction with virtual register use. Add the - // register as an implicit use on the use MI and update the register - // interval's spill weight to HUGE_VALF to prevent it from being - // spilled. + // register as an implicit use on the use MI and mark the register + // interval as unspillable. LiveInterval &ImpLi = getInterval(ImpUse); - ImpLi.weight = HUGE_VALF; + ImpLi.markNotSpillable(); MI->addOperand(MachineOperand::CreateReg(ImpUse, false, true)); } } @@ -1970,6 +1971,7 @@ addIntervalsForSpills(const LiveInterval &li, } handleSpilledImpDefs(li, vrm, rc, RetNewLIs); + normalizeSpillWeights(RetNewLIs); return RetNewLIs; } diff --git a/lib/CodeGen/LiveVariables.cpp b/lib/CodeGen/LiveVariables.cpp index 8a124dc..519990e 100644 --- a/lib/CodeGen/LiveVariables.cpp +++ b/lib/CodeGen/LiveVariables.cpp @@ -365,27 +365,7 @@ bool LiveVariables::HandlePhysRegKill(unsigned Reg, MachineInstr *MI) { } } - if (LastRefOrPartRef == PhysRegDef[Reg] && LastRefOrPartRef != MI) { - if (LastPartDef) - // The last partial def kills the register. - LastPartDef->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, - true/*IsImp*/, true/*IsKill*/)); - else { - MachineOperand *MO = - LastRefOrPartRef->findRegisterDefOperand(Reg, false, TRI); - bool NeedEC = MO->isEarlyClobber() && MO->getReg() != Reg; - // If the last reference is the last def, then it's not used at all. - // That is, unless we are currently processing the last reference itself. - LastRefOrPartRef->addRegisterDead(Reg, TRI, true); - if (NeedEC) { - // If we are adding a subreg def and the superreg def is marked early - // clobber, add an early clobber marker to the subreg def. - MO = LastRefOrPartRef->findRegisterDefOperand(Reg); - if (MO) - MO->setIsEarlyClobber(); - } - } - } else if (!PhysRegUse[Reg]) { + if (!PhysRegUse[Reg]) { // Partial uses. Mark register def dead and add implicit def of // sub-registers which are used. // EAX<dead> = op AL<imp-def> @@ -419,6 +399,26 @@ bool LiveVariables::HandlePhysRegKill(unsigned Reg, MachineInstr *MI) { for (const unsigned *SS = TRI->getSubRegisters(SubReg); *SS; ++SS) PartUses.erase(*SS); } + } else if (LastRefOrPartRef == PhysRegDef[Reg] && LastRefOrPartRef != MI) { + if (LastPartDef) + // The last partial def kills the register. + LastPartDef->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, + true/*IsImp*/, true/*IsKill*/)); + else { + MachineOperand *MO = + LastRefOrPartRef->findRegisterDefOperand(Reg, false, TRI); + bool NeedEC = MO->isEarlyClobber() && MO->getReg() != Reg; + // If the last reference is the last def, then it's not used at all. + // That is, unless we are currently processing the last reference itself. + LastRefOrPartRef->addRegisterDead(Reg, TRI, true); + if (NeedEC) { + // If we are adding a subreg def and the superreg def is marked early + // clobber, add an early clobber marker to the subreg def. + MO = LastRefOrPartRef->findRegisterDefOperand(Reg); + if (MO) + MO->setIsEarlyClobber(); + } + } } else LastRefOrPartRef->addRegisterKilled(Reg, TRI, true); return true; @@ -510,6 +510,7 @@ bool LiveVariables::runOnMachineFunction(MachineFunction &mf) { PHIVarInfo = new SmallVector<unsigned, 4>[MF->getNumBlockIDs()]; std::fill(PhysRegDef, PhysRegDef + NumRegs, (MachineInstr*)0); std::fill(PhysRegUse, PhysRegUse + NumRegs, (MachineInstr*)0); + PHIJoins.clear(); /// Get some space for a respectable number of registers. VirtRegInfo.resize(64); diff --git a/lib/CodeGen/MachineBasicBlock.cpp b/lib/CodeGen/MachineBasicBlock.cpp index 655a0bf..64134ce 100644 --- a/lib/CodeGen/MachineBasicBlock.cpp +++ b/lib/CodeGen/MachineBasicBlock.cpp @@ -143,36 +143,6 @@ MachineBasicBlock::iterator MachineBasicBlock::getFirstTerminator() { return I; } -/// isOnlyReachableViaFallthough - Return true if this basic block has -/// exactly one predecessor and the control transfer mechanism between -/// the predecessor and this block is a fall-through. -bool MachineBasicBlock::isOnlyReachableByFallthrough() const { - // If this is a landing pad, it isn't a fall through. If it has no preds, - // then nothing falls through to it. - if (isLandingPad() || pred_empty()) - return false; - - // If there isn't exactly one predecessor, it can't be a fall through. - const_pred_iterator PI = pred_begin(), PI2 = PI; - ++PI2; - if (PI2 != pred_end()) - return false; - - // The predecessor has to be immediately before this block. - const MachineBasicBlock *Pred = *PI; - - if (!Pred->isLayoutSuccessor(this)) - return false; - - // If the block is completely empty, then it definitely does fall through. - if (Pred->empty()) - return true; - - // Otherwise, check the last instruction. - const MachineInstr &LastInst = Pred->back(); - return !LastInst.getDesc().isBarrier(); -} - void MachineBasicBlock::dump() const { print(dbgs()); } diff --git a/lib/CodeGen/MachineCSE.cpp b/lib/CodeGen/MachineCSE.cpp new file mode 100644 index 0000000..b376e3d --- /dev/null +++ b/lib/CodeGen/MachineCSE.cpp @@ -0,0 +1,268 @@ +//===-- MachineCSE.cpp - Machine Common Subexpression Elimination Pass ----===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This pass performs global common subexpression elimination on machine +// instructions using a scoped hash table based value numbering scheme. It +// must be run while the machine function is still in SSA form. +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "machine-cse" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/MachineDominators.h" +#include "llvm/CodeGen/MachineInstr.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/Analysis/AliasAnalysis.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/ADT/ScopedHashTable.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Support/Debug.h" + +using namespace llvm; + +STATISTIC(NumCoalesces, "Number of copies coalesced"); +STATISTIC(NumCSEs, "Number of common subexpression eliminated"); + +namespace { + class MachineCSE : public MachineFunctionPass { + const TargetInstrInfo *TII; + const TargetRegisterInfo *TRI; + MachineRegisterInfo *MRI; + MachineDominatorTree *DT; + AliasAnalysis *AA; + public: + static char ID; // Pass identification + MachineCSE() : MachineFunctionPass(&ID), CurrVN(0) {} + + virtual bool runOnMachineFunction(MachineFunction &MF); + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.setPreservesCFG(); + MachineFunctionPass::getAnalysisUsage(AU); + AU.addRequired<AliasAnalysis>(); + AU.addRequired<MachineDominatorTree>(); + AU.addPreserved<MachineDominatorTree>(); + } + + private: + unsigned CurrVN; + ScopedHashTable<MachineInstr*, unsigned, MachineInstrExpressionTrait> VNT; + SmallVector<MachineInstr*, 64> Exps; + + bool PerformTrivialCoalescing(MachineInstr *MI, MachineBasicBlock *MBB); + bool isPhysDefTriviallyDead(unsigned Reg, + MachineBasicBlock::const_iterator I, + MachineBasicBlock::const_iterator E); + bool hasLivePhysRegDefUse(MachineInstr *MI, MachineBasicBlock *MBB); + bool isCSECandidate(MachineInstr *MI); + bool ProcessBlock(MachineDomTreeNode *Node); + }; +} // end anonymous namespace + +char MachineCSE::ID = 0; +static RegisterPass<MachineCSE> +X("machine-cse", "Machine Common Subexpression Elimination"); + +FunctionPass *llvm::createMachineCSEPass() { return new MachineCSE(); } + +bool MachineCSE::PerformTrivialCoalescing(MachineInstr *MI, + MachineBasicBlock *MBB) { + bool Changed = false; + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || !MO.isUse()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg || TargetRegisterInfo::isPhysicalRegister(Reg)) + continue; + if (!MRI->hasOneUse(Reg)) + // Only coalesce single use copies. This ensure the copy will be + // deleted. + continue; + MachineInstr *DefMI = MRI->getVRegDef(Reg); + if (DefMI->getParent() != MBB) + continue; + unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx; + if (TII->isMoveInstr(*DefMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) && + TargetRegisterInfo::isVirtualRegister(SrcReg) && + !SrcSubIdx && !DstSubIdx) { + MO.setReg(SrcReg); + DefMI->eraseFromParent(); + ++NumCoalesces; + Changed = true; + } + } + + return Changed; +} + +bool MachineCSE::isPhysDefTriviallyDead(unsigned Reg, + MachineBasicBlock::const_iterator I, + MachineBasicBlock::const_iterator E) { + unsigned LookAheadLeft = 5; + while (LookAheadLeft--) { + if (I == E) + // Reached end of block, register is obviously dead. + return true; + + if (I->isDebugValue()) + continue; + bool SeenDef = false; + for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = I->getOperand(i); + if (!MO.isReg() || !MO.getReg()) + continue; + if (!TRI->regsOverlap(MO.getReg(), Reg)) + continue; + if (MO.isUse()) + return false; + SeenDef = true; + } + if (SeenDef) + // See a def of Reg (or an alias) before encountering any use, it's + // trivially dead. + return true; + ++I; + } + return false; +} + +bool MachineCSE::hasLivePhysRegDefUse(MachineInstr *MI, MachineBasicBlock *MBB){ + unsigned PhysDef = 0; + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg()) + continue; + unsigned Reg = MO.getReg(); + if (!Reg) + continue; + if (TargetRegisterInfo::isPhysicalRegister(Reg)) { + if (MO.isUse()) + // Can't touch anything to read a physical register. + return true; + if (MO.isDead()) + // If the def is dead, it's ok. + continue; + // Ok, this is a physical register def that's not marked "dead". That's + // common since this pass is run before livevariables. We can scan + // forward a few instructions and check if it is obviously dead. + if (PhysDef) + // Multiple physical register defs. These are rare, forget about it. + return true; + PhysDef = Reg; + } + } + + if (PhysDef) { + MachineBasicBlock::iterator I = MI; I = llvm::next(I); + if (!isPhysDefTriviallyDead(PhysDef, I, MBB->end())) + return true; + } + return false; +} + +bool MachineCSE::isCSECandidate(MachineInstr *MI) { + // Ignore copies or instructions that read / write physical registers + // (except for dead defs of physical registers). + unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx; + if (TII->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx) || + MI->isExtractSubreg() || MI->isInsertSubreg() || MI->isSubregToReg()) + return false; + + // Ignore stuff that we obviously can't move. + const TargetInstrDesc &TID = MI->getDesc(); + if (TID.mayStore() || TID.isCall() || TID.isTerminator() || + TID.hasUnmodeledSideEffects()) + return false; + + if (TID.mayLoad()) { + // Okay, this instruction does a load. As a refinement, we allow the target + // to decide whether the loaded value is actually a constant. If so, we can + // actually use it as a load. + if (!MI->isInvariantLoad(AA)) + // FIXME: we should be able to hoist loads with no other side effects if + // there are no other instructions which can change memory in this loop. + // This is a trivial form of alias analysis. + return false; + } + return true; +} + +bool MachineCSE::ProcessBlock(MachineDomTreeNode *Node) { + bool Changed = false; + + ScopedHashTableScope<MachineInstr*, unsigned, + MachineInstrExpressionTrait> VNTS(VNT); + MachineBasicBlock *MBB = Node->getBlock(); + for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E; ) { + MachineInstr *MI = &*I; + ++I; + + if (!isCSECandidate(MI)) + continue; + + bool FoundCSE = VNT.count(MI); + if (!FoundCSE) { + // Look for trivial copy coalescing opportunities. + if (PerformTrivialCoalescing(MI, MBB)) + FoundCSE = VNT.count(MI); + } + // FIXME: commute commutable instructions? + + // If the instruction defines a physical register and the value *may* be + // used, then it's not safe to replace it with a common subexpression. + if (FoundCSE && hasLivePhysRegDefUse(MI, MBB)) + FoundCSE = false; + + if (!FoundCSE) { + VNT.insert(MI, CurrVN++); + Exps.push_back(MI); + continue; + } + + // Found a common subexpression, eliminate it. + unsigned CSVN = VNT.lookup(MI); + MachineInstr *CSMI = Exps[CSVN]; + DEBUG(dbgs() << "Examining: " << *MI); + DEBUG(dbgs() << "*** Found a common subexpression: " << *CSMI); + unsigned NumDefs = MI->getDesc().getNumDefs(); + for (unsigned i = 0, e = MI->getNumOperands(); NumDefs && i != e; ++i) { + MachineOperand &MO = MI->getOperand(i); + if (!MO.isReg() || !MO.isDef()) + continue; + unsigned OldReg = MO.getReg(); + unsigned NewReg = CSMI->getOperand(i).getReg(); + if (OldReg == NewReg) + continue; + assert(TargetRegisterInfo::isVirtualRegister(OldReg) && + TargetRegisterInfo::isVirtualRegister(NewReg) && + "Do not CSE physical register defs!"); + MRI->replaceRegWith(OldReg, NewReg); + --NumDefs; + } + MI->eraseFromParent(); + ++NumCSEs; + } + + // Recursively call ProcessBlock with childred. + const std::vector<MachineDomTreeNode*> &Children = Node->getChildren(); + for (unsigned i = 0, e = Children.size(); i != e; ++i) + Changed |= ProcessBlock(Children[i]); + + return Changed; +} + +bool MachineCSE::runOnMachineFunction(MachineFunction &MF) { + TII = MF.getTarget().getInstrInfo(); + TRI = MF.getTarget().getRegisterInfo(); + MRI = &MF.getRegInfo(); + DT = &getAnalysis<MachineDominatorTree>(); + AA = &getAnalysis<AliasAnalysis>(); + return ProcessBlock(DT->getRootNode()); +} diff --git a/lib/CodeGen/MachineFunction.cpp b/lib/CodeGen/MachineFunction.cpp index f141c56..4377d5b 100644 --- a/lib/CodeGen/MachineFunction.cpp +++ b/lib/CodeGen/MachineFunction.cpp @@ -95,6 +95,9 @@ MachineFunction::MachineFunction(Function *F, const TargetMachine &TM, MFInfo = 0; FrameInfo = new (Allocator.Allocate<MachineFrameInfo>()) MachineFrameInfo(*TM.getFrameInfo()); + if (Fn->hasFnAttr(Attribute::StackAlignment)) + FrameInfo->setMaxAlignment(Attribute::getStackAlignmentFromAttrs( + Fn->getAttributes().getFnAttributes())); ConstantPool = new (Allocator.Allocate<MachineConstantPool>()) MachineConstantPool(TM.getTargetData()); Alignment = TM.getTargetLowering()->getFunctionAlignment(F); diff --git a/lib/CodeGen/MachineInstr.cpp b/lib/CodeGen/MachineInstr.cpp index df61c74..e23670d 100644 --- a/lib/CodeGen/MachineInstr.cpp +++ b/lib/CodeGen/MachineInstr.cpp @@ -18,6 +18,7 @@ #include "llvm/Type.h" #include "llvm/Value.h" #include "llvm/Assembly/Writer.h" +#include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineMemOperand.h" #include "llvm/CodeGen/MachineRegisterInfo.h" @@ -305,7 +306,7 @@ void MachineOperand::print(raw_ostream &OS, const TargetMachine *TM) const { MachineMemOperand::MachineMemOperand(const Value *v, unsigned int f, int64_t o, uint64_t s, unsigned int a) : Offset(o), Size(s), V(v), - Flags((f & 7) | ((Log2_32(a) + 1) << 3)) { + Flags((f & ((1 << MOMaxBits) - 1)) | ((Log2_32(a) + 1) << MOMaxBits)) { assert(getBaseAlignment() == a && "Alignment is not a power of 2!"); assert((isLoad() || isStore()) && "Not a load/store!"); } @@ -327,7 +328,8 @@ void MachineMemOperand::refineAlignment(const MachineMemOperand *MMO) { if (MMO->getBaseAlignment() >= getBaseAlignment()) { // Update the alignment value. - Flags = (Flags & 7) | ((Log2_32(MMO->getBaseAlignment()) + 1) << 3); + Flags = (Flags & ((1 << MOMaxBits) - 1)) | + ((Log2_32(MMO->getBaseAlignment()) + 1) << MOMaxBits); // Also update the base and offset, because the new alignment may // not be applicable with the old ones. V = MMO->getValue(); @@ -700,6 +702,35 @@ void MachineInstr::addMemOperand(MachineFunction &MF, MemRefsEnd = NewMemRefsEnd; } +bool MachineInstr::isIdenticalTo(const MachineInstr *Other, + MICheckType Check) const { + // If opcodes or number of operands are not the same then the two + // instructions are obviously not identical. + if (Other->getOpcode() != getOpcode() || + Other->getNumOperands() != getNumOperands()) + return false; + + // Check operands to make sure they match. + for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { + const MachineOperand &MO = getOperand(i); + const MachineOperand &OMO = Other->getOperand(i); + // Clients may or may not want to ignore defs when testing for equality. + // For example, machine CSE pass only cares about finding common + // subexpressions, so it's safe to ignore virtual register defs. + if (Check != CheckDefs && MO.isReg() && MO.isDef()) { + if (Check == IgnoreDefs) + continue; + // Check == IgnoreVRegDefs + if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()) || + TargetRegisterInfo::isPhysicalRegister(OMO.getReg())) + if (MO.getReg() != OMO.getReg()) + return false; + } else if (!MO.isIdenticalTo(OMO)) + return false; + } + return true; +} + /// removeFromParent - This method unlinks 'this' from the containing basic /// block, and returns it, but does not delete it. MachineInstr *MachineInstr::removeFromParent() { @@ -958,8 +989,8 @@ void MachineInstr::copyPredicates(const MachineInstr *MI) { /// SawStore is set to true, it means that there is a store (or call) between /// the instruction's location and its intended destination. bool MachineInstr::isSafeToMove(const TargetInstrInfo *TII, - bool &SawStore, - AliasAnalysis *AA) const { + AliasAnalysis *AA, + bool &SawStore) const { // Ignore stuff that we obviously can't move. if (TID->mayStore() || TID->isCall()) { SawStore = true; @@ -984,11 +1015,11 @@ bool MachineInstr::isSafeToMove(const TargetInstrInfo *TII, /// isSafeToReMat - Return true if it's safe to rematerialize the specified /// instruction which defined the specified register instead of copying it. bool MachineInstr::isSafeToReMat(const TargetInstrInfo *TII, - unsigned DstReg, - AliasAnalysis *AA) const { + AliasAnalysis *AA, + unsigned DstReg) const { bool SawStore = false; if (!TII->isTriviallyReMaterializable(this, AA) || - !isSafeToMove(TII, SawStore, AA)) + !isSafeToMove(TII, AA, SawStore)) return false; for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { const MachineOperand &MO = getOperand(i); @@ -1324,3 +1355,48 @@ void MachineInstr::addRegisterDefined(unsigned IncomingReg, true /*IsDef*/, true /*IsImp*/)); } + +unsigned +MachineInstrExpressionTrait::getHashValue(const MachineInstr* const &MI) { + unsigned Hash = MI->getOpcode() * 37; + for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { + const MachineOperand &MO = MI->getOperand(i); + uint64_t Key = (uint64_t)MO.getType() << 32; + switch (MO.getType()) { + default: break; + case MachineOperand::MO_Register: + if (MO.isDef() && MO.getReg() && + TargetRegisterInfo::isVirtualRegister(MO.getReg())) + continue; // Skip virtual register defs. + Key |= MO.getReg(); + break; + case MachineOperand::MO_Immediate: + Key |= MO.getImm(); + break; + case MachineOperand::MO_FrameIndex: + case MachineOperand::MO_ConstantPoolIndex: + case MachineOperand::MO_JumpTableIndex: + Key |= MO.getIndex(); + break; + case MachineOperand::MO_MachineBasicBlock: + Key |= DenseMapInfo<void*>::getHashValue(MO.getMBB()); + break; + case MachineOperand::MO_GlobalAddress: + Key |= DenseMapInfo<void*>::getHashValue(MO.getGlobal()); + break; + case MachineOperand::MO_BlockAddress: + Key |= DenseMapInfo<void*>::getHashValue(MO.getBlockAddress()); + break; + } + Key += ~(Key << 32); + Key ^= (Key >> 22); + Key += ~(Key << 13); + Key ^= (Key >> 8); + Key += (Key << 3); + Key ^= (Key >> 15); + Key += ~(Key << 27); + Key ^= (Key >> 31); + Hash = (unsigned)Key + Hash * 37; + } + return Hash; +} diff --git a/lib/CodeGen/MachineLICM.cpp b/lib/CodeGen/MachineLICM.cpp index 92c84f3..0361694 100644 --- a/lib/CodeGen/MachineLICM.cpp +++ b/lib/CodeGen/MachineLICM.cpp @@ -252,32 +252,6 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { return false; } - DEBUG({ - dbgs() << "--- Checking if we can hoist " << I; - if (I.getDesc().getImplicitUses()) { - dbgs() << " * Instruction has implicit uses:\n"; - - const TargetRegisterInfo *TRI = TM->getRegisterInfo(); - for (const unsigned *ImpUses = I.getDesc().getImplicitUses(); - *ImpUses; ++ImpUses) - dbgs() << " -> " << TRI->getName(*ImpUses) << "\n"; - } - - if (I.getDesc().getImplicitDefs()) { - dbgs() << " * Instruction has implicit defines:\n"; - - const TargetRegisterInfo *TRI = TM->getRegisterInfo(); - for (const unsigned *ImpDefs = I.getDesc().getImplicitDefs(); - *ImpDefs; ++ImpDefs) - dbgs() << " -> " << TRI->getName(*ImpDefs) << "\n"; - } - }); - - if (I.getDesc().getImplicitDefs() || I.getDesc().getImplicitUses()) { - DEBUG(dbgs() << "Cannot hoist with implicit defines or uses\n"); - return false; - } - // The instruction is loop invariant if all of its operands are. for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { const MachineOperand &MO = I.getOperand(i); @@ -311,6 +285,10 @@ bool MachineLICM::IsLoopInvariantInst(MachineInstr &I) { } else if (!MO.isDead()) { // A def that isn't dead. We can't move it. return false; + } else if (CurLoop->getHeader()->isLiveIn(Reg)) { + // If the reg is live into the loop, we can't hoist an instruction + // which would clobber it. + return false; } } @@ -467,7 +445,7 @@ MachineLICM::LookForDuplicate(const MachineInstr *MI, std::vector<const MachineInstr*> &PrevMIs) { for (unsigned i = 0, e = PrevMIs.size(); i != e; ++i) { const MachineInstr *PrevMI = PrevMIs[i]; - if (TII->isIdentical(MI, PrevMI, RegInfo)) + if (TII->produceSameValue(MI, PrevMI)) return PrevMI; } return 0; @@ -480,9 +458,20 @@ bool MachineLICM::EliminateCSE(MachineInstr *MI, if (const MachineInstr *Dup = LookForDuplicate(MI, CI->second)) { DEBUG(dbgs() << "CSEing " << *MI << " with " << *Dup); + + // Replace virtual registers defined by MI by their counterparts defined + // by Dup. for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { const MachineOperand &MO = MI->getOperand(i); - if (MO.isReg() && MO.isDef()) + + // Physical registers may not differ here. + assert((!MO.isReg() || MO.getReg() == 0 || + !TargetRegisterInfo::isPhysicalRegister(MO.getReg()) || + MO.getReg() == Dup->getOperand(i).getReg()) && + "Instructions with different phys regs are not identical!"); + + if (MO.isReg() && MO.isDef() && + !TargetRegisterInfo::isPhysicalRegister(MO.getReg())) RegInfo->replaceRegWith(MO.getReg(), Dup->getOperand(i).getReg()); } MI->eraseFromParent(); diff --git a/lib/CodeGen/MachineModuleInfoImpls.cpp b/lib/CodeGen/MachineModuleInfoImpls.cpp index 8378906..39d2c75 100644 --- a/lib/CodeGen/MachineModuleInfoImpls.cpp +++ b/lib/CodeGen/MachineModuleInfoImpls.cpp @@ -22,7 +22,7 @@ using namespace llvm; // Out of line virtual method. void MachineModuleInfoMachO::Anchor() {} - +void MachineModuleInfoELF::Anchor() {} static int SortSymbolPair(const void *LHS, const void *RHS) { const MCSymbol *LHSS = @@ -34,10 +34,11 @@ static int SortSymbolPair(const void *LHS, const void *RHS) { /// GetSortedStubs - Return the entries from a DenseMap in a deterministic /// sorted orer. -MachineModuleInfoMachO::SymbolListTy -MachineModuleInfoMachO::GetSortedStubs(const DenseMap<MCSymbol*, - MCSymbol*> &Map) { - MachineModuleInfoMachO::SymbolListTy List(Map.begin(), Map.end()); +MachineModuleInfoImpl::SymbolListTy +MachineModuleInfoImpl::GetSortedStubs(const DenseMap<MCSymbol*, + MCSymbol*> &Map) { + MachineModuleInfoImpl::SymbolListTy List(Map.begin(), Map.end()); + if (!List.empty()) qsort(&List[0], List.size(), sizeof(List[0]), SortSymbolPair); return List; diff --git a/lib/CodeGen/MachineRegisterInfo.cpp b/lib/CodeGen/MachineRegisterInfo.cpp index b31973e..d9ab677 100644 --- a/lib/CodeGen/MachineRegisterInfo.cpp +++ b/lib/CodeGen/MachineRegisterInfo.cpp @@ -116,6 +116,19 @@ MachineInstr *MachineRegisterInfo::getVRegDef(unsigned Reg) const { return 0; } +bool MachineRegisterInfo::hasOneUse(unsigned RegNo) const { + use_iterator UI = use_begin(RegNo); + if (UI == use_end()) + return false; + return ++UI == use_end(); +} + +bool MachineRegisterInfo::hasOneNonDBGUse(unsigned RegNo) const { + use_nodbg_iterator UI = use_nodbg_begin(RegNo); + if (UI == use_nodbg_end()) + return false; + return ++UI == use_nodbg_end(); +} #ifndef NDEBUG void MachineRegisterInfo::dumpUses(unsigned Reg) const { diff --git a/lib/CodeGen/MachineSink.cpp b/lib/CodeGen/MachineSink.cpp index c391576..e47ba7c 100644 --- a/lib/CodeGen/MachineSink.cpp +++ b/lib/CodeGen/MachineSink.cpp @@ -72,8 +72,13 @@ bool MachineSinking::AllUsesDominatedByBlock(unsigned Reg, MachineBasicBlock *MBB) const { assert(TargetRegisterInfo::isVirtualRegister(Reg) && "Only makes sense for vregs"); - for (MachineRegisterInfo::use_iterator I = RegInfo->use_begin(Reg), - E = RegInfo->use_end(); I != E; ++I) { + // Ignoring debug uses is necessary so debug info doesn't affect the code. + // This may leave a referencing dbg_value in the original block, before + // the definition of the vreg. Dwarf generator handles this although the + // user might not get the right info at runtime. + for (MachineRegisterInfo::use_nodbg_iterator I = + RegInfo->use_nodbg_begin(Reg), + E = RegInfo->use_nodbg_end(); I != E; ++I) { // Determine the block of the use. MachineInstr *UseInst = &*I; MachineBasicBlock *UseBlock = UseInst->getParent(); @@ -135,7 +140,10 @@ bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) { ProcessedBegin = I == MBB.begin(); if (!ProcessedBegin) --I; - + + if (MI->isDebugValue()) + continue; + if (SinkInstruction(MI, SawStore)) ++NumSunk, MadeChange = true; @@ -149,7 +157,7 @@ bool MachineSinking::ProcessBlock(MachineBasicBlock &MBB) { /// instruction out of its current block into a successor. bool MachineSinking::SinkInstruction(MachineInstr *MI, bool &SawStore) { // Check if it's safe to move the instruction. - if (!MI->isSafeToMove(TII, SawStore, AA)) + if (!MI->isSafeToMove(TII, AA, SawStore)) return false; // FIXME: This should include support for sinking instructions within the diff --git a/lib/CodeGen/OptimizePHIs.cpp b/lib/CodeGen/OptimizePHIs.cpp new file mode 100644 index 0000000..2717d4d --- /dev/null +++ b/lib/CodeGen/OptimizePHIs.cpp @@ -0,0 +1,189 @@ +//===-- OptimizePHIs.cpp - Optimize machine instruction PHIs --------------===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This pass optimizes machine instruction PHIs to take advantage of +// opportunities created during DAG legalization. +// +//===----------------------------------------------------------------------===// + +#define DEBUG_TYPE "phi-opt" +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineInstr.h" +#include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Function.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/Statistic.h" +using namespace llvm; + +STATISTIC(NumPHICycles, "Number of PHI cycles replaced"); +STATISTIC(NumDeadPHICycles, "Number of dead PHI cycles"); + +namespace { + class OptimizePHIs : public MachineFunctionPass { + MachineRegisterInfo *MRI; + const TargetInstrInfo *TII; + + public: + static char ID; // Pass identification + OptimizePHIs() : MachineFunctionPass(&ID) {} + + virtual bool runOnMachineFunction(MachineFunction &MF); + + virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.setPreservesCFG(); + MachineFunctionPass::getAnalysisUsage(AU); + } + + private: + typedef SmallPtrSet<MachineInstr*, 16> InstrSet; + typedef SmallPtrSetIterator<MachineInstr*> InstrSetIterator; + + bool IsSingleValuePHICycle(MachineInstr *MI, unsigned &SingleValReg, + InstrSet &PHIsInCycle); + bool IsDeadPHICycle(MachineInstr *MI, InstrSet &PHIsInCycle); + bool OptimizeBB(MachineBasicBlock &MBB); + }; +} + +char OptimizePHIs::ID = 0; +static RegisterPass<OptimizePHIs> +X("opt-phis", "Optimize machine instruction PHIs"); + +FunctionPass *llvm::createOptimizePHIsPass() { return new OptimizePHIs(); } + +bool OptimizePHIs::runOnMachineFunction(MachineFunction &Fn) { + MRI = &Fn.getRegInfo(); + TII = Fn.getTarget().getInstrInfo(); + + // Find dead PHI cycles and PHI cycles that can be replaced by a single + // value. InstCombine does these optimizations, but DAG legalization may + // introduce new opportunities, e.g., when i64 values are split up for + // 32-bit targets. + bool Changed = false; + for (MachineFunction::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I) + Changed |= OptimizeBB(*I); + + return Changed; +} + +/// IsSingleValuePHICycle - Check if MI is a PHI where all the source operands +/// are copies of SingleValReg, possibly via copies through other PHIs. If +/// SingleValReg is zero on entry, it is set to the register with the single +/// non-copy value. PHIsInCycle is a set used to keep track of the PHIs that +/// have been scanned. +bool OptimizePHIs::IsSingleValuePHICycle(MachineInstr *MI, + unsigned &SingleValReg, + InstrSet &PHIsInCycle) { + assert(MI->isPHI() && "IsSingleValuePHICycle expects a PHI instruction"); + unsigned DstReg = MI->getOperand(0).getReg(); + + // See if we already saw this register. + if (!PHIsInCycle.insert(MI)) + return true; + + // Don't scan crazily complex things. + if (PHIsInCycle.size() == 16) + return false; + + // Scan the PHI operands. + for (unsigned i = 1; i != MI->getNumOperands(); i += 2) { + unsigned SrcReg = MI->getOperand(i).getReg(); + if (SrcReg == DstReg) + continue; + MachineInstr *SrcMI = MRI->getVRegDef(SrcReg); + + // Skip over register-to-register moves. + unsigned MvSrcReg, MvDstReg, SrcSubIdx, DstSubIdx; + if (SrcMI && + TII->isMoveInstr(*SrcMI, MvSrcReg, MvDstReg, SrcSubIdx, DstSubIdx) && + SrcSubIdx == 0 && DstSubIdx == 0 && + TargetRegisterInfo::isVirtualRegister(MvSrcReg)) + SrcMI = MRI->getVRegDef(MvSrcReg); + if (!SrcMI) + return false; + + if (SrcMI->isPHI()) { + if (!IsSingleValuePHICycle(SrcMI, SingleValReg, PHIsInCycle)) + return false; + } else { + // Fail if there is more than one non-phi/non-move register. + if (SingleValReg != 0) + return false; + SingleValReg = SrcReg; + } + } + return true; +} + +/// IsDeadPHICycle - Check if the register defined by a PHI is only used by +/// other PHIs in a cycle. +bool OptimizePHIs::IsDeadPHICycle(MachineInstr *MI, InstrSet &PHIsInCycle) { + assert(MI->isPHI() && "IsDeadPHICycle expects a PHI instruction"); + unsigned DstReg = MI->getOperand(0).getReg(); + assert(TargetRegisterInfo::isVirtualRegister(DstReg) && + "PHI destination is not a virtual register"); + + // See if we already saw this register. + if (!PHIsInCycle.insert(MI)) + return true; + + // Don't scan crazily complex things. + if (PHIsInCycle.size() == 16) + return false; + + for (MachineRegisterInfo::use_iterator I = MRI->use_begin(DstReg), + E = MRI->use_end(); I != E; ++I) { + MachineInstr *UseMI = &*I; + if (!UseMI->isPHI() || !IsDeadPHICycle(UseMI, PHIsInCycle)) + return false; + } + + return true; +} + +/// OptimizeBB - Remove dead PHI cycles and PHI cycles that can be replaced by +/// a single value. +bool OptimizePHIs::OptimizeBB(MachineBasicBlock &MBB) { + bool Changed = false; + for (MachineBasicBlock::iterator + MII = MBB.begin(), E = MBB.end(); MII != E; ) { + MachineInstr *MI = &*MII++; + if (!MI->isPHI()) + break; + + // Check for single-value PHI cycles. + unsigned SingleValReg = 0; + InstrSet PHIsInCycle; + if (IsSingleValuePHICycle(MI, SingleValReg, PHIsInCycle) && + SingleValReg != 0) { + MRI->replaceRegWith(MI->getOperand(0).getReg(), SingleValReg); + MI->eraseFromParent(); + ++NumPHICycles; + Changed = true; + continue; + } + + // Check for dead PHI cycles. + PHIsInCycle.clear(); + if (IsDeadPHICycle(MI, PHIsInCycle)) { + for (InstrSetIterator PI = PHIsInCycle.begin(), PE = PHIsInCycle.end(); + PI != PE; ++PI) { + MachineInstr *PhiMI = *PI; + if (&*MII == PhiMI) + ++MII; + PhiMI->eraseFromParent(); + } + ++NumDeadPHICycles; + Changed = true; + } + } + return Changed; +} diff --git a/lib/CodeGen/PBQP/HeuristicSolver.h b/lib/CodeGen/PBQP/HeuristicSolver.h index b48f548..bd18b52 100644 --- a/lib/CodeGen/PBQP/HeuristicSolver.h +++ b/lib/CodeGen/PBQP/HeuristicSolver.h @@ -18,7 +18,6 @@ #include "Graph.h" #include "Solution.h" -#include "llvm/Support/raw_ostream.h" #include <vector> #include <limits> @@ -230,7 +229,7 @@ namespace PBQP { } /// \brief Apply rule R1. - /// @param nItr Node iterator for node to apply R1 to. + /// @param xnItr Node iterator for node to apply R1 to. /// /// Node will be automatically pushed to the solver stack. void applyR1(Graph::NodeItr xnItr) { @@ -278,7 +277,7 @@ namespace PBQP { } /// \brief Apply rule R2. - /// @param nItr Node iterator for node to apply R2 to. + /// @param xnItr Node iterator for node to apply R2 to. /// /// Node will be automatically pushed to the solver stack. void applyR2(Graph::NodeItr xnItr) { @@ -494,14 +493,23 @@ namespace PBQP { bool tryNormaliseEdgeMatrix(Graph::EdgeItr &eItr) { + const PBQPNum infinity = std::numeric_limits<PBQPNum>::infinity(); + Matrix &edgeCosts = g.getEdgeCosts(eItr); Vector &uCosts = g.getNodeCosts(g.getEdgeNode1(eItr)), &vCosts = g.getNodeCosts(g.getEdgeNode2(eItr)); for (unsigned r = 0; r < edgeCosts.getRows(); ++r) { - PBQPNum rowMin = edgeCosts.getRowMin(r); + PBQPNum rowMin = infinity; + + for (unsigned c = 0; c < edgeCosts.getCols(); ++c) { + if (vCosts[c] != infinity && edgeCosts[r][c] < rowMin) + rowMin = edgeCosts[r][c]; + } + uCosts[r] += rowMin; - if (rowMin != std::numeric_limits<PBQPNum>::infinity()) { + + if (rowMin != infinity) { edgeCosts.subFromRow(r, rowMin); } else { @@ -510,9 +518,16 @@ namespace PBQP { } for (unsigned c = 0; c < edgeCosts.getCols(); ++c) { - PBQPNum colMin = edgeCosts.getColMin(c); + PBQPNum colMin = infinity; + + for (unsigned r = 0; r < edgeCosts.getRows(); ++r) { + if (uCosts[r] != infinity && edgeCosts[r][c] < colMin) + colMin = edgeCosts[r][c]; + } + vCosts[c] += colMin; - if (colMin != std::numeric_limits<PBQPNum>::infinity()) { + + if (colMin != infinity) { edgeCosts.subFromCol(c, colMin); } else { diff --git a/lib/CodeGen/PBQP/Heuristics/Briggs.h b/lib/CodeGen/PBQP/Heuristics/Briggs.h index c09ad74..30d34d9 100644 --- a/lib/CodeGen/PBQP/Heuristics/Briggs.h +++ b/lib/CodeGen/PBQP/Heuristics/Briggs.h @@ -128,14 +128,7 @@ namespace PBQP { /// selected for heuristic reduction instead. bool shouldOptimallyReduce(Graph::NodeItr nItr) { if (getSolver().getSolverDegree(nItr) < 3) { - if (getGraph().getNodeCosts(nItr)[0] != - std::numeric_limits<PBQPNum>::infinity()) { - return true; - } - // Otherwise we have an infinite spill cost node. - initializeNode(nItr); - NodeData &nd = getHeuristicNodeData(nItr); - return nd.isAllocable; + return true; } // else return false; diff --git a/lib/CodeGen/PHIElimination.cpp b/lib/CodeGen/PHIElimination.cpp index b740c68..8bbe0a7 100644 --- a/lib/CodeGen/PHIElimination.cpp +++ b/lib/CodeGen/PHIElimination.cpp @@ -55,8 +55,6 @@ void llvm::PHIElimination::getAnalysisUsage(AnalysisUsage &AU) const { bool llvm::PHIElimination::runOnMachineFunction(MachineFunction &Fn) { MRI = &Fn.getRegInfo(); - PHIDefs.clear(); - PHIKills.clear(); bool Changed = false; // Split critical edges to help the coalescer @@ -215,10 +213,6 @@ void llvm::PHIElimination::LowerAtomicPHINode( TII->copyRegToReg(MBB, AfterPHIsIt, DestReg, IncomingReg, RC, RC); } - // Record PHI def. - assert(!hasPHIDef(DestReg) && "Vreg has multiple phi-defs?"); - PHIDefs[DestReg] = &MBB; - // Update live variable information if there is any. LiveVariables *LV = getAnalysisIfAvailable<LiveVariables>(); if (LV) { @@ -229,6 +223,7 @@ void llvm::PHIElimination::LowerAtomicPHINode( // Increment use count of the newly created virtual register. VI.NumUses++; + LV->setPHIJoin(IncomingReg); // When we are reusing the incoming register, it may already have been // killed in this block. The old kill will also have been inserted at @@ -276,9 +271,6 @@ void llvm::PHIElimination::LowerAtomicPHINode( // path the PHI. MachineBasicBlock &opBlock = *MPhi->getOperand(i*2+2).getMBB(); - // Record the kill. - PHIKills[SrcReg].insert(&opBlock); - // If source is defined by an implicit def, there is no need to insert a // copy. MachineInstr *DefMI = MRI->getVRegDef(SrcReg); @@ -451,34 +443,3 @@ MachineBasicBlock *PHIElimination::SplitCriticalEdge(MachineBasicBlock *A, return NMBB; } - -unsigned -PHIElimination::PHINodeTraits::getHashValue(const MachineInstr *MI) { - if (!MI || MI==getEmptyKey() || MI==getTombstoneKey()) - return DenseMapInfo<MachineInstr*>::getHashValue(MI); - unsigned hash = 0; - for (unsigned ni = 1, ne = MI->getNumOperands(); ni != ne; ni += 2) - hash = hash*37 + DenseMapInfo<BBVRegPair>:: - getHashValue(BBVRegPair(MI->getOperand(ni+1).getMBB()->getNumber(), - MI->getOperand(ni).getReg())); - return hash; -} - -bool PHIElimination::PHINodeTraits::isEqual(const MachineInstr *LHS, - const MachineInstr *RHS) { - const MachineInstr *EmptyKey = getEmptyKey(); - const MachineInstr *TombstoneKey = getTombstoneKey(); - if (!LHS || !RHS || LHS==EmptyKey || RHS==EmptyKey || - LHS==TombstoneKey || RHS==TombstoneKey) - return LHS==RHS; - - unsigned ne = LHS->getNumOperands(); - if (ne != RHS->getNumOperands()) - return false; - // Ignore operand 0, the defined register. - for (unsigned ni = 1; ni != ne; ni += 2) - if (LHS->getOperand(ni).getReg() != RHS->getOperand(ni).getReg() || - LHS->getOperand(ni+1).getMBB() != RHS->getOperand(ni+1).getMBB()) - return false; - return true; -} diff --git a/lib/CodeGen/PHIElimination.h b/lib/CodeGen/PHIElimination.h index 895aaa4..7dedf03 100644 --- a/lib/CodeGen/PHIElimination.h +++ b/lib/CodeGen/PHIElimination.h @@ -22,17 +22,8 @@ namespace llvm { /// Lower PHI instructions to copies. class PHIElimination : public MachineFunctionPass { MachineRegisterInfo *MRI; // Machine register information - private: - - typedef SmallSet<MachineBasicBlock*, 4> PHIKillList; - typedef DenseMap<unsigned, PHIKillList> PHIKillMap; - typedef DenseMap<unsigned, MachineBasicBlock*> PHIDefMap; public: - - typedef PHIKillList::iterator phi_kill_iterator; - typedef PHIKillList::const_iterator const_phi_kill_iterator; - static char ID; // Pass identification, replacement for typeid PHIElimination() : MachineFunctionPass(&ID) {} @@ -40,38 +31,6 @@ namespace llvm { virtual void getAnalysisUsage(AnalysisUsage &AU) const; - /// Return true if the given vreg was defined by a PHI intsr prior to - /// lowering. - bool hasPHIDef(unsigned vreg) const { - return PHIDefs.count(vreg); - } - - /// Returns the block in which the PHI instruction which defined the - /// given vreg used to reside. - MachineBasicBlock* getPHIDefBlock(unsigned vreg) { - PHIDefMap::iterator phiDefItr = PHIDefs.find(vreg); - assert(phiDefItr != PHIDefs.end() && "vreg has no phi-def."); - return phiDefItr->second; - } - - /// Returns true if the given vreg was killed by a PHI instr. - bool hasPHIKills(unsigned vreg) const { - return PHIKills.count(vreg); - } - - /// Returns an iterator over the BasicBlocks which contained PHI - /// kills of this register prior to lowering. - phi_kill_iterator phiKillsBegin(unsigned vreg) { - PHIKillMap::iterator phiKillItr = PHIKills.find(vreg); - assert(phiKillItr != PHIKills.end() && "vreg has no phi-kills."); - return phiKillItr->second.begin(); - } - phi_kill_iterator phiKillsEnd(unsigned vreg) { - PHIKillMap::iterator phiKillItr = PHIKills.find(vreg); - assert(phiKillItr != PHIKills.end() && "vreg has no phi-kills."); - return phiKillItr->second.end(); - } - private: /// EliminatePHINodes - Eliminate phi nodes by inserting copy instructions /// in predecessor basic blocks. @@ -109,12 +68,29 @@ namespace llvm { // SkipPHIsAndLabels - Copies need to be inserted after phi nodes and // also after any exception handling labels: in landing pads execution // starts at the label, so any copies placed before it won't be executed! + // We also deal with DBG_VALUEs, which are a bit tricky: + // PHI + // DBG_VALUE + // LABEL + // Here the DBG_VALUE needs to be skipped, and if it refers to a PHI it + // needs to be annulled or, better, moved to follow the label, as well. + // PHI + // DBG_VALUE + // no label + // Here it is not a good idea to skip the DBG_VALUE. + // FIXME: For now we skip and annul all DBG_VALUEs, maximally simple and + // maximally stupid. MachineBasicBlock::iterator SkipPHIsAndLabels(MachineBasicBlock &MBB, MachineBasicBlock::iterator I) { // Rather than assuming that EH labels come before other kinds of labels, // just skip all labels. - while (I != MBB.end() && (I->isPHI() || I->isLabel())) + while (I != MBB.end() && + (I->isPHI() || I->isLabel() || I->isDebugValue())) { + if (I->isDebugValue() && I->getNumOperands()==3 && + I->getOperand(0).isReg()) + I->getOperand(0).setReg(0U); ++I; + } return I; } @@ -122,21 +98,13 @@ namespace llvm { typedef DenseMap<BBVRegPair, unsigned> VRegPHIUse; VRegPHIUse VRegPHIUseCount; - PHIDefMap PHIDefs; - PHIKillMap PHIKills; // Defs of PHI sources which are implicit_def. SmallPtrSet<MachineInstr*, 4> ImpDefs; - // Lowered PHI nodes may be reused. We provide special DenseMap traits to - // match PHI nodes with identical arguments. - struct PHINodeTraits : public DenseMapInfo<MachineInstr*> { - static unsigned getHashValue(const MachineInstr *PtrVal); - static bool isEqual(const MachineInstr *LHS, const MachineInstr *RHS); - }; - // Map reusable lowered PHI node -> incoming join register. - typedef DenseMap<MachineInstr*, unsigned, PHINodeTraits> LoweredPHIMap; + typedef DenseMap<MachineInstr*, unsigned, + MachineInstrExpressionTrait> LoweredPHIMap; LoweredPHIMap LoweredPHIs; }; diff --git a/lib/CodeGen/Passes.cpp b/lib/CodeGen/Passes.cpp index f67eb79..5ea2941 100644 --- a/lib/CodeGen/Passes.cpp +++ b/lib/CodeGen/Passes.cpp @@ -34,7 +34,7 @@ static cl::opt<RegisterRegAlloc::FunctionPassCtor, false, RegisterPassParser<RegisterRegAlloc> > RegAlloc("regalloc", cl::init(&createLinearScanRegisterAllocator), - cl::desc("Register allocator to use: (default = linearscan)")); + cl::desc("Register allocator to use (default=linearscan)")); //===---------------------------------------------------------------------===// diff --git a/lib/CodeGen/PostRASchedulerList.cpp b/lib/CodeGen/PostRASchedulerList.cpp index f43395f..424181c 100644 --- a/lib/CodeGen/PostRASchedulerList.cpp +++ b/lib/CodeGen/PostRASchedulerList.cpp @@ -460,6 +460,8 @@ void SchedulePostRATDList::FixupKills(MachineBasicBlock *MBB) { for (MachineBasicBlock::iterator I = MBB->end(), E = MBB->begin(); I != E; --Count) { MachineInstr *MI = --I; + if (MI->isDebugValue()) + continue; // Update liveness. Registers that are defed but not used in this // instruction are now dead. Mark register and all subregs as they diff --git a/lib/CodeGen/ProcessImplicitDefs.cpp b/lib/CodeGen/ProcessImplicitDefs.cpp index e3df2e4..d7179b3 100644 --- a/lib/CodeGen/ProcessImplicitDefs.cpp +++ b/lib/CodeGen/ProcessImplicitDefs.cpp @@ -205,10 +205,9 @@ bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) { // Process each use instruction once. for (MachineRegisterInfo::use_iterator UI = mri_->use_begin(Reg), UE = mri_->use_end(); UI != UE; ++UI) { - MachineInstr *RMI = &*UI; - MachineBasicBlock *RMBB = RMI->getParent(); - if (RMBB == MBB) + if (UI.getOperand().isUndef()) continue; + MachineInstr *RMI = &*UI; if (ModInsts.insert(RMI)) RUses.push_back(RMI); } diff --git a/lib/CodeGen/PrologEpilogInserter.cpp b/lib/CodeGen/PrologEpilogInserter.cpp index 036f59a..138e711 100644 --- a/lib/CodeGen/PrologEpilogInserter.cpp +++ b/lib/CodeGen/PrologEpilogInserter.cpp @@ -175,9 +175,10 @@ void PEI::calculateCallsInformation(MachineFunction &Fn) { MachineBasicBlock::iterator I = *i; // If call frames are not being included as part of the stack frame, and - // there is no dynamic allocation (therefore referencing frame slots off - // sp), leave the pseudo ops alone. We'll eliminate them later. - if (RegInfo->hasReservedCallFrame(Fn) || RegInfo->hasFP(Fn)) + // the target doesn't indicate otherwise, remove the call frame pseudos + // here. The sub/add sp instruction pairs are still inserted, but we don't + // need to track the SP adjustment for frame index elimination. + if (RegInfo->canSimplifyCallFramePseudos(Fn)) RegInfo->eliminateCallFramePseudoInstr(Fn, *I->getParent(), I); } } @@ -476,8 +477,6 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { // Loop over all of the stack objects, assigning sequential addresses... MachineFrameInfo *FFI = Fn.getFrameInfo(); - unsigned MaxAlign = 1; - // Start at the beginning of the local area. // The Offset is the distance from the stack top in the direction // of stack growth -- so it's always nonnegative. @@ -517,9 +516,6 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { Offset += FFI->getObjectSize(i); unsigned Align = FFI->getObjectAlignment(i); - // If the alignment of this object is greater than that of the stack, - // then increase the stack alignment to match. - MaxAlign = std::max(MaxAlign, Align); // Adjust to alignment boundary Offset = (Offset+Align-1)/Align*Align; @@ -529,9 +525,6 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { int MaxCSFI = MaxCSFrameIndex, MinCSFI = MinCSFrameIndex; for (int i = MaxCSFI; i >= MinCSFI ; --i) { unsigned Align = FFI->getObjectAlignment(i); - // If the alignment of this object is greater than that of the stack, - // then increase the stack alignment to match. - MaxAlign = std::max(MaxAlign, Align); // Adjust to alignment boundary Offset = (Offset+Align-1)/Align*Align; @@ -540,6 +533,8 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { } } + unsigned MaxAlign = FFI->getMaxAlignment(); + // Make sure the special register scavenging spill slot is closest to the // frame pointer if a frame pointer is required. const TargetRegisterInfo *RegInfo = Fn.getTarget().getRegisterInfo(); @@ -605,11 +600,6 @@ void PEI::calculateFrameObjectOffsets(MachineFunction &Fn) { // Update frame info to pretend that this is part of the stack... FFI->setStackSize(Offset - LocalAreaOffset); - - // Remember the required stack alignment in case targets need it to perform - // dynamic stack alignment. - if (MaxAlign > FFI->getMaxAlignment()) - FFI->setMaxAlignment(MaxAlign); } diff --git a/lib/CodeGen/PseudoSourceValue.cpp b/lib/CodeGen/PseudoSourceValue.cpp index 7fb3e6e..5e86e5a 100644 --- a/lib/CodeGen/PseudoSourceValue.cpp +++ b/lib/CodeGen/PseudoSourceValue.cpp @@ -18,19 +18,38 @@ #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/ManagedStatic.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/System/Mutex.h" #include <map> using namespace llvm; -static ManagedStatic<PseudoSourceValue[4]> PSVs; +namespace { +struct PSVGlobalsTy { + // PseudoSourceValues are immutable so don't need locking. + const PseudoSourceValue PSVs[4]; + sys::Mutex Lock; // Guards FSValues, but not the values inside it. + std::map<int, const PseudoSourceValue *> FSValues; + + PSVGlobalsTy() : PSVs() {} + ~PSVGlobalsTy() { + for (std::map<int, const PseudoSourceValue *>::iterator + I = FSValues.begin(), E = FSValues.end(); I != E; ++I) { + delete I->second; + } + } +}; + +static ManagedStatic<PSVGlobalsTy> PSVGlobals; + +} // anonymous namespace const PseudoSourceValue *PseudoSourceValue::getStack() -{ return &(*PSVs)[0]; } +{ return &PSVGlobals->PSVs[0]; } const PseudoSourceValue *PseudoSourceValue::getGOT() -{ return &(*PSVs)[1]; } +{ return &PSVGlobals->PSVs[1]; } const PseudoSourceValue *PseudoSourceValue::getJumpTable() -{ return &(*PSVs)[2]; } +{ return &PSVGlobals->PSVs[2]; } const PseudoSourceValue *PseudoSourceValue::getConstantPool() -{ return &(*PSVs)[3]; } +{ return &PSVGlobals->PSVs[3]; } static const char *const PSVNames[] = { "Stack", @@ -48,13 +67,13 @@ PseudoSourceValue::PseudoSourceValue(enum ValueTy Subclass) : Subclass) {} void PseudoSourceValue::printCustom(raw_ostream &O) const { - O << PSVNames[this - *PSVs]; + O << PSVNames[this - PSVGlobals->PSVs]; } -static ManagedStatic<std::map<int, const PseudoSourceValue *> > FSValues; - const PseudoSourceValue *PseudoSourceValue::getFixedStack(int FI) { - const PseudoSourceValue *&V = (*FSValues)[FI]; + PSVGlobalsTy &PG = *PSVGlobals; + sys::ScopedLock locked(PG.Lock); + const PseudoSourceValue *&V = PG.FSValues[FI]; if (!V) V = new FixedStackPseudoSourceValue(FI); return V; diff --git a/lib/CodeGen/RegAllocLinearScan.cpp b/lib/CodeGen/RegAllocLinearScan.cpp index 8e44a57..5c5a394 100644 --- a/lib/CodeGen/RegAllocLinearScan.cpp +++ b/lib/CodeGen/RegAllocLinearScan.cpp @@ -334,10 +334,6 @@ namespace { SmallVector<unsigned, 256> &inactiveCounts, bool SkipDGRegs); - /// assignVirt2StackSlot - assigns this virtual register to a - /// stack slot. returns the stack slot - int assignVirt2StackSlot(unsigned virtReg); - void ComputeRelatedRegClasses(); template <typename ItTy> diff --git a/lib/CodeGen/RegAllocLocal.cpp b/lib/CodeGen/RegAllocLocal.cpp index 4d2e3a3..04303cf 100644 --- a/lib/CodeGen/RegAllocLocal.cpp +++ b/lib/CodeGen/RegAllocLocal.cpp @@ -490,9 +490,12 @@ MachineInstr *RALocal::reloadVirtReg(MachineBasicBlock &MBB, MachineInstr *MI, // If the virtual register is already available, just update the instruction // and return. if (unsigned PR = getVirt2PhysRegMapSlot(VirtReg)) { - MarkPhysRegRecentlyUsed(PR); // Already have this value available! MI->getOperand(OpNum).setReg(PR); // Assign the input register - getVirtRegLastUse(VirtReg) = std::make_pair(MI, OpNum); + if (!MI->isDebugValue()) { + // Do not do these for DBG_VALUE as they can affect codegen. + MarkPhysRegRecentlyUsed(PR); // Already have this value available! + getVirtRegLastUse(VirtReg) = std::make_pair(MI, OpNum); + } return MI; } @@ -609,6 +612,8 @@ void RALocal::ComputeLocalLiveness(MachineBasicBlock& MBB) { DenseMap<unsigned, std::pair<MachineInstr*, unsigned> > LastUseDef; for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end(); I != E; ++I) { + if (I->isDebugValue()) + continue; for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { MachineOperand& MO = I->getOperand(i); // Uses don't trigger any flags, but we need to save @@ -691,7 +696,13 @@ void RALocal::ComputeLocalLiveness(MachineBasicBlock& MBB) { bool usedOutsideBlock = isPhysReg ? false : UsedInMultipleBlocks.test(MO.getReg() - TargetRegisterInfo::FirstVirtualRegister); - if (!isPhysReg && !usedOutsideBlock) + if (!isPhysReg && !usedOutsideBlock) { + // DBG_VALUE complicates this: if the only refs of a register outside + // this block are DBG_VALUE, we can't keep the reg live just for that, + // as it will cause the reg to be spilled at the end of this block when + // it wouldn't have been otherwise. Nullify the DBG_VALUEs when that + // happens. + bool UsedByDebugValueOnly = false; for (MachineRegisterInfo::reg_iterator UI = MRI.reg_begin(MO.getReg()), UE = MRI.reg_end(); UI != UE; ++UI) // Two cases: @@ -699,12 +710,26 @@ void RALocal::ComputeLocalLiveness(MachineBasicBlock& MBB) { // - used in the same block before it is defined (loop) if (UI->getParent() != &MBB || (MO.isDef() && UI.getOperand().isUse() && precedes(&*UI, MI))) { + if (UI->isDebugValue()) { + UsedByDebugValueOnly = true; + continue; + } + // A non-DBG_VALUE use means we can leave DBG_VALUE uses alone. UsedInMultipleBlocks.set(MO.getReg() - TargetRegisterInfo::FirstVirtualRegister); usedOutsideBlock = true; + UsedByDebugValueOnly = false; break; } - + if (UsedByDebugValueOnly) + for (MachineRegisterInfo::reg_iterator UI = MRI.reg_begin(MO.getReg()), + UE = MRI.reg_end(); UI != UE; ++UI) + if (UI->isDebugValue() && + (UI->getParent() != &MBB || + (MO.isDef() && precedes(&*UI, MI)))) + UI.getOperand().setReg(0U); + } + // Physical registers and those that are not live-out of the block // are killed/dead at their last use/def within this block. if (isPhysReg || !usedOutsideBlock) { diff --git a/lib/CodeGen/RegAllocPBQP.cpp b/lib/CodeGen/RegAllocPBQP.cpp index 2701faf..81cfd8f 100644 --- a/lib/CodeGen/RegAllocPBQP.cpp +++ b/lib/CodeGen/RegAllocPBQP.cpp @@ -57,7 +57,7 @@ using namespace llvm; static RegisterRegAlloc -registerPBQPRepAlloc("pbqp", "PBQP register allocator.", +registerPBQPRepAlloc("pbqp", "PBQP register allocator", llvm::createPBQPRegisterAllocator); static cl::opt<bool> @@ -867,10 +867,6 @@ bool PBQPRegAlloc::runOnMachineFunction(MachineFunction &MF) { // Find the vreg intervals in need of allocation. findVRegIntervalsToAlloc(); - // If there aren't any then we're done here. - if (vregIntervalsToAlloc.empty() && emptyVRegIntervals.empty()) - return true; - // If there are non-empty intervals allocate them using pbqp. if (!vregIntervalsToAlloc.empty()) { diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp index 56dd533..badf34e 100644 --- a/lib/CodeGen/ScheduleDAGInstrs.cpp +++ b/lib/CodeGen/ScheduleDAGInstrs.cpp @@ -72,7 +72,7 @@ static const Value *getUnderlyingObjectFromInt(const Value *V) { } else { return V; } - assert(isa<IntegerType>(V->getType()) && "Unexpected operand type!"); + assert(V->getType()->isIntegerTy() && "Unexpected operand type!"); } while (1); } @@ -87,7 +87,7 @@ static const Value *getUnderlyingObject(const Value *V) { break; const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0)); // If that succeeded in finding a pointer, continue the search. - if (!isa<PointerType>(O->getType())) + if (!O->getType()->isPointerTy()) break; V = O; } while (1); diff --git a/lib/CodeGen/SelectionDAG/Android.mk b/lib/CodeGen/SelectionDAG/Android.mk new file mode 100644 index 0000000..eb15a18 --- /dev/null +++ b/lib/CodeGen/SelectionDAG/Android.mk @@ -0,0 +1,48 @@ +LOCAL_PATH:= $(call my-dir) + +codegen_selectiondag_SRC_FILES := \ + CallingConvLower.cpp \ + DAGCombiner.cpp \ + FastISel.cpp \ + FunctionLoweringInfo.cpp \ + InstrEmitter.cpp \ + LegalizeDAG.cpp \ + LegalizeFloatTypes.cpp \ + LegalizeIntegerTypes.cpp \ + LegalizeTypes.cpp \ + LegalizeTypesGeneric.cpp \ + LegalizeVectorOps.cpp \ + LegalizeVectorTypes.cpp \ + ScheduleDAGFast.cpp \ + ScheduleDAGList.cpp \ + ScheduleDAGRRList.cpp \ + ScheduleDAGSDNodes.cpp \ + SelectionDAG.cpp \ + SelectionDAGBuilder.cpp \ + SelectionDAGISel.cpp \ + SelectionDAGPrinter.cpp \ + TargetLowering.cpp + +# For the host +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_selectiondag_SRC_FILES) + +LOCAL_MODULE:= libLLVMSelectionDAG + +include $(LLVM_HOST_BUILD_MK) +include $(LLVM_GEN_INTRINSICS_MK) +include $(BUILD_HOST_STATIC_LIBRARY) + +# For the device +# ===================================================== +include $(CLEAR_VARS) + +LOCAL_SRC_FILES := $(codegen_selectiondag_SRC_FILES) + +LOCAL_MODULE:= libLLVMSelectionDAG + +include $(LLVM_DEVICE_BUILD_MK) +include $(LLVM_GEN_INTRINSICS_MK) +include $(BUILD_STATIC_LIBRARY) diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp index 9189e71..3be6b43 100644 --- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -1064,7 +1064,7 @@ SDValue DAGCombiner::visitADD(SDNode *N) { if (VT.isInteger() && !VT.isVector()) { APInt LHSZero, LHSOne; APInt RHSZero, RHSOne; - APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()); + APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()); DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne); if (LHSZero.getBoolValue()) { @@ -1136,7 +1136,7 @@ SDValue DAGCombiner::visitADDC(SDNode *N) { // fold (addc a, b) -> (or a, b), CARRY_FALSE iff a and b share no bits. APInt LHSZero, LHSOne; APInt RHSZero, RHSOne; - APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()); + APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()); DAG.ComputeMaskedBits(N0, Mask, LHSZero, LHSOne); if (LHSZero.getBoolValue()) { @@ -1758,7 +1758,7 @@ SDValue DAGCombiner::visitAND(SDNode *N) { ConstantSDNode *N0C = dyn_cast<ConstantSDNode>(N0); ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1); EVT VT = N1.getValueType(); - unsigned BitWidth = VT.getSizeInBits(); + unsigned BitWidth = VT.getScalarType().getSizeInBits(); // fold vector ops if (VT.isVector()) { @@ -1786,7 +1786,7 @@ SDValue DAGCombiner::visitAND(SDNode *N) { SDValue RAND = ReassociateOps(ISD::AND, N->getDebugLoc(), N0, N1); if (RAND.getNode() != 0) return RAND; - // fold (and (or x, 0xFFFF), 0xFF) -> 0xFF + // fold (and (or x, C), D) -> D if (C & D) == D if (N1C && N0.getOpcode() == ISD::OR) if (ConstantSDNode *ORI = dyn_cast<ConstantSDNode>(N0.getOperand(1))) if ((ORI->getAPIntValue() & N1C->getAPIntValue()) == N1C->getAPIntValue()) @@ -1872,16 +1872,17 @@ SDValue DAGCombiner::visitAND(SDNode *N) { EVT MemVT = LN0->getMemoryVT(); // If we zero all the possible extended bits, then we can turn this into // a zextload if we are running before legalize or the operation is legal. - unsigned BitWidth = N1.getValueSizeInBits(); + unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, - BitWidth - MemVT.getSizeInBits())) && + BitWidth - MemVT.getScalarType().getSizeInBits())) && ((!LegalOperations && !LN0->isVolatile()) || TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT, LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), MemVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); AddToWorkList(N); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -1894,16 +1895,17 @@ SDValue DAGCombiner::visitAND(SDNode *N) { EVT MemVT = LN0->getMemoryVT(); // If we zero all the possible extended bits, then we can turn this into // a zextload if we are running before legalize or the operation is legal. - unsigned BitWidth = N1.getValueSizeInBits(); + unsigned BitWidth = N1.getValueType().getScalarType().getSizeInBits(); if (DAG.MaskedValueIsZero(N1, APInt::getHighBitsSet(BitWidth, - BitWidth - MemVT.getSizeInBits())) && + BitWidth - MemVT.getScalarType().getSizeInBits())) && ((!LegalOperations && !LN0->isVolatile()) || TLI.isLoadExtLegal(ISD::ZEXTLOAD, MemVT))) { SDValue ExtLoad = DAG.getExtLoad(ISD::ZEXTLOAD, N0.getDebugLoc(), VT, LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), MemVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); AddToWorkList(N); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -1935,7 +1937,8 @@ SDValue DAGCombiner::visitAND(SDNode *N) { DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy, LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), - ExtVT, LN0->isVolatile(), LN0->getAlignment()); + ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); AddToWorkList(N); CombineTo(LN0, NewLoad, NewLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -1970,7 +1973,8 @@ SDValue DAGCombiner::visitAND(SDNode *N) { DAG.getExtLoad(ISD::ZEXTLOAD, LN0->getDebugLoc(), LoadResultTy, LN0->getChain(), NewPtr, LN0->getSrcValue(), LN0->getSrcValueOffset(), - ExtVT, LN0->isVolatile(), Alignment); + ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), + Alignment); AddToWorkList(N); CombineTo(LN0, Load, Load.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -2021,13 +2025,15 @@ SDValue DAGCombiner::visitOR(SDNode *N) { if (ROR.getNode() != 0) return ROR; // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2) + // iff (c1 & c2) == 0. if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() && isa<ConstantSDNode>(N0.getOperand(1))) { ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1)); - return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, - DAG.getNode(ISD::OR, N0.getDebugLoc(), VT, - N0.getOperand(0), N1), - DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1)); + if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) + return DAG.getNode(ISD::AND, N->getDebugLoc(), VT, + DAG.getNode(ISD::OR, N0.getDebugLoc(), VT, + N0.getOperand(0), N1), + DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1)); } // fold (or (setcc x), (setcc y)) -> (setcc (or x, y)) if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){ @@ -2750,7 +2756,7 @@ SDValue DAGCombiner::visitSRL(SDNode *N) { if (N1C && N0.getOpcode() == ISD::CTLZ && N1C->getAPIntValue() == Log2_32(VT.getSizeInBits())) { APInt KnownZero, KnownOne; - APInt Mask = APInt::getAllOnesValue(VT.getSizeInBits()); + APInt Mask = APInt::getAllOnesValue(VT.getScalarType().getSizeInBits()); DAG.ComputeMaskedBits(N0.getOperand(0), Mask, KnownZero, KnownOne); // If any of the input bits are KnownOne, then the input couldn't be all @@ -3143,7 +3149,8 @@ SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), N0.getValueType(), - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(), ExtLoad); @@ -3185,7 +3192,8 @@ SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) { LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), MemVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), @@ -3315,7 +3323,8 @@ SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), N0.getValueType(), - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(), ExtLoad); @@ -3357,7 +3366,8 @@ SDValue DAGCombiner::visitZERO_EXTEND(SDNode *N) { LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), MemVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(), @@ -3471,7 +3481,8 @@ SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), N0.getValueType(), - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); SDValue Trunc = DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), N0.getValueType(), ExtLoad); @@ -3513,7 +3524,8 @@ SDValue DAGCombiner::visitANY_EXTEND(SDNode *N) { VT, LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), MemVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), DAG.getNode(ISD::TRUNCATE, N0.getDebugLoc(), @@ -3636,10 +3648,11 @@ SDValue DAGCombiner::ReduceLoadWidth(SDNode *N) { SDValue Load = (ExtType == ISD::NON_EXTLOAD) ? DAG.getLoad(VT, N0.getDebugLoc(), LN0->getChain(), NewPtr, LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff, - LN0->isVolatile(), NewAlign) + LN0->isVolatile(), LN0->isNonTemporal(), NewAlign) : DAG.getExtLoad(ExtType, N0.getDebugLoc(), VT, LN0->getChain(), NewPtr, LN0->getSrcValue(), LN0->getSrcValueOffset() + PtrOff, - ExtVT, LN0->isVolatile(), NewAlign); + ExtVT, LN0->isVolatile(), LN0->isNonTemporal(), + NewAlign); // Replace the old load's chain with the new load's chain. WorkListRemover DeadNodes(*this); @@ -3726,7 +3739,8 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -3742,7 +3756,8 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) { LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), EVT, - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), ExtLoad, ExtLoad.getValue(1)); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -3826,7 +3841,7 @@ SDValue DAGCombiner::CombineConsecutiveLoads(SDNode *N, EVT VT) { (!LegalOperations || TLI.isOperationLegal(ISD::LOAD, VT))) return DAG.getLoad(VT, N->getDebugLoc(), LD1->getChain(), LD1->getBasePtr(), LD1->getSrcValue(), - LD1->getSrcValueOffset(), false, Align); + LD1->getSrcValueOffset(), false, false, Align); } return SDValue(); @@ -3896,7 +3911,8 @@ SDValue DAGCombiner::visitBIT_CONVERT(SDNode *N) { SDValue Load = DAG.getLoad(VT, N->getDebugLoc(), LN0->getChain(), LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), - LN0->isVolatile(), OrigAlign); + LN0->isVolatile(), LN0->isNonTemporal(), + OrigAlign); AddToWorkList(N); CombineTo(N0.getNode(), DAG.getNode(ISD::BIT_CONVERT, N0.getDebugLoc(), @@ -4492,7 +4508,8 @@ SDValue DAGCombiner::visitFP_EXTEND(SDNode *N) { LN0->getBasePtr(), LN0->getSrcValue(), LN0->getSrcValueOffset(), N0.getValueType(), - LN0->isVolatile(), LN0->getAlignment()); + LN0->isVolatile(), LN0->isNonTemporal(), + LN0->getAlignment()); CombineTo(N, ExtLoad); CombineTo(N0.getNode(), DAG.getNode(ISD::FP_ROUND, N0.getDebugLoc(), @@ -4640,7 +4657,8 @@ SDValue DAGCombiner::visitBRCOND(SDNode *N) { DAG.DeleteNode(Trunc); } // Replace the uses of SRL with SETCC - DAG.ReplaceAllUsesOfValueWith(N1, SetCC); + WorkListRemover DeadNodes(*this); + DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes); removeFromWorkList(N1.getNode()); DAG.DeleteNode(N1.getNode()); return SDValue(N, 0); // Return N so it doesn't get rechecked! @@ -4648,6 +4666,56 @@ SDValue DAGCombiner::visitBRCOND(SDNode *N) { } } } + + // Transform br(xor(x, y)) -> br(x != y) + // Transform br(xor(xor(x,y), 1)) -> br (x == y) + if (N1.hasOneUse() && N1.getOpcode() == ISD::XOR) { + SDNode *TheXor = N1.getNode(); + SDValue Op0 = TheXor->getOperand(0); + SDValue Op1 = TheXor->getOperand(1); + if (Op0.getOpcode() == Op1.getOpcode()) { + // Avoid missing important xor optimizations. + SDValue Tmp = visitXOR(TheXor); + if (Tmp.getNode()) { + DEBUG(dbgs() << "\nReplacing.8 "; + TheXor->dump(&DAG); + dbgs() << "\nWith: "; + Tmp.getNode()->dump(&DAG); + dbgs() << '\n'); + WorkListRemover DeadNodes(*this); + DAG.ReplaceAllUsesOfValueWith(N1, Tmp, &DeadNodes); + removeFromWorkList(TheXor); + DAG.DeleteNode(TheXor); + return DAG.getNode(ISD::BRCOND, N->getDebugLoc(), + MVT::Other, Chain, Tmp, N2); + } + } + + if (Op0.getOpcode() != ISD::SETCC && Op1.getOpcode() != ISD::SETCC) { + bool Equal = false; + if (ConstantSDNode *RHSCI = dyn_cast<ConstantSDNode>(Op0)) + if (RHSCI->getAPIntValue() == 1 && Op0.hasOneUse() && + Op0.getOpcode() == ISD::XOR) { + TheXor = Op0.getNode(); + Equal = true; + } + + EVT SetCCVT = N1.getValueType(); + if (LegalTypes) + SetCCVT = TLI.getSetCCResultType(SetCCVT); + SDValue SetCC = DAG.getSetCC(TheXor->getDebugLoc(), + SetCCVT, + Op0, Op1, + Equal ? ISD::SETEQ : ISD::SETNE); + // Replace the uses of XOR with SETCC + WorkListRemover DeadNodes(*this); + DAG.ReplaceAllUsesOfValueWith(N1, SetCC, &DeadNodes); + removeFromWorkList(N1.getNode()); + DAG.DeleteNode(N1.getNode()); + return DAG.getNode(ISD::BRCOND, N->getDebugLoc(), + MVT::Other, Chain, SetCC, N2); + } + } return SDValue(); } @@ -4960,7 +5028,7 @@ SDValue DAGCombiner::visitLOAD(SDNode *N) { LD->getValueType(0), Chain, Ptr, LD->getSrcValue(), LD->getSrcValueOffset(), LD->getMemoryVT(), - LD->isVolatile(), Align); + LD->isVolatile(), LD->isNonTemporal(), Align); } } @@ -4997,7 +5065,7 @@ SDValue DAGCombiner::visitLOAD(SDNode *N) { assert(N->getValueType(2) == MVT::Other && "Malformed indexed loads?"); if (N->hasNUsesOfValue(0, 0) && N->hasNUsesOfValue(0, 1)) { SDValue Undef = DAG.getUNDEF(N->getValueType(0)); - DEBUG(dbgs() << "\nReplacing.6 "; + DEBUG(dbgs() << "\nReplacing.7 "; N->dump(&DAG); dbgs() << "\nWith: "; Undef.getNode()->dump(&DAG); @@ -5042,7 +5110,8 @@ SDValue DAGCombiner::visitLOAD(SDNode *N) { ReplLoad = DAG.getLoad(N->getValueType(0), LD->getDebugLoc(), BetterChain, Ptr, LD->getSrcValue(), LD->getSrcValueOffset(), - LD->isVolatile(), LD->getAlignment()); + LD->isVolatile(), LD->isNonTemporal(), + LD->getAlignment()); } else { ReplLoad = DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(), LD->getValueType(0), @@ -5050,6 +5119,7 @@ SDValue DAGCombiner::visitLOAD(SDNode *N) { LD->getSrcValueOffset(), LD->getMemoryVT(), LD->isVolatile(), + LD->isNonTemporal(), LD->getAlignment()); } @@ -5149,13 +5219,14 @@ SDValue DAGCombiner::ReduceLoadOpStoreWidth(SDNode *N) { SDValue NewLD = DAG.getLoad(NewVT, N0.getDebugLoc(), LD->getChain(), NewPtr, LD->getSrcValue(), LD->getSrcValueOffset(), - LD->isVolatile(), NewAlign); + LD->isVolatile(), LD->isNonTemporal(), + NewAlign); SDValue NewVal = DAG.getNode(Opc, Value.getDebugLoc(), NewVT, NewLD, DAG.getConstant(NewImm, NewVT)); SDValue NewST = DAG.getStore(Chain, N->getDebugLoc(), NewVal, NewPtr, ST->getSrcValue(), ST->getSrcValueOffset(), - false, NewAlign); + false, false, NewAlign); AddToWorkList(NewPtr.getNode()); AddToWorkList(NewLD.getNode()); @@ -5184,7 +5255,7 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { return DAG.getTruncStore(Chain, N->getDebugLoc(), Value, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), ST->getMemoryVT(), - ST->isVolatile(), Align); + ST->isVolatile(), ST->isNonTemporal(), Align); } } @@ -5201,7 +5272,8 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { TLI.isOperationLegalOrCustom(ISD::STORE, SVT))) return DAG.getStore(Chain, N->getDebugLoc(), Value.getOperand(0), Ptr, ST->getSrcValue(), - ST->getSrcValueOffset(), ST->isVolatile(), OrigAlign); + ST->getSrcValueOffset(), ST->isVolatile(), + ST->isNonTemporal(), OrigAlign); } // Turn 'store float 1.0, Ptr' -> 'store int 0x12345678, Ptr' @@ -5227,7 +5299,7 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { return DAG.getStore(Chain, N->getDebugLoc(), Tmp, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), ST->isVolatile(), - ST->getAlignment()); + ST->isNonTemporal(), ST->getAlignment()); } break; case MVT::f64: @@ -5239,7 +5311,7 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { return DAG.getStore(Chain, N->getDebugLoc(), Tmp, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), ST->isVolatile(), - ST->getAlignment()); + ST->isNonTemporal(), ST->getAlignment()); } else if (!ST->isVolatile() && TLI.isOperationLegalOrCustom(ISD::STORE, MVT::i32)) { // Many FP stores are not made apparent until after legalize, e.g. for @@ -5253,18 +5325,21 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { int SVOffset = ST->getSrcValueOffset(); unsigned Alignment = ST->getAlignment(); bool isVolatile = ST->isVolatile(); + bool isNonTemporal = ST->isNonTemporal(); SDValue St0 = DAG.getStore(Chain, ST->getDebugLoc(), Lo, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), - isVolatile, ST->getAlignment()); + isVolatile, isNonTemporal, + ST->getAlignment()); Ptr = DAG.getNode(ISD::ADD, N->getDebugLoc(), Ptr.getValueType(), Ptr, DAG.getConstant(4, Ptr.getValueType())); SVOffset += 4; Alignment = MinAlign(Alignment, 4U); SDValue St1 = DAG.getStore(Chain, ST->getDebugLoc(), Hi, Ptr, ST->getSrcValue(), - SVOffset, isVolatile, Alignment); + SVOffset, isVolatile, isNonTemporal, + Alignment); return DAG.getNode(ISD::TokenFactor, N->getDebugLoc(), MVT::Other, St0, St1); } @@ -5286,12 +5361,13 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { if (ST->isTruncatingStore()) { ReplStore = DAG.getTruncStore(BetterChain, N->getDebugLoc(), Value, Ptr, ST->getSrcValue(),ST->getSrcValueOffset(), - ST->getMemoryVT(), - ST->isVolatile(), ST->getAlignment()); + ST->getMemoryVT(), ST->isVolatile(), + ST->isNonTemporal(), ST->getAlignment()); } else { ReplStore = DAG.getStore(BetterChain, N->getDebugLoc(), Value, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), - ST->isVolatile(), ST->getAlignment()); + ST->isVolatile(), ST->isNonTemporal(), + ST->getAlignment()); } // Create token to keep both nodes around. @@ -5325,7 +5401,8 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { return DAG.getTruncStore(Chain, N->getDebugLoc(), Shorter, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), ST->getMemoryVT(), - ST->isVolatile(), ST->getAlignment()); + ST->isVolatile(), ST->isNonTemporal(), + ST->getAlignment()); // Otherwise, see if we can simplify the operation with // SimplifyDemandedBits, which only works if the value has a single use. @@ -5358,7 +5435,8 @@ SDValue DAGCombiner::visitSTORE(SDNode *N) { return DAG.getTruncStore(Chain, N->getDebugLoc(), Value.getOperand(0), Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), ST->getMemoryVT(), - ST->isVolatile(), ST->getAlignment()); + ST->isVolatile(), ST->isNonTemporal(), + ST->getAlignment()); } return ReduceLoadOpStoreWidth(N); @@ -5503,7 +5581,7 @@ SDValue DAGCombiner::visitEXTRACT_VECTOR_ELT(SDNode *N) { return DAG.getLoad(LVT, N->getDebugLoc(), LN0->getChain(), NewPtr, LN0->getSrcValue(), LN0->getSrcValueOffset(), - LN0->isVolatile(), Align); + LN0->isVolatile(), LN0->isNonTemporal(), Align); } return SDValue(); @@ -5883,6 +5961,7 @@ bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, LLD->getChain(), Addr, 0, 0, LLD->isVolatile(), + LLD->isNonTemporal(), LLD->getAlignment()); } else { Load = DAG.getExtLoad(LLD->getExtensionType(), @@ -5891,6 +5970,7 @@ bool DAGCombiner::SimplifySelectOps(SDNode *TheSelect, SDValue LHS, LLD->getChain(), Addr, 0, 0, LLD->getMemoryVT(), LLD->isVolatile(), + LLD->isNonTemporal(), LLD->getAlignment()); } @@ -5998,7 +6078,7 @@ SDValue DAGCombiner::SimplifySelectCC(DebugLoc DL, SDValue N0, SDValue N1, CstOffset); return DAG.getLoad(TV->getValueType(0), DL, DAG.getEntryNode(), CPIdx, PseudoSourceValue::getConstantPool(), 0, false, - Alignment); + false, Alignment); } } diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp index 35ef5b7..1d76c7c 100644 --- a/lib/CodeGen/SelectionDAG/FastISel.cpp +++ b/lib/CodeGen/SelectionDAG/FastISel.cpp @@ -350,6 +350,34 @@ bool FastISel::SelectCall(User *I) { (void)TargetSelectInstruction(cast<Instruction>(I)); return true; } + case Intrinsic::dbg_value: { + // This requires target support, but right now X86 is the only Fast target. + DbgValueInst *DI = cast<DbgValueInst>(I); + const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE); + Value *V = DI->getValue(); + if (!V) { + // Currently the optimizer can produce this; insert an undef to + // help debugging. Probably the optimizer should not do this. + BuildMI(MBB, DL, II).addReg(0U).addImm(DI->getOffset()). + addMetadata(DI->getVariable()); + } else if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { + BuildMI(MBB, DL, II).addImm(CI->getZExtValue()).addImm(DI->getOffset()). + addMetadata(DI->getVariable()); + } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) { + BuildMI(MBB, DL, II).addFPImm(CF).addImm(DI->getOffset()). + addMetadata(DI->getVariable()); + } else if (unsigned Reg = lookUpRegForValue(V)) { + BuildMI(MBB, DL, II).addReg(Reg, RegState::Debug).addImm(DI->getOffset()). + addMetadata(DI->getVariable()); + } else { + // We can't yet handle anything else here because it would require + // generating code, thus altering codegen because of debug info. + // Insert an undef so we can see what we dropped. + BuildMI(MBB, DL, II).addReg(0U).addImm(DI->getOffset()). + addMetadata(DI->getVariable()); + } + return true; + } case Intrinsic::eh_exception: { EVT VT = TLI.getValueType(I->getType()); switch (TLI.getOperationAction(ISD::EXCEPTIONADDR, VT)) { diff --git a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp index 02fe85d..625de11 100644 --- a/lib/CodeGen/SelectionDAG/InstrEmitter.cpp +++ b/lib/CodeGen/SelectionDAG/InstrEmitter.cpp @@ -15,6 +15,7 @@ #define DEBUG_TYPE "instr-emitter" #include "InstrEmitter.h" +#include "SDDbgValue.h" #include "llvm/CodeGen/MachineConstantPool.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineInstrBuilder.h" @@ -497,6 +498,56 @@ InstrEmitter::EmitCopyToRegClassNode(SDNode *Node, assert(isNew && "Node emitted out of order - early"); } +/// EmitDbgValue - Generate any debug info that refers to this Node. Constant +/// dbg_value is not handled here. +void +InstrEmitter::EmitDbgValue(SDNode *Node, + DenseMap<SDValue, unsigned> &VRBaseMap, + SDDbgValue *sd) { + if (!Node->getHasDebugValue()) + return; + if (!sd) + return; + unsigned VReg = getVR(SDValue(sd->getSDNode(), sd->getResNo()), VRBaseMap); + const TargetInstrDesc &II = TII->get(TargetOpcode::DBG_VALUE); + DebugLoc DL = sd->getDebugLoc(); + MachineInstr *MI; + if (VReg) { + MI = BuildMI(*MF, DL, II).addReg(VReg, RegState::Debug). + addImm(sd->getOffset()). + addMetadata(sd->getMDPtr()); + } else { + // Insert an Undef so we can see what we dropped. + MI = BuildMI(*MF, DL, II).addReg(0U).addImm(sd->getOffset()). + addMetadata(sd->getMDPtr()); + } + MBB->insert(InsertPos, MI); +} + +/// EmitDbgValue - Generate constant debug info. No SDNode is involved. +void +InstrEmitter::EmitDbgValue(SDDbgValue *sd) { + if (!sd) + return; + const TargetInstrDesc &II = TII->get(TargetOpcode::DBG_VALUE); + DebugLoc DL = sd->getDebugLoc(); + MachineInstr *MI; + Value *V = sd->getConst(); + if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) { + MI = BuildMI(*MF, DL, II).addImm(CI->getZExtValue()). + addImm(sd->getOffset()). + addMetadata(sd->getMDPtr()); + } else if (ConstantFP *CF = dyn_cast<ConstantFP>(V)) { + MI = BuildMI(*MF, DL, II).addFPImm(CF).addImm(sd->getOffset()). + addMetadata(sd->getMDPtr()); + } else { + // Insert an Undef so we can see what we dropped. + MI = BuildMI(*MF, DL, II).addReg(0U).addImm(sd->getOffset()). + addMetadata(sd->getMDPtr()); + } + MBB->insert(InsertPos, MI); +} + /// EmitNode - Generate machine code for a node and needed dependencies. /// void InstrEmitter::EmitNode(SDNode *Node, bool IsClone, bool IsCloned, diff --git a/lib/CodeGen/SelectionDAG/InstrEmitter.h b/lib/CodeGen/SelectionDAG/InstrEmitter.h index 91817e4..4fe9f19 100644 --- a/lib/CodeGen/SelectionDAG/InstrEmitter.h +++ b/lib/CodeGen/SelectionDAG/InstrEmitter.h @@ -23,6 +23,7 @@ namespace llvm { class TargetInstrDesc; +class SDDbgValue; class InstrEmitter { MachineFunction *MF; @@ -97,6 +98,16 @@ public: /// MachineInstr. static unsigned CountOperands(SDNode *Node); + /// EmitDbgValue - Generate any debug info that refers to this Node. Constant + /// dbg_value is not handled here. + void EmitDbgValue(SDNode *Node, + DenseMap<SDValue, unsigned> &VRBaseMap, + SDDbgValue* sd); + + + /// EmitDbgValue - Generate a constant DBG_VALUE. No node is involved. + void EmitDbgValue(SDDbgValue* sd); + /// EmitNode - Generate machine code for a node and needed dependencies. /// void EmitNode(SDNode *Node, bool IsClone, bool IsCloned, diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp index 78e6e4e..f498263 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp @@ -377,9 +377,10 @@ static SDValue ExpandConstantFP(ConstantFPSDNode *CFP, bool UseCP, return DAG.getExtLoad(ISD::EXTLOAD, dl, OrigVT, DAG.getEntryNode(), CPIdx, PseudoSourceValue::getConstantPool(), - 0, VT, false, Alignment); + 0, VT, false, false, Alignment); return DAG.getLoad(OrigVT, dl, DAG.getEntryNode(), CPIdx, - PseudoSourceValue::getConstantPool(), 0, false, Alignment); + PseudoSourceValue::getConstantPool(), 0, false, false, + Alignment); } /// ExpandUnalignedStore - Expands an unaligned store to 2 half-size stores. @@ -402,7 +403,8 @@ SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG, // FIXME: Does not handle truncating floating point stores! SDValue Result = DAG.getNode(ISD::BIT_CONVERT, dl, intVT, Val); return DAG.getStore(Chain, dl, Result, Ptr, ST->getSrcValue(), - SVOffset, ST->isVolatile(), Alignment); + SVOffset, ST->isVolatile(), ST->isNonTemporal(), + Alignment); } else { // Do a (aligned) store to a stack slot, then copy from the stack slot // to the final destination using (unaligned) integer loads and stores. @@ -418,7 +420,8 @@ SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG, // Perform the original store, only redirected to the stack slot. SDValue Store = DAG.getTruncStore(Chain, dl, - Val, StackPtr, NULL, 0, StoredVT); + Val, StackPtr, NULL, 0, StoredVT, + false, false, 0); SDValue Increment = DAG.getConstant(RegBytes, TLI.getPointerTy()); SmallVector<SDValue, 8> Stores; unsigned Offset = 0; @@ -426,11 +429,12 @@ SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG, // Do all but one copies using the full register width. for (unsigned i = 1; i < NumRegs; i++) { // Load one integer register's worth from the stack slot. - SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr, NULL, 0); + SDValue Load = DAG.getLoad(RegVT, dl, Store, StackPtr, NULL, 0, + false, false, 0); // Store it to the final location. Remember the store. Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, Ptr, ST->getSrcValue(), SVOffset + Offset, - ST->isVolatile(), + ST->isVolatile(), ST->isNonTemporal(), MinAlign(ST->getAlignment(), Offset))); // Increment the pointers. Offset += RegBytes; @@ -446,11 +450,12 @@ SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG, // Load from the stack slot. SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Store, StackPtr, - NULL, 0, MemVT); + NULL, 0, MemVT, false, false, 0); Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, Ptr, ST->getSrcValue(), SVOffset + Offset, MemVT, ST->isVolatile(), + ST->isNonTemporal(), MinAlign(ST->getAlignment(), Offset))); // The order of the stores doesn't matter - say it with a TokenFactor. return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0], @@ -474,13 +479,14 @@ SDValue ExpandUnalignedStore(StoreSDNode *ST, SelectionDAG &DAG, SDValue Store1, Store2; Store1 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Lo:Hi, Ptr, ST->getSrcValue(), SVOffset, NewStoredVT, - ST->isVolatile(), Alignment); + ST->isVolatile(), ST->isNonTemporal(), Alignment); Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, DAG.getConstant(IncrementSize, TLI.getPointerTy())); Alignment = MinAlign(Alignment, IncrementSize); Store2 = DAG.getTruncStore(Chain, dl, TLI.isLittleEndian()?Hi:Lo, Ptr, ST->getSrcValue(), SVOffset + IncrementSize, - NewStoredVT, ST->isVolatile(), Alignment); + NewStoredVT, ST->isVolatile(), ST->isNonTemporal(), + Alignment); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Store1, Store2); } @@ -502,7 +508,7 @@ SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG, // then bitconvert to floating point or vector. SDValue newLoad = DAG.getLoad(intVT, dl, Chain, Ptr, LD->getSrcValue(), SVOffset, LD->isVolatile(), - LD->getAlignment()); + LD->isNonTemporal(), LD->getAlignment()); SDValue Result = DAG.getNode(ISD::BIT_CONVERT, dl, LoadedVT, newLoad); if (VT.isFloatingPoint() && LoadedVT != VT) Result = DAG.getNode(ISD::FP_EXTEND, dl, VT, Result); @@ -530,10 +536,11 @@ SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG, // Load one integer register's worth from the original location. SDValue Load = DAG.getLoad(RegVT, dl, Chain, Ptr, LD->getSrcValue(), SVOffset + Offset, LD->isVolatile(), + LD->isNonTemporal(), MinAlign(LD->getAlignment(), Offset)); // Follow the load with a store to the stack slot. Remember the store. Stores.push_back(DAG.getStore(Load.getValue(1), dl, Load, StackPtr, - NULL, 0)); + NULL, 0, false, false, 0)); // Increment the pointers. Offset += RegBytes; Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment); @@ -546,12 +553,13 @@ SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG, SDValue Load = DAG.getExtLoad(ISD::EXTLOAD, dl, RegVT, Chain, Ptr, LD->getSrcValue(), SVOffset + Offset, MemVT, LD->isVolatile(), + LD->isNonTemporal(), MinAlign(LD->getAlignment(), Offset)); // Follow the load with a store to the stack slot. Remember the store. // On big-endian machines this requires a truncating store to ensure // that the bits end up in the right place. Stores.push_back(DAG.getTruncStore(Load.getValue(1), dl, Load, StackPtr, - NULL, 0, MemVT)); + NULL, 0, MemVT, false, false, 0)); // The order of the stores doesn't matter - say it with a TokenFactor. SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Stores[0], @@ -559,7 +567,7 @@ SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG, // Finally, perform the original load only redirected to the stack slot. Load = DAG.getExtLoad(LD->getExtensionType(), dl, VT, TF, StackBase, - NULL, 0, LoadedVT); + NULL, 0, LoadedVT, false, false, 0); // Callers expect a MERGE_VALUES node. SDValue Ops[] = { Load, TF }; @@ -588,20 +596,22 @@ SDValue ExpandUnalignedLoad(LoadSDNode *LD, SelectionDAG &DAG, SDValue Lo, Hi; if (TLI.isLittleEndian()) { Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getSrcValue(), - SVOffset, NewLoadedVT, LD->isVolatile(), Alignment); + SVOffset, NewLoadedVT, LD->isVolatile(), + LD->isNonTemporal(), Alignment); Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, DAG.getConstant(IncrementSize, TLI.getPointerTy())); Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getSrcValue(), SVOffset + IncrementSize, NewLoadedVT, LD->isVolatile(), - MinAlign(Alignment, IncrementSize)); + LD->isNonTemporal(), MinAlign(Alignment, IncrementSize)); } else { Hi = DAG.getExtLoad(HiExtType, dl, VT, Chain, Ptr, LD->getSrcValue(), - SVOffset, NewLoadedVT, LD->isVolatile(), Alignment); + SVOffset, NewLoadedVT, LD->isVolatile(), + LD->isNonTemporal(), Alignment); Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, DAG.getConstant(IncrementSize, TLI.getPointerTy())); Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, VT, Chain, Ptr, LD->getSrcValue(), SVOffset + IncrementSize, NewLoadedVT, LD->isVolatile(), - MinAlign(Alignment, IncrementSize)); + LD->isNonTemporal(), MinAlign(Alignment, IncrementSize)); } // aggregate the two parts @@ -643,7 +653,8 @@ PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx, // Store the vector. SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Tmp1, StackPtr, - PseudoSourceValue::getFixedStack(SPFI), 0); + PseudoSourceValue::getFixedStack(SPFI), 0, + false, false, 0); // Truncate or zero extend offset to target pointer type. unsigned CastOpc = IdxVT.bitsGT(PtrVT) ? ISD::TRUNCATE : ISD::ZERO_EXTEND; @@ -654,10 +665,12 @@ PerformInsertVectorEltInMemory(SDValue Vec, SDValue Val, SDValue Idx, SDValue StackPtr2 = DAG.getNode(ISD::ADD, dl, IdxVT, Tmp3, StackPtr); // Store the scalar value. Ch = DAG.getTruncStore(Ch, dl, Tmp2, StackPtr2, - PseudoSourceValue::getFixedStack(SPFI), 0, EltVT); + PseudoSourceValue::getFixedStack(SPFI), 0, EltVT, + false, false, 0); // Load the updated vector. return DAG.getLoad(VT, dl, Ch, StackPtr, - PseudoSourceValue::getFixedStack(SPFI), 0); + PseudoSourceValue::getFixedStack(SPFI), 0, + false, false, 0); } @@ -702,6 +715,7 @@ SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) { int SVOffset = ST->getSrcValueOffset(); unsigned Alignment = ST->getAlignment(); bool isVolatile = ST->isVolatile(); + bool isNonTemporal = ST->isNonTemporal(); DebugLoc dl = ST->getDebugLoc(); if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(ST->getValue())) { if (CFP->getValueType(0) == MVT::f32 && @@ -710,14 +724,14 @@ SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) { bitcastToAPInt().zextOrTrunc(32), MVT::i32); return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), - SVOffset, isVolatile, Alignment); + SVOffset, isVolatile, isNonTemporal, Alignment); } else if (CFP->getValueType(0) == MVT::f64) { // If this target supports 64-bit registers, do a single 64-bit store. if (getTypeAction(MVT::i64) == Legal) { Tmp3 = DAG.getConstant(CFP->getValueAPF().bitcastToAPInt(). zextOrTrunc(64), MVT::i64); return DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), - SVOffset, isVolatile, Alignment); + SVOffset, isVolatile, isNonTemporal, Alignment); } else if (getTypeAction(MVT::i32) == Legal && !ST->isVolatile()) { // Otherwise, if the target supports 32-bit registers, use 2 32-bit // stores. If the target supports neither 32- nor 64-bits, this @@ -728,11 +742,11 @@ SDValue SelectionDAGLegalize::OptimizeFloatStore(StoreSDNode* ST) { if (TLI.isBigEndian()) std::swap(Lo, Hi); Lo = DAG.getStore(Tmp1, dl, Lo, Tmp2, ST->getSrcValue(), - SVOffset, isVolatile, Alignment); + SVOffset, isVolatile, isNonTemporal, Alignment); Tmp2 = DAG.getNode(ISD::ADD, dl, Tmp2.getValueType(), Tmp2, DAG.getIntPtrConstant(4)); Hi = DAG.getStore(Tmp1, dl, Hi, Tmp2, ST->getSrcValue(), SVOffset+4, - isVolatile, MinAlign(Alignment, 4U)); + isVolatile, isNonTemporal, MinAlign(Alignment, 4U)); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); } @@ -1108,7 +1122,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Tmp1 = DAG.getLoad(NVT, dl, Tmp1, Tmp2, LD->getSrcValue(), LD->getSrcValueOffset(), - LD->isVolatile(), LD->getAlignment()); + LD->isVolatile(), LD->isNonTemporal(), + LD->getAlignment()); Tmp3 = LegalizeOp(DAG.getNode(ISD::BIT_CONVERT, dl, VT, Tmp1)); Tmp4 = LegalizeOp(Tmp1.getValue(1)); break; @@ -1125,6 +1140,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { int SVOffset = LD->getSrcValueOffset(); unsigned Alignment = LD->getAlignment(); bool isVolatile = LD->isVolatile(); + bool isNonTemporal = LD->isNonTemporal(); if (SrcWidth != SrcVT.getStoreSizeInBits() && // Some targets pretend to have an i1 loading operation, and actually @@ -1150,7 +1166,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Result = DAG.getExtLoad(NewExtType, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), SVOffset, - NVT, isVolatile, Alignment); + NVT, isVolatile, isNonTemporal, Alignment); Ch = Result.getValue(1); // The chain. @@ -1187,7 +1203,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), SVOffset, RoundVT, isVolatile, - Alignment); + isNonTemporal, Alignment); // Load the remaining ExtraWidth bits. IncrementSize = RoundWidth / 8; @@ -1195,7 +1211,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { DAG.getIntPtrConstant(IncrementSize)); Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), SVOffset + IncrementSize, - ExtraVT, isVolatile, + ExtraVT, isVolatile, isNonTemporal, MinAlign(Alignment, IncrementSize)); // Build a factor node to remember that this load is independent of the @@ -1215,7 +1231,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { // Load the top RoundWidth bits. Hi = DAG.getExtLoad(ExtType, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), SVOffset, RoundVT, isVolatile, - Alignment); + isNonTemporal, Alignment); // Load the remaining ExtraWidth bits. IncrementSize = RoundWidth / 8; @@ -1224,7 +1240,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), SVOffset + IncrementSize, - ExtraVT, isVolatile, + ExtraVT, isVolatile, isNonTemporal, MinAlign(Alignment, IncrementSize)); // Build a factor node to remember that this load is independent of the @@ -1284,7 +1300,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { (SrcVT == MVT::f64 && Node->getValueType(0) == MVT::f128)) { SDValue Load = DAG.getLoad(SrcVT, dl, Tmp1, Tmp2, LD->getSrcValue(), LD->getSrcValueOffset(), - LD->isVolatile(), LD->getAlignment()); + LD->isVolatile(), LD->isNonTemporal(), + LD->getAlignment()); Result = DAG.getNode(ISD::FP_EXTEND, dl, Node->getValueType(0), Load); Tmp1 = LegalizeOp(Result); // Relegalize new nodes. @@ -1297,7 +1314,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Result = DAG.getExtLoad(ISD::EXTLOAD, dl, Node->getValueType(0), Tmp1, Tmp2, LD->getSrcValue(), LD->getSrcValueOffset(), SrcVT, - LD->isVolatile(), LD->getAlignment()); + LD->isVolatile(), LD->isNonTemporal(), + LD->getAlignment()); SDValue ValRes; if (ExtType == ISD::SEXTLOAD) ValRes = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, @@ -1325,6 +1343,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { int SVOffset = ST->getSrcValueOffset(); unsigned Alignment = ST->getAlignment(); bool isVolatile = ST->isVolatile(); + bool isNonTemporal = ST->isNonTemporal(); if (!ST->isTruncatingStore()) { if (SDNode *OptStore = OptimizeFloatStore(ST).getNode()) { @@ -1361,7 +1380,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { TLI.getTypeToPromoteTo(ISD::STORE, VT), Tmp3); Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), SVOffset, isVolatile, - Alignment); + isNonTemporal, Alignment); break; } break; @@ -1379,7 +1398,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { EVT NVT = EVT::getIntegerVT(*DAG.getContext(), StVT.getStoreSizeInBits()); Tmp3 = DAG.getZeroExtendInReg(Tmp3, dl, StVT); Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), - SVOffset, NVT, isVolatile, Alignment); + SVOffset, NVT, isVolatile, isNonTemporal, + Alignment); } else if (StWidth & (StWidth - 1)) { // If not storing a power-of-2 number of bits, expand as two stores. assert(!StVT.isVector() && "Unsupported truncstore!"); @@ -1399,7 +1419,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { // Store the bottom RoundWidth bits. Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), SVOffset, RoundVT, - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); // Store the remaining ExtraWidth bits. IncrementSize = RoundWidth / 8; @@ -1409,6 +1429,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { DAG.getConstant(RoundWidth, TLI.getShiftAmountTy())); Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getSrcValue(), SVOffset + IncrementSize, ExtraVT, isVolatile, + isNonTemporal, MinAlign(Alignment, IncrementSize)); } else { // Big endian - avoid unaligned stores. @@ -1417,7 +1438,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { Hi = DAG.getNode(ISD::SRL, dl, Tmp3.getValueType(), Tmp3, DAG.getConstant(ExtraWidth, TLI.getShiftAmountTy())); Hi = DAG.getTruncStore(Tmp1, dl, Hi, Tmp2, ST->getSrcValue(), - SVOffset, RoundVT, isVolatile, Alignment); + SVOffset, RoundVT, isVolatile, isNonTemporal, + Alignment); // Store the remaining ExtraWidth bits. IncrementSize = RoundWidth / 8; @@ -1425,6 +1447,7 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { DAG.getIntPtrConstant(IncrementSize)); Lo = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), SVOffset + IncrementSize, ExtraVT, isVolatile, + isNonTemporal, MinAlign(Alignment, IncrementSize)); } @@ -1457,7 +1480,8 @@ SDValue SelectionDAGLegalize::LegalizeOp(SDValue Op) { assert(isTypeLegal(StVT) && "Do not know how to expand this store!"); Tmp3 = DAG.getNode(ISD::TRUNCATE, dl, StVT, Tmp3); Result = DAG.getStore(Tmp1, dl, Tmp3, Tmp2, ST->getSrcValue(), - SVOffset, isVolatile, Alignment); + SVOffset, isVolatile, isNonTemporal, + Alignment); break; } } @@ -1484,7 +1508,8 @@ SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) { DebugLoc dl = Op.getDebugLoc(); // Store the value to a temporary stack slot, then LOAD the returned part. SDValue StackPtr = DAG.CreateStackTemporary(Vec.getValueType()); - SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0); + SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0, + false, false, 0); // Add the offset to the index. unsigned EltSize = @@ -1500,10 +1525,12 @@ SDValue SelectionDAGLegalize::ExpandExtractFromVectorThroughStack(SDValue Op) { StackPtr = DAG.getNode(ISD::ADD, dl, Idx.getValueType(), Idx, StackPtr); if (Op.getValueType().isVector()) - return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, NULL, 0); + return DAG.getLoad(Op.getValueType(), dl, Ch, StackPtr, NULL, 0, + false, false, 0); else return DAG.getExtLoad(ISD::EXTLOAD, dl, Op.getValueType(), Ch, StackPtr, - NULL, 0, Vec.getValueType().getVectorElementType()); + NULL, 0, Vec.getValueType().getVectorElementType(), + false, false, 0); } SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { @@ -1512,7 +1539,6 @@ SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { // the result as a vector. // Create the stack frame object. EVT VT = Node->getValueType(0); - EVT OpVT = Node->getOperand(0).getValueType(); EVT EltVT = VT.getVectorElementType(); DebugLoc dl = Node->getDebugLoc(); SDValue FIPtr = DAG.CreateStackTemporary(VT); @@ -1532,13 +1558,16 @@ SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { SDValue Idx = DAG.getConstant(Offset, FIPtr.getValueType()); Idx = DAG.getNode(ISD::ADD, dl, FIPtr.getValueType(), FIPtr, Idx); - // If EltVT smaller than OpVT, only store the bits necessary. - if (!OpVT.isVector() && EltVT.bitsLT(OpVT)) { + // If the destination vector element type is narrower than the source + // element type, only store the bits necessary. + if (EltVT.bitsLT(Node->getOperand(i).getValueType().getScalarType())) { Stores.push_back(DAG.getTruncStore(DAG.getEntryNode(), dl, - Node->getOperand(i), Idx, SV, Offset, EltVT)); + Node->getOperand(i), Idx, SV, Offset, + EltVT, false, false, 0)); } else Stores.push_back(DAG.getStore(DAG.getEntryNode(), dl, - Node->getOperand(i), Idx, SV, Offset)); + Node->getOperand(i), Idx, SV, Offset, + false, false, 0)); } SDValue StoreChain; @@ -1549,7 +1578,7 @@ SDValue SelectionDAGLegalize::ExpandVectorBuildThroughStack(SDNode* Node) { StoreChain = DAG.getEntryNode(); // Result is a load from the stack slot. - return DAG.getLoad(VT, dl, StoreChain, FIPtr, SV, 0); + return DAG.getLoad(VT, dl, StoreChain, FIPtr, SV, 0, false, false, 0); } SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) { @@ -1572,12 +1601,14 @@ SDValue SelectionDAGLegalize::ExpandFCOPYSIGN(SDNode* Node) { SDValue StackPtr = DAG.CreateStackTemporary(Tmp2.getValueType()); SDValue StorePtr = StackPtr, LoadPtr = StackPtr; SDValue Ch = - DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StorePtr, NULL, 0); + DAG.getStore(DAG.getEntryNode(), dl, Tmp2, StorePtr, NULL, 0, + false, false, 0); if (Tmp2.getValueType() == MVT::f64 && TLI.isLittleEndian()) LoadPtr = DAG.getNode(ISD::ADD, dl, StackPtr.getValueType(), LoadPtr, DAG.getIntPtrConstant(4)); SignBit = DAG.getExtLoad(ISD::SEXTLOAD, dl, TLI.getPointerTy(), - Ch, LoadPtr, NULL, 0, MVT::i32); + Ch, LoadPtr, NULL, 0, MVT::i32, + false, false, 0); } SignBit = DAG.getSetCC(dl, TLI.getSetCCResultType(SignBit.getValueType()), @@ -1701,20 +1732,21 @@ SDValue SelectionDAGLegalize::EmitStackConvert(SDValue SrcOp, if (SrcSize > SlotSize) Store = DAG.getTruncStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, - SV, 0, SlotVT, false, SrcAlign); + SV, 0, SlotVT, false, false, SrcAlign); else { assert(SrcSize == SlotSize && "Invalid store"); Store = DAG.getStore(DAG.getEntryNode(), dl, SrcOp, FIPtr, - SV, 0, false, SrcAlign); + SV, 0, false, false, SrcAlign); } // Result is a load from the stack slot. if (SlotSize == DestSize) - return DAG.getLoad(DestVT, dl, Store, FIPtr, SV, 0, false, DestAlign); + return DAG.getLoad(DestVT, dl, Store, FIPtr, SV, 0, false, false, + DestAlign); assert(SlotSize < DestSize && "Unknown extension!"); return DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, Store, FIPtr, SV, 0, SlotVT, - false, DestAlign); + false, false, DestAlign); } SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) { @@ -1729,9 +1761,11 @@ SDValue SelectionDAGLegalize::ExpandSCALAR_TO_VECTOR(SDNode *Node) { SDValue Ch = DAG.getTruncStore(DAG.getEntryNode(), dl, Node->getOperand(0), StackPtr, PseudoSourceValue::getFixedStack(SPFI), 0, - Node->getValueType(0).getVectorElementType()); + Node->getValueType(0).getVectorElementType(), + false, false, 0); return DAG.getLoad(Node->getValueType(0), dl, Ch, StackPtr, - PseudoSourceValue::getFixedStack(SPFI), 0); + PseudoSourceValue::getFixedStack(SPFI), 0, + false, false, 0); } @@ -1805,7 +1839,7 @@ SDValue SelectionDAGLegalize::ExpandBUILD_VECTOR(SDNode *Node) { unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment(); return DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx, PseudoSourceValue::getConstantPool(), 0, - false, Alignment); + false, false, Alignment); } if (!MoreThanTwoValues) { @@ -1865,8 +1899,7 @@ SDValue SelectionDAGLegalize::ExpandLibCall(RTLIB::Libcall LC, SDNode *Node, TLI.LowerCallTo(InChain, RetTy, isSigned, !isSigned, false, false, 0, TLI.getLibcallCallingConv(LC), false, /*isReturnValueUsed=*/true, - Callee, Args, DAG, - Node->getDebugLoc(), DAG.GetOrdering(Node)); + Callee, Args, DAG, Node->getDebugLoc()); // Legalize the call sequence, starting with the chain. This will advance // the LastCALLSEQ_END to the legalized version of the CALLSEQ_END node that @@ -1943,13 +1976,16 @@ SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, } // store the lo of the constructed double - based on integer input SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, - Op0Mapped, Lo, NULL, 0); + Op0Mapped, Lo, NULL, 0, + false, false, 0); // initial hi portion of constructed double SDValue InitialHi = DAG.getConstant(0x43300000u, MVT::i32); // store the hi of the constructed double - biased exponent - SDValue Store2=DAG.getStore(Store1, dl, InitialHi, Hi, NULL, 0); + SDValue Store2=DAG.getStore(Store1, dl, InitialHi, Hi, NULL, 0, + false, false, 0); // load the constructed double - SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot, NULL, 0); + SDValue Load = DAG.getLoad(MVT::f64, dl, Store2, StackSlot, NULL, 0, + false, false, 0); // FP constant to bias correct the final result SDValue Bias = DAG.getConstantFP(isSigned ? BitsToDouble(0x4330000080000000ULL) : @@ -1972,6 +2008,31 @@ SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, return Result; } assert(!isSigned && "Legalize cannot Expand SINT_TO_FP for i64 yet"); + + // Implementation of unsigned i64 to f64 following the algorithm in + // __floatundidf in compiler_rt. This implementation has the advantage + // of performing rounding correctly, both in the default rounding mode + // and in all alternate rounding modes. + // TODO: Generalize this for use with other types. + if (Op0.getValueType() == MVT::i64 && DestVT == MVT::f64) { + SDValue TwoP52 = + DAG.getConstant(UINT64_C(0x4330000000000000), MVT::i64); + SDValue TwoP84PlusTwoP52 = + DAG.getConstantFP(BitsToDouble(UINT64_C(0x4530000000100000)), MVT::f64); + SDValue TwoP84 = + DAG.getConstant(UINT64_C(0x4530000000000000), MVT::i64); + + SDValue Lo = DAG.getZeroExtendInReg(Op0, dl, MVT::i32); + SDValue Hi = DAG.getNode(ISD::SRL, dl, MVT::i64, Op0, + DAG.getConstant(32, MVT::i64)); + SDValue LoOr = DAG.getNode(ISD::OR, dl, MVT::i64, Lo, TwoP52); + SDValue HiOr = DAG.getNode(ISD::OR, dl, MVT::i64, Hi, TwoP84); + SDValue LoFlt = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, LoOr); + SDValue HiFlt = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, HiOr); + SDValue HiSub = DAG.getNode(ISD::FSUB, dl, MVT::f64, HiFlt, TwoP84PlusTwoP52); + return DAG.getNode(ISD::FADD, dl, MVT::f64, LoFlt, HiSub); + } + SDValue Tmp1 = DAG.getNode(ISD::SINT_TO_FP, dl, DestVT, Op0); SDValue SignSet = DAG.getSetCC(dl, TLI.getSetCCResultType(Op0.getValueType()), @@ -2004,13 +2065,13 @@ SDValue SelectionDAGLegalize::ExpandLegalINT_TO_FP(bool isSigned, if (DestVT == MVT::f32) FudgeInReg = DAG.getLoad(MVT::f32, dl, DAG.getEntryNode(), CPIdx, PseudoSourceValue::getConstantPool(), 0, - false, Alignment); + false, false, Alignment); else { FudgeInReg = LegalizeOp(DAG.getExtLoad(ISD::EXTLOAD, dl, DestVT, DAG.getEntryNode(), CPIdx, PseudoSourceValue::getConstantPool(), 0, - MVT::f32, false, Alignment)); + MVT::f32, false, false, Alignment)); } return DAG.getNode(ISD::FADD, dl, DestVT, Tmp1, FudgeInReg); @@ -2271,7 +2332,7 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node, false, false, false, false, 0, CallingConv::C, false, /*isReturnValueUsed=*/true, DAG.getExternalSymbol("abort", TLI.getPointerTy()), - Args, DAG, dl, DAG.GetOrdering(Node)); + Args, DAG, dl); Results.push_back(CallResult.second); break; } @@ -2350,16 +2411,19 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node, EVT VT = Node->getValueType(0); Tmp1 = Node->getOperand(0); Tmp2 = Node->getOperand(1); - SDValue VAList = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2, V, 0); + SDValue VAList = DAG.getLoad(TLI.getPointerTy(), dl, Tmp1, Tmp2, V, 0, + false, false, 0); // Increment the pointer, VAList, to the next vaarg Tmp3 = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(), VAList, DAG.getConstant(TLI.getTargetData()-> getTypeAllocSize(VT.getTypeForEVT(*DAG.getContext())), TLI.getPointerTy())); // Store the incremented VAList to the legalized pointer - Tmp3 = DAG.getStore(VAList.getValue(1), dl, Tmp3, Tmp2, V, 0); + Tmp3 = DAG.getStore(VAList.getValue(1), dl, Tmp3, Tmp2, V, 0, + false, false, 0); // Load the actual argument out of the pointer VAList - Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, NULL, 0)); + Results.push_back(DAG.getLoad(VT, dl, Tmp3, VAList, NULL, 0, + false, false, 0)); Results.push_back(Results[0].getValue(1)); break; } @@ -2369,8 +2433,9 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node, const Value *VD = cast<SrcValueSDNode>(Node->getOperand(3))->getValue(); const Value *VS = cast<SrcValueSDNode>(Node->getOperand(4))->getValue(); Tmp1 = DAG.getLoad(TLI.getPointerTy(), dl, Node->getOperand(0), - Node->getOperand(2), VS, 0); - Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1), VD, 0); + Node->getOperand(2), VS, 0, false, false, 0); + Tmp1 = DAG.getStore(Tmp1.getValue(1), dl, Tmp1, Node->getOperand(1), VD, 0, + false, false, 0); Results.push_back(Tmp1); break; } @@ -2827,7 +2892,8 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node, EVT MemVT = EVT::getIntegerVT(*DAG.getContext(), EntrySize * 8); SDValue LD = DAG.getExtLoad(ISD::SEXTLOAD, dl, PTy, Chain, Addr, - PseudoSourceValue::getJumpTable(), 0, MemVT); + PseudoSourceValue::getJumpTable(), 0, MemVT, + false, false, 0); Addr = LD; if (TLI.getTargetMachine().getRelocationModel() == Reloc::PIC_) { // For PIC, the sequence is: diff --git a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp index 4f0fce7..35a7c7c 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeFloatTypes.cpp @@ -444,7 +444,7 @@ SDValue DAGTypeLegalizer::SoftenFloatRes_LOAD(SDNode *N) { NewL = DAG.getLoad(L->getAddressingMode(), dl, L->getExtensionType(), NVT, L->getChain(), L->getBasePtr(), L->getOffset(), L->getSrcValue(), L->getSrcValueOffset(), NVT, - L->isVolatile(), L->getAlignment()); + L->isVolatile(), L->isNonTemporal(), L->getAlignment()); // Legalized the chain result - switch anything that used the old chain to // use the new one. ReplaceValueWith(SDValue(N, 1), NewL.getValue(1)); @@ -456,8 +456,8 @@ SDValue DAGTypeLegalizer::SoftenFloatRes_LOAD(SDNode *N) { L->getMemoryVT(), L->getChain(), L->getBasePtr(), L->getOffset(), L->getSrcValue(), L->getSrcValueOffset(), - L->getMemoryVT(), - L->isVolatile(), L->getAlignment()); + L->getMemoryVT(), L->isVolatile(), + L->isNonTemporal(), L->getAlignment()); // Legalized the chain result - switch anything that used the old chain to // use the new one. ReplaceValueWith(SDValue(N, 1), NewL.getValue(1)); @@ -755,7 +755,8 @@ SDValue DAGTypeLegalizer::SoftenFloatOp_STORE(SDNode *N, unsigned OpNo) { return DAG.getStore(ST->getChain(), dl, Val, ST->getBasePtr(), ST->getSrcValue(), ST->getSrcValueOffset(), - ST->isVolatile(), ST->getAlignment()); + ST->isVolatile(), ST->isNonTemporal(), + ST->getAlignment()); } @@ -1073,8 +1074,8 @@ void DAGTypeLegalizer::ExpandFloatRes_LOAD(SDNode *N, SDValue &Lo, Hi = DAG.getExtLoad(LD->getExtensionType(), dl, NVT, Chain, Ptr, LD->getSrcValue(), LD->getSrcValueOffset(), - LD->getMemoryVT(), - LD->isVolatile(), LD->getAlignment()); + LD->getMemoryVT(), LD->isVolatile(), + LD->isNonTemporal(), LD->getAlignment()); // Remember the chain. Chain = Hi.getValue(1); @@ -1382,6 +1383,6 @@ SDValue DAGTypeLegalizer::ExpandFloatOp_STORE(SDNode *N, unsigned OpNo) { return DAG.getTruncStore(Chain, N->getDebugLoc(), Hi, Ptr, ST->getSrcValue(), ST->getSrcValueOffset(), - ST->getMemoryVT(), - ST->isVolatile(), ST->getAlignment()); + ST->getMemoryVT(), ST->isVolatile(), + ST->isNonTemporal(), ST->getAlignment()); } diff --git a/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp index 9932cf4..81f28ad 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp @@ -359,7 +359,7 @@ SDValue DAGTypeLegalizer::PromoteIntRes_LOAD(LoadSDNode *N) { SDValue Res = DAG.getExtLoad(ExtType, dl, NVT, N->getChain(), N->getBasePtr(), N->getSrcValue(), N->getSrcValueOffset(), N->getMemoryVT(), N->isVolatile(), - N->getAlignment()); + N->isNonTemporal(), N->getAlignment()); // Legalized the chain result - switch anything that used the old chain to // use the new one. @@ -873,6 +873,7 @@ SDValue DAGTypeLegalizer::PromoteIntOp_STORE(StoreSDNode *N, unsigned OpNo){ int SVOffset = N->getSrcValueOffset(); unsigned Alignment = N->getAlignment(); bool isVolatile = N->isVolatile(); + bool isNonTemporal = N->isNonTemporal(); DebugLoc dl = N->getDebugLoc(); SDValue Val = GetPromotedInteger(N->getValue()); // Get promoted value. @@ -880,7 +881,7 @@ SDValue DAGTypeLegalizer::PromoteIntOp_STORE(StoreSDNode *N, unsigned OpNo){ // Truncate the value and store the result. return DAG.getTruncStore(Ch, dl, Val, Ptr, N->getSrcValue(), SVOffset, N->getMemoryVT(), - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); } SDValue DAGTypeLegalizer::PromoteIntOp_TRUNCATE(SDNode *N) { @@ -1079,8 +1080,8 @@ ExpandShiftWithKnownAmountBit(SDNode *N, SDValue &Lo, SDValue &Hi) { SDValue Amt = N->getOperand(1); EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); EVT ShTy = Amt.getValueType(); - unsigned ShBits = ShTy.getSizeInBits(); - unsigned NVTBits = NVT.getSizeInBits(); + unsigned ShBits = ShTy.getScalarType().getSizeInBits(); + unsigned NVTBits = NVT.getScalarType().getSizeInBits(); assert(isPowerOf2_32(NVTBits) && "Expanded integer type size not a power of two!"); DebugLoc dl = N->getDebugLoc(); @@ -1500,6 +1501,7 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, int SVOffset = N->getSrcValueOffset(); unsigned Alignment = N->getAlignment(); bool isVolatile = N->isVolatile(); + bool isNonTemporal = N->isNonTemporal(); DebugLoc dl = N->getDebugLoc(); assert(NVT.isByteSized() && "Expanded type not byte sized!"); @@ -1508,7 +1510,7 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, EVT MemVT = N->getMemoryVT(); Lo = DAG.getExtLoad(ExtType, dl, NVT, Ch, Ptr, N->getSrcValue(), SVOffset, - MemVT, isVolatile, Alignment); + MemVT, isVolatile, isNonTemporal, Alignment); // Remember the chain. Ch = Lo.getValue(1); @@ -1530,7 +1532,7 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, } else if (TLI.isLittleEndian()) { // Little-endian - low bits are at low addresses. Lo = DAG.getLoad(NVT, dl, Ch, Ptr, N->getSrcValue(), SVOffset, - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); unsigned ExcessBits = N->getMemoryVT().getSizeInBits() - NVT.getSizeInBits(); @@ -1542,7 +1544,8 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, DAG.getIntPtrConstant(IncrementSize)); Hi = DAG.getExtLoad(ExtType, dl, NVT, Ch, Ptr, N->getSrcValue(), SVOffset+IncrementSize, NEVT, - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); // Build a factor node to remember that this load is independent of the // other one. @@ -1560,7 +1563,7 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, Hi = DAG.getExtLoad(ExtType, dl, NVT, Ch, Ptr, N->getSrcValue(), SVOffset, EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits() - ExcessBits), - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); // Increment the pointer to the other half. Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, @@ -1569,7 +1572,8 @@ void DAGTypeLegalizer::ExpandIntRes_LOAD(LoadSDNode *N, Lo = DAG.getExtLoad(ISD::ZEXTLOAD, dl, NVT, Ch, Ptr, N->getSrcValue(), SVOffset+IncrementSize, EVT::getIntegerVT(*DAG.getContext(), ExcessBits), - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); // Build a factor node to remember that this load is independent of the // other one. @@ -2212,6 +2216,7 @@ SDValue DAGTypeLegalizer::ExpandIntOp_STORE(StoreSDNode *N, unsigned OpNo) { int SVOffset = N->getSrcValueOffset(); unsigned Alignment = N->getAlignment(); bool isVolatile = N->isVolatile(); + bool isNonTemporal = N->isNonTemporal(); DebugLoc dl = N->getDebugLoc(); SDValue Lo, Hi; @@ -2220,13 +2225,14 @@ SDValue DAGTypeLegalizer::ExpandIntOp_STORE(StoreSDNode *N, unsigned OpNo) { if (N->getMemoryVT().bitsLE(NVT)) { GetExpandedInteger(N->getValue(), Lo, Hi); return DAG.getTruncStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, - N->getMemoryVT(), isVolatile, Alignment); + N->getMemoryVT(), isVolatile, isNonTemporal, + Alignment); } else if (TLI.isLittleEndian()) { // Little-endian - low bits are at low addresses. GetExpandedInteger(N->getValue(), Lo, Hi); Lo = DAG.getStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); unsigned ExcessBits = N->getMemoryVT().getSizeInBits() - NVT.getSizeInBits(); @@ -2238,7 +2244,8 @@ SDValue DAGTypeLegalizer::ExpandIntOp_STORE(StoreSDNode *N, unsigned OpNo) { DAG.getIntPtrConstant(IncrementSize)); Hi = DAG.getTruncStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset+IncrementSize, NEVT, - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); } else { // Big-endian - high bits are at low addresses. Favor aligned stores at @@ -2264,7 +2271,8 @@ SDValue DAGTypeLegalizer::ExpandIntOp_STORE(StoreSDNode *N, unsigned OpNo) { // Store both the high bits and maybe some of the low bits. Hi = DAG.getTruncStore(Ch, dl, Hi, Ptr, N->getSrcValue(), - SVOffset, HiVT, isVolatile, Alignment); + SVOffset, HiVT, isVolatile, isNonTemporal, + Alignment); // Increment the pointer to the other half. Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, @@ -2273,7 +2281,8 @@ SDValue DAGTypeLegalizer::ExpandIntOp_STORE(StoreSDNode *N, unsigned OpNo) { Lo = DAG.getTruncStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset+IncrementSize, EVT::getIntegerVT(*DAG.getContext(), ExcessBits), - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); } } @@ -2341,7 +2350,7 @@ SDValue DAGTypeLegalizer::ExpandIntOp_UINT_TO_FP(SDNode *N) { // FIXME: Avoid the extend by constructing the right constant pool? SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, DstVT, DAG.getEntryNode(), FudgePtr, NULL, 0, MVT::f32, - false, Alignment); + false, false, Alignment); return DAG.getNode(ISD::FADD, dl, DstVT, SignedConv, Fudge); } diff --git a/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp index 37f36a3..f3e7ca4 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeTypes.cpp @@ -871,9 +871,10 @@ SDValue DAGTypeLegalizer::CreateStackStoreLoad(SDValue Op, // the source and destination types. SDValue StackPtr = DAG.CreateStackTemporary(Op.getValueType(), DestVT); // Emit a store to the stack slot. - SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op, StackPtr, NULL, 0); + SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op, StackPtr, NULL, 0, + false, false, 0); // Result is a load from the stack slot. - return DAG.getLoad(DestVT, dl, Store, StackPtr, NULL, 0); + return DAG.getLoad(DestVT, dl, Store, StackPtr, NULL, 0, false, false, 0); } /// CustomLowerNode - Replace the node's results with custom code provided @@ -1033,8 +1034,7 @@ SDValue DAGTypeLegalizer::MakeLibCall(RTLIB::Libcall LC, EVT RetVT, TLI.LowerCallTo(DAG.getEntryNode(), RetTy, isSigned, !isSigned, false, false, 0, TLI.getLibcallCallingConv(LC), false, /*isReturnValueUsed=*/true, - Callee, Args, DAG, dl, - DAG.GetOrdering(DAG.getEntryNode().getNode())); + Callee, Args, DAG, dl); return CallInfo.first; } diff --git a/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp b/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp index a1b6ced..5e83b4b 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeTypesGeneric.cpp @@ -122,10 +122,11 @@ void DAGTypeLegalizer::ExpandRes_BIT_CONVERT(SDNode *N, SDValue &Lo, const Value *SV = PseudoSourceValue::getFixedStack(SPFI); // Emit a store to the stack slot. - SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, SV, 0); + SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, InOp, StackPtr, SV, 0, + false, false, 0); // Load the first half from the stack slot. - Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, 0); + Lo = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, 0, false, false, 0); // Increment the pointer to the other half. unsigned IncrementSize = NOutVT.getSizeInBits() / 8; @@ -134,7 +135,7 @@ void DAGTypeLegalizer::ExpandRes_BIT_CONVERT(SDNode *N, SDValue &Lo, // Load the second half from the stack slot. Hi = DAG.getLoad(NOutVT, dl, Store, StackPtr, SV, IncrementSize, false, - MinAlign(Alignment, IncrementSize)); + false, MinAlign(Alignment, IncrementSize)); // Handle endianness of the load. if (TLI.isBigEndian()) @@ -205,11 +206,12 @@ void DAGTypeLegalizer::ExpandRes_NormalLoad(SDNode *N, SDValue &Lo, int SVOffset = LD->getSrcValueOffset(); unsigned Alignment = LD->getAlignment(); bool isVolatile = LD->isVolatile(); + bool isNonTemporal = LD->isNonTemporal(); assert(NVT.isByteSized() && "Expanded type not byte sized!"); Lo = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getSrcValue(), SVOffset, - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); // Increment the pointer to the other half. unsigned IncrementSize = NVT.getSizeInBits() / 8; @@ -217,7 +219,8 @@ void DAGTypeLegalizer::ExpandRes_NormalLoad(SDNode *N, SDValue &Lo, DAG.getIntPtrConstant(IncrementSize)); Hi = DAG.getLoad(NVT, dl, Chain, Ptr, LD->getSrcValue(), SVOffset+IncrementSize, - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); // Build a factor node to remember that this load is independent of the // other one. @@ -383,6 +386,7 @@ SDValue DAGTypeLegalizer::ExpandOp_NormalStore(SDNode *N, unsigned OpNo) { int SVOffset = St->getSrcValueOffset(); unsigned Alignment = St->getAlignment(); bool isVolatile = St->isVolatile(); + bool isNonTemporal = St->isNonTemporal(); assert(NVT.isByteSized() && "Expanded type not byte sized!"); unsigned IncrementSize = NVT.getSizeInBits() / 8; @@ -394,14 +398,15 @@ SDValue DAGTypeLegalizer::ExpandOp_NormalStore(SDNode *N, unsigned OpNo) { std::swap(Lo, Hi); Lo = DAG.getStore(Chain, dl, Lo, Ptr, St->getSrcValue(), SVOffset, - isVolatile, Alignment); + isVolatile, isNonTemporal, Alignment); Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, DAG.getIntPtrConstant(IncrementSize)); assert(isTypeLegal(Ptr.getValueType()) && "Pointers must be legal!"); Hi = DAG.getStore(Chain, dl, Hi, Ptr, St->getSrcValue(), SVOffset + IncrementSize, - isVolatile, MinAlign(Alignment, IncrementSize)); + isVolatile, isNonTemporal, + MinAlign(Alignment, IncrementSize)); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); } diff --git a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp index bf95bb5..8363c3a 100644 --- a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp +++ b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp @@ -172,7 +172,8 @@ SDValue DAGTypeLegalizer::ScalarizeVecRes_LOAD(LoadSDNode *N) { DAG.getUNDEF(N->getBasePtr().getValueType()), N->getSrcValue(), N->getSrcValueOffset(), N->getMemoryVT().getVectorElementType(), - N->isVolatile(), N->getOriginalAlignment()); + N->isVolatile(), N->isNonTemporal(), + N->getOriginalAlignment()); // Legalized the chain result - switch anything that used the old chain to // use the new one. @@ -366,11 +367,13 @@ SDValue DAGTypeLegalizer::ScalarizeVecOp_STORE(StoreSDNode *N, unsigned OpNo){ N->getBasePtr(), N->getSrcValue(), N->getSrcValueOffset(), N->getMemoryVT().getVectorElementType(), - N->isVolatile(), N->getAlignment()); + N->isVolatile(), N->isNonTemporal(), + N->getAlignment()); return DAG.getStore(N->getChain(), dl, GetScalarizedVector(N->getOperand(1)), N->getBasePtr(), N->getSrcValue(), N->getSrcValueOffset(), - N->isVolatile(), N->getOriginalAlignment()); + N->isVolatile(), N->isNonTemporal(), + N->getOriginalAlignment()); } @@ -696,17 +699,20 @@ void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, EVT VecVT = Vec.getValueType(); EVT EltVT = VecVT.getVectorElementType(); SDValue StackPtr = DAG.CreateStackTemporary(VecVT); - SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0); + SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, NULL, 0, + false, false, 0); // Store the new element. This may be larger than the vector element type, // so use a truncating store. SDValue EltPtr = GetVectorElementPointer(StackPtr, EltVT, Idx); unsigned Alignment = TLI.getTargetData()->getPrefTypeAlignment(VecVT.getTypeForEVT(*DAG.getContext())); - Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, NULL, 0, EltVT); + Store = DAG.getTruncStore(Store, dl, Elt, EltPtr, NULL, 0, EltVT, + false, false, 0); // Load the Lo part from the stack slot. - Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, NULL, 0); + Lo = DAG.getLoad(Lo.getValueType(), dl, Store, StackPtr, NULL, 0, + false, false, 0); // Increment the pointer to the other part. unsigned IncrementSize = Lo.getValueType().getSizeInBits() / 8; @@ -715,7 +721,7 @@ void DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, // Load the Hi part from the stack slot. Hi = DAG.getLoad(Hi.getValueType(), dl, Store, StackPtr, NULL, 0, false, - MinAlign(Alignment, IncrementSize)); + false, MinAlign(Alignment, IncrementSize)); } void DAGTypeLegalizer::SplitVecRes_SCALAR_TO_VECTOR(SDNode *N, SDValue &Lo, @@ -743,19 +749,20 @@ void DAGTypeLegalizer::SplitVecRes_LOAD(LoadSDNode *LD, SDValue &Lo, EVT MemoryVT = LD->getMemoryVT(); unsigned Alignment = LD->getOriginalAlignment(); bool isVolatile = LD->isVolatile(); + bool isNonTemporal = LD->isNonTemporal(); EVT LoMemVT, HiMemVT; GetSplitDestVTs(MemoryVT, LoMemVT, HiMemVT); Lo = DAG.getLoad(ISD::UNINDEXED, dl, ExtType, LoVT, Ch, Ptr, Offset, - SV, SVOffset, LoMemVT, isVolatile, Alignment); + SV, SVOffset, LoMemVT, isVolatile, isNonTemporal, Alignment); unsigned IncrementSize = LoMemVT.getSizeInBits()/8; Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, DAG.getIntPtrConstant(IncrementSize)); SVOffset += IncrementSize; Hi = DAG.getLoad(ISD::UNINDEXED, dl, ExtType, HiVT, Ch, Ptr, Offset, - SV, SVOffset, HiMemVT, isVolatile, Alignment); + SV, SVOffset, HiMemVT, isVolatile, isNonTemporal, Alignment); // Build a factor node to remember that this load is independent of the // other one. @@ -1086,12 +1093,13 @@ SDValue DAGTypeLegalizer::SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N) { SDValue StackPtr = DAG.CreateStackTemporary(VecVT); int SPFI = cast<FrameIndexSDNode>(StackPtr.getNode())->getIndex(); const Value *SV = PseudoSourceValue::getFixedStack(SPFI); - SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, SV, 0); + SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Vec, StackPtr, SV, 0, + false, false, 0); // Load back the required element. StackPtr = GetVectorElementPointer(StackPtr, EltVT, Idx); return DAG.getExtLoad(ISD::EXTLOAD, dl, N->getValueType(0), Store, StackPtr, - SV, 0, EltVT); + SV, 0, EltVT, false, false, 0); } SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { @@ -1106,6 +1114,7 @@ SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { EVT MemoryVT = N->getMemoryVT(); unsigned Alignment = N->getOriginalAlignment(); bool isVol = N->isVolatile(); + bool isNT = N->isNonTemporal(); SDValue Lo, Hi; GetSplitVector(N->getOperand(1), Lo, Hi); @@ -1116,10 +1125,10 @@ SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { if (isTruncating) Lo = DAG.getTruncStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, - LoMemVT, isVol, Alignment); + LoMemVT, isVol, isNT, Alignment); else Lo = DAG.getStore(Ch, dl, Lo, Ptr, N->getSrcValue(), SVOffset, - isVol, Alignment); + isVol, isNT, Alignment); // Increment the pointer to the other half. Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, @@ -1128,10 +1137,10 @@ SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) { if (isTruncating) Hi = DAG.getTruncStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset, - HiMemVT, isVol, Alignment); + HiMemVT, isVol, isNT, Alignment); else Hi = DAG.getStore(Ch, dl, Hi, Ptr, N->getSrcValue(), SVOffset, - isVol, Alignment); + isVol, isNT, Alignment); return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo, Hi); } @@ -1242,10 +1251,96 @@ void DAGTypeLegalizer::WidenVectorResult(SDNode *N, unsigned ResNo) { SDValue DAGTypeLegalizer::WidenVecRes_Binary(SDNode *N) { // Binary op widening. + unsigned Opcode = N->getOpcode(); + DebugLoc dl = N->getDebugLoc(); EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0)); - SDValue InOp1 = GetWidenedVector(N->getOperand(0)); - SDValue InOp2 = GetWidenedVector(N->getOperand(1)); - return DAG.getNode(N->getOpcode(), N->getDebugLoc(), WidenVT, InOp1, InOp2); + EVT WidenEltVT = WidenVT.getVectorElementType(); + EVT VT = WidenVT; + unsigned NumElts = VT.getVectorNumElements(); + while (!TLI.isTypeLegal(VT) && NumElts != 1) { + NumElts = NumElts / 2; + VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); + } + + if (NumElts != 1 && !TLI.canOpTrap(N->getOpcode(), VT)) { + // Operation doesn't trap so just widen as normal. + SDValue InOp1 = GetWidenedVector(N->getOperand(0)); + SDValue InOp2 = GetWidenedVector(N->getOperand(1)); + return DAG.getNode(N->getOpcode(), dl, WidenVT, InOp1, InOp2); + } else if (NumElts == 1) { + // No legal vector version so unroll the vector operation and then widen. + return DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()); + } else { + // Since the operation can trap, apply operation on the original vector. + SDValue InOp1 = GetWidenedVector(N->getOperand(0)); + SDValue InOp2 = GetWidenedVector(N->getOperand(1)); + unsigned CurNumElts = N->getValueType(0).getVectorNumElements(); + + SmallVector<SDValue, 16> ConcatOps(CurNumElts); + unsigned ConcatEnd = 0; // Current ConcatOps index. + unsigned Idx = 0; // Current Idx into input vectors. + while (CurNumElts != 0) { + while (CurNumElts >= NumElts) { + SDValue EOp1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp1, + DAG.getIntPtrConstant(Idx)); + SDValue EOp2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, InOp2, + DAG.getIntPtrConstant(Idx)); + ConcatOps[ConcatEnd++] = DAG.getNode(Opcode, dl, VT, EOp1, EOp2); + Idx += NumElts; + CurNumElts -= NumElts; + } + EVT PrevVecVT = VT; + do { + NumElts = NumElts / 2; + VT = EVT::getVectorVT(*DAG.getContext(), WidenEltVT, NumElts); + } while (!TLI.isTypeLegal(VT) && NumElts != 1); + + if (NumElts == 1) { + // Since we are using concat vector, build a vector from the scalar ops. + SDValue VecOp = DAG.getUNDEF(PrevVecVT); + for (unsigned i = 0; i != CurNumElts; ++i, ++Idx) { + SDValue EOp1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, + InOp1, DAG.getIntPtrConstant(Idx)); + SDValue EOp2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, WidenEltVT, + InOp2, DAG.getIntPtrConstant(Idx)); + VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, PrevVecVT, VecOp, + DAG.getNode(Opcode, dl, WidenEltVT, EOp1, EOp2), + DAG.getIntPtrConstant(i)); + } + CurNumElts = 0; + ConcatOps[ConcatEnd++] = VecOp; + } + } + + // Check to see if we have a single operation with the widen type. + if (ConcatEnd == 1) { + VT = ConcatOps[0].getValueType(); + if (VT == WidenVT) + return ConcatOps[0]; + } + + // Rebuild vector to one with the widen type + Idx = ConcatEnd - 1; + while (Idx != 0) { + VT = ConcatOps[Idx--].getValueType(); + while (Idx != 0 && ConcatOps[Idx].getValueType() == VT) + --Idx; + if (Idx != 0) { + VT = ConcatOps[Idx].getValueType(); + ConcatOps[Idx+1] = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, + &ConcatOps[Idx+1], ConcatEnd - Idx - 1); + ConcatEnd = Idx + 2; + } + } + + unsigned NumOps = WidenVT.getVectorNumElements()/VT.getVectorNumElements(); + if (NumOps != ConcatEnd ) { + SDValue UndefVal = DAG.getUNDEF(VT); + for (unsigned j = ConcatEnd; j < NumOps; ++j) + ConcatOps[j] = UndefVal; + } + return DAG.getNode(ISD::CONCAT_VECTORS, dl, WidenVT, &ConcatOps[0], NumOps); + } } SDValue DAGTypeLegalizer::WidenVecRes_Convert(SDNode *N) { @@ -2042,6 +2137,7 @@ SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain, int SVOffset = LD->getSrcValueOffset(); unsigned Align = LD->getAlignment(); bool isVolatile = LD->isVolatile(); + bool isNonTemporal = LD->isNonTemporal(); const Value *SV = LD->getSrcValue(); int LdWidth = LdVT.getSizeInBits(); @@ -2052,7 +2148,7 @@ SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain, EVT NewVT = FindMemType(DAG, TLI, LdWidth, WidenVT, LdAlign, WidthDiff); int NewVTWidth = NewVT.getSizeInBits(); SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, SV, SVOffset, - isVolatile, Align); + isVolatile, isNonTemporal, Align); LdChain.push_back(LdOp.getValue(1)); // Check if we can load the element with one instruction @@ -2099,7 +2195,7 @@ SDValue DAGTypeLegalizer::GenWidenVectorLoads(SmallVector<SDValue, 16>& LdChain, SDValue LdOp = DAG.getLoad(NewVT, dl, Chain, BasePtr, SV, SVOffset+Offset, isVolatile, - MinAlign(Align, Increment)); + isNonTemporal, MinAlign(Align, Increment)); LdChain.push_back(LdOp.getValue(1)); LdOps.push_back(LdOp); @@ -2173,6 +2269,7 @@ DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVector<SDValue, 16>& LdChain, int SVOffset = LD->getSrcValueOffset(); unsigned Align = LD->getAlignment(); bool isVolatile = LD->isVolatile(); + bool isNonTemporal = LD->isNonTemporal(); const Value *SV = LD->getSrcValue(); EVT EltVT = WidenVT.getVectorElementType(); @@ -2184,14 +2281,15 @@ DAGTypeLegalizer::GenWidenVectorExtLoads(SmallVector<SDValue, 16>& LdChain, SmallVector<SDValue, 16> Ops(WidenNumElts); unsigned Increment = LdEltVT.getSizeInBits() / 8; Ops[0] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, BasePtr, SV, SVOffset, - LdEltVT, isVolatile, Align); + LdEltVT, isVolatile, isNonTemporal, Align); LdChain.push_back(Ops[0].getValue(1)); unsigned i = 0, Offset = Increment; for (i=1; i < NumElts; ++i, Offset += Increment) { SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, DAG.getIntPtrConstant(Offset)); Ops[i] = DAG.getExtLoad(ExtType, dl, EltVT, Chain, NewBasePtr, SV, - SVOffset + Offset, LdEltVT, isVolatile, Align); + SVOffset + Offset, LdEltVT, isVolatile, + isNonTemporal, Align); LdChain.push_back(Ops[i].getValue(1)); } @@ -2215,6 +2313,7 @@ void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain, int SVOffset = ST->getSrcValueOffset(); unsigned Align = ST->getAlignment(); bool isVolatile = ST->isVolatile(); + bool isNonTemporal = ST->isNonTemporal(); SDValue ValOp = GetWidenedVector(ST->getValue()); DebugLoc dl = ST->getDebugLoc(); @@ -2240,6 +2339,7 @@ void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain, DAG.getIntPtrConstant(Idx)); StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, SV, SVOffset + Offset, isVolatile, + isNonTemporal, MinAlign(Align, Offset))); StWidth -= NewVTWidth; Offset += Increment; @@ -2258,8 +2358,8 @@ void DAGTypeLegalizer::GenWidenVectorStores(SmallVector<SDValue, 16>& StChain, SDValue EOp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, NewVT, VecOp, DAG.getIntPtrConstant(Idx++)); StChain.push_back(DAG.getStore(Chain, dl, EOp, BasePtr, SV, - SVOffset + Offset, isVolatile, - MinAlign(Align, Offset))); + SVOffset + Offset, isVolatile, + isNonTemporal, MinAlign(Align, Offset))); StWidth -= NewVTWidth; Offset += Increment; BasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), BasePtr, @@ -2282,6 +2382,7 @@ DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVector<SDValue, 16>& StChain, int SVOffset = ST->getSrcValueOffset(); unsigned Align = ST->getAlignment(); bool isVolatile = ST->isVolatile(); + bool isNonTemporal = ST->isNonTemporal(); SDValue ValOp = GetWidenedVector(ST->getValue()); DebugLoc dl = ST->getDebugLoc(); @@ -2304,7 +2405,7 @@ DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVector<SDValue, 16>& StChain, DAG.getIntPtrConstant(0)); StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, BasePtr, SV, SVOffset, StEltVT, - isVolatile, Align)); + isVolatile, isNonTemporal, Align)); unsigned Offset = Increment; for (unsigned i=1; i < NumElts; ++i, Offset += Increment) { SDValue NewBasePtr = DAG.getNode(ISD::ADD, dl, BasePtr.getValueType(), @@ -2313,7 +2414,8 @@ DAGTypeLegalizer::GenWidenVectorTruncStores(SmallVector<SDValue, 16>& StChain, DAG.getIntPtrConstant(0)); StChain.push_back(DAG.getTruncStore(Chain, dl, EOp, NewBasePtr, SV, SVOffset + Offset, StEltVT, - isVolatile, MinAlign(Align, Offset))); + isVolatile, isNonTemporal, + MinAlign(Align, Offset))); } } diff --git a/lib/CodeGen/SelectionDAG/SDDbgValue.h b/lib/CodeGen/SelectionDAG/SDDbgValue.h new file mode 100644 index 0000000..9e15fc9 --- /dev/null +++ b/lib/CodeGen/SelectionDAG/SDDbgValue.h @@ -0,0 +1,67 @@ +//===-- llvm/CodeGen/SDDbgValue.h - SD dbg_value handling--------*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file declares the SDDbgValue class. +// +//===----------------------------------------------------------------------===// + +#ifndef LLVM_CODEGEN_SDDBGVALUE_H +#define LLVM_CODEGEN_SDDBGVALUE_H + +#include "llvm/ADT/SmallVector.h" +#include "llvm/Support/DebugLoc.h" + +namespace llvm { + +class MDNode; +class SDNode; +class Value; + +/// SDDbgValue - Holds the information from a dbg_value node through SDISel. +/// Either Const or Node is nonzero, but not both. +/// We do not use SDValue here to avoid including its header. + +class SDDbgValue { + SDNode *Node; // valid for non-constants + unsigned ResNo; // valid for non-constants + Value *Const; // valid for constants + MDNode *mdPtr; + uint64_t Offset; + DebugLoc DL; +public: + // Constructor for non-constants. + SDDbgValue(MDNode *mdP, SDNode *N, unsigned R, uint64_t off, DebugLoc dl) : + Node(N), ResNo(R), Const(0), mdPtr(mdP), Offset(off), DL(dl) {} + + // Constructor for constants. + SDDbgValue(MDNode *mdP, Value *C, uint64_t off, DebugLoc dl) : Node(0), + ResNo(0), Const(C), mdPtr(mdP), Offset(off), DL(dl) {} + + // Returns the MDNode pointer. + MDNode *getMDPtr() { return mdPtr; } + + // Returns the SDNode* (valid for non-constants only). + SDNode *getSDNode() { assert (!Const); return Node; } + + // Returns the ResNo (valid for non-constants only). + unsigned getResNo() { assert (!Const); return ResNo; } + + // Returns the Value* for a constant (invalid for non-constants). + Value *getConst() { assert (!Node); return Const; } + + // Returns the offset. + uint64_t getOffset() { return Offset; } + + // Returns the DebugLoc. + DebugLoc getDebugLoc() { return DL; } +}; + +} // end llvm namespace + +#endif diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp index b51c61b..06e7b8c 100644 --- a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp +++ b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp @@ -218,8 +218,20 @@ void ScheduleDAGSDNodes::BuildSchedUnits() { // Check to see if the scheduler cares about latencies. bool UnitLatencies = ForceUnitLatencies(); - for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(), - E = DAG->allnodes_end(); NI != E; ++NI) { + // Add all nodes in depth first order. + SmallVector<SDNode*, 64> Worklist; + SmallPtrSet<SDNode*, 64> Visited; + Worklist.push_back(DAG->getRoot().getNode()); + Visited.insert(DAG->getRoot().getNode()); + + while (!Worklist.empty()) { + SDNode *NI = Worklist.pop_back_val(); + + // Add all operands to the worklist unless they've already been added. + for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i) + if (Visited.insert(NI->getOperand(i).getNode())) + Worklist.push_back(NI->getOperand(i).getNode()); + if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate. continue; diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp index 6122a2a..746d4e2 100644 --- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp +++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp @@ -468,18 +468,20 @@ static void AddNodeIDNode(FoldingSetNodeID &ID, const SDNode *N) { } /// encodeMemSDNodeFlags - Generic routine for computing a value for use in -/// the CSE map that carries volatility, indexing mode, and +/// the CSE map that carries volatility, temporalness, indexing mode, and /// extension/truncation information. /// static inline unsigned -encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile) { +encodeMemSDNodeFlags(int ConvType, ISD::MemIndexedMode AM, bool isVolatile, + bool isNonTemporal) { assert((ConvType & 3) == ConvType && "ConvType may not require more than 2 bits!"); assert((AM & 7) == AM && "AM may not require more than 3 bits!"); return ConvType | (AM << 2) | - (isVolatile << 5); + (isVolatile << 5) | + (isNonTemporal << 6); } //===----------------------------------------------------------------------===// @@ -829,6 +831,7 @@ void SelectionDAG::clear() { EntryNode.UseList = 0; AllNodes.push_back(&EntryNode); Root = getEntryNode(); + delete Ordering; Ordering = new SDNodeOrdering(); } @@ -859,14 +862,14 @@ SDValue SelectionDAG::getZeroExtendInReg(SDValue Op, DebugLoc DL, EVT VT) { /// getNOT - Create a bitwise NOT operation as (XOR Val, -1). /// SDValue SelectionDAG::getNOT(DebugLoc DL, SDValue Val, EVT VT) { - EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; + EVT EltVT = VT.getScalarType(); SDValue NegOne = getConstant(APInt::getAllOnesValue(EltVT.getSizeInBits()), VT); return getNode(ISD::XOR, DL, VT, Val, NegOne); } SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) { - EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; + EVT EltVT = VT.getScalarType(); assert((EltVT.getSizeInBits() >= 64 || (uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) && "getConstant with a uint64_t value that doesn't fit in the type!"); @@ -880,7 +883,7 @@ SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) { SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) { assert(VT.isInteger() && "Cannot create FP integer constant!"); - EVT EltVT = VT.isVector() ? VT.getVectorElementType() : VT; + EVT EltVT = VT.getScalarType(); assert(Val.getBitWidth() == EltVT.getSizeInBits() && "APInt size does not match type size!"); @@ -923,8 +926,7 @@ SDValue SelectionDAG::getConstantFP(const APFloat& V, EVT VT, bool isTarget) { SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){ assert(VT.isFloatingPoint() && "Cannot create integer FP constant!"); - EVT EltVT = - VT.isVector() ? VT.getVectorElementType() : VT; + EVT EltVT = VT.getScalarType(); // Do the map lookup using the actual bit pattern for the floating point // value, so that we don't have problems with 0.0 comparing equal to -0.0, and @@ -958,8 +960,7 @@ SDValue SelectionDAG::getConstantFP(const ConstantFP& V, EVT VT, bool isTarget){ } SDValue SelectionDAG::getConstantFP(double Val, EVT VT, bool isTarget) { - EVT EltVT = - VT.isVector() ? VT.getVectorElementType() : VT; + EVT EltVT = VT.getScalarType(); if (EltVT==MVT::f32) return getConstantFP(APFloat((float)Val), VT, isTarget); else @@ -1344,7 +1345,7 @@ SDValue SelectionDAG::getBlockAddress(BlockAddress *BA, EVT VT, } SDValue SelectionDAG::getSrcValue(const Value *V) { - assert((!V || isa<PointerType>(V->getType())) && + assert((!V || V->getType()->isPointerTy()) && "SrcValue is not a pointer?"); FoldingSetNodeID ID; @@ -2232,6 +2233,29 @@ bool SelectionDAG::isKnownNeverNaN(SDValue Op) const { return false; } +bool SelectionDAG::isKnownNeverZero(SDValue Op) const { + // If the value is a constant, we can obviously see if it is a zero or not. + if (const ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Op)) + return !C->isZero(); + + // TODO: Recognize more cases here. + + return false; +} + +bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const { + // Check the obvious case. + if (A == B) return true; + + // For for negative and positive zero. + if (const ConstantFPSDNode *CA = dyn_cast<ConstantFPSDNode>(A)) + if (const ConstantFPSDNode *CB = dyn_cast<ConstantFPSDNode>(B)) + if (CA->isZero() && CB->isZero()) return true; + + // Otherwise they may not be equal. + return false; +} + bool SelectionDAG::isVerifiedDebugInfoDesc(SDValue Op) const { GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); if (!GA) return false; @@ -3080,8 +3104,7 @@ SDValue SelectionDAG::getStackArgumentTokenFactor(SDValue Chain) { /// operand. static SDValue getMemsetValue(SDValue Value, EVT VT, SelectionDAG &DAG, DebugLoc dl) { - unsigned NumBits = VT.isVector() ? - VT.getVectorElementType().getSizeInBits() : VT.getSizeInBits(); + unsigned NumBits = VT.getScalarType().getSizeInBits(); if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) { APInt Val = APInt(NumBits, C->getZExtValue() & 255); unsigned Shift = 8; @@ -3185,7 +3208,7 @@ bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps, bool isSrcConst = isa<ConstantSDNode>(Src); EVT VT = TLI.getOptimalMemOpType(Size, Align, isSrcConst, isSrcStr, DAG); bool AllowUnalign = TLI.allowsUnalignedMemoryAccesses(VT); - if (VT != MVT::iAny) { + if (VT != MVT::Other) { const Type *Ty = VT.getTypeForEVT(*DAG.getContext()); unsigned NewAlign = (unsigned) TLI.getTargetData()->getABITypeAlignment(Ty); // If source is a string constant, this will require an unaligned load. @@ -3193,14 +3216,14 @@ bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps, if (Dst.getOpcode() != ISD::FrameIndex) { // Can't change destination alignment. It requires a unaligned store. if (AllowUnalign) - VT = MVT::iAny; + VT = MVT::Other; } else { int FI = cast<FrameIndexSDNode>(Dst)->getIndex(); MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); if (MFI->isFixedObjectIndex(FI)) { // Can't change destination alignment. It requires a unaligned store. if (AllowUnalign) - VT = MVT::iAny; + VT = MVT::Other; } else { // Give the stack frame object a larger alignment if needed. if (MFI->getObjectAlignment(FI) < NewAlign) @@ -3211,7 +3234,7 @@ bool MeetsMaxMemopRequirement(std::vector<EVT> &MemOps, } } - if (VT == MVT::iAny) { + if (VT == MVT::Other) { if (TLI.allowsUnalignedMemoryAccesses(MVT::i64)) { VT = MVT::i64; } else { @@ -3299,7 +3322,7 @@ static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, Value = getMemsetStringVal(VT, dl, DAG, TLI, Str, SrcOff); Store = DAG.getStore(Chain, dl, Value, getMemBasePlusOffset(Dst, DstOff, DAG), - DstSV, DstSVOff + DstOff, false, DstAlign); + DstSV, DstSVOff + DstOff, false, false, DstAlign); } else { // The type might not be legal for the target. This should only happen // if the type is smaller than a legal type, as on PPC, so the right @@ -3310,10 +3333,11 @@ static SDValue getMemcpyLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, assert(NVT.bitsGE(VT)); Value = DAG.getExtLoad(ISD::EXTLOAD, dl, NVT, Chain, getMemBasePlusOffset(Src, SrcOff, DAG), - SrcSV, SrcSVOff + SrcOff, VT, false, Align); + SrcSV, SrcSVOff + SrcOff, VT, false, false, Align); Store = DAG.getTruncStore(Chain, dl, Value, - getMemBasePlusOffset(Dst, DstOff, DAG), - DstSV, DstSVOff + DstOff, VT, false, DstAlign); + getMemBasePlusOffset(Dst, DstOff, DAG), + DstSV, DstSVOff + DstOff, VT, false, false, + DstAlign); } OutChains.push_back(Store); SrcOff += VTSize; @@ -3358,7 +3382,7 @@ static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, Value = DAG.getLoad(VT, dl, Chain, getMemBasePlusOffset(Src, SrcOff, DAG), - SrcSV, SrcSVOff + SrcOff, false, Align); + SrcSV, SrcSVOff + SrcOff, false, false, Align); LoadValues.push_back(Value); LoadChains.push_back(Value.getValue(1)); SrcOff += VTSize; @@ -3373,7 +3397,7 @@ static SDValue getMemmoveLoadsAndStores(SelectionDAG &DAG, DebugLoc dl, Store = DAG.getStore(Chain, dl, LoadValues[i], getMemBasePlusOffset(Dst, DstOff, DAG), - DstSV, DstSVOff + DstOff, false, DstAlign); + DstSV, DstSVOff + DstOff, false, false, DstAlign); OutChains.push_back(Store); DstOff += VTSize; } @@ -3408,7 +3432,7 @@ static SDValue getMemsetStores(SelectionDAG &DAG, DebugLoc dl, SDValue Value = getMemsetValue(Src, VT, DAG, dl); SDValue Store = DAG.getStore(Chain, dl, Value, getMemBasePlusOffset(Dst, DstOff, DAG), - DstSV, DstSVOff + DstOff); + DstSV, DstSVOff + DstOff, false, false, 0); OutChains.push_back(Store); DstOff += VTSize; } @@ -3472,7 +3496,7 @@ SDValue SelectionDAG::getMemcpy(SDValue Chain, DebugLoc dl, SDValue Dst, /*isReturnValueUsed=*/false, getExternalSymbol(TLI.getLibcallName(RTLIB::MEMCPY), TLI.getPointerTy()), - Args, *this, dl, GetOrdering(Chain.getNode())); + Args, *this, dl); return CallResult.second; } @@ -3521,7 +3545,7 @@ SDValue SelectionDAG::getMemmove(SDValue Chain, DebugLoc dl, SDValue Dst, /*isReturnValueUsed=*/false, getExternalSymbol(TLI.getLibcallName(RTLIB::MEMMOVE), TLI.getPointerTy()), - Args, *this, dl, GetOrdering(Chain.getNode())); + Args, *this, dl); return CallResult.second; } @@ -3580,7 +3604,7 @@ SDValue SelectionDAG::getMemset(SDValue Chain, DebugLoc dl, SDValue Dst, /*isReturnValueUsed=*/false, getExternalSymbol(TLI.getLibcallName(RTLIB::MEMSET), TLI.getPointerTy()), - Args, *this, dl, GetOrdering(Chain.getNode())); + Args, *this, dl); return CallResult.second; } @@ -3788,7 +3812,8 @@ SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, ISD::LoadExtType ExtType, EVT VT, SDValue Chain, SDValue Ptr, SDValue Offset, const Value *SV, int SVOffset, EVT MemVT, - bool isVolatile, unsigned Alignment) { + bool isVolatile, bool isNonTemporal, + unsigned Alignment) { if (Alignment == 0) // Ensure that codegen never sees alignment 0 Alignment = getEVTAlignment(VT); @@ -3802,6 +3827,8 @@ SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, unsigned Flags = MachineMemOperand::MOLoad; if (isVolatile) Flags |= MachineMemOperand::MOVolatile; + if (isNonTemporal) + Flags |= MachineMemOperand::MONonTemporal; MachineMemOperand *MMO = MF.getMachineMemOperand(SV, Flags, SVOffset, MemVT.getStoreSize(), Alignment); @@ -3840,7 +3867,8 @@ SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, FoldingSetNodeID ID; AddNodeIDNode(ID, ISD::LOAD, VTs, Ops, 3); ID.AddInteger(MemVT.getRawBits()); - ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile())); + ID.AddInteger(encodeMemSDNodeFlags(ExtType, AM, MMO->isVolatile(), + MMO->isNonTemporal())); void *IP = 0; if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { cast<LoadSDNode>(E)->refineAlignment(MMO); @@ -3856,20 +3884,22 @@ SelectionDAG::getLoad(ISD::MemIndexedMode AM, DebugLoc dl, SDValue SelectionDAG::getLoad(EVT VT, DebugLoc dl, SDValue Chain, SDValue Ptr, const Value *SV, int SVOffset, - bool isVolatile, unsigned Alignment) { + bool isVolatile, bool isNonTemporal, + unsigned Alignment) { SDValue Undef = getUNDEF(Ptr.getValueType()); return getLoad(ISD::UNINDEXED, dl, ISD::NON_EXTLOAD, VT, Chain, Ptr, Undef, - SV, SVOffset, VT, isVolatile, Alignment); + SV, SVOffset, VT, isVolatile, isNonTemporal, Alignment); } SDValue SelectionDAG::getExtLoad(ISD::LoadExtType ExtType, DebugLoc dl, EVT VT, SDValue Chain, SDValue Ptr, const Value *SV, int SVOffset, EVT MemVT, - bool isVolatile, unsigned Alignment) { + bool isVolatile, bool isNonTemporal, + unsigned Alignment) { SDValue Undef = getUNDEF(Ptr.getValueType()); return getLoad(ISD::UNINDEXED, dl, ExtType, VT, Chain, Ptr, Undef, - SV, SVOffset, MemVT, isVolatile, Alignment); + SV, SVOffset, MemVT, isVolatile, isNonTemporal, Alignment); } SDValue @@ -3881,12 +3911,13 @@ SelectionDAG::getIndexedLoad(SDValue OrigLoad, DebugLoc dl, SDValue Base, return getLoad(AM, dl, LD->getExtensionType(), OrigLoad.getValueType(), LD->getChain(), Base, Offset, LD->getSrcValue(), LD->getSrcValueOffset(), LD->getMemoryVT(), - LD->isVolatile(), LD->getAlignment()); + LD->isVolatile(), LD->isNonTemporal(), LD->getAlignment()); } SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, SDValue Ptr, const Value *SV, int SVOffset, - bool isVolatile, unsigned Alignment) { + bool isVolatile, bool isNonTemporal, + unsigned Alignment) { if (Alignment == 0) // Ensure that codegen never sees alignment 0 Alignment = getEVTAlignment(Val.getValueType()); @@ -3900,6 +3931,8 @@ SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, unsigned Flags = MachineMemOperand::MOStore; if (isVolatile) Flags |= MachineMemOperand::MOVolatile; + if (isNonTemporal) + Flags |= MachineMemOperand::MONonTemporal; MachineMemOperand *MMO = MF.getMachineMemOperand(SV, Flags, SVOffset, Val.getValueType().getStoreSize(), Alignment); @@ -3916,7 +3949,8 @@ SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, FoldingSetNodeID ID; AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); ID.AddInteger(VT.getRawBits()); - ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile())); + ID.AddInteger(encodeMemSDNodeFlags(false, ISD::UNINDEXED, MMO->isVolatile(), + MMO->isNonTemporal())); void *IP = 0; if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { cast<StoreSDNode>(E)->refineAlignment(MMO); @@ -3932,7 +3966,8 @@ SDValue SelectionDAG::getStore(SDValue Chain, DebugLoc dl, SDValue Val, SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, SDValue Ptr, const Value *SV, int SVOffset, EVT SVT, - bool isVolatile, unsigned Alignment) { + bool isVolatile, bool isNonTemporal, + unsigned Alignment) { if (Alignment == 0) // Ensure that codegen never sees alignment 0 Alignment = getEVTAlignment(SVT); @@ -3946,6 +3981,8 @@ SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, unsigned Flags = MachineMemOperand::MOStore; if (isVolatile) Flags |= MachineMemOperand::MOVolatile; + if (isNonTemporal) + Flags |= MachineMemOperand::MONonTemporal; MachineMemOperand *MMO = MF.getMachineMemOperand(SV, Flags, SVOffset, SVT.getStoreSize(), Alignment); @@ -3976,7 +4013,8 @@ SDValue SelectionDAG::getTruncStore(SDValue Chain, DebugLoc dl, SDValue Val, FoldingSetNodeID ID; AddNodeIDNode(ID, ISD::STORE, VTs, Ops, 4); ID.AddInteger(SVT.getRawBits()); - ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile())); + ID.AddInteger(encodeMemSDNodeFlags(true, ISD::UNINDEXED, MMO->isVolatile(), + MMO->isNonTemporal())); void *IP = 0; if (SDNode *E = CSEMap.FindNodeOrInsertPos(ID, IP)) { cast<StoreSDNode>(E)->refineAlignment(MMO); @@ -4535,91 +4573,13 @@ SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, SDNode *SelectionDAG::SelectNodeTo(SDNode *N, unsigned MachineOpc, SDVTList VTs, const SDValue *Ops, unsigned NumOps) { - return MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT) { - SDVTList VTs = getVTList(VT); - return MorphNodeTo(N, Opc, VTs, 0, 0); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT, SDValue Op1) { - SDVTList VTs = getVTList(VT); - SDValue Ops[] = { Op1 }; - return MorphNodeTo(N, Opc, VTs, Ops, 1); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT, SDValue Op1, - SDValue Op2) { - SDVTList VTs = getVTList(VT); - SDValue Ops[] = { Op1, Op2 }; - return MorphNodeTo(N, Opc, VTs, Ops, 2); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT, SDValue Op1, - SDValue Op2, SDValue Op3) { - SDVTList VTs = getVTList(VT); - SDValue Ops[] = { Op1, Op2, Op3 }; - return MorphNodeTo(N, Opc, VTs, Ops, 3); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT, const SDValue *Ops, - unsigned NumOps) { - SDVTList VTs = getVTList(VT); - return MorphNodeTo(N, Opc, VTs, Ops, NumOps); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2, const SDValue *Ops, - unsigned NumOps) { - SDVTList VTs = getVTList(VT1, VT2); - return MorphNodeTo(N, Opc, VTs, Ops, NumOps); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2) { - SDVTList VTs = getVTList(VT1, VT2); - return MorphNodeTo(N, Opc, VTs, (SDValue *)0, 0); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2, EVT VT3, - const SDValue *Ops, unsigned NumOps) { - SDVTList VTs = getVTList(VT1, VT2, VT3); - return MorphNodeTo(N, Opc, VTs, Ops, NumOps); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2, - SDValue Op1) { - SDVTList VTs = getVTList(VT1, VT2); - SDValue Ops[] = { Op1 }; - return MorphNodeTo(N, Opc, VTs, Ops, 1); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2, - SDValue Op1, SDValue Op2) { - SDVTList VTs = getVTList(VT1, VT2); - SDValue Ops[] = { Op1, Op2 }; - return MorphNodeTo(N, Opc, VTs, Ops, 2); -} - -SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, - EVT VT1, EVT VT2, - SDValue Op1, SDValue Op2, - SDValue Op3) { - SDVTList VTs = getVTList(VT1, VT2); - SDValue Ops[] = { Op1, Op2, Op3 }; - return MorphNodeTo(N, Opc, VTs, Ops, 3); + N = MorphNodeTo(N, ~MachineOpc, VTs, Ops, NumOps); + // Reset the NodeID to -1. + N->setNodeId(-1); + return N; } -/// MorphNodeTo - These *mutate* the specified node to have the specified +/// MorphNodeTo - This *mutates* the specified node to have the specified /// return type, opcode, and operands. /// /// Note that MorphNodeTo returns the resultant node. If there is already a @@ -4695,12 +4655,14 @@ SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc, // Delete any nodes that are still dead after adding the uses for the // new operands. - SmallVector<SDNode *, 16> DeadNodes; - for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(), - E = DeadNodeSet.end(); I != E; ++I) - if ((*I)->use_empty()) - DeadNodes.push_back(*I); - RemoveDeadNodes(DeadNodes); + if (!DeadNodeSet.empty()) { + SmallVector<SDNode *, 16> DeadNodes; + for (SmallPtrSet<SDNode *, 16>::iterator I = DeadNodeSet.begin(), + E = DeadNodeSet.end(); I != E; ++I) + if ((*I)->use_empty()) + DeadNodes.push_back(*I); + RemoveDeadNodes(DeadNodes); + } if (IP) CSEMap.InsertNode(N, IP); // Memoize the new node. @@ -4907,6 +4869,43 @@ SDNode *SelectionDAG::getNodeIfExists(unsigned Opcode, SDVTList VTList, return NULL; } +namespace { + +/// RAUWUpdateListener - Helper for ReplaceAllUsesWith - When the node +/// pointed to by a use iterator is deleted, increment the use iterator +/// so that it doesn't dangle. +/// +/// This class also manages a "downlink" DAGUpdateListener, to forward +/// messages to ReplaceAllUsesWith's callers. +/// +class RAUWUpdateListener : public SelectionDAG::DAGUpdateListener { + SelectionDAG::DAGUpdateListener *DownLink; + SDNode::use_iterator &UI; + SDNode::use_iterator &UE; + + virtual void NodeDeleted(SDNode *N, SDNode *E) { + // Increment the iterator as needed. + while (UI != UE && N == *UI) + ++UI; + + // Then forward the message. + if (DownLink) DownLink->NodeDeleted(N, E); + } + + virtual void NodeUpdated(SDNode *N) { + // Just forward the message. + if (DownLink) DownLink->NodeUpdated(N); + } + +public: + RAUWUpdateListener(SelectionDAG::DAGUpdateListener *dl, + SDNode::use_iterator &ui, + SDNode::use_iterator &ue) + : DownLink(dl), UI(ui), UE(ue) {} +}; + +} + /// ReplaceAllUsesWith - Modify anything using 'From' to use 'To' instead. /// This can cause recursive merging of nodes in the DAG. /// @@ -4927,6 +4926,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To, // is replaced by To, we don't want to replace of all its users with To // too. See PR3018 for more info. SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); + RAUWUpdateListener Listener(UpdateListener, UI, UE); while (UI != UE) { SDNode *User = *UI; @@ -4945,7 +4945,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDValue FromN, SDValue To, // Now that we have modified User, add it back to the CSE maps. If it // already exists there, recursively merge the results together. - AddModifiedNodeToCSEMaps(User, UpdateListener); + AddModifiedNodeToCSEMaps(User, &Listener); } } @@ -4971,6 +4971,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To, // Iterate over just the existing users of From. See the comments in // the ReplaceAllUsesWith above. SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); + RAUWUpdateListener Listener(UpdateListener, UI, UE); while (UI != UE) { SDNode *User = *UI; @@ -4989,7 +4990,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDNode *From, SDNode *To, // Now that we have modified User, add it back to the CSE maps. If it // already exists there, recursively merge the results together. - AddModifiedNodeToCSEMaps(User, UpdateListener); + AddModifiedNodeToCSEMaps(User, &Listener); } } @@ -5007,6 +5008,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDNode *From, // Iterate over just the existing users of From. See the comments in // the ReplaceAllUsesWith above. SDNode::use_iterator UI = From->use_begin(), UE = From->use_end(); + RAUWUpdateListener Listener(UpdateListener, UI, UE); while (UI != UE) { SDNode *User = *UI; @@ -5026,7 +5028,7 @@ void SelectionDAG::ReplaceAllUsesWith(SDNode *From, // Now that we have modified User, add it back to the CSE maps. If it // already exists there, recursively merge the results together. - AddModifiedNodeToCSEMaps(User, UpdateListener); + AddModifiedNodeToCSEMaps(User, &Listener); } } @@ -5048,6 +5050,7 @@ void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To, // the ReplaceAllUsesWith above. SDNode::use_iterator UI = From.getNode()->use_begin(), UE = From.getNode()->use_end(); + RAUWUpdateListener Listener(UpdateListener, UI, UE); while (UI != UE) { SDNode *User = *UI; bool UserRemovedFromCSEMaps = false; @@ -5083,7 +5086,7 @@ void SelectionDAG::ReplaceAllUsesOfValueWith(SDValue From, SDValue To, // Now that we have modified User, add it back to the CSE maps. If it // already exists there, recursively merge the results together. - AddModifiedNodeToCSEMaps(User, UpdateListener); + AddModifiedNodeToCSEMaps(User, &Listener); } } @@ -5280,8 +5283,11 @@ GlobalAddressSDNode::GlobalAddressSDNode(unsigned Opc, const GlobalValue *GA, MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, EVT memvt, MachineMemOperand *mmo) : SDNode(Opc, dl, VTs), MemoryVT(memvt), MMO(mmo) { - SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile()); + SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), + MMO->isNonTemporal()); assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); + assert(isNonTemporal() == MMO->isNonTemporal() && + "Non-temporal encoding error!"); assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); } @@ -5290,7 +5296,8 @@ MemSDNode::MemSDNode(unsigned Opc, DebugLoc dl, SDVTList VTs, MachineMemOperand *mmo) : SDNode(Opc, dl, VTs, Ops, NumOps), MemoryVT(memvt), MMO(mmo) { - SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile()); + SubclassData = encodeMemSDNodeFlags(0, ISD::UNINDEXED, MMO->isVolatile(), + MMO->isNonTemporal()); assert(isVolatile() == MMO->isVolatile() && "Volatile encoding error!"); assert(memvt.getStoreSize() == MMO->getSize() && "Size mismatch!"); } @@ -5459,15 +5466,15 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const { if (const TargetInstrInfo *TII = G->getTarget().getInstrInfo()) if (getMachineOpcode() < TII->getNumOpcodes()) return TII->get(getMachineOpcode()).getName(); - return "<<Unknown Machine Node>>"; + return "<<Unknown Machine Node #" + utostr(getOpcode()) + ">>"; } if (G) { const TargetLowering &TLI = G->getTargetLoweringInfo(); const char *Name = TLI.getTargetNodeName(getOpcode()); if (Name) return Name; - return "<<Unknown Target Node>>"; + return "<<Unknown Target Node #" + utostr(getOpcode()) + ">>"; } - return "<<Unknown Node>>"; + return "<<Unknown Node #" + utostr(getOpcode()) + ">>"; #ifndef NDEBUG case ISD::DELETED_NODE: @@ -5904,6 +5911,9 @@ void SDNode::print_details(raw_ostream &OS, const SelectionDAG *G) const { if (G) if (unsigned Order = G->GetOrdering(this)) OS << " [ORD=" << Order << ']'; + + if (getNodeId() != -1) + OS << " [ID=" << getNodeId() << ']'; } void SDNode::print(raw_ostream &OS, const SelectionDAG *G) const { @@ -6292,31 +6302,37 @@ bool ShuffleVectorSDNode::isSplatMask(const int *Mask, EVT VT) { return true; } +#ifdef XDEBUG static void checkForCyclesHelper(const SDNode *N, - std::set<const SDNode *> &visited) { - if (visited.find(N) != visited.end()) { + SmallPtrSet<const SDNode*, 32> &Visited, + SmallPtrSet<const SDNode*, 32> &Checked) { + // If this node has already been checked, don't check it again. + if (Checked.count(N)) + return; + + // If a node has already been visited on this depth-first walk, reject it as + // a cycle. + if (!Visited.insert(N)) { dbgs() << "Offending node:\n"; N->dumprFull(); - assert(0 && "Detected cycle in SelectionDAG"); + errs() << "Detected cycle in SelectionDAG\n"; + abort(); } - - std::set<const SDNode*>::iterator i; - bool inserted; - - tie(i, inserted) = visited.insert(N); - assert(inserted && "Missed cycle"); - - for(unsigned i = 0; i < N->getNumOperands(); ++i) { - checkForCyclesHelper(N->getOperand(i).getNode(), visited); - } - visited.erase(i); + + for(unsigned i = 0, e = N->getNumOperands(); i != e; ++i) + checkForCyclesHelper(N->getOperand(i).getNode(), Visited, Checked); + + Checked.insert(N); + Visited.erase(N); } +#endif void llvm::checkForCycles(const llvm::SDNode *N) { #ifdef XDEBUG assert(N && "Checking nonexistant SDNode"); - std::set<const SDNode *> visited; - checkForCyclesHelper(N, visited); + SmallPtrSet<const SDNode*, 32> visited; + SmallPtrSet<const SDNode*, 32> checked; + checkForCyclesHelper(N, visited, checked); #endif } diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp index de17f90..05be9a1 100644 --- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp +++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp @@ -155,7 +155,7 @@ namespace { /// this value and returns the result as a ValueVTs value. This uses /// Chain/Flag as the input and updates them for the output Chain/Flag. /// If the Flag pointer is NULL, no flag is used. - SDValue getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, unsigned Order, + SDValue getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, SDValue &Chain, SDValue *Flag) const; /// getCopyToRegs - Emit a series of CopyToReg nodes that copies the @@ -163,14 +163,14 @@ namespace { /// Chain/Flag as the input and updates them for the output Chain/Flag. /// If the Flag pointer is NULL, no flag is used. void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, - unsigned Order, SDValue &Chain, SDValue *Flag) const; + SDValue &Chain, SDValue *Flag) const; /// AddInlineAsmOperands - Add this value to the specified inlineasm node /// operand list. This adds the code marker, matching input operand index /// (if applicable), and includes the number of values added into it. void AddInlineAsmOperands(unsigned Code, bool HasMatching, unsigned MatchingIdx, - SelectionDAG &DAG, unsigned Order, + SelectionDAG &DAG, std::vector<SDValue> &Ops) const; }; } @@ -180,7 +180,7 @@ namespace { /// larger then ValueVT then AssertOp can be used to specify whether the extra /// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT /// (ISD::AssertSext). -static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, +static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, const SDValue *Parts, unsigned NumParts, EVT PartVT, EVT ValueVT, ISD::NodeType AssertOp = ISD::DELETED_NODE) { @@ -205,9 +205,9 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2); if (RoundParts > 2) { - Lo = getCopyFromParts(DAG, dl, Order, Parts, RoundParts / 2, + Lo = getCopyFromParts(DAG, dl, Parts, RoundParts / 2, PartVT, HalfVT); - Hi = getCopyFromParts(DAG, dl, Order, Parts + RoundParts / 2, + Hi = getCopyFromParts(DAG, dl, Parts + RoundParts / 2, RoundParts / 2, PartVT, HalfVT); } else { Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]); @@ -223,7 +223,7 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, // Assemble the trailing non-power-of-2 part. unsigned OddParts = NumParts - RoundParts; EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits); - Hi = getCopyFromParts(DAG, dl, Order, + Hi = getCopyFromParts(DAG, dl, Parts + RoundParts, OddParts, PartVT, OddVT); // Combine the round and odd parts. @@ -259,7 +259,7 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, // If the register was not expanded, truncate or copy the value, // as appropriate. for (unsigned i = 0; i != NumParts; ++i) - Ops[i] = getCopyFromParts(DAG, dl, Order, &Parts[i], 1, + Ops[i] = getCopyFromParts(DAG, dl, &Parts[i], 1, PartVT, IntermediateVT); } else if (NumParts > 0) { // If the intermediate type was expanded, build the intermediate @@ -268,7 +268,7 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, "Must expand into a divisible number of parts!"); unsigned Factor = NumParts / NumIntermediates; for (unsigned i = 0; i != NumIntermediates; ++i) - Ops[i] = getCopyFromParts(DAG, dl, Order, &Parts[i * Factor], Factor, + Ops[i] = getCopyFromParts(DAG, dl, &Parts[i * Factor], Factor, PartVT, IntermediateVT); } @@ -292,7 +292,7 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, assert(ValueVT.isFloatingPoint() && PartVT.isInteger() && !PartVT.isVector() && "Unexpected split"); EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits()); - Val = getCopyFromParts(DAG, dl, Order, Parts, NumParts, PartVT, IntVT); + Val = getCopyFromParts(DAG, dl, Parts, NumParts, PartVT, IntVT); } } @@ -349,7 +349,7 @@ static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, /// getCopyToParts - Create a series of nodes that contain the specified value /// split into legal parts. If the parts contain more bits than Val, then, for /// integers, ExtendKind can be used to specify how to generate the extra bits. -static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, +static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, SDValue Val, SDValue *Parts, unsigned NumParts, EVT PartVT, ISD::NodeType ExtendKind = ISD::ANY_EXTEND) { @@ -417,7 +417,7 @@ static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, SDValue OddVal = DAG.getNode(ISD::SRL, dl, ValueVT, Val, DAG.getConstant(RoundBits, TLI.getPointerTy())); - getCopyToParts(DAG, dl, Order, OddVal, Parts + RoundParts, + getCopyToParts(DAG, dl, OddVal, Parts + RoundParts, OddParts, PartVT); if (TLI.isBigEndian()) @@ -514,7 +514,7 @@ static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, // If the register was not expanded, promote or copy the value, // as appropriate. for (unsigned i = 0; i != NumParts; ++i) - getCopyToParts(DAG, dl, Order, Ops[i], &Parts[i], 1, PartVT); + getCopyToParts(DAG, dl, Ops[i], &Parts[i], 1, PartVT); } else if (NumParts > 0) { // If the intermediate type was expanded, split each the value into // legal parts. @@ -522,7 +522,7 @@ static void getCopyToParts(SelectionDAG &DAG, DebugLoc dl, unsigned Order, "Must expand into a divisible number of parts!"); unsigned Factor = NumParts / NumIntermediates; for (unsigned i = 0; i != NumIntermediates; ++i) - getCopyToParts(DAG, dl, Order, Ops[i], &Parts[i*Factor], Factor, PartVT); + getCopyToParts(DAG, dl, Ops[i], &Parts[i*Factor], Factor, PartVT); } } @@ -680,7 +680,7 @@ SDValue SelectionDAGBuilder::getValue(const Value *V) { getCurDebugLoc()); } - if (isa<StructType>(C->getType()) || isa<ArrayType>(C->getType())) { + if (C->getType()->isStructTy() || C->getType()->isArrayTy()) { assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) && "Unknown struct or array constant!"); @@ -747,8 +747,7 @@ SDValue SelectionDAGBuilder::getValue(const Value *V) { RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType()); SDValue Chain = DAG.getEntryNode(); - return RFV.getCopyFromRegs(DAG, getCurDebugLoc(), - SDNodeOrder, Chain, NULL); + return RFV.getCopyFromRegs(DAG, getCurDebugLoc(), Chain, NULL); } /// Get the EVTs and ArgFlags collections that represent the legalized return @@ -844,19 +843,17 @@ void SelectionDAGBuilder::visitRet(ReturnInst &I) { Chains[i] = DAG.getStore(Chain, getCurDebugLoc(), SDValue(RetOp.getNode(), RetOp.getResNo() + i), - Add, NULL, Offsets[i], false, 0); + Add, NULL, Offsets[i], false, false, 0); } Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other, &Chains[0], NumValues); - } else { - for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) { - SmallVector<EVT, 4> ValueVTs; - ComputeValueVTs(TLI, I.getOperand(i)->getType(), ValueVTs); - unsigned NumValues = ValueVTs.size(); - if (NumValues == 0) continue; - - SDValue RetOp = getValue(I.getOperand(i)); + } else if (I.getNumOperands() != 0) { + SmallVector<EVT, 4> ValueVTs; + ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs); + unsigned NumValues = ValueVTs.size(); + if (NumValues) { + SDValue RetOp = getValue(I.getOperand(0)); for (unsigned j = 0, f = NumValues; j != f; ++j) { EVT VT = ValueVTs[j]; @@ -881,7 +878,7 @@ void SelectionDAGBuilder::visitRet(ReturnInst &I) { unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT); EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT); SmallVector<SDValue, 4> Parts(NumParts); - getCopyToParts(DAG, getCurDebugLoc(), SDNodeOrder, + getCopyToParts(DAG, getCurDebugLoc(), SDValue(RetOp.getNode(), RetOp.getResNo() + j), &Parts[0], NumParts, PartVT, ExtendKind); @@ -1973,7 +1970,7 @@ size_t SelectionDAGBuilder::Clusterify(CaseVector& Cases, if (Cases.size() >= 2) // Must recompute end() each iteration because it may be // invalidated by erase if we hold on to it - for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) { + for (CaseItr TmpBegin = Cases.begin(), I = TmpBegin, J = ++TmpBegin; J != Cases.end(); ) { const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue(); const APInt& currentValue = cast<ConstantInt>(I->High)->getValue(); MachineBasicBlock* nextBB = J->BB; @@ -2062,9 +2059,15 @@ void SelectionDAGBuilder::visitSwitch(SwitchInst &SI) { } void SelectionDAGBuilder::visitIndirectBr(IndirectBrInst &I) { - // Update machine-CFG edges. + // Update machine-CFG edges with unique successors. + SmallVector<BasicBlock*, 32> succs; + succs.reserve(I.getNumSuccessors()); for (unsigned i = 0, e = I.getNumSuccessors(); i != e; ++i) - CurMBB->addSuccessor(FuncInfo.MBBMap[I.getSuccessor(i)]); + succs.push_back(I.getSuccessor(i)); + array_pod_sort(succs.begin(), succs.end()); + succs.erase(std::unique(succs.begin(), succs.end()), succs.end()); + for (unsigned i = 0, e = succs.size(); i != e; ++i) + CurMBB->addSuccessor(FuncInfo.MBBMap[succs[i]]); DAG.setRoot(DAG.getNode(ISD::BRIND, getCurDebugLoc(), MVT::Other, getControlRoot(), @@ -2074,7 +2077,7 @@ void SelectionDAGBuilder::visitIndirectBr(IndirectBrInst &I) { void SelectionDAGBuilder::visitFSub(User &I) { // -0.0 - X --> fneg const Type *Ty = I.getType(); - if (isa<VectorType>(Ty)) { + if (Ty->isVectorTy()) { if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) { const VectorType *DestTy = cast<VectorType>(I.getType()); const Type *ElTy = DestTy->getElementType(); @@ -2111,7 +2114,7 @@ void SelectionDAGBuilder::visitBinary(User &I, unsigned OpCode) { void SelectionDAGBuilder::visitShift(User &I, unsigned Opcode) { SDValue Op1 = getValue(I.getOperand(0)); SDValue Op2 = getValue(I.getOperand(1)); - if (!isa<VectorType>(I.getType()) && + if (!I.getType()->isVectorTy() && Op2.getValueType() != TLI.getShiftAmountTy()) { // If the operand is smaller than the shift count type, promote it. EVT PTy = TLI.getPointerTy(); @@ -2699,7 +2702,9 @@ void SelectionDAGBuilder::visitLoad(LoadInst &I) { SDValue Ptr = getValue(SV); const Type *Ty = I.getType(); + bool isVolatile = I.isVolatile(); + bool isNonTemporal = I.getMetadata("nontemporal") != 0; unsigned Alignment = I.getAlignment(); SmallVector<EVT, 4> ValueVTs; @@ -2731,7 +2736,8 @@ void SelectionDAGBuilder::visitLoad(LoadInst &I) { PtrVT, Ptr, DAG.getConstant(Offsets[i], PtrVT)); SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root, - A, SV, Offsets[i], isVolatile, Alignment); + A, SV, Offsets[i], isVolatile, + isNonTemporal, Alignment); Values[i] = L; Chains[i] = L.getValue(1); @@ -2772,6 +2778,7 @@ void SelectionDAGBuilder::visitStore(StoreInst &I) { SmallVector<SDValue, 4> Chains(NumValues); EVT PtrVT = Ptr.getValueType(); bool isVolatile = I.isVolatile(); + bool isNonTemporal = I.getMetadata("nontemporal") != 0; unsigned Alignment = I.getAlignment(); for (unsigned i = 0; i != NumValues; ++i) { @@ -2779,7 +2786,8 @@ void SelectionDAGBuilder::visitStore(StoreInst &I) { DAG.getConstant(Offsets[i], PtrVT)); Chains[i] = DAG.getStore(Root, getCurDebugLoc(), SDValue(Src.getNode(), Src.getResNo() + i), - Add, PtrV, Offsets[i], isVolatile, Alignment); + Add, PtrV, Offsets[i], isVolatile, + isNonTemporal, Alignment); } DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), @@ -2879,7 +2887,7 @@ void SelectionDAGBuilder::visitTargetIntrinsic(CallInst &I, /// /// where Op is the hexidecimal representation of floating point value. static SDValue -GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl, unsigned Order) { +GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl) { SDValue t1 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, DAG.getConstant(0x007fffff, MVT::i32)); SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1, @@ -2894,7 +2902,7 @@ GetSignificand(SelectionDAG &DAG, SDValue Op, DebugLoc dl, unsigned Order) { /// where Op is the hexidecimal representation of floating point value. static SDValue GetExponent(SelectionDAG &DAG, SDValue Op, const TargetLowering &TLI, - DebugLoc dl, unsigned Order) { + DebugLoc dl) { SDValue t0 = DAG.getNode(ISD::AND, dl, MVT::i32, Op, DAG.getConstant(0x7f800000, MVT::i32)); SDValue t1 = DAG.getNode(ISD::SRL, dl, MVT::i32, t0, @@ -3078,13 +3086,13 @@ SelectionDAGBuilder::visitLog(CallInst &I) { SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); // Scale the exponent by log(2) [0.69314718f]. - SDValue Exp = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder); + SDValue Exp = GetExponent(DAG, Op1, TLI, dl); SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, getF32Constant(DAG, 0x3f317218)); // Get the significand and build it into a floating-point number with // exponent of 1. - SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder); + SDValue X = GetSignificand(DAG, Op1, dl); if (LimitFloatPrecision <= 6) { // For floating-point precision of 6: @@ -3188,11 +3196,11 @@ SelectionDAGBuilder::visitLog2(CallInst &I) { SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); // Get the exponent. - SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder); + SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl); // Get the significand and build it into a floating-point number with // exponent of 1. - SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder); + SDValue X = GetSignificand(DAG, Op1, dl); // Different possible minimax approximations of significand in // floating-point for various degrees of accuracy over [1,2]. @@ -3297,13 +3305,13 @@ SelectionDAGBuilder::visitLog10(CallInst &I) { SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); // Scale the exponent by log10(2) [0.30102999f]. - SDValue Exp = GetExponent(DAG, Op1, TLI, dl, SDNodeOrder); + SDValue Exp = GetExponent(DAG, Op1, TLI, dl); SDValue LogOfExponent = DAG.getNode(ISD::FMUL, dl, MVT::f32, Exp, getF32Constant(DAG, 0x3e9a209a)); // Get the significand and build it into a floating-point number with // exponent of 1. - SDValue X = GetSignificand(DAG, Op1, dl, SDNodeOrder); + SDValue X = GetSignificand(DAG, Op1, dl); if (LimitFloatPrecision <= 6) { // For floating-point precision of 6: @@ -4058,7 +4066,7 @@ SelectionDAGBuilder::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) { // Store the stack protector onto the stack. Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN, PseudoSourceValue::getFixedStack(FI), - 0, true); + 0, true, false, 0); setValue(&I, Res); DAG.setRoot(Res); return 0; @@ -4276,8 +4284,8 @@ isInTailCallPosition(CallSite CS, Attributes CalleeRetAttr, // Check for a truly no-op bitcast. if (isa<BitCastInst>(U) && (U->getOperand(0)->getType() == U->getType() || - (isa<PointerType>(U->getOperand(0)->getType()) && - isa<PointerType>(U->getType())))) + (U->getOperand(0)->getType()->isPointerTy() && + U->getType()->isPointerTy()))) continue; // Otherwise it's not a true no-op. return false; @@ -4385,7 +4393,7 @@ void SelectionDAGBuilder::LowerCallTo(CallSite CS, SDValue Callee, CS.getCallingConv(), isTailCall, !CS.getInstruction()->use_empty(), - Callee, Args, DAG, getCurDebugLoc(), SDNodeOrder); + Callee, Args, DAG, getCurDebugLoc()); assert((isTailCall || Result.second.getNode()) && "Non-null chain expected with non-tail call!"); assert((Result.second.getNode() || !Result.first.getNode()) && @@ -4410,7 +4418,7 @@ void SelectionDAGBuilder::LowerCallTo(CallSite CS, SDValue Callee, DemoteStackSlot, DAG.getConstant(Offsets[i], PtrVT)); SDValue L = DAG.getLoad(OutVTs[i], getCurDebugLoc(), Result.second, - Add, NULL, Offsets[i], false, 1); + Add, NULL, Offsets[i], false, false, 1); Values[i] = L; Chains[i] = L.getValue(1); } @@ -4433,7 +4441,7 @@ void SelectionDAGBuilder::LowerCallTo(CallSite CS, SDValue Callee, unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT); SDValue ReturnValue = - getCopyFromParts(DAG, getCurDebugLoc(), SDNodeOrder, &Values[CurReg], NumRegs, + getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs, RegisterVT, VT, AssertOp); ReturnValues.push_back(ReturnValue); CurReg += NumRegs; @@ -4512,7 +4520,8 @@ static SDValue getMemCmpLoad(Value *PtrVal, MVT LoadVT, const Type *LoadTy, SDValue Ptr = Builder.getValue(PtrVal); SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root, Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/, - false /*volatile*/, 1 /* align=1 */); + false /*volatile*/, + false /*nontemporal*/, 1 /* align=1 */); if (!ConstantMemory) Builder.PendingLoads.push_back(LoadVal.getValue(1)); @@ -4529,9 +4538,9 @@ bool SelectionDAGBuilder::visitMemCmpCall(CallInst &I) { return false; Value *LHS = I.getOperand(1), *RHS = I.getOperand(2); - if (!isa<PointerType>(LHS->getType()) || !isa<PointerType>(RHS->getType()) || - !isa<IntegerType>(I.getOperand(3)->getType()) || - !isa<IntegerType>(I.getType())) + if (!LHS->getType()->isPointerTy() || !RHS->getType()->isPointerTy() || + !I.getOperand(3)->getType()->isIntegerTy() || + !I.getType()->isIntegerTy()) return false; ConstantInt *Size = dyn_cast<ConstantInt>(I.getOperand(3)); @@ -4625,7 +4634,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { StringRef Name = F->getName(); if (Name == "copysign" || Name == "copysignf") { if (I.getNumOperands() == 3 && // Basic sanity checks. - I.getOperand(1)->getType()->isFloatingPoint() && + I.getOperand(1)->getType()->isFloatingPointTy() && I.getType() == I.getOperand(1)->getType() && I.getType() == I.getOperand(2)->getType()) { SDValue LHS = getValue(I.getOperand(1)); @@ -4636,7 +4645,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { } } else if (Name == "fabs" || Name == "fabsf" || Name == "fabsl") { if (I.getNumOperands() == 2 && // Basic sanity checks. - I.getOperand(1)->getType()->isFloatingPoint() && + I.getOperand(1)->getType()->isFloatingPointTy() && I.getType() == I.getOperand(1)->getType()) { SDValue Tmp = getValue(I.getOperand(1)); setValue(&I, DAG.getNode(ISD::FABS, getCurDebugLoc(), @@ -4645,7 +4654,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { } } else if (Name == "sin" || Name == "sinf" || Name == "sinl") { if (I.getNumOperands() == 2 && // Basic sanity checks. - I.getOperand(1)->getType()->isFloatingPoint() && + I.getOperand(1)->getType()->isFloatingPointTy() && I.getType() == I.getOperand(1)->getType() && I.onlyReadsMemory()) { SDValue Tmp = getValue(I.getOperand(1)); @@ -4655,7 +4664,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { } } else if (Name == "cos" || Name == "cosf" || Name == "cosl") { if (I.getNumOperands() == 2 && // Basic sanity checks. - I.getOperand(1)->getType()->isFloatingPoint() && + I.getOperand(1)->getType()->isFloatingPointTy() && I.getType() == I.getOperand(1)->getType() && I.onlyReadsMemory()) { SDValue Tmp = getValue(I.getOperand(1)); @@ -4665,7 +4674,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { } } else if (Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") { if (I.getNumOperands() == 2 && // Basic sanity checks. - I.getOperand(1)->getType()->isFloatingPoint() && + I.getOperand(1)->getType()->isFloatingPointTy() && I.getType() == I.getOperand(1)->getType() && I.onlyReadsMemory()) { SDValue Tmp = getValue(I.getOperand(1)); @@ -4699,8 +4708,7 @@ void SelectionDAGBuilder::visitCall(CallInst &I) { /// Chain/Flag as the input and updates them for the output Chain/Flag. /// If the Flag pointer is NULL, no flag is used. SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, - unsigned Order, SDValue &Chain, - SDValue *Flag) const { + SDValue &Chain, SDValue *Flag) const { // Assemble the legal parts into the final values. SmallVector<SDValue, 4> Values(ValueVTs.size()); SmallVector<SDValue, 8> Parts; @@ -4765,7 +4773,7 @@ SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, Parts[i] = P; } - Values[Value] = getCopyFromParts(DAG, dl, Order, Parts.begin(), + Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(), NumRegs, RegisterVT, ValueVT); Part += NumRegs; Parts.clear(); @@ -4781,8 +4789,7 @@ SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG, DebugLoc dl, /// Chain/Flag as the input and updates them for the output Chain/Flag. /// If the Flag pointer is NULL, no flag is used. void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, - unsigned Order, SDValue &Chain, - SDValue *Flag) const { + SDValue &Chain, SDValue *Flag) const { // Get the list of the values's legal parts. unsigned NumRegs = Regs.size(); SmallVector<SDValue, 8> Parts(NumRegs); @@ -4791,7 +4798,7 @@ void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, unsigned NumParts = TLI->getNumRegisters(*DAG.getContext(), ValueVT); EVT RegisterVT = RegVTs[Value]; - getCopyToParts(DAG, dl, Order, + getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value), &Parts[Part], NumParts, RegisterVT); Part += NumParts; @@ -4832,7 +4839,7 @@ void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl, /// values added into it. void RegsForValue::AddInlineAsmOperands(unsigned Code, bool HasMatching,unsigned MatchingIdx, - SelectionDAG &DAG, unsigned Order, + SelectionDAG &DAG, std::vector<SDValue> &Ops) const { assert(Regs.size() < (1 << 13) && "Too many inline asm outputs!"); unsigned Flag = Code | (Regs.size() << 3); @@ -5330,7 +5337,8 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false); SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy()); Chain = DAG.getStore(Chain, getCurDebugLoc(), - OpInfo.CallOperand, StackSlot, NULL, 0); + OpInfo.CallOperand, StackSlot, NULL, 0, + false, false, 0); OpInfo.CallOperand = StackSlot; } @@ -5421,7 +5429,7 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { 2 /* REGDEF */ , false, 0, - DAG, SDNodeOrder, + DAG, AsmNodeOperands); break; } @@ -5469,10 +5477,10 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { // Use the produced MatchedRegs object to MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), - SDNodeOrder, Chain, &Flag); + Chain, &Flag); MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, true, OpInfo.getMatchedOperand(), - DAG, SDNodeOrder, AsmNodeOperands); + DAG, AsmNodeOperands); break; } else { assert(((OpFlag & 7) == 4) && "Unknown matching constraint!"); @@ -5533,11 +5541,10 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { } OpInfo.AssignedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(), - SDNodeOrder, Chain, &Flag); + Chain, &Flag); OpInfo.AssignedRegs.AddInlineAsmOperands(1/*REGUSE*/, false, 0, - DAG, SDNodeOrder, - AsmNodeOperands); + DAG, AsmNodeOperands); break; } case InlineAsm::isClobber: { @@ -5545,7 +5552,7 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { // allocator is aware that the physreg got clobbered. if (!OpInfo.AssignedRegs.Regs.empty()) OpInfo.AssignedRegs.AddInlineAsmOperands(6 /* EARLYCLOBBER REGDEF */, - false, 0, DAG, SDNodeOrder, + false, 0, DAG, AsmNodeOperands); break; } @@ -5565,7 +5572,7 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { // and set it as the value of the call. if (!RetValRegs.Regs.empty()) { SDValue Val = RetValRegs.getCopyFromRegs(DAG, getCurDebugLoc(), - SDNodeOrder, Chain, &Flag); + Chain, &Flag); // FIXME: Why don't we do this for inline asms with MRVs? if (CS.getType()->isSingleValueType() && CS.getType()->isSized()) { @@ -5605,7 +5612,7 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { RegsForValue &OutRegs = IndirectStoresToEmit[i].first; Value *Ptr = IndirectStoresToEmit[i].second; SDValue OutVal = OutRegs.getCopyFromRegs(DAG, getCurDebugLoc(), - SDNodeOrder, Chain, &Flag); + Chain, &Flag); StoresToEmit.push_back(std::make_pair(OutVal, Ptr)); } @@ -5616,7 +5623,8 @@ void SelectionDAGBuilder::visitInlineAsm(CallSite CS) { SDValue Val = DAG.getStore(Chain, getCurDebugLoc(), StoresToEmit[i].first, getValue(StoresToEmit[i].second), - StoresToEmit[i].second, 0); + StoresToEmit[i].second, 0, + false, false, 0); OutChains.push_back(Val); } @@ -5669,8 +5677,7 @@ TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy, CallingConv::ID CallConv, bool isTailCall, bool isReturnValueUsed, SDValue Callee, - ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl, - unsigned Order) { + ArgListTy &Args, SelectionDAG &DAG, DebugLoc dl) { // Handle all of the outgoing arguments. SmallVector<ISD::OutputArg, 32> Outs; for (unsigned i = 0, e = Args.size(); i != e; ++i) { @@ -5721,7 +5728,7 @@ TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy, else if (Args[i].isZExt) ExtendKind = ISD::ZERO_EXTEND; - getCopyToParts(DAG, dl, Order, Op, &Parts[0], NumParts, + getCopyToParts(DAG, dl, Op, &Parts[0], NumParts, PartVT, ExtendKind); for (unsigned j = 0; j != NumParts; ++j) { @@ -5800,7 +5807,7 @@ TargetLowering::LowerCallTo(SDValue Chain, const Type *RetTy, EVT RegisterVT = getRegisterType(RetTy->getContext(), VT); unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT); - ReturnValues.push_back(getCopyFromParts(DAG, dl, Order, &InVals[CurReg], + ReturnValues.push_back(getCopyFromParts(DAG, dl, &InVals[CurReg], NumRegs, RegisterVT, VT, AssertOp)); CurReg += NumRegs; @@ -5840,7 +5847,7 @@ void SelectionDAGBuilder::CopyValueToVirtualRegister(Value *V, unsigned Reg) { RegsForValue RFV(V->getContext(), TLI, Reg, V->getType()); SDValue Chain = DAG.getEntryNode(); - RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), SDNodeOrder, Chain, 0); + RFV.getCopyToRegs(Op, DAG, getCurDebugLoc(), Chain, 0); PendingExports.push_back(Chain); } @@ -5966,7 +5973,7 @@ void SelectionDAGISel::LowerArguments(BasicBlock *LLVMBB) { EVT VT = ValueVTs[0]; EVT RegVT = TLI.getRegisterType(*CurDAG->getContext(), VT); ISD::NodeType AssertOp = ISD::DELETED_NODE; - SDValue ArgValue = getCopyFromParts(DAG, dl, 0, &InVals[0], 1, + SDValue ArgValue = getCopyFromParts(DAG, dl, &InVals[0], 1, RegVT, VT, AssertOp); MachineFunction& MF = SDB->DAG.getMachineFunction(); @@ -6000,7 +6007,7 @@ void SelectionDAGISel::LowerArguments(BasicBlock *LLVMBB) { else if (F.paramHasAttr(Idx, Attribute::ZExt)) AssertOp = ISD::AssertZext; - ArgValues.push_back(getCopyFromParts(DAG, dl, 0, &InVals[i], + ArgValues.push_back(getCopyFromParts(DAG, dl, &InVals[i], NumParts, PartVT, VT, AssertOp)); } diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp index da2e6e4..05f9f1f 100644 --- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp +++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp @@ -56,9 +56,12 @@ #include "llvm/Support/MathExtras.h" #include "llvm/Support/Timer.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/ADT/Statistic.h" #include <algorithm> using namespace llvm; +STATISTIC(NumFastIselFailures, "Number of instructions fast isel failed on"); + static cl::opt<bool> EnableFastISelVerbose("fast-isel-verbose", cl::Hidden, cl::desc("Enable verbose messages in the \"fast\" " @@ -723,9 +726,9 @@ void SelectionDAGISel::CodeGenAndEmitDAG() { // code to the MachineBasicBlock. if (TimePassesIsEnabled) { NamedRegionTimer T("Instruction Selection", GroupName); - InstructionSelect(); + DoInstructionSelection(); } else { - InstructionSelect(); + DoInstructionSelection(); } DEBUG(dbgs() << "Selected selection DAG:\n"); @@ -765,6 +768,66 @@ void SelectionDAGISel::CodeGenAndEmitDAG() { DEBUG(BB->dump()); } +void SelectionDAGISel::DoInstructionSelection() { + DEBUG(errs() << "===== Instruction selection begins:\n"); + + PreprocessISelDAG(); + + // Select target instructions for the DAG. + { + // Number all nodes with a topological order and set DAGSize. + DAGSize = CurDAG->AssignTopologicalOrder(); + + // Create a dummy node (which is not added to allnodes), that adds + // a reference to the root node, preventing it from being deleted, + // and tracking any changes of the root. + HandleSDNode Dummy(CurDAG->getRoot()); + ISelPosition = SelectionDAG::allnodes_iterator(CurDAG->getRoot().getNode()); + ++ISelPosition; + + // The AllNodes list is now topological-sorted. Visit the + // nodes by starting at the end of the list (the root of the + // graph) and preceding back toward the beginning (the entry + // node). + while (ISelPosition != CurDAG->allnodes_begin()) { + SDNode *Node = --ISelPosition; + // Skip dead nodes. DAGCombiner is expected to eliminate all dead nodes, + // but there are currently some corner cases that it misses. Also, this + // makes it theoretically possible to disable the DAGCombiner. + if (Node->use_empty()) + continue; + + SDNode *ResNode = Select(Node); + + // FIXME: This is pretty gross. 'Select' should be changed to not return + // anything at all and this code should be nuked with a tactical strike. + + // If node should not be replaced, continue with the next one. + if (ResNode == Node || Node->getOpcode() == ISD::DELETED_NODE) + continue; + // Replace node. + if (ResNode) + ReplaceUses(Node, ResNode); + + // If after the replacement this node is not used any more, + // remove this dead node. + if (Node->use_empty()) { // Don't delete EntryToken, etc. + ISelUpdater ISU(ISelPosition); + CurDAG->RemoveDeadNode(Node, &ISU); + } + } + + CurDAG->setRoot(Dummy.getValue()); + } + DEBUG(errs() << "===== Instruction selection ends:\n"); + + PostprocessISelDAG(); + + // FIXME: This shouldn't be needed, remove it. + CurDAG->RemoveDeadNodes(); +} + + void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, MachineFunction &MF, MachineModuleInfo *MMI, @@ -870,6 +933,7 @@ void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, // feed PHI nodes in successor blocks. if (isa<TerminatorInst>(BI)) if (!HandlePHINodesInSuccessorBlocksFast(LLVMBB, FastIS)) { + ++NumFastIselFailures; ResetDebugLoc(SDB, FastIS); if (EnableFastISelVerbose || EnableFastISelAbort) { dbgs() << "FastISel miss: "; @@ -894,6 +958,7 @@ void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, // Then handle certain instructions as single-LLVM-Instruction blocks. if (isa<CallInst>(BI)) { + ++NumFastIselFailures; if (EnableFastISelVerbose || EnableFastISelAbort) { dbgs() << "FastISel missed call: "; BI->dump(); @@ -923,6 +988,7 @@ void SelectionDAGISel::SelectAllBasicBlocks(Function &Fn, // Otherwise, give up on FastISel for the rest of the block. // For now, be a little lenient about non-branch terminators. if (!isa<TerminatorInst>(BI) || isa<BranchInst>(BI)) { + ++NumFastIselFailures; if (EnableFastISelVerbose || EnableFastISelAbort) { dbgs() << "FastISel miss: "; BI->dump(); @@ -972,6 +1038,8 @@ SelectionDAGISel::FinishBasicBlock() { MachineInstr *PHI = SDB->PHINodesToUpdate[i].first; assert(PHI->isPHI() && "This is not a machine PHI node that we are updating!"); + if (!BB->isSuccessor(PHI->getParent())) + continue; PHI->addOperand(MachineOperand::CreateReg(SDB->PHINodesToUpdate[i].second, false)); PHI->addOperand(MachineOperand::CreateMBB(BB)); @@ -1316,13 +1384,29 @@ static SDNode *findFlagUse(SDNode *N) { /// This function recursively traverses up the operand chain, ignoring /// certain nodes. static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse, - SDNode *Root, - SmallPtrSet<SDNode*, 16> &Visited) { - if (Use->getNodeId() < Def->getNodeId() || - !Visited.insert(Use)) + SDNode *Root, SmallPtrSet<SDNode*, 16> &Visited, + bool IgnoreChains) { + // The NodeID's are given uniques ID's where a node ID is guaranteed to be + // greater than all of its (recursive) operands. If we scan to a point where + // 'use' is smaller than the node we're scanning for, then we know we will + // never find it. + // + // The Use may be -1 (unassigned) if it is a newly allocated node. This can + // happen because we scan down to newly selected nodes in the case of flag + // uses. + if ((Use->getNodeId() < Def->getNodeId() && Use->getNodeId() != -1)) + return false; + + // Don't revisit nodes if we already scanned it and didn't fail, we know we + // won't fail if we scan it again. + if (!Visited.insert(Use)) return false; for (unsigned i = 0, e = Use->getNumOperands(); i != e; ++i) { + // Ignore chain uses, they are validated by HandleMergeInputChains. + if (Use->getOperand(i).getValueType() == MVT::Other && IgnoreChains) + continue; + SDNode *N = Use->getOperand(i).getNode(); if (N == Def) { if (Use == ImmedUse || Use == Root) @@ -1332,32 +1416,24 @@ static bool findNonImmUse(SDNode *Use, SDNode* Def, SDNode *ImmedUse, } // Traverse up the operand chain. - if (findNonImmUse(N, Def, ImmedUse, Root, Visited)) + if (findNonImmUse(N, Def, ImmedUse, Root, Visited, IgnoreChains)) return true; } return false; } -/// isNonImmUse - Start searching from Root up the DAG to check is Def can -/// be reached. Return true if that's the case. However, ignore direct uses -/// by ImmedUse (which would be U in the example illustrated in -/// IsLegalAndProfitableToFold) and by Root (which can happen in the store -/// case). -/// FIXME: to be really generic, we should allow direct use by any node -/// that is being folded. But realisticly since we only fold loads which -/// have one non-chain use, we only need to watch out for load/op/store -/// and load/op/cmp case where the root (store / cmp) may reach the load via -/// its chain operand. -static inline bool isNonImmUse(SDNode *Root, SDNode *Def, SDNode *ImmedUse) { - SmallPtrSet<SDNode*, 16> Visited; - return findNonImmUse(Root, Def, ImmedUse, Root, Visited); +/// IsProfitableToFold - Returns true if it's profitable to fold the specific +/// operand node N of U during instruction selection that starts at Root. +bool SelectionDAGISel::IsProfitableToFold(SDValue N, SDNode *U, + SDNode *Root) const { + if (OptLevel == CodeGenOpt::None) return false; + return N.hasOneUse(); } -/// IsLegalAndProfitableToFold - Returns true if the specific operand node N of -/// U can be folded during instruction selection that starts at Root and -/// folding N is profitable. -bool SelectionDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U, - SDNode *Root) const { +/// IsLegalToFold - Returns true if the specific operand node N of +/// U can be folded during instruction selection that starts at Root. +bool SelectionDAGISel::IsLegalToFold(SDValue N, SDNode *U, SDNode *Root, + bool IgnoreChains) const { if (OptLevel == CodeGenOpt::None) return false; // If Root use can somehow reach N through a path that that doesn't contain @@ -1402,6 +1478,8 @@ bool SelectionDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U, // Fold. But since Fold and FU are flagged together, this will create // a cycle in the scheduling graph. + // If the node has flags, walk down the graph to the "lowest" node in the + // flagged set. EVT VT = Root->getValueType(Root->getNumValues()-1); while (VT == MVT::Flag) { SDNode *FU = findFlagUse(Root); @@ -1409,9 +1487,17 @@ bool SelectionDAGISel::IsLegalAndProfitableToFold(SDNode *N, SDNode *U, break; Root = FU; VT = Root->getValueType(Root->getNumValues()-1); + + // If our query node has a flag result with a use, we've walked up it. If + // the user (which has already been selected) has a chain or indirectly uses + // the chain, our WalkChainUsers predicate will not consider it. Because of + // this, we cannot ignore chains in this predicate. + IgnoreChains = false; } + - return !isNonImmUse(Root, N, U); + SmallPtrSet<SDNode*, 16> Visited; + return !findNonImmUse(Root, N.getNode(), U, Root, Visited, IgnoreChains); } SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) { @@ -1423,6 +1509,7 @@ SDNode *SelectionDAGISel::Select_INLINEASM(SDNode *N) { VTs.push_back(MVT::Flag); SDValue New = CurDAG->getNode(ISD::INLINEASM, N->getDebugLoc(), VTs, &Ops[0], Ops.size()); + New->setNodeId(-1); return New.getNode(); } @@ -1438,25 +1525,1219 @@ SDNode *SelectionDAGISel::Select_EH_LABEL(SDNode *N) { MVT::Other, Tmp, Chain); } +/// GetVBR - decode a vbr encoding whose top bit is set. +ALWAYS_INLINE static uint64_t +GetVBR(uint64_t Val, const unsigned char *MatcherTable, unsigned &Idx) { + assert(Val >= 128 && "Not a VBR"); + Val &= 127; // Remove first vbr bit. + + unsigned Shift = 7; + uint64_t NextBits; + do { + NextBits = MatcherTable[Idx++]; + Val |= (NextBits&127) << Shift; + Shift += 7; + } while (NextBits & 128); + + return Val; +} + + +/// UpdateChainsAndFlags - When a match is complete, this method updates uses of +/// interior flag and chain results to use the new flag and chain results. +void SelectionDAGISel:: +UpdateChainsAndFlags(SDNode *NodeToMatch, SDValue InputChain, + const SmallVectorImpl<SDNode*> &ChainNodesMatched, + SDValue InputFlag, + const SmallVectorImpl<SDNode*> &FlagResultNodesMatched, + bool isMorphNodeTo) { + SmallVector<SDNode*, 4> NowDeadNodes; + + ISelUpdater ISU(ISelPosition); + + // Now that all the normal results are replaced, we replace the chain and + // flag results if present. + if (!ChainNodesMatched.empty()) { + assert(InputChain.getNode() != 0 && + "Matched input chains but didn't produce a chain"); + // Loop over all of the nodes we matched that produced a chain result. + // Replace all the chain results with the final chain we ended up with. + for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { + SDNode *ChainNode = ChainNodesMatched[i]; + + // If this node was already deleted, don't look at it. + if (ChainNode->getOpcode() == ISD::DELETED_NODE) + continue; + + // Don't replace the results of the root node if we're doing a + // MorphNodeTo. + if (ChainNode == NodeToMatch && isMorphNodeTo) + continue; + + SDValue ChainVal = SDValue(ChainNode, ChainNode->getNumValues()-1); + if (ChainVal.getValueType() == MVT::Flag) + ChainVal = ChainVal.getValue(ChainVal->getNumValues()-2); + assert(ChainVal.getValueType() == MVT::Other && "Not a chain?"); + CurDAG->ReplaceAllUsesOfValueWith(ChainVal, InputChain, &ISU); + + // If the node became dead, delete it. + if (ChainNode->use_empty()) + NowDeadNodes.push_back(ChainNode); + } + } + + // If the result produces a flag, update any flag results in the matched + // pattern with the flag result. + if (InputFlag.getNode() != 0) { + // Handle any interior nodes explicitly marked. + for (unsigned i = 0, e = FlagResultNodesMatched.size(); i != e; ++i) { + SDNode *FRN = FlagResultNodesMatched[i]; + + // If this node was already deleted, don't look at it. + if (FRN->getOpcode() == ISD::DELETED_NODE) + continue; + + assert(FRN->getValueType(FRN->getNumValues()-1) == MVT::Flag && + "Doesn't have a flag result"); + CurDAG->ReplaceAllUsesOfValueWith(SDValue(FRN, FRN->getNumValues()-1), + InputFlag, &ISU); + + // If the node became dead, delete it. + if (FRN->use_empty()) + NowDeadNodes.push_back(FRN); + } + } + + if (!NowDeadNodes.empty()) + CurDAG->RemoveDeadNodes(NowDeadNodes, &ISU); + + DEBUG(errs() << "ISEL: Match complete!\n"); +} + +enum ChainResult { + CR_Simple, + CR_InducesCycle, + CR_LeadsToInteriorNode +}; + +/// WalkChainUsers - Walk down the users of the specified chained node that is +/// part of the pattern we're matching, looking at all of the users we find. +/// This determines whether something is an interior node, whether we have a +/// non-pattern node in between two pattern nodes (which prevent folding because +/// it would induce a cycle) and whether we have a TokenFactor node sandwiched +/// between pattern nodes (in which case the TF becomes part of the pattern). +/// +/// The walk we do here is guaranteed to be small because we quickly get down to +/// already selected nodes "below" us. +static ChainResult +WalkChainUsers(SDNode *ChainedNode, + SmallVectorImpl<SDNode*> &ChainedNodesInPattern, + SmallVectorImpl<SDNode*> &InteriorChainedNodes) { + ChainResult Result = CR_Simple; + + for (SDNode::use_iterator UI = ChainedNode->use_begin(), + E = ChainedNode->use_end(); UI != E; ++UI) { + // Make sure the use is of the chain, not some other value we produce. + if (UI.getUse().getValueType() != MVT::Other) continue; + + SDNode *User = *UI; + + // If we see an already-selected machine node, then we've gone beyond the + // pattern that we're selecting down into the already selected chunk of the + // DAG. + if (User->isMachineOpcode() || + User->getOpcode() == ISD::HANDLENODE) // Root of the graph. + continue; + + if (User->getOpcode() == ISD::CopyToReg || + User->getOpcode() == ISD::CopyFromReg || + User->getOpcode() == ISD::INLINEASM) { + // If their node ID got reset to -1 then they've already been selected. + // Treat them like a MachineOpcode. + if (User->getNodeId() == -1) + continue; + } + + // If we have a TokenFactor, we handle it specially. + if (User->getOpcode() != ISD::TokenFactor) { + // If the node isn't a token factor and isn't part of our pattern, then it + // must be a random chained node in between two nodes we're selecting. + // This happens when we have something like: + // x = load ptr + // call + // y = x+4 + // store y -> ptr + // Because we structurally match the load/store as a read/modify/write, + // but the call is chained between them. We cannot fold in this case + // because it would induce a cycle in the graph. + if (!std::count(ChainedNodesInPattern.begin(), + ChainedNodesInPattern.end(), User)) + return CR_InducesCycle; + + // Otherwise we found a node that is part of our pattern. For example in: + // x = load ptr + // y = x+4 + // store y -> ptr + // This would happen when we're scanning down from the load and see the + // store as a user. Record that there is a use of ChainedNode that is + // part of the pattern and keep scanning uses. + Result = CR_LeadsToInteriorNode; + InteriorChainedNodes.push_back(User); + continue; + } + + // If we found a TokenFactor, there are two cases to consider: first if the + // TokenFactor is just hanging "below" the pattern we're matching (i.e. no + // uses of the TF are in our pattern) we just want to ignore it. Second, + // the TokenFactor can be sandwiched in between two chained nodes, like so: + // [Load chain] + // ^ + // | + // [Load] + // ^ ^ + // | \ DAG's like cheese + // / \ do you? + // / | + // [TokenFactor] [Op] + // ^ ^ + // | | + // \ / + // \ / + // [Store] + // + // In this case, the TokenFactor becomes part of our match and we rewrite it + // as a new TokenFactor. + // + // To distinguish these two cases, do a recursive walk down the uses. + switch (WalkChainUsers(User, ChainedNodesInPattern, InteriorChainedNodes)) { + case CR_Simple: + // If the uses of the TokenFactor are just already-selected nodes, ignore + // it, it is "below" our pattern. + continue; + case CR_InducesCycle: + // If the uses of the TokenFactor lead to nodes that are not part of our + // pattern that are not selected, folding would turn this into a cycle, + // bail out now. + return CR_InducesCycle; + case CR_LeadsToInteriorNode: + break; // Otherwise, keep processing. + } + + // Okay, we know we're in the interesting interior case. The TokenFactor + // is now going to be considered part of the pattern so that we rewrite its + // uses (it may have uses that are not part of the pattern) with the + // ultimate chain result of the generated code. We will also add its chain + // inputs as inputs to the ultimate TokenFactor we create. + Result = CR_LeadsToInteriorNode; + ChainedNodesInPattern.push_back(User); + InteriorChainedNodes.push_back(User); + continue; + } + + return Result; +} + +/// HandleMergeInputChains - This implements the OPC_EmitMergeInputChains +/// operation for when the pattern matched at least one node with a chains. The +/// input vector contains a list of all of the chained nodes that we match. We +/// must determine if this is a valid thing to cover (i.e. matching it won't +/// induce cycles in the DAG) and if so, creating a TokenFactor node. that will +/// be used as the input node chain for the generated nodes. +static SDValue +HandleMergeInputChains(SmallVectorImpl<SDNode*> &ChainNodesMatched, + SelectionDAG *CurDAG) { + // Walk all of the chained nodes we've matched, recursively scanning down the + // users of the chain result. This adds any TokenFactor nodes that are caught + // in between chained nodes to the chained and interior nodes list. + SmallVector<SDNode*, 3> InteriorChainedNodes; + for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { + if (WalkChainUsers(ChainNodesMatched[i], ChainNodesMatched, + InteriorChainedNodes) == CR_InducesCycle) + return SDValue(); // Would induce a cycle. + } + + // Okay, we have walked all the matched nodes and collected TokenFactor nodes + // that we are interested in. Form our input TokenFactor node. + SmallVector<SDValue, 3> InputChains; + for (unsigned i = 0, e = ChainNodesMatched.size(); i != e; ++i) { + // Add the input chain of this node to the InputChains list (which will be + // the operands of the generated TokenFactor) if it's not an interior node. + SDNode *N = ChainNodesMatched[i]; + if (N->getOpcode() != ISD::TokenFactor) { + if (std::count(InteriorChainedNodes.begin(),InteriorChainedNodes.end(),N)) + continue; + + // Otherwise, add the input chain. + SDValue InChain = ChainNodesMatched[i]->getOperand(0); + assert(InChain.getValueType() == MVT::Other && "Not a chain"); + InputChains.push_back(InChain); + continue; + } + + // If we have a token factor, we want to add all inputs of the token factor + // that are not part of the pattern we're matching. + for (unsigned op = 0, e = N->getNumOperands(); op != e; ++op) { + if (!std::count(ChainNodesMatched.begin(), ChainNodesMatched.end(), + N->getOperand(op).getNode())) + InputChains.push_back(N->getOperand(op)); + } + } + + SDValue Res; + if (InputChains.size() == 1) + return InputChains[0]; + return CurDAG->getNode(ISD::TokenFactor, ChainNodesMatched[0]->getDebugLoc(), + MVT::Other, &InputChains[0], InputChains.size()); +} + +/// MorphNode - Handle morphing a node in place for the selector. +SDNode *SelectionDAGISel:: +MorphNode(SDNode *Node, unsigned TargetOpc, SDVTList VTList, + const SDValue *Ops, unsigned NumOps, unsigned EmitNodeInfo) { + // It is possible we're using MorphNodeTo to replace a node with no + // normal results with one that has a normal result (or we could be + // adding a chain) and the input could have flags and chains as well. + // In this case we need to shifting the operands down. + // FIXME: This is a horrible hack and broken in obscure cases, no worse + // than the old isel though. We should sink this into MorphNodeTo. + int OldFlagResultNo = -1, OldChainResultNo = -1; + + unsigned NTMNumResults = Node->getNumValues(); + if (Node->getValueType(NTMNumResults-1) == MVT::Flag) { + OldFlagResultNo = NTMNumResults-1; + if (NTMNumResults != 1 && + Node->getValueType(NTMNumResults-2) == MVT::Other) + OldChainResultNo = NTMNumResults-2; + } else if (Node->getValueType(NTMNumResults-1) == MVT::Other) + OldChainResultNo = NTMNumResults-1; + + // Call the underlying SelectionDAG routine to do the transmogrification. Note + // that this deletes operands of the old node that become dead. + SDNode *Res = CurDAG->MorphNodeTo(Node, ~TargetOpc, VTList, Ops, NumOps); + + // MorphNodeTo can operate in two ways: if an existing node with the + // specified operands exists, it can just return it. Otherwise, it + // updates the node in place to have the requested operands. + if (Res == Node) { + // If we updated the node in place, reset the node ID. To the isel, + // this should be just like a newly allocated machine node. + Res->setNodeId(-1); + } + + unsigned ResNumResults = Res->getNumValues(); + // Move the flag if needed. + if ((EmitNodeInfo & OPFL_FlagOutput) && OldFlagResultNo != -1 && + (unsigned)OldFlagResultNo != ResNumResults-1) + CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldFlagResultNo), + SDValue(Res, ResNumResults-1)); + + if ((EmitNodeInfo & OPFL_FlagOutput) != 0) + --ResNumResults; + + // Move the chain reference if needed. + if ((EmitNodeInfo & OPFL_Chain) && OldChainResultNo != -1 && + (unsigned)OldChainResultNo != ResNumResults-1) + CurDAG->ReplaceAllUsesOfValueWith(SDValue(Node, OldChainResultNo), + SDValue(Res, ResNumResults-1)); + + // Otherwise, no replacement happened because the node already exists. Replace + // Uses of the old node with the new one. + if (Res != Node) + CurDAG->ReplaceAllUsesWith(Node, Res); + + return Res; +} + +/// CheckPatternPredicate - Implements OP_CheckPatternPredicate. +ALWAYS_INLINE static bool +CheckSame(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, const SmallVectorImpl<SDValue> &RecordedNodes) { + // Accept if it is exactly the same as a previously recorded node. + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + return N == RecordedNodes[RecNo]; +} + +/// CheckPatternPredicate - Implements OP_CheckPatternPredicate. +ALWAYS_INLINE static bool +CheckPatternPredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SelectionDAGISel &SDISel) { + return SDISel.CheckPatternPredicate(MatcherTable[MatcherIndex++]); +} + +/// CheckNodePredicate - Implements OP_CheckNodePredicate. +ALWAYS_INLINE static bool +CheckNodePredicate(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SelectionDAGISel &SDISel, SDNode *N) { + return SDISel.CheckNodePredicate(N, MatcherTable[MatcherIndex++]); +} + +ALWAYS_INLINE static bool +CheckOpcode(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDNode *N) { + return N->getOpcode() == MatcherTable[MatcherIndex++]; +} + +ALWAYS_INLINE static bool +CheckType(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, const TargetLowering &TLI) { + MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + if (N.getValueType() == VT) return true; + + // Handle the case when VT is iPTR. + return VT == MVT::iPTR && N.getValueType() == TLI.getPointerTy(); +} + +ALWAYS_INLINE static bool +CheckChildType(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, const TargetLowering &TLI, + unsigned ChildNo) { + if (ChildNo >= N.getNumOperands()) + return false; // Match fails if out of range child #. + return ::CheckType(MatcherTable, MatcherIndex, N.getOperand(ChildNo), TLI); +} + + +ALWAYS_INLINE static bool +CheckCondCode(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N) { + return cast<CondCodeSDNode>(N)->get() == + (ISD::CondCode)MatcherTable[MatcherIndex++]; +} + +ALWAYS_INLINE static bool +CheckValueType(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, const TargetLowering &TLI) { + MVT::SimpleValueType VT = (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + if (cast<VTSDNode>(N)->getVT() == VT) + return true; + + // Handle the case when VT is iPTR. + return VT == MVT::iPTR && cast<VTSDNode>(N)->getVT() == TLI.getPointerTy(); +} + +ALWAYS_INLINE static bool +CheckInteger(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N) { + int64_t Val = MatcherTable[MatcherIndex++]; + if (Val & 128) + Val = GetVBR(Val, MatcherTable, MatcherIndex); + + ConstantSDNode *C = dyn_cast<ConstantSDNode>(N); + return C != 0 && C->getSExtValue() == Val; +} + +ALWAYS_INLINE static bool +CheckAndImm(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, SelectionDAGISel &SDISel) { + int64_t Val = MatcherTable[MatcherIndex++]; + if (Val & 128) + Val = GetVBR(Val, MatcherTable, MatcherIndex); + + if (N->getOpcode() != ISD::AND) return false; + + ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); + return C != 0 && SDISel.CheckAndMask(N.getOperand(0), C, Val); +} + +ALWAYS_INLINE static bool +CheckOrImm(const unsigned char *MatcherTable, unsigned &MatcherIndex, + SDValue N, SelectionDAGISel &SDISel) { + int64_t Val = MatcherTable[MatcherIndex++]; + if (Val & 128) + Val = GetVBR(Val, MatcherTable, MatcherIndex); + + if (N->getOpcode() != ISD::OR) return false; + + ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1)); + return C != 0 && SDISel.CheckOrMask(N.getOperand(0), C, Val); +} + +/// IsPredicateKnownToFail - If we know how and can do so without pushing a +/// scope, evaluate the current node. If the current predicate is known to +/// fail, set Result=true and return anything. If the current predicate is +/// known to pass, set Result=false and return the MatcherIndex to continue +/// with. If the current predicate is unknown, set Result=false and return the +/// MatcherIndex to continue with. +static unsigned IsPredicateKnownToFail(const unsigned char *Table, + unsigned Index, SDValue N, + bool &Result, SelectionDAGISel &SDISel, + SmallVectorImpl<SDValue> &RecordedNodes){ + switch (Table[Index++]) { + default: + Result = false; + return Index-1; // Could not evaluate this predicate. + case SelectionDAGISel::OPC_CheckSame: + Result = !::CheckSame(Table, Index, N, RecordedNodes); + return Index; + case SelectionDAGISel::OPC_CheckPatternPredicate: + Result = !::CheckPatternPredicate(Table, Index, SDISel); + return Index; + case SelectionDAGISel::OPC_CheckPredicate: + Result = !::CheckNodePredicate(Table, Index, SDISel, N.getNode()); + return Index; + case SelectionDAGISel::OPC_CheckOpcode: + Result = !::CheckOpcode(Table, Index, N.getNode()); + return Index; + case SelectionDAGISel::OPC_CheckType: + Result = !::CheckType(Table, Index, N, SDISel.TLI); + return Index; + case SelectionDAGISel::OPC_CheckChild0Type: + case SelectionDAGISel::OPC_CheckChild1Type: + case SelectionDAGISel::OPC_CheckChild2Type: + case SelectionDAGISel::OPC_CheckChild3Type: + case SelectionDAGISel::OPC_CheckChild4Type: + case SelectionDAGISel::OPC_CheckChild5Type: + case SelectionDAGISel::OPC_CheckChild6Type: + case SelectionDAGISel::OPC_CheckChild7Type: + Result = !::CheckChildType(Table, Index, N, SDISel.TLI, + Table[Index-1] - SelectionDAGISel::OPC_CheckChild0Type); + return Index; + case SelectionDAGISel::OPC_CheckCondCode: + Result = !::CheckCondCode(Table, Index, N); + return Index; + case SelectionDAGISel::OPC_CheckValueType: + Result = !::CheckValueType(Table, Index, N, SDISel.TLI); + return Index; + case SelectionDAGISel::OPC_CheckInteger: + Result = !::CheckInteger(Table, Index, N); + return Index; + case SelectionDAGISel::OPC_CheckAndImm: + Result = !::CheckAndImm(Table, Index, N, SDISel); + return Index; + case SelectionDAGISel::OPC_CheckOrImm: + Result = !::CheckOrImm(Table, Index, N, SDISel); + return Index; + } +} + + +struct MatchScope { + /// FailIndex - If this match fails, this is the index to continue with. + unsigned FailIndex; + + /// NodeStack - The node stack when the scope was formed. + SmallVector<SDValue, 4> NodeStack; + + /// NumRecordedNodes - The number of recorded nodes when the scope was formed. + unsigned NumRecordedNodes; + + /// NumMatchedMemRefs - The number of matched memref entries. + unsigned NumMatchedMemRefs; + + /// InputChain/InputFlag - The current chain/flag + SDValue InputChain, InputFlag; + + /// HasChainNodesMatched - True if the ChainNodesMatched list is non-empty. + bool HasChainNodesMatched, HasFlagResultNodesMatched; +}; + +SDNode *SelectionDAGISel:: +SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable, + unsigned TableSize) { + // FIXME: Should these even be selected? Handle these cases in the caller? + switch (NodeToMatch->getOpcode()) { + default: + break; + case ISD::EntryToken: // These nodes remain the same. + case ISD::BasicBlock: + case ISD::Register: + case ISD::HANDLENODE: + case ISD::TargetConstant: + case ISD::TargetConstantFP: + case ISD::TargetConstantPool: + case ISD::TargetFrameIndex: + case ISD::TargetExternalSymbol: + case ISD::TargetBlockAddress: + case ISD::TargetJumpTable: + case ISD::TargetGlobalTLSAddress: + case ISD::TargetGlobalAddress: + case ISD::TokenFactor: + case ISD::CopyFromReg: + case ISD::CopyToReg: + NodeToMatch->setNodeId(-1); // Mark selected. + return 0; + case ISD::AssertSext: + case ISD::AssertZext: + CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, 0), + NodeToMatch->getOperand(0)); + return 0; + case ISD::INLINEASM: return Select_INLINEASM(NodeToMatch); + case ISD::EH_LABEL: return Select_EH_LABEL(NodeToMatch); + case ISD::UNDEF: return Select_UNDEF(NodeToMatch); + } + + assert(!NodeToMatch->isMachineOpcode() && "Node already selected!"); + + // Set up the node stack with NodeToMatch as the only node on the stack. + SmallVector<SDValue, 8> NodeStack; + SDValue N = SDValue(NodeToMatch, 0); + NodeStack.push_back(N); + + // MatchScopes - Scopes used when matching, if a match failure happens, this + // indicates where to continue checking. + SmallVector<MatchScope, 8> MatchScopes; + + // RecordedNodes - This is the set of nodes that have been recorded by the + // state machine. + SmallVector<SDValue, 8> RecordedNodes; + + // MatchedMemRefs - This is the set of MemRef's we've seen in the input + // pattern. + SmallVector<MachineMemOperand*, 2> MatchedMemRefs; + + // These are the current input chain and flag for use when generating nodes. + // Various Emit operations change these. For example, emitting a copytoreg + // uses and updates these. + SDValue InputChain, InputFlag; + + // ChainNodesMatched - If a pattern matches nodes that have input/output + // chains, the OPC_EmitMergeInputChains operation is emitted which indicates + // which ones they are. The result is captured into this list so that we can + // update the chain results when the pattern is complete. + SmallVector<SDNode*, 3> ChainNodesMatched; + SmallVector<SDNode*, 3> FlagResultNodesMatched; + + DEBUG(errs() << "ISEL: Starting pattern match on root node: "; + NodeToMatch->dump(CurDAG); + errs() << '\n'); + + // Determine where to start the interpreter. Normally we start at opcode #0, + // but if the state machine starts with an OPC_SwitchOpcode, then we + // accelerate the first lookup (which is guaranteed to be hot) with the + // OpcodeOffset table. + unsigned MatcherIndex = 0; + + if (!OpcodeOffset.empty()) { + // Already computed the OpcodeOffset table, just index into it. + if (N.getOpcode() < OpcodeOffset.size()) + MatcherIndex = OpcodeOffset[N.getOpcode()]; + DEBUG(errs() << " Initial Opcode index to " << MatcherIndex << "\n"); + + } else if (MatcherTable[0] == OPC_SwitchOpcode) { + // Otherwise, the table isn't computed, but the state machine does start + // with an OPC_SwitchOpcode instruction. Populate the table now, since this + // is the first time we're selecting an instruction. + unsigned Idx = 1; + while (1) { + // Get the size of this case. + unsigned CaseSize = MatcherTable[Idx++]; + if (CaseSize & 128) + CaseSize = GetVBR(CaseSize, MatcherTable, Idx); + if (CaseSize == 0) break; + + // Get the opcode, add the index to the table. + unsigned Opc = MatcherTable[Idx++]; + if (Opc >= OpcodeOffset.size()) + OpcodeOffset.resize((Opc+1)*2); + OpcodeOffset[Opc] = Idx; + Idx += CaseSize; + } + + // Okay, do the lookup for the first opcode. + if (N.getOpcode() < OpcodeOffset.size()) + MatcherIndex = OpcodeOffset[N.getOpcode()]; + } + + while (1) { + assert(MatcherIndex < TableSize && "Invalid index"); + BuiltinOpcodes Opcode = (BuiltinOpcodes)MatcherTable[MatcherIndex++]; + switch (Opcode) { + case OPC_Scope: { + // Okay, the semantics of this operation are that we should push a scope + // then evaluate the first child. However, pushing a scope only to have + // the first check fail (which then pops it) is inefficient. If we can + // determine immediately that the first check (or first several) will + // immediately fail, don't even bother pushing a scope for them. + unsigned FailIndex; + + while (1) { + unsigned NumToSkip = MatcherTable[MatcherIndex++]; + if (NumToSkip & 128) + NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex); + // Found the end of the scope with no match. + if (NumToSkip == 0) { + FailIndex = 0; + break; + } + + FailIndex = MatcherIndex+NumToSkip; + + // If we can't evaluate this predicate without pushing a scope (e.g. if + // it is a 'MoveParent') or if the predicate succeeds on this node, we + // push the scope and evaluate the full predicate chain. + bool Result; + MatcherIndex = IsPredicateKnownToFail(MatcherTable, MatcherIndex, N, + Result, *this, RecordedNodes); + if (!Result) + break; + + DEBUG(errs() << " Skipped scope entry at index " << MatcherIndex + << " continuing at " << FailIndex << "\n"); + + + // Otherwise, we know that this case of the Scope is guaranteed to fail, + // move to the next case. + MatcherIndex = FailIndex; + } + + // If the whole scope failed to match, bail. + if (FailIndex == 0) break; + + // Push a MatchScope which indicates where to go if the first child fails + // to match. + MatchScope NewEntry; + NewEntry.FailIndex = FailIndex; + NewEntry.NodeStack.append(NodeStack.begin(), NodeStack.end()); + NewEntry.NumRecordedNodes = RecordedNodes.size(); + NewEntry.NumMatchedMemRefs = MatchedMemRefs.size(); + NewEntry.InputChain = InputChain; + NewEntry.InputFlag = InputFlag; + NewEntry.HasChainNodesMatched = !ChainNodesMatched.empty(); + NewEntry.HasFlagResultNodesMatched = !FlagResultNodesMatched.empty(); + MatchScopes.push_back(NewEntry); + continue; + } + case OPC_RecordNode: + // Remember this node, it may end up being an operand in the pattern. + RecordedNodes.push_back(N); + continue; + + case OPC_RecordChild0: case OPC_RecordChild1: + case OPC_RecordChild2: case OPC_RecordChild3: + case OPC_RecordChild4: case OPC_RecordChild5: + case OPC_RecordChild6: case OPC_RecordChild7: { + unsigned ChildNo = Opcode-OPC_RecordChild0; + if (ChildNo >= N.getNumOperands()) + break; // Match fails if out of range child #. + + RecordedNodes.push_back(N->getOperand(ChildNo)); + continue; + } + case OPC_RecordMemRef: + MatchedMemRefs.push_back(cast<MemSDNode>(N)->getMemOperand()); + continue; + + case OPC_CaptureFlagInput: + // If the current node has an input flag, capture it in InputFlag. + if (N->getNumOperands() != 0 && + N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) + InputFlag = N->getOperand(N->getNumOperands()-1); + continue; + + case OPC_MoveChild: { + unsigned ChildNo = MatcherTable[MatcherIndex++]; + if (ChildNo >= N.getNumOperands()) + break; // Match fails if out of range child #. + N = N.getOperand(ChildNo); + NodeStack.push_back(N); + continue; + } + + case OPC_MoveParent: + // Pop the current node off the NodeStack. + NodeStack.pop_back(); + assert(!NodeStack.empty() && "Node stack imbalance!"); + N = NodeStack.back(); + continue; + + case OPC_CheckSame: + if (!::CheckSame(MatcherTable, MatcherIndex, N, RecordedNodes)) break; + continue; + case OPC_CheckPatternPredicate: + if (!::CheckPatternPredicate(MatcherTable, MatcherIndex, *this)) break; + continue; + case OPC_CheckPredicate: + if (!::CheckNodePredicate(MatcherTable, MatcherIndex, *this, + N.getNode())) + break; + continue; + case OPC_CheckComplexPat: { + unsigned CPNum = MatcherTable[MatcherIndex++]; + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckComplexPat"); + if (!CheckComplexPattern(NodeToMatch, RecordedNodes[RecNo], CPNum, + RecordedNodes)) + break; + continue; + } + case OPC_CheckOpcode: + if (!::CheckOpcode(MatcherTable, MatcherIndex, N.getNode())) break; + continue; + + case OPC_CheckType: + if (!::CheckType(MatcherTable, MatcherIndex, N, TLI)) break; + continue; + + case OPC_SwitchOpcode: { + unsigned CurNodeOpcode = N.getOpcode(); + unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart; + unsigned CaseSize; + while (1) { + // Get the size of this case. + CaseSize = MatcherTable[MatcherIndex++]; + if (CaseSize & 128) + CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex); + if (CaseSize == 0) break; + + // If the opcode matches, then we will execute this case. + if (CurNodeOpcode == MatcherTable[MatcherIndex++]) + break; + + // Otherwise, skip over this case. + MatcherIndex += CaseSize; + } + + // If no cases matched, bail out. + if (CaseSize == 0) break; + + // Otherwise, execute the case we found. + DEBUG(errs() << " OpcodeSwitch from " << SwitchStart + << " to " << MatcherIndex << "\n"); + continue; + } + + case OPC_SwitchType: { + MVT::SimpleValueType CurNodeVT = N.getValueType().getSimpleVT().SimpleTy; + unsigned SwitchStart = MatcherIndex-1; (void)SwitchStart; + unsigned CaseSize; + while (1) { + // Get the size of this case. + CaseSize = MatcherTable[MatcherIndex++]; + if (CaseSize & 128) + CaseSize = GetVBR(CaseSize, MatcherTable, MatcherIndex); + if (CaseSize == 0) break; + + MVT::SimpleValueType CaseVT = + (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + if (CaseVT == MVT::iPTR) + CaseVT = TLI.getPointerTy().SimpleTy; + + // If the VT matches, then we will execute this case. + if (CurNodeVT == CaseVT) + break; + + // Otherwise, skip over this case. + MatcherIndex += CaseSize; + } + + // If no cases matched, bail out. + if (CaseSize == 0) break; + + // Otherwise, execute the case we found. + DEBUG(errs() << " TypeSwitch[" << EVT(CurNodeVT).getEVTString() + << "] from " << SwitchStart << " to " << MatcherIndex<<'\n'); + continue; + } + case OPC_CheckChild0Type: case OPC_CheckChild1Type: + case OPC_CheckChild2Type: case OPC_CheckChild3Type: + case OPC_CheckChild4Type: case OPC_CheckChild5Type: + case OPC_CheckChild6Type: case OPC_CheckChild7Type: + if (!::CheckChildType(MatcherTable, MatcherIndex, N, TLI, + Opcode-OPC_CheckChild0Type)) + break; + continue; + case OPC_CheckCondCode: + if (!::CheckCondCode(MatcherTable, MatcherIndex, N)) break; + continue; + case OPC_CheckValueType: + if (!::CheckValueType(MatcherTable, MatcherIndex, N, TLI)) break; + continue; + case OPC_CheckInteger: + if (!::CheckInteger(MatcherTable, MatcherIndex, N)) break; + continue; + case OPC_CheckAndImm: + if (!::CheckAndImm(MatcherTable, MatcherIndex, N, *this)) break; + continue; + case OPC_CheckOrImm: + if (!::CheckOrImm(MatcherTable, MatcherIndex, N, *this)) break; + continue; + + case OPC_CheckFoldableChainNode: { + assert(NodeStack.size() != 1 && "No parent node"); + // Verify that all intermediate nodes between the root and this one have + // a single use. + bool HasMultipleUses = false; + for (unsigned i = 1, e = NodeStack.size()-1; i != e; ++i) + if (!NodeStack[i].hasOneUse()) { + HasMultipleUses = true; + break; + } + if (HasMultipleUses) break; + + // Check to see that the target thinks this is profitable to fold and that + // we can fold it without inducing cycles in the graph. + if (!IsProfitableToFold(N, NodeStack[NodeStack.size()-2].getNode(), + NodeToMatch) || + !IsLegalToFold(N, NodeStack[NodeStack.size()-2].getNode(), + NodeToMatch, true/*We validate our own chains*/)) + break; + + continue; + } + case OPC_EmitInteger: { + MVT::SimpleValueType VT = + (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + int64_t Val = MatcherTable[MatcherIndex++]; + if (Val & 128) + Val = GetVBR(Val, MatcherTable, MatcherIndex); + RecordedNodes.push_back(CurDAG->getTargetConstant(Val, VT)); + continue; + } + case OPC_EmitRegister: { + MVT::SimpleValueType VT = + (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + unsigned RegNo = MatcherTable[MatcherIndex++]; + RecordedNodes.push_back(CurDAG->getRegister(RegNo, VT)); + continue; + } + + case OPC_EmitConvertToTarget: { + // Convert from IMM/FPIMM to target version. + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + SDValue Imm = RecordedNodes[RecNo]; + + if (Imm->getOpcode() == ISD::Constant) { + int64_t Val = cast<ConstantSDNode>(Imm)->getZExtValue(); + Imm = CurDAG->getTargetConstant(Val, Imm.getValueType()); + } else if (Imm->getOpcode() == ISD::ConstantFP) { + const ConstantFP *Val=cast<ConstantFPSDNode>(Imm)->getConstantFPValue(); + Imm = CurDAG->getTargetConstantFP(*Val, Imm.getValueType()); + } + + RecordedNodes.push_back(Imm); + continue; + } + + case OPC_EmitMergeInputChains: { + assert(InputChain.getNode() == 0 && + "EmitMergeInputChains should be the first chain producing node"); + // This node gets a list of nodes we matched in the input that have + // chains. We want to token factor all of the input chains to these nodes + // together. However, if any of the input chains is actually one of the + // nodes matched in this pattern, then we have an intra-match reference. + // Ignore these because the newly token factored chain should not refer to + // the old nodes. + unsigned NumChains = MatcherTable[MatcherIndex++]; + assert(NumChains != 0 && "Can't TF zero chains"); + + assert(ChainNodesMatched.empty() && + "Should only have one EmitMergeInputChains per match"); + + // Read all of the chained nodes. + for (unsigned i = 0; i != NumChains; ++i) { + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + ChainNodesMatched.push_back(RecordedNodes[RecNo].getNode()); + + // FIXME: What if other value results of the node have uses not matched + // by this pattern? + if (ChainNodesMatched.back() != NodeToMatch && + !RecordedNodes[RecNo].hasOneUse()) { + ChainNodesMatched.clear(); + break; + } + } + + // If the inner loop broke out, the match fails. + if (ChainNodesMatched.empty()) + break; + + // Merge the input chains if they are not intra-pattern references. + InputChain = HandleMergeInputChains(ChainNodesMatched, CurDAG); + + if (InputChain.getNode() == 0) + break; // Failed to merge. + + continue; + } + + case OPC_EmitCopyToReg: { + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + unsigned DestPhysReg = MatcherTable[MatcherIndex++]; + + if (InputChain.getNode() == 0) + InputChain = CurDAG->getEntryNode(); + + InputChain = CurDAG->getCopyToReg(InputChain, NodeToMatch->getDebugLoc(), + DestPhysReg, RecordedNodes[RecNo], + InputFlag); + + InputFlag = InputChain.getValue(1); + continue; + } + + case OPC_EmitNodeXForm: { + unsigned XFormNo = MatcherTable[MatcherIndex++]; + unsigned RecNo = MatcherTable[MatcherIndex++]; + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + RecordedNodes.push_back(RunSDNodeXForm(RecordedNodes[RecNo], XFormNo)); + continue; + } + + case OPC_EmitNode: + case OPC_MorphNodeTo: { + uint16_t TargetOpc = MatcherTable[MatcherIndex++]; + TargetOpc |= (unsigned short)MatcherTable[MatcherIndex++] << 8; + unsigned EmitNodeInfo = MatcherTable[MatcherIndex++]; + // Get the result VT list. + unsigned NumVTs = MatcherTable[MatcherIndex++]; + SmallVector<EVT, 4> VTs; + for (unsigned i = 0; i != NumVTs; ++i) { + MVT::SimpleValueType VT = + (MVT::SimpleValueType)MatcherTable[MatcherIndex++]; + if (VT == MVT::iPTR) VT = TLI.getPointerTy().SimpleTy; + VTs.push_back(VT); + } + + if (EmitNodeInfo & OPFL_Chain) + VTs.push_back(MVT::Other); + if (EmitNodeInfo & OPFL_FlagOutput) + VTs.push_back(MVT::Flag); + + // This is hot code, so optimize the two most common cases of 1 and 2 + // results. + SDVTList VTList; + if (VTs.size() == 1) + VTList = CurDAG->getVTList(VTs[0]); + else if (VTs.size() == 2) + VTList = CurDAG->getVTList(VTs[0], VTs[1]); + else + VTList = CurDAG->getVTList(VTs.data(), VTs.size()); + + // Get the operand list. + unsigned NumOps = MatcherTable[MatcherIndex++]; + SmallVector<SDValue, 8> Ops; + for (unsigned i = 0; i != NumOps; ++i) { + unsigned RecNo = MatcherTable[MatcherIndex++]; + if (RecNo & 128) + RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex); + + assert(RecNo < RecordedNodes.size() && "Invalid EmitNode"); + Ops.push_back(RecordedNodes[RecNo]); + } + + // If there are variadic operands to add, handle them now. + if (EmitNodeInfo & OPFL_VariadicInfo) { + // Determine the start index to copy from. + unsigned FirstOpToCopy = getNumFixedFromVariadicInfo(EmitNodeInfo); + FirstOpToCopy += (EmitNodeInfo & OPFL_Chain) ? 1 : 0; + assert(NodeToMatch->getNumOperands() >= FirstOpToCopy && + "Invalid variadic node"); + // Copy all of the variadic operands, not including a potential flag + // input. + for (unsigned i = FirstOpToCopy, e = NodeToMatch->getNumOperands(); + i != e; ++i) { + SDValue V = NodeToMatch->getOperand(i); + if (V.getValueType() == MVT::Flag) break; + Ops.push_back(V); + } + } + + // If this has chain/flag inputs, add them. + if (EmitNodeInfo & OPFL_Chain) + Ops.push_back(InputChain); + if ((EmitNodeInfo & OPFL_FlagInput) && InputFlag.getNode() != 0) + Ops.push_back(InputFlag); + + // Create the node. + SDNode *Res = 0; + if (Opcode != OPC_MorphNodeTo) { + // If this is a normal EmitNode command, just create the new node and + // add the results to the RecordedNodes list. + Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(), + VTList, Ops.data(), Ops.size()); + + // Add all the non-flag/non-chain results to the RecordedNodes list. + for (unsigned i = 0, e = VTs.size(); i != e; ++i) { + if (VTs[i] == MVT::Other || VTs[i] == MVT::Flag) break; + RecordedNodes.push_back(SDValue(Res, i)); + } + + } else { + Res = MorphNode(NodeToMatch, TargetOpc, VTList, Ops.data(), Ops.size(), + EmitNodeInfo); + } + + // If the node had chain/flag results, update our notion of the current + // chain and flag. + if (EmitNodeInfo & OPFL_FlagOutput) { + InputFlag = SDValue(Res, VTs.size()-1); + if (EmitNodeInfo & OPFL_Chain) + InputChain = SDValue(Res, VTs.size()-2); + } else if (EmitNodeInfo & OPFL_Chain) + InputChain = SDValue(Res, VTs.size()-1); + + // If the OPFL_MemRefs flag is set on this node, slap all of the + // accumulated memrefs onto it. + // + // FIXME: This is vastly incorrect for patterns with multiple outputs + // instructions that access memory and for ComplexPatterns that match + // loads. + if (EmitNodeInfo & OPFL_MemRefs) { + MachineSDNode::mmo_iterator MemRefs = + MF->allocateMemRefsArray(MatchedMemRefs.size()); + std::copy(MatchedMemRefs.begin(), MatchedMemRefs.end(), MemRefs); + cast<MachineSDNode>(Res) + ->setMemRefs(MemRefs, MemRefs + MatchedMemRefs.size()); + } + + DEBUG(errs() << " " + << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created") + << " node: "; Res->dump(CurDAG); errs() << "\n"); + + // If this was a MorphNodeTo then we're completely done! + if (Opcode == OPC_MorphNodeTo) { + // Update chain and flag uses. + UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched, + InputFlag, FlagResultNodesMatched, true); + return Res; + } + + continue; + } + + case OPC_MarkFlagResults: { + unsigned NumNodes = MatcherTable[MatcherIndex++]; + + // Read and remember all the flag-result nodes. + for (unsigned i = 0; i != NumNodes; ++i) { + unsigned RecNo = MatcherTable[MatcherIndex++]; + if (RecNo & 128) + RecNo = GetVBR(RecNo, MatcherTable, MatcherIndex); + + assert(RecNo < RecordedNodes.size() && "Invalid CheckSame"); + FlagResultNodesMatched.push_back(RecordedNodes[RecNo].getNode()); + } + continue; + } + + case OPC_CompleteMatch: { + // The match has been completed, and any new nodes (if any) have been + // created. Patch up references to the matched dag to use the newly + // created nodes. + unsigned NumResults = MatcherTable[MatcherIndex++]; + + for (unsigned i = 0; i != NumResults; ++i) { + unsigned ResSlot = MatcherTable[MatcherIndex++]; + if (ResSlot & 128) + ResSlot = GetVBR(ResSlot, MatcherTable, MatcherIndex); + + assert(ResSlot < RecordedNodes.size() && "Invalid CheckSame"); + SDValue Res = RecordedNodes[ResSlot]; + + // FIXME2: Eliminate this horrible hack by fixing the 'Gen' program + // after (parallel) on input patterns are removed. This would also + // allow us to stop encoding #results in OPC_CompleteMatch's table + // entry. + if (NodeToMatch->getNumValues() <= i || + NodeToMatch->getValueType(i) == MVT::Other || + NodeToMatch->getValueType(i) == MVT::Flag) + break; + assert((NodeToMatch->getValueType(i) == Res.getValueType() || + NodeToMatch->getValueType(i) == MVT::iPTR || + Res.getValueType() == MVT::iPTR || + NodeToMatch->getValueType(i).getSizeInBits() == + Res.getValueType().getSizeInBits()) && + "invalid replacement"); + CurDAG->ReplaceAllUsesOfValueWith(SDValue(NodeToMatch, i), Res); + } + + // If the root node defines a flag, add it to the flag nodes to update + // list. + if (NodeToMatch->getValueType(NodeToMatch->getNumValues()-1) == MVT::Flag) + FlagResultNodesMatched.push_back(NodeToMatch); + + // Update chain and flag uses. + UpdateChainsAndFlags(NodeToMatch, InputChain, ChainNodesMatched, + InputFlag, FlagResultNodesMatched, false); + + assert(NodeToMatch->use_empty() && + "Didn't replace all uses of the node?"); + + // FIXME: We just return here, which interacts correctly with SelectRoot + // above. We should fix this to not return an SDNode* anymore. + return 0; + } + } + + // If the code reached this point, then the match failed. See if there is + // another child to try in the current 'Scope', otherwise pop it until we + // find a case to check. + while (1) { + if (MatchScopes.empty()) { + CannotYetSelect(NodeToMatch); + return 0; + } + + // Restore the interpreter state back to the point where the scope was + // formed. + MatchScope &LastScope = MatchScopes.back(); + RecordedNodes.resize(LastScope.NumRecordedNodes); + NodeStack.clear(); + NodeStack.append(LastScope.NodeStack.begin(), LastScope.NodeStack.end()); + N = NodeStack.back(); + + DEBUG(errs() << " Match failed at index " << MatcherIndex + << " continuing at " << LastScope.FailIndex << "\n"); + + if (LastScope.NumMatchedMemRefs != MatchedMemRefs.size()) + MatchedMemRefs.resize(LastScope.NumMatchedMemRefs); + MatcherIndex = LastScope.FailIndex; + + InputChain = LastScope.InputChain; + InputFlag = LastScope.InputFlag; + if (!LastScope.HasChainNodesMatched) + ChainNodesMatched.clear(); + if (!LastScope.HasFlagResultNodesMatched) + FlagResultNodesMatched.clear(); + + // Check to see what the offset is at the new MatcherIndex. If it is zero + // we have reached the end of this scope, otherwise we have another child + // in the current scope to try. + unsigned NumToSkip = MatcherTable[MatcherIndex++]; + if (NumToSkip & 128) + NumToSkip = GetVBR(NumToSkip, MatcherTable, MatcherIndex); + + // If we have another child in this scope to match, update FailIndex and + // try it. + if (NumToSkip != 0) { + LastScope.FailIndex = MatcherIndex+NumToSkip; + break; + } + + // End of this scope, pop it and try the next child in the containing + // scope. + MatchScopes.pop_back(); + } + } +} + + + void SelectionDAGISel::CannotYetSelect(SDNode *N) { std::string msg; raw_string_ostream Msg(msg); Msg << "Cannot yet select: "; - N->printrFull(Msg, CurDAG); + + if (N->getOpcode() != ISD::INTRINSIC_W_CHAIN && + N->getOpcode() != ISD::INTRINSIC_WO_CHAIN && + N->getOpcode() != ISD::INTRINSIC_VOID) { + N->printrFull(Msg, CurDAG); + } else { + bool HasInputChain = N->getOperand(0).getValueType() == MVT::Other; + unsigned iid = + cast<ConstantSDNode>(N->getOperand(HasInputChain))->getZExtValue(); + if (iid < Intrinsic::num_intrinsics) + Msg << "intrinsic %" << Intrinsic::getName((Intrinsic::ID)iid); + else if (const TargetIntrinsicInfo *TII = TM.getIntrinsicInfo()) + Msg << "target intrinsic %" << TII->getName(iid); + else + Msg << "unknown intrinsic #" << iid; + } llvm_report_error(Msg.str()); } -void SelectionDAGISel::CannotYetSelectIntrinsic(SDNode *N) { - dbgs() << "Cannot yet select: "; - unsigned iid = - cast<ConstantSDNode>(N->getOperand(N->getOperand(0).getValueType() == - MVT::Other))->getZExtValue(); - if (iid < Intrinsic::num_intrinsics) - llvm_report_error("Cannot yet select: intrinsic %" + - Intrinsic::getName((Intrinsic::ID)iid)); - else if (const TargetIntrinsicInfo *tii = TM.getIntrinsicInfo()) - llvm_report_error(Twine("Cannot yet select: target intrinsic %") + - tii->getName(iid)); -} - char SelectionDAGISel::ID = 0; diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp index d74ec7e..8d0d884 100644 --- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp +++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp @@ -28,6 +28,7 @@ #include "llvm/ADT/STLExtras.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/MathExtras.h" +#include <ctype.h> using namespace llvm; namespace llvm { @@ -540,6 +541,24 @@ TargetLowering::~TargetLowering() { delete &TLOF; } +/// canOpTrap - Returns true if the operation can trap for the value type. +/// VT must be a legal type. +bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const { + assert(isTypeLegal(VT)); + switch (Op) { + default: + return false; + case ISD::FDIV: + case ISD::FREM: + case ISD::SDIV: + case ISD::UDIV: + case ISD::SREM: + case ISD::UREM: + return true; + } +} + + static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, unsigned &NumIntermediates, EVT &RegisterVT, @@ -1423,8 +1442,10 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op, case ISD::TRUNCATE: { // Simplify the input, using demanded bit information, and compute the known // zero/one bits live out. + unsigned OperandBitWidth = + Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); APInt TruncMask = NewMask; - TruncMask.zext(Op.getOperand(0).getValueSizeInBits()); + TruncMask.zext(OperandBitWidth); if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, KnownZero, KnownOne, TLO, Depth+1)) return true; @@ -1435,15 +1456,14 @@ bool TargetLowering::SimplifyDemandedBits(SDValue Op, // on the known demanded bits. if (Op.getOperand(0).getNode()->hasOneUse()) { SDValue In = Op.getOperand(0); - unsigned InBitWidth = In.getValueSizeInBits(); switch (In.getOpcode()) { default: break; case ISD::SRL: // Shrink SRL by a constant if none of the high bits shifted in are // demanded. if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){ - APInt HighBits = APInt::getHighBitsSet(InBitWidth, - InBitWidth - BitWidth); + APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, + OperandBitWidth - BitWidth); HighBits = HighBits.lshr(ShAmt->getZExtValue()); HighBits.trunc(BitWidth); @@ -1589,7 +1609,7 @@ static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { // Fall back to ComputeMaskedBits to catch other known cases. EVT OpVT = Val.getValueType(); - unsigned BitWidth = OpVT.getSizeInBits(); + unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); APInt Mask = APInt::getAllOnesValue(BitWidth); APInt KnownZero, KnownOne; DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne); @@ -1698,7 +1718,7 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, Lod->getSrcValue(), Lod->getSrcValueOffset() + bestOffset, - false, NewAlign); + false, false, NewAlign); return DAG.getSetCC(dl, VT, DAG.getNode(ISD::AND, dl, newVT, NewLoad, DAG.getConstant(bestMask.trunc(bestWidth), @@ -1757,7 +1777,7 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, break; // todo, be more careful with signed comparisons } } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && - (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { + (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); EVT ExtDstTy = N0.getValueType(); @@ -1791,22 +1811,21 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, Cond); } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { - // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC - if (N0.getOpcode() == ISD::SETCC) { + if (N0.getOpcode() == ISD::SETCC && + isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); if (TrueWhenTrue) - return N0; - + return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); // Invert the condition. ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); CC = ISD::getSetCCInverse(CC, N0.getOperand(0).getValueType().isInteger()); return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); } - + if ((N0.getOpcode() == ISD::XOR || - (N0.getOpcode() == ISD::AND && + (N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::XOR && N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && isa<ConstantSDNode>(N0.getOperand(1)) && @@ -1829,9 +1848,36 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, N0.getOperand(0).getOperand(0), N0.getOperand(1)); } + return DAG.getSetCC(dl, VT, Val, N1, Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); } + } else if (N1C->getAPIntValue() == 1 && + (VT == MVT::i1 || + getBooleanContents() == ZeroOrOneBooleanContent)) { + SDValue Op0 = N0; + if (Op0.getOpcode() == ISD::TRUNCATE) + Op0 = Op0.getOperand(0); + + if ((Op0.getOpcode() == ISD::XOR) && + Op0.getOperand(0).getOpcode() == ISD::SETCC && + Op0.getOperand(1).getOpcode() == ISD::SETCC) { + // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) + Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; + return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), + Cond); + } else if (Op0.getOpcode() == ISD::AND && + isa<ConstantSDNode>(Op0.getOperand(1)) && + cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { + // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. + if (Op0.getValueType() != VT) + Op0 = DAG.getNode(ISD::AND, dl, VT, + DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), + DAG.getConstant(1, VT)); + return DAG.getSetCC(dl, VT, Op0, + DAG.getConstant(0, Op0.getValueType()), + Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); + } } } diff --git a/lib/CodeGen/SimpleRegisterCoalescing.cpp b/lib/CodeGen/SimpleRegisterCoalescing.cpp index 1d9bda4..ce72b2f 100644 --- a/lib/CodeGen/SimpleRegisterCoalescing.cpp +++ b/lib/CodeGen/SimpleRegisterCoalescing.cpp @@ -662,7 +662,7 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt, if (!tii_->isTriviallyReMaterializable(DefMI, AA)) return false; bool SawStore = false; - if (!DefMI->isSafeToMove(tii_, SawStore, AA)) + if (!DefMI->isSafeToMove(tii_, AA, SawStore)) return false; if (TID.getNumDefs() != 1) return false; @@ -702,7 +702,8 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt, for (const unsigned* SR = tri_->getSubRegisters(DstReg); *SR; ++SR) { if (!li_->hasInterval(*SR)) continue; - DLR = li_->getInterval(*SR).getLiveRangeContaining(DefIdx); + const LiveRange *DLR = + li_->getInterval(*SR).getLiveRangeContaining(DefIdx); if (DLR && DLR->valno->getCopy() == CopyMI) DLR->valno->setCopy(0); } @@ -741,9 +742,21 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt, NewMI->addOperand(MO); if (MO.isDef() && li_->hasInterval(MO.getReg())) { unsigned Reg = MO.getReg(); - DLR = li_->getInterval(Reg).getLiveRangeContaining(DefIdx); + const LiveRange *DLR = + li_->getInterval(Reg).getLiveRangeContaining(DefIdx); if (DLR && DLR->valno->getCopy() == CopyMI) DLR->valno->setCopy(0); + // Handle subregs as well + if (TargetRegisterInfo::isPhysicalRegister(Reg)) { + for (const unsigned* SR = tri_->getSubRegisters(Reg); *SR; ++SR) { + if (!li_->hasInterval(*SR)) + continue; + const LiveRange *DLR = + li_->getInterval(*SR).getLiveRangeContaining(DefIdx); + if (DLR && DLR->valno->getCopy() == CopyMI) + DLR->valno->setCopy(0); + } + } } } @@ -752,6 +765,7 @@ bool SimpleRegisterCoalescing::ReMaterializeTrivialDef(LiveInterval &SrcInt, CopyMI->eraseFromParent(); ReMatCopies.insert(CopyMI); ReMatDefs.insert(DefMI); + DEBUG(dbgs() << "Remat: " << *NewMI); ++NumReMats; return true; } @@ -771,11 +785,16 @@ SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg, SubIdx = 0; } + // Copy the register use-list before traversing it. We may be adding operands + // and invalidating pointers. + SmallVector<std::pair<MachineInstr*, unsigned>, 32> reglist; for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(SrcReg), - E = mri_->reg_end(); I != E; ) { - MachineOperand &O = I.getOperand(); - MachineInstr *UseMI = &*I; - ++I; + E = mri_->reg_end(); I != E; ++I) + reglist.push_back(std::make_pair(&*I, I.getOperandNo())); + + for (unsigned N=0; N != reglist.size(); ++N) { + MachineInstr *UseMI = reglist[N].first; + MachineOperand &O = UseMI->getOperand(reglist[N].second); unsigned OldSubIdx = O.getSubReg(); if (DstIsPhys) { unsigned UseDstReg = DstReg; @@ -796,6 +815,19 @@ SimpleRegisterCoalescing::UpdateRegDefsUses(unsigned SrcReg, unsigned DstReg, O.setReg(UseDstReg); O.setSubReg(0); + if (OldSubIdx) { + // Def and kill of subregister of a virtual register actually defs and + // kills the whole register. Add imp-defs and imp-kills as needed. + if (O.isDef()) { + if(O.isDead()) + UseMI->addRegisterDead(DstReg, tri_, true); + else + UseMI->addRegisterDefined(DstReg, tri_); + } else if (!O.isUndef() && + (O.isKill() || + UseMI->isRegTiedToDefOperand(&O-&UseMI->getOperand(0)))) + UseMI->addRegisterKilled(DstReg, tri_, true); + } continue; } @@ -1148,12 +1180,14 @@ SimpleRegisterCoalescing::isWinToJoinCrossClass(unsigned LargeReg, LiveInterval &SmallInt = li_->getInterval(SmallReg); unsigned LargeSize = li_->getApproximateInstructionCount(LargeInt); unsigned SmallSize = li_->getApproximateInstructionCount(SmallInt); - if (SmallSize > Threshold || LargeSize > Threshold) - if ((float)std::distance(mri_->use_nodbg_begin(SmallReg), - mri_->use_nodbg_end()) / SmallSize < - (float)std::distance(mri_->use_nodbg_begin(LargeReg), - mri_->use_nodbg_end()) / LargeSize) + if (LargeSize > Threshold) { + unsigned SmallUses = std::distance(mri_->use_nodbg_begin(SmallReg), + mri_->use_nodbg_end()); + unsigned LargeUses = std::distance(mri_->use_nodbg_begin(LargeReg), + mri_->use_nodbg_end()); + if (SmallUses*LargeSize < LargeUses*SmallSize) return false; + } return true; } @@ -1173,6 +1207,8 @@ SimpleRegisterCoalescing::HasIncompatibleSubRegDefUse(MachineInstr *CopyMI, for (MachineRegisterInfo::reg_iterator I = mri_->reg_begin(VirtReg), E = mri_->reg_end(); I != E; ++I) { MachineOperand &O = I.getOperand(); + if (O.isDebug()) + continue; MachineInstr *MI = &*I; if (MI == CopyMI || JoinedCopies.count(MI)) continue; @@ -1559,7 +1595,10 @@ bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) { (isExtSubReg || DstRC->isASubClass()) && !isWinToJoinCrossClass(LargeReg, SmallReg, allocatableRCRegs_[NewRC].count())) { - DEBUG(dbgs() << "\tSrc/Dest are different register classes.\n"); + DEBUG(dbgs() << "\tSrc/Dest are different register classes: " + << SrcRC->getName() << "/" + << DstRC->getName() << " -> " + << NewRC->getName() << ".\n"); // Allow the coalescer to try again in case either side gets coalesced to // a physical register that's compatible with the other side. e.g. // r1024 = MOV32to32_ r1025 @@ -1680,6 +1719,7 @@ bool SimpleRegisterCoalescing::JoinCopy(CopyRec &TheCopy, bool &Again) { (AdjustCopiesBackFrom(SrcInt, DstInt, CopyMI) || RemoveCopyByCommutingDef(SrcInt, DstInt, CopyMI))) { JoinedCopies.insert(CopyMI); + DEBUG(dbgs() << "Trivial!\n"); return true; } @@ -1839,7 +1879,7 @@ static unsigned ComputeUltimateVN(VNInfo *VNI, // If the VN has already been computed, just return it. if (ThisValNoAssignments[VN] >= 0) return ThisValNoAssignments[VN]; -// assert(ThisValNoAssignments[VN] != -2 && "Cyclic case?"); + assert(ThisValNoAssignments[VN] != -2 && "Cyclic value numbers"); // If this val is not a copy from the other val, then it must be a new value // number in the destination. diff --git a/lib/CodeGen/SjLjEHPrepare.cpp b/lib/CodeGen/SjLjEHPrepare.cpp index 8d4d1b2..059e8d6 100644 --- a/lib/CodeGen/SjLjEHPrepare.cpp +++ b/lib/CodeGen/SjLjEHPrepare.cpp @@ -44,7 +44,6 @@ namespace { const Type *FunctionContextTy; Constant *RegisterFn; Constant *UnregisterFn; - Constant *ResumeFn; Constant *BuiltinSetjmpFn; Constant *FrameAddrFn; Constant *LSDAAddrFn; @@ -67,8 +66,8 @@ namespace { } private: - void markInvokeCallSite(InvokeInst *II, unsigned InvokeNo, - Value *CallSite, + void insertCallSiteStore(Instruction *I, int Number, Value *CallSite); + void markInvokeCallSite(InvokeInst *II, int InvokeNo, Value *CallSite, SwitchInst *CatchSwitch); void splitLiveRangesLiveAcrossInvokes(SmallVector<InvokeInst*,16> &Invokes); bool insertSjLjEHSupport(Function &F); @@ -107,11 +106,6 @@ bool SjLjEHPass::doInitialization(Module &M) { Type::getVoidTy(M.getContext()), PointerType::getUnqual(FunctionContextTy), (Type *)0); - ResumeFn = - M.getOrInsertFunction("_Unwind_SjLj_Resume", - Type::getVoidTy(M.getContext()), - VoidPtrTy, - (Type *)0); FrameAddrFn = Intrinsic::getDeclaration(&M, Intrinsic::frameaddress); BuiltinSetjmpFn = Intrinsic::getDeclaration(&M, Intrinsic::eh_sjlj_setjmp); LSDAAddrFn = Intrinsic::getDeclaration(&M, Intrinsic::eh_sjlj_lsda); @@ -123,12 +117,22 @@ bool SjLjEHPass::doInitialization(Module &M) { return true; } +/// insertCallSiteStore - Insert a store of the call-site value to the +/// function context +void SjLjEHPass::insertCallSiteStore(Instruction *I, int Number, + Value *CallSite) { + ConstantInt *CallSiteNoC = ConstantInt::get(Type::getInt32Ty(I->getContext()), + Number); + // Insert a store of the call-site number + new StoreInst(CallSiteNoC, CallSite, true, I); // volatile +} + /// markInvokeCallSite - Insert code to mark the call_site for this invoke -void SjLjEHPass::markInvokeCallSite(InvokeInst *II, unsigned InvokeNo, +void SjLjEHPass::markInvokeCallSite(InvokeInst *II, int InvokeNo, Value *CallSite, SwitchInst *CatchSwitch) { ConstantInt *CallSiteNoC= ConstantInt::get(Type::getInt32Ty(II->getContext()), - InvokeNo); + InvokeNo); // The runtime comes back to the dispatcher with the call_site - 1 in // the context. Odd, but there it is. ConstantInt *SwitchValC = ConstantInt::get(Type::getInt32Ty(II->getContext()), @@ -145,8 +149,11 @@ void SjLjEHPass::markInvokeCallSite(InvokeInst *II, unsigned InvokeNo, } } - // Insert a store of the invoke num before the invoke - new StoreInst(CallSiteNoC, CallSite, true, II); // volatile + // Insert the store of the call site value + insertCallSiteStore(II, InvokeNo, CallSite); + + // Record the call site value for the back end so it stays associated with + // the invoke. CallInst::Create(CallSiteFn, CallSiteNoC, "", II); // Add a switch case to our unwind block. @@ -272,8 +279,8 @@ bool SjLjEHPass::insertSjLjEHSupport(Function &F) { SmallVector<InvokeInst*,16> Invokes; // Look through the terminators of the basic blocks to find invokes, returns - // and unwinds - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) + // and unwinds. + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) { if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { // Remember all return instructions in case we insert an invoke into this // function. @@ -283,6 +290,7 @@ bool SjLjEHPass::insertSjLjEHSupport(Function &F) { } else if (UnwindInst *UI = dyn_cast<UnwindInst>(BB->getTerminator())) { Unwinds.push_back(UI); } + } // If we don't have any invokes or unwinds, there's nothing to do. if (Unwinds.empty() && Invokes.empty()) return false; @@ -478,24 +486,21 @@ bool SjLjEHPass::insertSjLjEHSupport(Function &F) { for (unsigned i = 0, e = Invokes.size(); i != e; ++i) markInvokeCallSite(Invokes[i], i+1, CallSite, DispatchSwitch); - // The front end has likely added calls to _Unwind_Resume. We need - // to find those calls and mark the call_site as -1 immediately prior. - // resume is a noreturn function, so any block that has a call to it - // should end in an 'unreachable' instruction with the call immediately - // prior. That's how we'll search. - // ??? There's got to be a better way. this is fugly. - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) - if ((dyn_cast<UnreachableInst>(BB->getTerminator()))) { - BasicBlock::iterator I = BB->getTerminator(); - // Check the previous instruction and see if it's a resume call - if (I == BB->begin()) continue; - if (CallInst *CI = dyn_cast<CallInst>(--I)) { - if (CI->getCalledFunction() == ResumeFn) { - Value *NegativeOne = Constant::getAllOnesValue(Int32Ty); - new StoreInst(NegativeOne, CallSite, true, I); // volatile - } + // Mark call instructions that aren't nounwind as no-action + // (call_site == -1). Skip the entry block, as prior to then, no function + // context has been created for this function and any unexpected exceptions + // thrown will go directly to the caller's context, which is what we want + // anyway, so no need to do anything here. + for (Function::iterator BB = F.begin(), E = F.end(); ++BB != E;) { + for (BasicBlock::iterator I = BB->begin(), end = BB->end(); I != end; ++I) + if (CallInst *CI = dyn_cast<CallInst>(I)) { + // Ignore calls to the EH builtins (eh.selector, eh.exception) + Constant *Callee = CI->getCalledFunction(); + if (Callee != SelectorFn && Callee != ExceptionFn + && !CI->doesNotThrow()) + insertCallSiteStore(CI, -1, CallSite); } - } + } // Replace all unwinds with a branch to the unwind handler. // ??? Should this ever happen with sjlj exceptions? diff --git a/lib/CodeGen/StackProtector.cpp b/lib/CodeGen/StackProtector.cpp index 48bb5af..8a6a727 100644 --- a/lib/CodeGen/StackProtector.cpp +++ b/lib/CodeGen/StackProtector.cpp @@ -113,7 +113,7 @@ bool StackProtector::RequiresStackProtector() const { if (const ArrayType *AT = dyn_cast<ArrayType>(AI->getAllocatedType())) { // We apparently only care about character arrays. - if (!AT->getElementType()->isInteger(8)) + if (!AT->getElementType()->isIntegerTy(8)) continue; // If an array has more than SSPBufferSize bytes of allocated space, diff --git a/lib/CodeGen/TailDuplication.cpp b/lib/CodeGen/TailDuplication.cpp index 9ab4058..3223e53 100644 --- a/lib/CodeGen/TailDuplication.cpp +++ b/lib/CodeGen/TailDuplication.cpp @@ -403,26 +403,45 @@ TailDuplicatePass::UpdateSuccessorsPHIs(MachineBasicBlock *FromBB, bool isDead, II->RemoveOperand(i); } } - II->RemoveOperand(Idx+1); - II->RemoveOperand(Idx); - } + } else + Idx = 0; + + // If Idx is set, the operands at Idx and Idx+1 must be removed. + // We reuse the location to avoid expensive RemoveOperand calls. + DenseMap<unsigned,AvailableValsTy>::iterator LI=SSAUpdateVals.find(Reg); if (LI != SSAUpdateVals.end()) { // This register is defined in the tail block. for (unsigned j = 0, ee = LI->second.size(); j != ee; ++j) { MachineBasicBlock *SrcBB = LI->second[j].first; unsigned SrcReg = LI->second[j].second; - II->addOperand(MachineOperand::CreateReg(SrcReg, false)); - II->addOperand(MachineOperand::CreateMBB(SrcBB)); + if (Idx != 0) { + II->getOperand(Idx).setReg(SrcReg); + II->getOperand(Idx+1).setMBB(SrcBB); + Idx = 0; + } else { + II->addOperand(MachineOperand::CreateReg(SrcReg, false)); + II->addOperand(MachineOperand::CreateMBB(SrcBB)); + } } } else { // Live in tail block, must also be live in predecessors. for (unsigned j = 0, ee = TDBBs.size(); j != ee; ++j) { MachineBasicBlock *SrcBB = TDBBs[j]; - II->addOperand(MachineOperand::CreateReg(Reg, false)); - II->addOperand(MachineOperand::CreateMBB(SrcBB)); + if (Idx != 0) { + II->getOperand(Idx).setReg(Reg); + II->getOperand(Idx+1).setMBB(SrcBB); + Idx = 0; + } else { + II->addOperand(MachineOperand::CreateReg(Reg, false)); + II->addOperand(MachineOperand::CreateMBB(SrcBB)); + } } } + if (Idx != 0) { + II->RemoveOperand(Idx+1); + II->RemoveOperand(Idx); + } } } } diff --git a/lib/CodeGen/TargetInstrInfoImpl.cpp b/lib/CodeGen/TargetInstrInfoImpl.cpp index a0fccab..e9e998f 100644 --- a/lib/CodeGen/TargetInstrInfoImpl.cpp +++ b/lib/CodeGen/TargetInstrInfoImpl.cpp @@ -150,6 +150,11 @@ void TargetInstrInfoImpl::reMaterialize(MachineBasicBlock &MBB, MBB.insert(I, MI); } +bool TargetInstrInfoImpl::produceSameValue(const MachineInstr *MI0, + const MachineInstr *MI1) const { + return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs); +} + MachineInstr *TargetInstrInfoImpl::duplicate(MachineInstr *Orig, MachineFunction &MF) const { assert(!Orig->getDesc().isNotDuplicable() && @@ -157,37 +162,6 @@ MachineInstr *TargetInstrInfoImpl::duplicate(MachineInstr *Orig, return MF.CloneMachineInstr(Orig); } -bool -TargetInstrInfoImpl::isIdentical(const MachineInstr *MI, - const MachineInstr *Other, - const MachineRegisterInfo *MRI) const { - if (MI->getOpcode() != Other->getOpcode() || - MI->getNumOperands() != Other->getNumOperands()) - return false; - - for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { - const MachineOperand &MO = MI->getOperand(i); - const MachineOperand &OMO = Other->getOperand(i); - if (MO.isReg() && MO.isDef()) { - assert(OMO.isReg() && OMO.isDef()); - unsigned Reg = MO.getReg(); - if (TargetRegisterInfo::isPhysicalRegister(Reg)) { - if (Reg != OMO.getReg()) - return false; - } else if (MRI->getRegClass(MO.getReg()) != - MRI->getRegClass(OMO.getReg())) - return false; - - continue; - } - - if (!MO.isIdenticalTo(OMO)) - return false; - } - - return true; -} - unsigned TargetInstrInfoImpl::GetFunctionSizeInBytes(const MachineFunction &MF) const { unsigned FnSize = 0; diff --git a/lib/CodeGen/TargetLoweringObjectFileImpl.cpp b/lib/CodeGen/TargetLoweringObjectFileImpl.cpp new file mode 100644 index 0000000..d127f53 --- /dev/null +++ b/lib/CodeGen/TargetLoweringObjectFileImpl.cpp @@ -0,0 +1,902 @@ +//===-- llvm/CodeGen/TargetLoweringObjectFileImpl.cpp - Object File Info --===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file implements classes used to handle lowerings specific to common +// object file formats. +// +//===----------------------------------------------------------------------===// + +#include "llvm/CodeGen/TargetLoweringObjectFileImpl.h" +#include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" +#include "llvm/Function.h" +#include "llvm/GlobalVariable.h" +#include "llvm/CodeGen/MachineModuleInfoImpls.h" +#include "llvm/MC/MCContext.h" +#include "llvm/MC/MCExpr.h" +#include "llvm/MC/MCSectionMachO.h" +#include "llvm/MC/MCSectionELF.h" +#include "llvm/MC/MCSymbol.h" +#include "llvm/Target/Mangler.h" +#include "llvm/Target/TargetData.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetOptions.h" +#include "llvm/Support/Dwarf.h" +#include "llvm/Support/ErrorHandling.h" +#include "llvm/Support/raw_ostream.h" +#include "llvm/ADT/SmallString.h" +#include "llvm/ADT/StringExtras.h" +using namespace llvm; +using namespace dwarf; + +//===----------------------------------------------------------------------===// +// ELF +//===----------------------------------------------------------------------===// +typedef StringMap<const MCSectionELF*> ELFUniqueMapTy; + +TargetLoweringObjectFileELF::~TargetLoweringObjectFileELF() { + // If we have the section uniquing map, free it. + delete (ELFUniqueMapTy*)UniquingMap; +} + +const MCSection *TargetLoweringObjectFileELF:: +getELFSection(StringRef Section, unsigned Type, unsigned Flags, + SectionKind Kind, bool IsExplicit) const { + if (UniquingMap == 0) + UniquingMap = new ELFUniqueMapTy(); + ELFUniqueMapTy &Map = *(ELFUniqueMapTy*)UniquingMap; + + // Do the lookup, if we have a hit, return it. + const MCSectionELF *&Entry = Map[Section]; + if (Entry) return Entry; + + return Entry = MCSectionELF::Create(Section, Type, Flags, Kind, IsExplicit, + getContext()); +} + +void TargetLoweringObjectFileELF::Initialize(MCContext &Ctx, + const TargetMachine &TM) { + if (UniquingMap != 0) + ((ELFUniqueMapTy*)UniquingMap)->clear(); + TargetLoweringObjectFile::Initialize(Ctx, TM); + + BSSSection = + getELFSection(".bss", MCSectionELF::SHT_NOBITS, + MCSectionELF::SHF_WRITE | MCSectionELF::SHF_ALLOC, + SectionKind::getBSS()); + + TextSection = + getELFSection(".text", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_EXECINSTR | MCSectionELF::SHF_ALLOC, + SectionKind::getText()); + + DataSection = + getELFSection(".data", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_WRITE | MCSectionELF::SHF_ALLOC, + SectionKind::getDataRel()); + + ReadOnlySection = + getELFSection(".rodata", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC, + SectionKind::getReadOnly()); + + TLSDataSection = + getELFSection(".tdata", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_TLS | + MCSectionELF::SHF_WRITE, SectionKind::getThreadData()); + + TLSBSSSection = + getELFSection(".tbss", MCSectionELF::SHT_NOBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_TLS | + MCSectionELF::SHF_WRITE, SectionKind::getThreadBSS()); + + DataRelSection = + getELFSection(".data.rel", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getDataRel()); + + DataRelLocalSection = + getELFSection(".data.rel.local", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getDataRelLocal()); + + DataRelROSection = + getELFSection(".data.rel.ro", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getReadOnlyWithRel()); + + DataRelROLocalSection = + getELFSection(".data.rel.ro.local", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getReadOnlyWithRelLocal()); + + MergeableConst4Section = + getELFSection(".rodata.cst4", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_MERGE, + SectionKind::getMergeableConst4()); + + MergeableConst8Section = + getELFSection(".rodata.cst8", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_MERGE, + SectionKind::getMergeableConst8()); + + MergeableConst16Section = + getELFSection(".rodata.cst16", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_MERGE, + SectionKind::getMergeableConst16()); + + StaticCtorSection = + getELFSection(".ctors", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getDataRel()); + + StaticDtorSection = + getELFSection(".dtors", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getDataRel()); + + // Exception Handling Sections. + + // FIXME: We're emitting LSDA info into a readonly section on ELF, even though + // it contains relocatable pointers. In PIC mode, this is probably a big + // runtime hit for C++ apps. Either the contents of the LSDA need to be + // adjusted or this should be a data section. + LSDASection = + getELFSection(".gcc_except_table", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC, SectionKind::getReadOnly()); + EHFrameSection = + getELFSection(".eh_frame", MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | MCSectionELF::SHF_WRITE, + SectionKind::getDataRel()); + + // Debug Info Sections. + DwarfAbbrevSection = + getELFSection(".debug_abbrev", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfInfoSection = + getELFSection(".debug_info", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfLineSection = + getELFSection(".debug_line", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfFrameSection = + getELFSection(".debug_frame", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfPubNamesSection = + getELFSection(".debug_pubnames", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfPubTypesSection = + getELFSection(".debug_pubtypes", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfStrSection = + getELFSection(".debug_str", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfLocSection = + getELFSection(".debug_loc", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfARangesSection = + getELFSection(".debug_aranges", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfRangesSection = + getELFSection(".debug_ranges", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); + DwarfMacroInfoSection = + getELFSection(".debug_macinfo", MCSectionELF::SHT_PROGBITS, 0, + SectionKind::getMetadata()); +} + + +static SectionKind +getELFKindForNamedSection(StringRef Name, SectionKind K) { + if (Name.empty() || Name[0] != '.') return K; + + // Some lame default implementation based on some magic section names. + if (Name == ".bss" || + Name.startswith(".bss.") || + Name.startswith(".gnu.linkonce.b.") || + Name.startswith(".llvm.linkonce.b.") || + Name == ".sbss" || + Name.startswith(".sbss.") || + Name.startswith(".gnu.linkonce.sb.") || + Name.startswith(".llvm.linkonce.sb.")) + return SectionKind::getBSS(); + + if (Name == ".tdata" || + Name.startswith(".tdata.") || + Name.startswith(".gnu.linkonce.td.") || + Name.startswith(".llvm.linkonce.td.")) + return SectionKind::getThreadData(); + + if (Name == ".tbss" || + Name.startswith(".tbss.") || + Name.startswith(".gnu.linkonce.tb.") || + Name.startswith(".llvm.linkonce.tb.")) + return SectionKind::getThreadBSS(); + + return K; +} + + +static unsigned getELFSectionType(StringRef Name, SectionKind K) { + + if (Name == ".init_array") + return MCSectionELF::SHT_INIT_ARRAY; + + if (Name == ".fini_array") + return MCSectionELF::SHT_FINI_ARRAY; + + if (Name == ".preinit_array") + return MCSectionELF::SHT_PREINIT_ARRAY; + + if (K.isBSS() || K.isThreadBSS()) + return MCSectionELF::SHT_NOBITS; + + return MCSectionELF::SHT_PROGBITS; +} + + +static unsigned +getELFSectionFlags(SectionKind K) { + unsigned Flags = 0; + + if (!K.isMetadata()) + Flags |= MCSectionELF::SHF_ALLOC; + + if (K.isText()) + Flags |= MCSectionELF::SHF_EXECINSTR; + + if (K.isWriteable()) + Flags |= MCSectionELF::SHF_WRITE; + + if (K.isThreadLocal()) + Flags |= MCSectionELF::SHF_TLS; + + // K.isMergeableConst() is left out to honour PR4650 + if (K.isMergeableCString() || K.isMergeableConst4() || + K.isMergeableConst8() || K.isMergeableConst16()) + Flags |= MCSectionELF::SHF_MERGE; + + if (K.isMergeableCString()) + Flags |= MCSectionELF::SHF_STRINGS; + + return Flags; +} + + +const MCSection *TargetLoweringObjectFileELF:: +getExplicitSectionGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + StringRef SectionName = GV->getSection(); + + // Infer section flags from the section name if we can. + Kind = getELFKindForNamedSection(SectionName, Kind); + + return getELFSection(SectionName, + getELFSectionType(SectionName, Kind), + getELFSectionFlags(Kind), Kind, true); +} + +static const char *getSectionPrefixForUniqueGlobal(SectionKind Kind) { + if (Kind.isText()) return ".gnu.linkonce.t."; + if (Kind.isReadOnly()) return ".gnu.linkonce.r."; + + if (Kind.isThreadData()) return ".gnu.linkonce.td."; + if (Kind.isThreadBSS()) return ".gnu.linkonce.tb."; + + if (Kind.isDataNoRel()) return ".gnu.linkonce.d."; + if (Kind.isDataRelLocal()) return ".gnu.linkonce.d.rel.local."; + if (Kind.isDataRel()) return ".gnu.linkonce.d.rel."; + if (Kind.isReadOnlyWithRelLocal()) return ".gnu.linkonce.d.rel.ro.local."; + + assert(Kind.isReadOnlyWithRel() && "Unknown section kind"); + return ".gnu.linkonce.d.rel.ro."; +} + +const MCSection *TargetLoweringObjectFileELF:: +SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + + // If this global is linkonce/weak and the target handles this by emitting it + // into a 'uniqued' section name, create and return the section now. + if (GV->isWeakForLinker() && !Kind.isCommon() && !Kind.isBSS()) { + const char *Prefix = getSectionPrefixForUniqueGlobal(Kind); + SmallString<128> Name; + Name.append(Prefix, Prefix+strlen(Prefix)); + Mang->getNameWithPrefix(Name, GV, false); + return getELFSection(Name.str(), getELFSectionType(Name.str(), Kind), + getELFSectionFlags(Kind), Kind); + } + + if (Kind.isText()) return TextSection; + + if (Kind.isMergeable1ByteCString() || + Kind.isMergeable2ByteCString() || + Kind.isMergeable4ByteCString()) { + + // We also need alignment here. + // FIXME: this is getting the alignment of the character, not the + // alignment of the global! + unsigned Align = + TM.getTargetData()->getPreferredAlignment(cast<GlobalVariable>(GV)); + + const char *SizeSpec = ".rodata.str1."; + if (Kind.isMergeable2ByteCString()) + SizeSpec = ".rodata.str2."; + else if (Kind.isMergeable4ByteCString()) + SizeSpec = ".rodata.str4."; + else + assert(Kind.isMergeable1ByteCString() && "unknown string width"); + + + std::string Name = SizeSpec + utostr(Align); + return getELFSection(Name, MCSectionELF::SHT_PROGBITS, + MCSectionELF::SHF_ALLOC | + MCSectionELF::SHF_MERGE | + MCSectionELF::SHF_STRINGS, + Kind); + } + + if (Kind.isMergeableConst()) { + if (Kind.isMergeableConst4() && MergeableConst4Section) + return MergeableConst4Section; + if (Kind.isMergeableConst8() && MergeableConst8Section) + return MergeableConst8Section; + if (Kind.isMergeableConst16() && MergeableConst16Section) + return MergeableConst16Section; + return ReadOnlySection; // .const + } + + if (Kind.isReadOnly()) return ReadOnlySection; + + if (Kind.isThreadData()) return TLSDataSection; + if (Kind.isThreadBSS()) return TLSBSSSection; + + // Note: we claim that common symbols are put in BSSSection, but they are + // really emitted with the magic .comm directive, which creates a symbol table + // entry but not a section. + if (Kind.isBSS() || Kind.isCommon()) return BSSSection; + + if (Kind.isDataNoRel()) return DataSection; + if (Kind.isDataRelLocal()) return DataRelLocalSection; + if (Kind.isDataRel()) return DataRelSection; + if (Kind.isReadOnlyWithRelLocal()) return DataRelROLocalSection; + + assert(Kind.isReadOnlyWithRel() && "Unknown section kind"); + return DataRelROSection; +} + +/// getSectionForConstant - Given a mergeable constant with the +/// specified size and relocation information, return a section that it +/// should be placed in. +const MCSection *TargetLoweringObjectFileELF:: +getSectionForConstant(SectionKind Kind) const { + if (Kind.isMergeableConst4() && MergeableConst4Section) + return MergeableConst4Section; + if (Kind.isMergeableConst8() && MergeableConst8Section) + return MergeableConst8Section; + if (Kind.isMergeableConst16() && MergeableConst16Section) + return MergeableConst16Section; + if (Kind.isReadOnly()) + return ReadOnlySection; + + if (Kind.isReadOnlyWithRelLocal()) return DataRelROLocalSection; + assert(Kind.isReadOnlyWithRel() && "Unknown section kind"); + return DataRelROSection; +} + +const MCExpr *TargetLoweringObjectFileELF:: +getSymbolForDwarfGlobalReference(const GlobalValue *GV, Mangler *Mang, + MachineModuleInfo *MMI, unsigned Encoding) const { + + if (Encoding & dwarf::DW_EH_PE_indirect) { + MachineModuleInfoELF &ELFMMI = MMI->getObjFileInfo<MachineModuleInfoELF>(); + + SmallString<128> Name; + Mang->getNameWithPrefix(Name, GV, true); + Name += ".DW.stub"; + + // Add information about the stub reference to ELFMMI so that the stub + // gets emitted by the asmprinter. + MCSymbol *Sym = getContext().GetOrCreateSymbol(Name.str()); + MCSymbol *&StubSym = ELFMMI.getGVStubEntry(Sym); + if (StubSym == 0) { + Name.clear(); + Mang->getNameWithPrefix(Name, GV, false); + StubSym = getContext().GetOrCreateSymbol(Name.str()); + } + + return TargetLoweringObjectFile:: + getSymbolForDwarfReference(Sym, MMI, + Encoding & ~dwarf::DW_EH_PE_indirect); + } + + return TargetLoweringObjectFile:: + getSymbolForDwarfGlobalReference(GV, Mang, MMI, Encoding); +} + +//===----------------------------------------------------------------------===// +// MachO +//===----------------------------------------------------------------------===// + +typedef StringMap<const MCSectionMachO*> MachOUniqueMapTy; + +TargetLoweringObjectFileMachO::~TargetLoweringObjectFileMachO() { + // If we have the MachO uniquing map, free it. + delete (MachOUniqueMapTy*)UniquingMap; +} + + +const MCSectionMachO *TargetLoweringObjectFileMachO:: +getMachOSection(StringRef Segment, StringRef Section, + unsigned TypeAndAttributes, + unsigned Reserved2, SectionKind Kind) const { + // We unique sections by their segment/section pair. The returned section + // may not have the same flags as the requested section, if so this should be + // diagnosed by the client as an error. + + // Create the map if it doesn't already exist. + if (UniquingMap == 0) + UniquingMap = new MachOUniqueMapTy(); + MachOUniqueMapTy &Map = *(MachOUniqueMapTy*)UniquingMap; + + // Form the name to look up. + SmallString<64> Name; + Name += Segment; + Name.push_back(','); + Name += Section; + + // Do the lookup, if we have a hit, return it. + const MCSectionMachO *&Entry = Map[Name.str()]; + if (Entry) return Entry; + + // Otherwise, return a new section. + return Entry = MCSectionMachO::Create(Segment, Section, TypeAndAttributes, + Reserved2, Kind, getContext()); +} + + +void TargetLoweringObjectFileMachO::Initialize(MCContext &Ctx, + const TargetMachine &TM) { + if (UniquingMap != 0) + ((MachOUniqueMapTy*)UniquingMap)->clear(); + TargetLoweringObjectFile::Initialize(Ctx, TM); + + TextSection // .text + = getMachOSection("__TEXT", "__text", + MCSectionMachO::S_ATTR_PURE_INSTRUCTIONS, + SectionKind::getText()); + DataSection // .data + = getMachOSection("__DATA", "__data", 0, SectionKind::getDataRel()); + + CStringSection // .cstring + = getMachOSection("__TEXT", "__cstring", MCSectionMachO::S_CSTRING_LITERALS, + SectionKind::getMergeable1ByteCString()); + UStringSection + = getMachOSection("__TEXT","__ustring", 0, + SectionKind::getMergeable2ByteCString()); + FourByteConstantSection // .literal4 + = getMachOSection("__TEXT", "__literal4", MCSectionMachO::S_4BYTE_LITERALS, + SectionKind::getMergeableConst4()); + EightByteConstantSection // .literal8 + = getMachOSection("__TEXT", "__literal8", MCSectionMachO::S_8BYTE_LITERALS, + SectionKind::getMergeableConst8()); + + // ld_classic doesn't support .literal16 in 32-bit mode, and ld64 falls back + // to using it in -static mode. + SixteenByteConstantSection = 0; + if (TM.getRelocationModel() != Reloc::Static && + TM.getTargetData()->getPointerSize() == 32) + SixteenByteConstantSection = // .literal16 + getMachOSection("__TEXT", "__literal16",MCSectionMachO::S_16BYTE_LITERALS, + SectionKind::getMergeableConst16()); + + ReadOnlySection // .const + = getMachOSection("__TEXT", "__const", 0, SectionKind::getReadOnly()); + + TextCoalSection + = getMachOSection("__TEXT", "__textcoal_nt", + MCSectionMachO::S_COALESCED | + MCSectionMachO::S_ATTR_PURE_INSTRUCTIONS, + SectionKind::getText()); + ConstTextCoalSection + = getMachOSection("__TEXT", "__const_coal", MCSectionMachO::S_COALESCED, + SectionKind::getText()); + ConstDataCoalSection + = getMachOSection("__DATA","__const_coal", MCSectionMachO::S_COALESCED, + SectionKind::getText()); + ConstDataSection // .const_data + = getMachOSection("__DATA", "__const", 0, + SectionKind::getReadOnlyWithRel()); + DataCoalSection + = getMachOSection("__DATA","__datacoal_nt", MCSectionMachO::S_COALESCED, + SectionKind::getDataRel()); + DataCommonSection + = getMachOSection("__DATA","__common", MCSectionMachO::S_ZEROFILL, + SectionKind::getBSS()); + DataBSSSection + = getMachOSection("__DATA","__bss", MCSectionMachO::S_ZEROFILL, + SectionKind::getBSS()); + + + LazySymbolPointerSection + = getMachOSection("__DATA", "__la_symbol_ptr", + MCSectionMachO::S_LAZY_SYMBOL_POINTERS, + SectionKind::getMetadata()); + NonLazySymbolPointerSection + = getMachOSection("__DATA", "__nl_symbol_ptr", + MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS, + SectionKind::getMetadata()); + + if (TM.getRelocationModel() == Reloc::Static) { + StaticCtorSection + = getMachOSection("__TEXT", "__constructor", 0,SectionKind::getDataRel()); + StaticDtorSection + = getMachOSection("__TEXT", "__destructor", 0, SectionKind::getDataRel()); + } else { + StaticCtorSection + = getMachOSection("__DATA", "__mod_init_func", + MCSectionMachO::S_MOD_INIT_FUNC_POINTERS, + SectionKind::getDataRel()); + StaticDtorSection + = getMachOSection("__DATA", "__mod_term_func", + MCSectionMachO::S_MOD_TERM_FUNC_POINTERS, + SectionKind::getDataRel()); + } + + // Exception Handling. + LSDASection = getMachOSection("__DATA", "__gcc_except_tab", 0, + SectionKind::getDataRel()); + EHFrameSection = + getMachOSection("__TEXT", "__eh_frame", + MCSectionMachO::S_COALESCED | + MCSectionMachO::S_ATTR_NO_TOC | + MCSectionMachO::S_ATTR_STRIP_STATIC_SYMS | + MCSectionMachO::S_ATTR_LIVE_SUPPORT, + SectionKind::getReadOnly()); + + // Debug Information. + DwarfAbbrevSection = + getMachOSection("__DWARF", "__debug_abbrev", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfInfoSection = + getMachOSection("__DWARF", "__debug_info", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfLineSection = + getMachOSection("__DWARF", "__debug_line", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfFrameSection = + getMachOSection("__DWARF", "__debug_frame", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfPubNamesSection = + getMachOSection("__DWARF", "__debug_pubnames", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfPubTypesSection = + getMachOSection("__DWARF", "__debug_pubtypes", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfStrSection = + getMachOSection("__DWARF", "__debug_str", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfLocSection = + getMachOSection("__DWARF", "__debug_loc", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfARangesSection = + getMachOSection("__DWARF", "__debug_aranges", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfRangesSection = + getMachOSection("__DWARF", "__debug_ranges", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfMacroInfoSection = + getMachOSection("__DWARF", "__debug_macinfo", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); + DwarfDebugInlineSection = + getMachOSection("__DWARF", "__debug_inlined", MCSectionMachO::S_ATTR_DEBUG, + SectionKind::getMetadata()); +} + +const MCSection *TargetLoweringObjectFileMachO:: +getExplicitSectionGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + // Parse the section specifier and create it if valid. + StringRef Segment, Section; + unsigned TAA, StubSize; + std::string ErrorCode = + MCSectionMachO::ParseSectionSpecifier(GV->getSection(), Segment, Section, + TAA, StubSize); + if (!ErrorCode.empty()) { + // If invalid, report the error with llvm_report_error. + llvm_report_error("Global variable '" + GV->getNameStr() + + "' has an invalid section specifier '" + GV->getSection()+ + "': " + ErrorCode + "."); + // Fall back to dropping it into the data section. + return DataSection; + } + + // Get the section. + const MCSectionMachO *S = + getMachOSection(Segment, Section, TAA, StubSize, Kind); + + // Okay, now that we got the section, verify that the TAA & StubSize agree. + // If the user declared multiple globals with different section flags, we need + // to reject it here. + if (S->getTypeAndAttributes() != TAA || S->getStubSize() != StubSize) { + // If invalid, report the error with llvm_report_error. + llvm_report_error("Global variable '" + GV->getNameStr() + + "' section type or attributes does not match previous" + " section specifier"); + } + + return S; +} + +const MCSection *TargetLoweringObjectFileMachO:: +SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + assert(!Kind.isThreadLocal() && "Darwin doesn't support TLS"); + + if (Kind.isText()) + return GV->isWeakForLinker() ? TextCoalSection : TextSection; + + // If this is weak/linkonce, put this in a coalescable section, either in text + // or data depending on if it is writable. + if (GV->isWeakForLinker()) { + if (Kind.isReadOnly()) + return ConstTextCoalSection; + return DataCoalSection; + } + + // FIXME: Alignment check should be handled by section classifier. + if (Kind.isMergeable1ByteCString() && + TM.getTargetData()->getPreferredAlignment(cast<GlobalVariable>(GV)) < 32) + return CStringSection; + + // Do not put 16-bit arrays in the UString section if they have an + // externally visible label, this runs into issues with certain linker + // versions. + if (Kind.isMergeable2ByteCString() && !GV->hasExternalLinkage() && + TM.getTargetData()->getPreferredAlignment(cast<GlobalVariable>(GV)) < 32) + return UStringSection; + + if (Kind.isMergeableConst()) { + if (Kind.isMergeableConst4()) + return FourByteConstantSection; + if (Kind.isMergeableConst8()) + return EightByteConstantSection; + if (Kind.isMergeableConst16() && SixteenByteConstantSection) + return SixteenByteConstantSection; + } + + // Otherwise, if it is readonly, but not something we can specially optimize, + // just drop it in .const. + if (Kind.isReadOnly()) + return ReadOnlySection; + + // If this is marked const, put it into a const section. But if the dynamic + // linker needs to write to it, put it in the data segment. + if (Kind.isReadOnlyWithRel()) + return ConstDataSection; + + // Put zero initialized globals with strong external linkage in the + // DATA, __common section with the .zerofill directive. + if (Kind.isBSSExtern()) + return DataCommonSection; + + // Put zero initialized globals with local linkage in __DATA,__bss directive + // with the .zerofill directive (aka .lcomm). + if (Kind.isBSSLocal()) + return DataBSSSection; + + // Otherwise, just drop the variable in the normal data section. + return DataSection; +} + +const MCSection * +TargetLoweringObjectFileMachO::getSectionForConstant(SectionKind Kind) const { + // If this constant requires a relocation, we have to put it in the data + // segment, not in the text segment. + if (Kind.isDataRel() || Kind.isReadOnlyWithRel()) + return ConstDataSection; + + if (Kind.isMergeableConst4()) + return FourByteConstantSection; + if (Kind.isMergeableConst8()) + return EightByteConstantSection; + if (Kind.isMergeableConst16() && SixteenByteConstantSection) + return SixteenByteConstantSection; + return ReadOnlySection; // .const +} + +/// shouldEmitUsedDirectiveFor - This hook allows targets to selectively decide +/// not to emit the UsedDirective for some symbols in llvm.used. +// FIXME: REMOVE this (rdar://7071300) +bool TargetLoweringObjectFileMachO:: +shouldEmitUsedDirectiveFor(const GlobalValue *GV, Mangler *Mang) const { + /// On Darwin, internally linked data beginning with "L" or "l" does not have + /// the directive emitted (this occurs in ObjC metadata). + if (!GV) return false; + + // Check whether the mangled name has the "Private" or "LinkerPrivate" prefix. + if (GV->hasLocalLinkage() && !isa<Function>(GV)) { + // FIXME: ObjC metadata is currently emitted as internal symbols that have + // \1L and \0l prefixes on them. Fix them to be Private/LinkerPrivate and + // this horrible hack can go away. + SmallString<64> Name; + Mang->getNameWithPrefix(Name, GV, false); + if (Name[0] == 'L' || Name[0] == 'l') + return false; + } + + return true; +} + +const MCExpr *TargetLoweringObjectFileMachO:: +getSymbolForDwarfGlobalReference(const GlobalValue *GV, Mangler *Mang, + MachineModuleInfo *MMI, unsigned Encoding) const { + // The mach-o version of this method defaults to returning a stub reference. + + if (Encoding & DW_EH_PE_indirect) { + MachineModuleInfoMachO &MachOMMI = + MMI->getObjFileInfo<MachineModuleInfoMachO>(); + + SmallString<128> Name; + Mang->getNameWithPrefix(Name, GV, true); + Name += "$non_lazy_ptr"; + + // Add information about the stub reference to MachOMMI so that the stub + // gets emitted by the asmprinter. + MCSymbol *Sym = getContext().GetOrCreateSymbol(Name.str()); + MCSymbol *&StubSym = MachOMMI.getGVStubEntry(Sym); + if (StubSym == 0) { + Name.clear(); + Mang->getNameWithPrefix(Name, GV, false); + StubSym = getContext().GetOrCreateSymbol(Name.str()); + } + + return TargetLoweringObjectFile:: + getSymbolForDwarfReference(Sym, MMI, + Encoding & ~dwarf::DW_EH_PE_indirect); + } + + return TargetLoweringObjectFile:: + getSymbolForDwarfGlobalReference(GV, Mang, MMI, Encoding); +} + +unsigned TargetLoweringObjectFileMachO::getPersonalityEncoding() const { + return DW_EH_PE_indirect | DW_EH_PE_pcrel | DW_EH_PE_sdata4; +} + +unsigned TargetLoweringObjectFileMachO::getLSDAEncoding() const { + return DW_EH_PE_pcrel; +} + +unsigned TargetLoweringObjectFileMachO::getFDEEncoding() const { + return DW_EH_PE_pcrel; +} + +unsigned TargetLoweringObjectFileMachO::getTTypeEncoding() const { + return DW_EH_PE_absptr; +} + +//===----------------------------------------------------------------------===// +// COFF +//===----------------------------------------------------------------------===// + +typedef StringMap<const MCSectionCOFF*> COFFUniqueMapTy; + +TargetLoweringObjectFileCOFF::~TargetLoweringObjectFileCOFF() { + delete (COFFUniqueMapTy*)UniquingMap; +} + + +const MCSection *TargetLoweringObjectFileCOFF:: +getCOFFSection(StringRef Name, bool isDirective, SectionKind Kind) const { + // Create the map if it doesn't already exist. + if (UniquingMap == 0) + UniquingMap = new MachOUniqueMapTy(); + COFFUniqueMapTy &Map = *(COFFUniqueMapTy*)UniquingMap; + + // Do the lookup, if we have a hit, return it. + const MCSectionCOFF *&Entry = Map[Name]; + if (Entry) return Entry; + + return Entry = MCSectionCOFF::Create(Name, isDirective, Kind, getContext()); +} + +void TargetLoweringObjectFileCOFF::Initialize(MCContext &Ctx, + const TargetMachine &TM) { + if (UniquingMap != 0) + ((COFFUniqueMapTy*)UniquingMap)->clear(); + TargetLoweringObjectFile::Initialize(Ctx, TM); + TextSection = getCOFFSection("\t.text", true, SectionKind::getText()); + DataSection = getCOFFSection("\t.data", true, SectionKind::getDataRel()); + StaticCtorSection = + getCOFFSection(".ctors", false, SectionKind::getDataRel()); + StaticDtorSection = + getCOFFSection(".dtors", false, SectionKind::getDataRel()); + + // FIXME: We're emitting LSDA info into a readonly section on COFF, even + // though it contains relocatable pointers. In PIC mode, this is probably a + // big runtime hit for C++ apps. Either the contents of the LSDA need to be + // adjusted or this should be a data section. + LSDASection = + getCOFFSection(".gcc_except_table", false, SectionKind::getReadOnly()); + EHFrameSection = + getCOFFSection(".eh_frame", false, SectionKind::getDataRel()); + + // Debug info. + // FIXME: Don't use 'directive' mode here. + DwarfAbbrevSection = + getCOFFSection("\t.section\t.debug_abbrev,\"dr\"", + true, SectionKind::getMetadata()); + DwarfInfoSection = + getCOFFSection("\t.section\t.debug_info,\"dr\"", + true, SectionKind::getMetadata()); + DwarfLineSection = + getCOFFSection("\t.section\t.debug_line,\"dr\"", + true, SectionKind::getMetadata()); + DwarfFrameSection = + getCOFFSection("\t.section\t.debug_frame,\"dr\"", + true, SectionKind::getMetadata()); + DwarfPubNamesSection = + getCOFFSection("\t.section\t.debug_pubnames,\"dr\"", + true, SectionKind::getMetadata()); + DwarfPubTypesSection = + getCOFFSection("\t.section\t.debug_pubtypes,\"dr\"", + true, SectionKind::getMetadata()); + DwarfStrSection = + getCOFFSection("\t.section\t.debug_str,\"dr\"", + true, SectionKind::getMetadata()); + DwarfLocSection = + getCOFFSection("\t.section\t.debug_loc,\"dr\"", + true, SectionKind::getMetadata()); + DwarfARangesSection = + getCOFFSection("\t.section\t.debug_aranges,\"dr\"", + true, SectionKind::getMetadata()); + DwarfRangesSection = + getCOFFSection("\t.section\t.debug_ranges,\"dr\"", + true, SectionKind::getMetadata()); + DwarfMacroInfoSection = + getCOFFSection("\t.section\t.debug_macinfo,\"dr\"", + true, SectionKind::getMetadata()); +} + +const MCSection *TargetLoweringObjectFileCOFF:: +getExplicitSectionGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + return getCOFFSection(GV->getSection(), false, Kind); +} + +static const char *getCOFFSectionPrefixForUniqueGlobal(SectionKind Kind) { + if (Kind.isText()) + return ".text$linkonce"; + if (Kind.isWriteable()) + return ".data$linkonce"; + return ".rdata$linkonce"; +} + + +const MCSection *TargetLoweringObjectFileCOFF:: +SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind, + Mangler *Mang, const TargetMachine &TM) const { + assert(!Kind.isThreadLocal() && "Doesn't support TLS"); + + // If this global is linkonce/weak and the target handles this by emitting it + // into a 'uniqued' section name, create and return the section now. + if (GV->isWeakForLinker()) { + const char *Prefix = getCOFFSectionPrefixForUniqueGlobal(Kind); + SmallString<128> Name(Prefix, Prefix+strlen(Prefix)); + Mang->getNameWithPrefix(Name, GV, false); + return getCOFFSection(Name.str(), false, Kind); + } + + if (Kind.isText()) + return getTextSection(); + + return getDataSection(); +} + diff --git a/lib/CodeGen/TwoAddressInstructionPass.cpp b/lib/CodeGen/TwoAddressInstructionPass.cpp index 6c7c1a1..c840b39 100644 --- a/lib/CodeGen/TwoAddressInstructionPass.cpp +++ b/lib/CodeGen/TwoAddressInstructionPass.cpp @@ -160,7 +160,7 @@ bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB, MachineBasicBlock::iterator OldPos) { // Check if it's safe to move this instruction. bool SeenStore = true; // Be conservative. - if (!MI->isSafeToMove(TII, SeenStore, AA)) + if (!MI->isSafeToMove(TII, AA, SeenStore)) return false; unsigned DefReg = 0; @@ -213,6 +213,9 @@ bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB, unsigned NumVisited = 0; for (MachineBasicBlock::iterator I = llvm::next(OldPos); I != KillPos; ++I) { MachineInstr *OtherMI = I; + // DBG_VALUE cannot be counted against the limit. + if (OtherMI->isDebugValue()) + continue; if (NumVisited > 30) // FIXME: Arbitrary limit to reduce compile time cost. return false; ++NumVisited; @@ -451,13 +454,10 @@ MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB, const TargetInstrInfo *TII, bool &IsCopy, unsigned &DstReg, bool &IsDstPhys) { - MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg); - if (UI == MRI->use_end()) - return 0; - MachineInstr &UseMI = *UI; - if (++UI != MRI->use_end()) - // More than one use. + if (!MRI->hasOneNonDBGUse(Reg)) + // None or more than one use. return 0; + MachineInstr &UseMI = *MRI->use_nodbg_begin(Reg); if (UseMI.getParent() != MBB) return 0; unsigned SrcReg; @@ -923,6 +923,10 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) { for (MachineBasicBlock::iterator mi = mbbi->begin(), me = mbbi->end(); mi != me; ) { MachineBasicBlock::iterator nmi = llvm::next(mi); + if (mi->isDebugValue()) { + mi = nmi; + continue; + } const TargetInstrDesc &TID = mi->getDesc(); bool FirstTied = true; @@ -1021,7 +1025,7 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) { // copying it. if (DefMI && DefMI->getDesc().isAsCheapAsAMove() && - DefMI->isSafeToReMat(TII, regB, AA) && + DefMI->isSafeToReMat(TII, AA, regB) && isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist)){ DEBUG(dbgs() << "2addr: REMATTING : " << *DefMI << "\n"); unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg(); diff --git a/lib/CodeGen/VirtRegMap.cpp b/lib/CodeGen/VirtRegMap.cpp index 5956b61..ed02696 100644 --- a/lib/CodeGen/VirtRegMap.cpp +++ b/lib/CodeGen/VirtRegMap.cpp @@ -261,19 +261,21 @@ bool VirtRegMap::FindUnusedRegisters(LiveIntervals* LIs) { void VirtRegMap::print(raw_ostream &OS, const Module* M) const { const TargetRegisterInfo* TRI = MF->getTarget().getRegisterInfo(); + const MachineRegisterInfo &MRI = MF->getRegInfo(); OS << "********** REGISTER MAP **********\n"; for (unsigned i = TargetRegisterInfo::FirstVirtualRegister, e = MF->getRegInfo().getLastVirtReg(); i <= e; ++i) { if (Virt2PhysMap[i] != (unsigned)VirtRegMap::NO_PHYS_REG) OS << "[reg" << i << " -> " << TRI->getName(Virt2PhysMap[i]) - << "]\n"; + << "] " << MRI.getRegClass(i)->getName() << "\n"; } for (unsigned i = TargetRegisterInfo::FirstVirtualRegister, e = MF->getRegInfo().getLastVirtReg(); i <= e; ++i) if (Virt2StackSlotMap[i] != VirtRegMap::NO_STACK_SLOT) - OS << "[reg" << i << " -> fi#" << Virt2StackSlotMap[i] << "]\n"; + OS << "[reg" << i << " -> fi#" << Virt2StackSlotMap[i] + << "] " << MRI.getRegClass(i)->getName() << "\n"; OS << '\n'; } diff --git a/lib/CodeGen/VirtRegRewriter.cpp b/lib/CodeGen/VirtRegRewriter.cpp index ce62594..7aa0a91 100644 --- a/lib/CodeGen/VirtRegRewriter.cpp +++ b/lib/CodeGen/VirtRegRewriter.cpp @@ -46,7 +46,7 @@ namespace { static cl::opt<RewriterName> RewriterOpt("rewriter", - cl::desc("Rewriter to use: (default: local)"), + cl::desc("Rewriter to use (default=local)"), cl::Prefix, cl::values(clEnumVal(local, "local rewriter"), clEnumVal(trivial, "trivial rewriter"), @@ -62,6 +62,7 @@ VirtRegRewriter::~VirtRegRewriter() {} /// substitutePhysReg - Replace virtual register in MachineOperand with a /// physical register. Do the right thing with the sub-register index. +/// Note that operands may be added, so the MO reference is no longer valid. static void substitutePhysReg(MachineOperand &MO, unsigned Reg, const TargetRegisterInfo &TRI) { if (unsigned SubIdx = MO.getSubReg()) { @@ -123,14 +124,15 @@ struct TrivialRewriter : public VirtRegRewriter { continue; unsigned pReg = VRM.getPhys(reg); mri->setPhysRegUsed(pReg); - for (MachineRegisterInfo::reg_iterator regItr = mri->reg_begin(reg), - regEnd = mri->reg_end(); regItr != regEnd;) { - MachineOperand &mop = regItr.getOperand(); - assert(mop.isReg() && mop.getReg() == reg && "reg_iterator broken?"); - ++regItr; - substitutePhysReg(mop, pReg, *tri); - changed = true; - } + // Copy the register use-list before traversing it. + SmallVector<std::pair<MachineInstr*, unsigned>, 32> reglist; + for (MachineRegisterInfo::reg_iterator I = mri->reg_begin(reg), + E = mri->reg_end(); I != E; ++I) + reglist.push_back(std::make_pair(&*I, I.getOperandNo())); + for (unsigned N=0; N != reglist.size(); ++N) + substitutePhysReg(reglist[N].first->getOperand(reglist[N].second), + pReg, *tri); + changed |= !reglist.empty(); } } @@ -1850,19 +1852,18 @@ private: KilledMIRegs.clear(); for (unsigned j = 0, e = VirtUseOps.size(); j != e; ++j) { unsigned i = VirtUseOps[j]; - MachineOperand &MO = MI.getOperand(i); - unsigned VirtReg = MO.getReg(); + unsigned VirtReg = MI.getOperand(i).getReg(); assert(TargetRegisterInfo::isVirtualRegister(VirtReg) && "Not a virtual register?"); - unsigned SubIdx = MO.getSubReg(); + unsigned SubIdx = MI.getOperand(i).getSubReg(); if (VRM.isAssignedReg(VirtReg)) { // This virtual register was assigned a physreg! unsigned Phys = VRM.getPhys(VirtReg); RegInfo->setPhysRegUsed(Phys); - if (MO.isDef()) + if (MI.getOperand(i).isDef()) ReusedOperands.markClobbered(Phys); - substitutePhysReg(MO, Phys, *TRI); + substitutePhysReg(MI.getOperand(i), Phys, *TRI); if (VRM.isImplicitlyDefined(VirtReg)) // FIXME: Is this needed? BuildMI(MBB, &MI, MI.getDebugLoc(), @@ -1871,10 +1872,10 @@ private: } // This virtual register is now known to be a spilled value. - if (!MO.isUse()) + if (!MI.getOperand(i).isUse()) continue; // Handle defs in the loop below (handle use&def here though) - bool AvoidReload = MO.isUndef(); + bool AvoidReload = MI.getOperand(i).isUndef(); // Check if it is defined by an implicit def. It should not be spilled. // Note, this is for correctness reason. e.g. // 8 %reg1024<def> = IMPLICIT_DEF |