//===----------------------------------------------------------------------===//
#include "NVPTXAsmPrinter.h"
+#include "InstPrinter/NVPTXInstPrinter.h"
#include "MCTargetDesc/NVPTXMCAsmInfo.h"
#include "NVPTX.h"
#include "NVPTXInstrInfo.h"
#include "NVPTXMCExpr.h"
+#include "NVPTXMachineFunctionInfo.h"
#include "NVPTXRegisterInfo.h"
#include "NVPTXTargetMachine.h"
#include "NVPTXUtilities.h"
-#include "InstPrinter/NVPTXInstPrinter.h"
#include "cl_common_defines.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Analysis/ConstantFolding.h"
-#include "llvm/Assembly/Writer.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
-#include "llvm/DebugInfo.h"
+#include "llvm/IR/DebugInfo.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Mangler.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/Operator.h"
#include "llvm/MC/MCStreamer.h"
#include "llvm/Support/Path.h"
#include "llvm/Support/TargetRegistry.h"
#include "llvm/Support/TimeValue.h"
-#include "llvm/Target/Mangler.h"
#include "llvm/Target/TargetLoweringObjectFile.h"
+#include "llvm/Transforms/Utils/UnrollLoop.h"
#include <sstream>
using namespace llvm;
return;
// Do we have a circular dependency?
- if (Visiting.count(GV))
+ if (!Visiting.insert(GV).second)
report_fatal_error("Circular dependency found in global variable set");
- // Start visiting this global
- Visiting.insert(GV);
-
// Make sure we visit all dependents first
DenseSet<const GlobalVariable *> Others;
for (unsigned i = 0, e = GV->getNumOperands(); i != e; ++i)
}
}
-// @TODO: This is a copy from AsmPrinter.cpp. The function is static, so we
-// cannot just link to the existing version.
-/// LowerConstant - Lower the specified LLVM Constant to an MCExpr.
-///
-using namespace nvptx;
-const MCExpr *nvptx::LowerConstant(const Constant *CV, AsmPrinter &AP) {
- MCContext &Ctx = AP.OutContext;
-
- if (CV->isNullValue() || isa<UndefValue>(CV))
- return MCConstantExpr::Create(0, Ctx);
-
- if (const ConstantInt *CI = dyn_cast<ConstantInt>(CV))
- return MCConstantExpr::Create(CI->getZExtValue(), Ctx);
-
- if (const GlobalValue *GV = dyn_cast<GlobalValue>(CV))
- return MCSymbolRefExpr::Create(AP.getSymbol(GV), Ctx);
-
- if (const BlockAddress *BA = dyn_cast<BlockAddress>(CV))
- return MCSymbolRefExpr::Create(AP.GetBlockAddressSymbol(BA), Ctx);
-
- const ConstantExpr *CE = dyn_cast<ConstantExpr>(CV);
- if (CE == 0)
- llvm_unreachable("Unknown constant value to lower!");
-
- switch (CE->getOpcode()) {
- default:
- // If the code isn't optimized, there may be outstanding folding
- // opportunities. Attempt to fold the expression using DataLayout as a
- // last resort before giving up.
- if (Constant *C = ConstantFoldConstantExpression(CE, AP.TM.getDataLayout()))
- if (C != CE)
- return LowerConstant(C, AP);
-
- // Otherwise report the problem to the user.
- {
- std::string S;
- raw_string_ostream OS(S);
- OS << "Unsupported expression in static initializer: ";
- WriteAsOperand(OS, CE, /*PrintType=*/ false,
- !AP.MF ? 0 : AP.MF->getFunction()->getParent());
- report_fatal_error(OS.str());
- }
- case Instruction::GetElementPtr: {
- const DataLayout &TD = *AP.TM.getDataLayout();
- // Generate a symbolic expression for the byte address
- APInt OffsetAI(TD.getPointerSizeInBits(), 0);
- cast<GEPOperator>(CE)->accumulateConstantOffset(TD, OffsetAI);
-
- const MCExpr *Base = LowerConstant(CE->getOperand(0), AP);
- if (!OffsetAI)
- return Base;
-
- int64_t Offset = OffsetAI.getSExtValue();
- return MCBinaryExpr::CreateAdd(Base, MCConstantExpr::Create(Offset, Ctx),
- Ctx);
- }
-
- case Instruction::Trunc:
- // We emit the value and depend on the assembler to truncate the generated
- // expression properly. This is important for differences between
- // blockaddress labels. Since the two labels are in the same function, it
- // is reasonable to treat their delta as a 32-bit value.
- // FALL THROUGH.
- case Instruction::BitCast:
- return LowerConstant(CE->getOperand(0), AP);
-
- case Instruction::IntToPtr: {
- const DataLayout &TD = *AP.TM.getDataLayout();
- // Handle casts to pointers by changing them into casts to the appropriate
- // integer type. This promotes constant folding and simplifies this code.
- Constant *Op = CE->getOperand(0);
- Op = ConstantExpr::getIntegerCast(Op, TD.getIntPtrType(CV->getContext()),
- false /*ZExt*/);
- return LowerConstant(Op, AP);
- }
-
- case Instruction::PtrToInt: {
- const DataLayout &TD = *AP.TM.getDataLayout();
- // Support only foldable casts to/from pointers that can be eliminated by
- // changing the pointer to the appropriately sized integer type.
- Constant *Op = CE->getOperand(0);
- Type *Ty = CE->getType();
-
- const MCExpr *OpExpr = LowerConstant(Op, AP);
-
- // We can emit the pointer value into this slot if the slot is an
- // integer slot equal to the size of the pointer.
- if (TD.getTypeAllocSize(Ty) == TD.getTypeAllocSize(Op->getType()))
- return OpExpr;
-
- // Otherwise the pointer is smaller than the resultant integer, mask off
- // the high bits so we are sure to get a proper truncation if the input is
- // a constant expr.
- unsigned InBits = TD.getTypeAllocSizeInBits(Op->getType());
- const MCExpr *MaskExpr =
- MCConstantExpr::Create(~0ULL >> (64 - InBits), Ctx);
- return MCBinaryExpr::CreateAnd(OpExpr, MaskExpr, Ctx);
- }
-
- // The MC library also has a right-shift operator, but it isn't consistently
- // signed or unsigned between different targets.
- case Instruction::Add:
- case Instruction::Sub:
- case Instruction::Mul:
- case Instruction::SDiv:
- case Instruction::SRem:
- case Instruction::Shl:
- case Instruction::And:
- case Instruction::Or:
- case Instruction::Xor: {
- const MCExpr *LHS = LowerConstant(CE->getOperand(0), AP);
- const MCExpr *RHS = LowerConstant(CE->getOperand(1), AP);
- switch (CE->getOpcode()) {
- default:
- llvm_unreachable("Unknown binary operator constant cast expr");
- case Instruction::Add:
- return MCBinaryExpr::CreateAdd(LHS, RHS, Ctx);
- case Instruction::Sub:
- return MCBinaryExpr::CreateSub(LHS, RHS, Ctx);
- case Instruction::Mul:
- return MCBinaryExpr::CreateMul(LHS, RHS, Ctx);
- case Instruction::SDiv:
- return MCBinaryExpr::CreateDiv(LHS, RHS, Ctx);
- case Instruction::SRem:
- return MCBinaryExpr::CreateMod(LHS, RHS, Ctx);
- case Instruction::Shl:
- return MCBinaryExpr::CreateShl(LHS, RHS, Ctx);
- case Instruction::And:
- return MCBinaryExpr::CreateAnd(LHS, RHS, Ctx);
- case Instruction::Or:
- return MCBinaryExpr::CreateOr(LHS, RHS, Ctx);
- case Instruction::Xor:
- return MCBinaryExpr::CreateXor(LHS, RHS, Ctx);
- }
- }
- }
-}
-
void NVPTXAsmPrinter::emitLineNumberAsDotLoc(const MachineInstr &MI) {
if (!EmitLineNumbers)
return;
void NVPTXAsmPrinter::EmitInstruction(const MachineInstr *MI) {
SmallString<128> Str;
raw_svector_ostream OS(Str);
- if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
+ if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA)
emitLineNumberAsDotLoc(*MI);
MCInst Inst;
lowerToMCInst(MI, Inst);
- OutStreamer.EmitInstruction(Inst);
+ EmitToStreamer(OutStreamer, Inst);
+}
+
+// Handle symbol backtracking for targets that do not support image handles
+bool NVPTXAsmPrinter::lowerImageHandleOperand(const MachineInstr *MI,
+ unsigned OpNo, MCOperand &MCOp) {
+ const MachineOperand &MO = MI->getOperand(OpNo);
+ const MCInstrDesc &MCID = MI->getDesc();
+
+ if (MCID.TSFlags & NVPTXII::IsTexFlag) {
+ // This is a texture fetch, so operand 4 is a texref and operand 5 is
+ // a samplerref
+ if (OpNo == 4 && MO.isImm()) {
+ lowerImageHandleSymbol(MO.getImm(), MCOp);
+ return true;
+ }
+ if (OpNo == 5 && MO.isImm() && !(MCID.TSFlags & NVPTXII::IsTexModeUnifiedFlag)) {
+ lowerImageHandleSymbol(MO.getImm(), MCOp);
+ return true;
+ }
+
+ return false;
+ } else if (MCID.TSFlags & NVPTXII::IsSuldMask) {
+ unsigned VecSize =
+ 1 << (((MCID.TSFlags & NVPTXII::IsSuldMask) >> NVPTXII::IsSuldShift) - 1);
+
+ // For a surface load of vector size N, the Nth operand will be the surfref
+ if (OpNo == VecSize && MO.isImm()) {
+ lowerImageHandleSymbol(MO.getImm(), MCOp);
+ return true;
+ }
+
+ return false;
+ } else if (MCID.TSFlags & NVPTXII::IsSustFlag) {
+ // This is a surface store, so operand 0 is a surfref
+ if (OpNo == 0 && MO.isImm()) {
+ lowerImageHandleSymbol(MO.getImm(), MCOp);
+ return true;
+ }
+
+ return false;
+ } else if (MCID.TSFlags & NVPTXII::IsSurfTexQueryFlag) {
+ // This is a query, so operand 1 is a surfref/texref
+ if (OpNo == 1 && MO.isImm()) {
+ lowerImageHandleSymbol(MO.getImm(), MCOp);
+ return true;
+ }
+
+ return false;
+ }
+
+ return false;
+}
+
+void NVPTXAsmPrinter::lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp) {
+ // Ewwww
+ TargetMachine &TM = const_cast<TargetMachine&>(MF->getTarget());
+ NVPTXTargetMachine &nvTM = static_cast<NVPTXTargetMachine&>(TM);
+ const NVPTXMachineFunctionInfo *MFI = MF->getInfo<NVPTXMachineFunctionInfo>();
+ const char *Sym = MFI->getImageHandleSymbol(Index);
+ std::string *SymNamePtr =
+ nvTM.getManagedStrPool()->getManagedString(Sym);
+ MCOp = GetSymbolRef(OutContext.GetOrCreateSymbol(
+ StringRef(SymNamePtr->c_str())));
}
void NVPTXAsmPrinter::lowerToMCInst(const MachineInstr *MI, MCInst &OutMI) {
OutMI.setOpcode(MI->getOpcode());
-
// Special: Do not mangle symbol operand of CALL_PROTOTYPE
if (MI->getOpcode() == NVPTX::CALL_PROTOTYPE) {
const MachineOperand &MO = MI->getOperand(0);
- OutMI.addOperand(GetSymbolRef(MO,
+ OutMI.addOperand(GetSymbolRef(
OutContext.GetOrCreateSymbol(Twine(MO.getSymbolName()))));
return;
}
const MachineOperand &MO = MI->getOperand(i);
MCOperand MCOp;
+ if (!nvptxSubtarget->hasImageHandles()) {
+ if (lowerImageHandleOperand(MI, i, MCOp)) {
+ OutMI.addOperand(MCOp);
+ continue;
+ }
+ }
+
if (lowerOperand(MO, MCOp))
OutMI.addOperand(MCOp);
}
MO.getMBB()->getSymbol(), OutContext));
break;
case MachineOperand::MO_ExternalSymbol:
- MCOp = GetSymbolRef(MO, GetExternalSymbolSymbol(MO.getSymbolName()));
+ MCOp = GetSymbolRef(GetExternalSymbolSymbol(MO.getSymbolName()));
break;
case MachineOperand::MO_GlobalAddress:
- MCOp = GetSymbolRef(MO, getSymbol(MO.getGlobal()));
+ MCOp = GetSymbolRef(getSymbol(MO.getGlobal()));
break;
case MachineOperand::MO_FPImmediate: {
const ConstantFP *Cnt = MO.getFPImm();
}
}
-MCOperand NVPTXAsmPrinter::GetSymbolRef(const MachineOperand &MO,
- const MCSymbol *Symbol) {
+MCOperand NVPTXAsmPrinter::GetSymbolRef(const MCSymbol *Symbol) {
const MCExpr *Expr;
Expr = MCSymbolRefExpr::Create(Symbol, MCSymbolRefExpr::VK_None,
OutContext);
void NVPTXAsmPrinter::printReturnValStr(const Function *F, raw_ostream &O) {
const DataLayout *TD = TM.getDataLayout();
- const TargetLowering *TLI = TM.getTargetLowering();
+ const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
Type *Ty = F->getReturnType();
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
if (Ty->getTypeID() == Type::VoidTyID)
return;
O << " (";
if (isABI) {
- if (Ty->isPrimitiveType() || Ty->isIntegerTy()) {
+ if (Ty->isFloatingPointTy() || Ty->isIntegerTy()) {
unsigned size = 0;
if (const IntegerType *ITy = dyn_cast<IntegerType>(Ty)) {
size = ITy->getBitWidth();
} else if (isa<PointerType>(Ty)) {
O << ".param .b" << TLI->getPointerTy().getSizeInBits()
<< " func_retval0";
- } else {
- if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
- SmallVector<EVT, 16> vtparts;
- ComputeValueVTs(*TLI, Ty, vtparts);
- unsigned totalsz = 0;
- for (unsigned i = 0, e = vtparts.size(); i != e; ++i) {
- unsigned elems = 1;
- EVT elemtype = vtparts[i];
- if (vtparts[i].isVector()) {
- elems = vtparts[i].getVectorNumElements();
- elemtype = vtparts[i].getVectorElementType();
- }
- for (unsigned j = 0, je = elems; j != je; ++j) {
- unsigned sz = elemtype.getSizeInBits();
- if (elemtype.isInteger() && (sz < 8))
- sz = 8;
- totalsz += sz / 8;
- }
- }
- unsigned retAlignment = 0;
- if (!llvm::getAlign(*F, 0, retAlignment))
- retAlignment = TD->getABITypeAlignment(Ty);
- O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
- << "]";
- } else
- assert(false && "Unknown return type");
- }
+ } else if ((Ty->getTypeID() == Type::StructTyID) || isa<VectorType>(Ty)) {
+ unsigned totalsz = TD->getTypeAllocSize(Ty);
+ unsigned retAlignment = 0;
+ if (!llvm::getAlign(*F, 0, retAlignment))
+ retAlignment = TD->getABITypeAlignment(Ty);
+ O << ".param .align " << retAlignment << " .b8 func_retval0[" << totalsz
+ << "]";
+ } else
+ llvm_unreachable("Unknown return type");
} else {
SmallVector<EVT, 16> vtparts;
ComputeValueVTs(*TLI, Ty, vtparts);
printReturnValStr(F, O);
}
+// Return true if MBB is the header of a loop marked with
+// llvm.loop.unroll.disable.
+// TODO: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
+bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
+ const MachineBasicBlock &MBB) const {
+ MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
+ // TODO: isLoopHeader() should take "const MachineBasicBlock *".
+ // We insert .pragma "nounroll" only to the loop header.
+ if (!LI.isLoopHeader(const_cast<MachineBasicBlock *>(&MBB)))
+ return false;
+
+ // llvm.loop.unroll.disable is marked on the back edges of a loop. Therefore,
+ // we iterate through each back edge of the loop with header MBB, and check
+ // whether its metadata contains llvm.loop.unroll.disable.
+ for (auto I = MBB.pred_begin(); I != MBB.pred_end(); ++I) {
+ const MachineBasicBlock *PMBB = *I;
+ if (LI.getLoopFor(PMBB) != LI.getLoopFor(&MBB)) {
+ // Edges from other loops to MBB are not back edges.
+ continue;
+ }
+ if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
+ if (MDNode *LoopID = PBB->getTerminator()->getMetadata("llvm.loop")) {
+ if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
+ return true;
+ }
+ }
+ }
+ return false;
+}
+
+void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
+ AsmPrinter::EmitBasicBlockStart(MBB);
+ if (isLoopHeaderOfNoUnroll(MBB))
+ OutStreamer.EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
+}
+
void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
SmallString<128> Str;
raw_svector_ostream O(Str);
void NVPTXAsmPrinter::emitImplicitDef(const MachineInstr *MI) const {
unsigned RegNo = MI->getOperand(0).getReg();
- const TargetRegisterInfo *TRI = TM.getRegisterInfo();
+ const TargetRegisterInfo *TRI = nvptxSubtarget->getRegisterInfo();
if (TRI->isVirtualRegister(RegNo)) {
OutStreamer.AddComment(Twine("implicit-def: ") +
getVirtualRegisterName(RegNo));
} else {
OutStreamer.AddComment(Twine("implicit-def: ") +
- TM.getRegisterInfo()->getName(RegNo));
+ nvptxSubtarget->getRegisterInfo()->getName(RegNo));
}
OutStreamer.AddBlankLine();
}
return true;
}
- for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
- ui != ue; ++ui) {
- const Constant *C = dyn_cast<Constant>(*ui);
- if (usedInGlobalVarDef(C))
- return true;
- }
+ for (const User *U : C->users())
+ if (const Constant *C = dyn_cast<Constant>(U))
+ if (usedInGlobalVarDef(C))
+ return true;
+
return false;
}
return false;
}
- if (const MDNode *md = dyn_cast<MDNode>(U))
- if (md->hasName() && ((md->getName().str() == "llvm.dbg.gv") ||
- (md->getName().str() == "llvm.dbg.sp")))
- return true;
-
- for (User::const_use_iterator ui = U->use_begin(), ue = U->use_end();
- ui != ue; ++ui) {
- if (usedInOneFunc(*ui, oneFunc) == false)
+ for (const User *UU : U->users())
+ if (usedInOneFunc(UU, oneFunc) == false)
return false;
- }
+
return true;
}
if (Pty->getAddressSpace() != llvm::ADDRESS_SPACE_SHARED)
return false;
- const Function *oneFunc = 0;
+ const Function *oneFunc = nullptr;
bool flag = usedInOneFunc(gv, oneFunc);
if (flag == false)
static bool useFuncSeen(const Constant *C,
llvm::DenseMap<const Function *, bool> &seenMap) {
- for (Value::const_use_iterator ui = C->use_begin(), ue = C->use_end();
- ui != ue; ++ui) {
- if (const Constant *cu = dyn_cast<Constant>(*ui)) {
+ for (const User *U : C->users()) {
+ if (const Constant *cu = dyn_cast<Constant>(U)) {
if (useFuncSeen(cu, seenMap))
return true;
- } else if (const Instruction *I = dyn_cast<Instruction>(*ui)) {
+ } else if (const Instruction *I = dyn_cast<Instruction>(U)) {
const BasicBlock *bb = I->getParent();
if (!bb)
continue;
emitDeclaration(F, O);
continue;
}
- for (Value::const_use_iterator iter = F->use_begin(),
- iterEnd = F->use_end();
- iter != iterEnd; ++iter) {
- if (const Constant *C = dyn_cast<Constant>(*iter)) {
+ for (const User *U : F->users()) {
+ if (const Constant *C = dyn_cast<Constant>(U)) {
if (usedInGlobalVarDef(C)) {
// The use is in the initialization of a global variable
// that is a function pointer, so print a declaration
}
}
- if (!isa<Instruction>(*iter))
+ if (!isa<Instruction>(U))
continue;
- const Instruction *instr = cast<Instruction>(*iter);
+ const Instruction *instr = cast<Instruction>(U);
const BasicBlock *bb = instr->getParent();
if (!bb)
continue;
DbgFinder.processModule(M);
unsigned i = 1;
- for (DebugInfoFinder::iterator I = DbgFinder.compile_unit_begin(),
- E = DbgFinder.compile_unit_end();
- I != E; ++I) {
- DICompileUnit DIUnit(*I);
+ for (DICompileUnit DIUnit : DbgFinder.compile_units()) {
StringRef Filename(DIUnit.getFilename());
StringRef Dirname(DIUnit.getDirectory());
SmallString<128> FullPathName = Dirname;
++i;
}
- for (DebugInfoFinder::iterator I = DbgFinder.subprogram_begin(),
- E = DbgFinder.subprogram_end();
- I != E; ++I) {
- DISubprogram SP(*I);
+ for (DISubprogram SP : DbgFinder.subprograms()) {
StringRef Filename(SP.getFilename());
StringRef Dirname(SP.getDirectory());
SmallString<128> FullPathName = Dirname;
}
bool NVPTXAsmPrinter::doInitialization(Module &M) {
+ // Construct a default subtarget off of the TargetMachine defaults. The
+ // rest of NVPTX isn't friendly to change subtargets per function and
+ // so the default TargetMachine will have all of the options.
+ StringRef TT = TM.getTargetTriple();
+ StringRef CPU = TM.getTargetCPU();
+ StringRef FS = TM.getTargetFeatureString();
+ const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
+ const NVPTXSubtarget STI(TT, CPU, FS, NTM);
SmallString<128> Str1;
raw_svector_ostream OS1(Str1);
const_cast<TargetLoweringObjectFile &>(getObjFileLowering())
.Initialize(OutContext, TM);
- Mang = new Mangler(&TM);
+ Mang = new Mangler(TM.getDataLayout());
// Emit header before any dwarf directives are emitted below.
- emitHeader(M, OS1);
+ emitHeader(M, OS1, STI);
OutStreamer.EmitRawText(OS1.str());
// Already commented out
OutStreamer.AddBlankLine();
}
- if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)
+ // If we're not NVCL we're CUDA, go ahead and emit filenames.
+ if (Triple(TM.getTargetTriple()).getOS() != Triple::NVCL)
recordAndEmitFilenames(M);
GlobalsEmitted = false;
OutStreamer.EmitRawText(OS2.str());
}
-void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O) {
+void NVPTXAsmPrinter::emitHeader(Module &M, raw_ostream &O,
+ const NVPTXSubtarget &STI) {
O << "//\n";
O << "// Generated by LLVM NVPTX Back-End\n";
O << "//\n";
O << "\n";
- unsigned PTXVersion = nvptxSubtarget.getPTXVersion();
+ unsigned PTXVersion = STI.getPTXVersion();
O << ".version " << (PTXVersion / 10) << "." << (PTXVersion % 10) << "\n";
O << ".target ";
- O << nvptxSubtarget.getTargetName();
+ O << STI.getTargetName();
- if (nvptxSubtarget.getDrvInterface() == NVPTX::NVCL)
+ const NVPTXTargetMachine &NTM = static_cast<const NVPTXTargetMachine &>(TM);
+ if (NTM.getDrvInterface() == NVPTX::NVCL)
O << ", texmode_independent";
- if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
- if (!nvptxSubtarget.hasDouble())
+ else {
+ if (!STI.hasDouble())
O << ", map_f64_to_f32";
}
O << "\n";
O << ".address_size ";
- if (nvptxSubtarget.is64Bit())
+ if (NTM.is64Bit())
O << "64";
else
O << "32";
}
bool NVPTXAsmPrinter::doFinalization(Module &M) {
-
// If we did not emit any functions, then the global declarations have not
// yet been emitted.
if (!GlobalsEmitted) {
for (i = 0; i < n; i++)
global_list.insert(global_list.end(), gv_array[i]);
+ clearAnnotationCache(&M);
+
delete[] gv_array;
return ret;
// external global variable with init -> .visible
// external without init -> .extern
// appending -> not allowed, assert.
+// for any linkage other than
+// internal, private, linker_private,
+// linker_private_weak, linker_private_weak_def_auto,
+// we emit -> .weak.
void NVPTXAsmPrinter::emitLinkageDirective(const GlobalValue *V,
raw_ostream &O) {
- if (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA) {
+ if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() == NVPTX::CUDA) {
if (V->hasExternalLinkage()) {
if (isa<GlobalVariable>(V)) {
const GlobalVariable *GVar = cast<GlobalVariable>(V);
msg.append(V->getName().str());
msg.append("has unsupported appending linkage type");
llvm_unreachable(msg.c_str());
+ } else if (!V->hasInternalLinkage() &&
+ !V->hasPrivateLinkage()) {
+ O << ".weak ";
}
}
}
// Skip meta data
if (GVar->hasSection()) {
- if (GVar->getSection() == "llvm.metadata")
+ if (GVar->getSection() == StringRef("llvm.metadata"))
return;
}
+ // Skip LLVM intrinsic global variables
+ if (GVar->getName().startswith("llvm.") ||
+ GVar->getName().startswith("nvvm."))
+ return;
+
const DataLayout *TD = TM.getDataLayout();
// GlobalVariables are always constant pointers themselves.
O << ".visible ";
else
O << ".extern ";
+ } else if (GVar->hasLinkOnceLinkage() || GVar->hasWeakLinkage() ||
+ GVar->hasAvailableExternallyLinkage() ||
+ GVar->hasCommonLinkage()) {
+ O << ".weak ";
}
if (llvm::isTexture(*GVar)) {
if (llvm::isSampler(*GVar)) {
O << ".global .samplerref " << llvm::getSamplerName(*GVar);
- const Constant *Initializer = NULL;
+ const Constant *Initializer = nullptr;
if (GVar->hasInitializer())
Initializer = GVar->getInitializer();
- const ConstantInt *CI = NULL;
+ const ConstantInt *CI = nullptr;
if (Initializer)
CI = dyn_cast<ConstantInt>(Initializer);
if (CI) {
O << "linear";
break;
case 2:
- assert(0 && "Anisotropic filtering is not supported");
+ llvm_unreachable("Anisotropic filtering is not supported");
default:
O << "nearest";
break;
return;
}
- const Function *demotedFunc = 0;
+ const Function *demotedFunc = nullptr;
if (!processDemoted && canDemoteGlobalVar(GVar, demotedFunc)) {
O << "// " << GVar->getName().str() << " has been demoted\n";
if (localDecls.find(demotedFunc) != localDecls.end())
O << ".";
emitPTXAddressSpace(PTy->getAddressSpace(), O);
+
+ if (isManaged(*GVar)) {
+ O << " .attribute(.managed)";
+ }
+
if (GVar->getAlignment() == 0)
O << " .align " << (int) TD->getPrefTypeAlignment(ETy);
else
O << " .align " << GVar->getAlignment();
- if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
+ if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
O << " .";
// Special case: ABI requires that we use .u8 for predicates
if (ETy->isIntegerTy(1))
// Ptx allows variable initilization only for constant and global state
// spaces.
- if (((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
- (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) &&
- GVar->hasInitializer()) {
- const Constant *Initializer = GVar->getInitializer();
- if (!Initializer->isNullValue()) {
- O << " = ";
- printScalarConstant(Initializer, O);
+ if (GVar->hasInitializer()) {
+ if ((PTy->getAddressSpace() == llvm::ADDRESS_SPACE_GLOBAL) ||
+ (PTy->getAddressSpace() == llvm::ADDRESS_SPACE_CONST)) {
+ const Constant *Initializer = GVar->getInitializer();
+ // 'undef' is treated as there is no value spefied.
+ if (!Initializer->isNullValue() && !isa<UndefValue>(Initializer)) {
+ O << " = ";
+ printScalarConstant(Initializer, O);
+ }
+ } else {
+ // The frontend adds zero-initializer to variables that don't have an
+ // initial value, so skip warning for this case.
+ if (!GVar->getInitializer()->isNullValue()) {
+ std::string warnMsg = "initial value of '" + GVar->getName().str() +
+ "' is not allowed in addrspace(" +
+ llvm::utostr_32(PTy->getAddressSpace()) + ")";
+ report_fatal_error(warnMsg.c_str());
+ }
}
}
} else {
AggBuffer aggBuffer(ElementSize, O, *this);
bufferAggregateConstant(Initializer, &aggBuffer);
if (aggBuffer.numSymbols) {
- if (nvptxSubtarget.is64Bit()) {
+ if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit()) {
O << " .u64 " << *getSymbol(GVar) << "[";
O << ElementSize / 8;
} else {
}
break;
default:
- assert(0 && "type not supported yet");
+ llvm_unreachable("type not supported yet");
}
}
case Type::DoubleTyID:
return "f64";
case Type::PointerTyID:
- if (nvptxSubtarget.is64Bit())
+ if (static_cast<const NVPTXTargetMachine &>(TM).is64Bit())
if (useB4PTR)
return "b64";
else
return "u32";
}
llvm_unreachable("unexpected type");
- return NULL;
+ return nullptr;
}
void NVPTXAsmPrinter::emitPTXGlobalVariable(const GlobalVariable *GVar,
else
O << " .align " << GVar->getAlignment();
- if (ETy->isPrimitiveType() || ETy->isIntegerTy() || isa<PointerType>(ETy)) {
+ if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
O << " .";
O << getPTXFundamentalTypeStr(ETy);
O << " ";
O << "]";
break;
default:
- assert(0 && "type not supported yet");
+ llvm_unreachable("type not supported yet");
}
return;
}
static unsigned int getOpenCLAlignment(const DataLayout *TD, Type *Ty) {
- if (Ty->isPrimitiveType() || Ty->isIntegerTy() || isa<PointerType>(Ty))
+ if (Ty->isSingleValueType())
return TD->getPrefTypeAlignment(Ty);
const ArrayType *ATy = dyn_cast<ArrayType>(Ty);
if (ATy)
return getOpenCLAlignment(TD, ATy->getElementType());
- const VectorType *VTy = dyn_cast<VectorType>(Ty);
- if (VTy) {
- Type *ETy = VTy->getElementType();
- unsigned int numE = VTy->getNumElements();
- unsigned int alignE = TD->getPrefTypeAlignment(ETy);
- if (numE == 3)
- return 4 * alignE;
- else
- return numE * alignE;
- }
-
const StructType *STy = dyn_cast<StructType>(Ty);
if (STy) {
unsigned int alignStruct = 1;
void NVPTXAsmPrinter::printParamName(Function::const_arg_iterator I,
int paramIndex, raw_ostream &O) {
- if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
- (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA))
- O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
- else {
- std::string argName = I->getName();
- const char *p = argName.c_str();
- while (*p) {
- if (*p == '.')
- O << "_";
- else
- O << *p;
- p++;
- }
- }
+ O << *getSymbol(I->getParent()) << "_param_" << paramIndex;
}
void NVPTXAsmPrinter::printParamName(int paramIndex, raw_ostream &O) {
- Function::const_arg_iterator I, E;
- int i = 0;
-
- if ((nvptxSubtarget.getDrvInterface() == NVPTX::NVCL) ||
- (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA)) {
- O << *CurrentFnSym << "_param_" << paramIndex;
- return;
- }
-
- for (I = F->arg_begin(), E = F->arg_end(); I != E; ++I, i++) {
- if (i == paramIndex) {
- printParamName(I, paramIndex, O);
- return;
- }
- }
- llvm_unreachable("paramIndex out of bound");
+ O << *CurrentFnSym << "_param_" << paramIndex;
}
void NVPTXAsmPrinter::emitFunctionParamList(const Function *F, raw_ostream &O) {
const DataLayout *TD = TM.getDataLayout();
const AttributeSet &PAL = F->getAttributes();
- const TargetLowering *TLI = TM.getTargetLowering();
+ const TargetLowering *TLI = nvptxSubtarget->getTargetLowering();
Function::const_arg_iterator I, E;
unsigned paramIndex = 0;
bool first = true;
bool isKernelFunc = llvm::isKernelFunction(*F);
- bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
+ bool isABI = (nvptxSubtarget->getSmVersion() >= 20);
MVT thePointerTy = TLI->getPointerTy();
O << "(\n";
first = false;
// Handle image/sampler parameters
- if (llvm::isSampler(*I) || llvm::isImage(*I)) {
- if (llvm::isImage(*I)) {
- std::string sname = I->getName();
- if (llvm::isImageWriteOnly(*I))
- O << "\t.param .surfref " << *getSymbol(F) << "_param_"
- << paramIndex;
- else // Default image is read_only
- O << "\t.param .texref " << *getSymbol(F) << "_param_"
- << paramIndex;
- } else // Should be llvm::isSampler(*I)
- O << "\t.param .samplerref " << *getSymbol(F) << "_param_"
- << paramIndex;
- continue;
+ if (isKernelFunction(*F)) {
+ if (isSampler(*I) || isImage(*I)) {
+ if (isImage(*I)) {
+ std::string sname = I->getName();
+ if (isImageWriteOnly(*I) || isImageReadWrite(*I)) {
+ if (nvptxSubtarget->hasImageHandles())
+ O << "\t.param .u64 .ptr .surfref ";
+ else
+ O << "\t.param .surfref ";
+ O << *CurrentFnSym << "_param_" << paramIndex;
+ }
+ else { // Default image is read_only
+ if (nvptxSubtarget->hasImageHandles())
+ O << "\t.param .u64 .ptr .texref ";
+ else
+ O << "\t.param .texref ";
+ O << *CurrentFnSym << "_param_" << paramIndex;
+ }
+ } else {
+ if (nvptxSubtarget->hasImageHandles())
+ O << "\t.param .u64 .ptr .samplerref ";
+ else
+ O << "\t.param .samplerref ";
+ O << *CurrentFnSym << "_param_" << paramIndex;
+ }
+ continue;
+ }
}
if (PAL.hasAttribute(paramIndex + 1, Attribute::ByVal) == false) {
- if (Ty->isVectorTy()) {
- // Just print .param .b8 .align <a> .param[size];
+ if (Ty->isAggregateType() || Ty->isVectorTy()) {
+ // Just print .param .align <a> .b8 .param[size];
// <a> = PAL.getparamalignment
// size = typeallocsize of element type
unsigned align = PAL.getParamAlignment(paramIndex + 1);
// Special handling for pointer arguments to kernel
O << "\t.param .u" << thePointerTy.getSizeInBits() << " ";
- if (nvptxSubtarget.getDrvInterface() != NVPTX::CUDA) {
+ if (static_cast<NVPTXTargetMachine &>(TM).getDrvInterface() !=
+ NVPTX::CUDA) {
Type *ETy = PTy->getElementType();
int addrSpace = PTy->getAddressSpace();
switch (addrSpace) {
continue;
}
// Non-kernel function, just print .param .b<size> for ABI
- // and .reg .b<size> for non ABY
+ // and .reg .b<size> for non-ABI
unsigned sz = 0;
if (isa<IntegerType>(Ty)) {
sz = cast<IntegerType>(Ty)->getBitWidth();
Type *ETy = PTy->getElementType();
if (isABI || isKernelFunc) {
- // Just print .param .b8 .align <a> .param[size];
+ // Just print .param .align <a> .b8 .param[size];
// <a> = PAL.getparamalignment
// size = typeallocsize of element type
unsigned align = PAL.getParamAlignment(paramIndex + 1);
// Map the global virtual register number to a register class specific
// virtual register number starting from 1 with that class.
- const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
//unsigned numRegClasses = TRI->getNumRegClasses();
// Emit the Fake Stack Object
if (NumBytes) {
O << "\t.local .align " << MFI->getMaxAlignment() << " .b8 \t" << DEPOTNAME
<< getFunctionNumber() << "[" << NumBytes << "];\n";
- if (nvptxSubtarget.is64Bit()) {
+ if (static_cast<const NVPTXTargetMachine &>(MF.getTarget()).is64Bit()) {
O << "\t.reg .b64 \t%SP;\n";
O << "\t.reg .b64 \t%SPL;\n";
} else {
// O << "\t.reg .s16 %rc<" << NVPTXNumRegisters << ">;\n";
// O << "\t.reg .s16 %rs<" << NVPTXNumRegisters << ">;\n";
// O << "\t.reg .s32 %r<" << NVPTXNumRegisters << ">;\n";
- // O << "\t.reg .s64 %rl<" << NVPTXNumRegisters << ">;\n";
+ // O << "\t.reg .s64 %rd<" << NVPTXNumRegisters << ">;\n";
// O << "\t.reg .f32 %f<" << NVPTXNumRegisters << ">;\n";
- // O << "\t.reg .f64 %fl<" << NVPTXNumRegisters << ">;\n";
+ // O << "\t.reg .f64 %fd<" << NVPTXNumRegisters << ">;\n";
// Emit declaration of the virtual registers or 'physical' registers for
// each register class
return;
}
if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {
- O << *getSymbol(GVar);
+ PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
+ bool IsNonGenericPointer = false;
+ if (PTy && PTy->getAddressSpace() != 0) {
+ IsNonGenericPointer = true;
+ }
+ if (EmitGeneric && !isa<Function>(CPV) && !IsNonGenericPointer) {
+ O << "generic(";
+ O << *getSymbol(GVar);
+ O << ")";
+ } else {
+ O << *getSymbol(GVar);
+ }
return;
}
if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
const Value *v = Cexpr->stripPointerCasts();
+ PointerType *PTy = dyn_cast<PointerType>(Cexpr->getType());
+ bool IsNonGenericPointer = false;
+ if (PTy && PTy->getAddressSpace() != 0) {
+ IsNonGenericPointer = true;
+ }
if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
- O << *getSymbol(GVar);
+ if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
+ O << "generic(";
+ O << *getSymbol(GVar);
+ O << ")";
+ } else {
+ O << *getSymbol(GVar);
+ }
return;
} else {
- O << *LowerConstant(CPV, *this);
+ O << *lowerConstant(CPV);
return;
}
}
}
LineReader *NVPTXAsmPrinter::getReader(std::string filename) {
- if (reader == NULL) {
+ if (!reader) {
reader = new LineReader(filename);
}
projects / oota-llvm.git / blobdiff