raw_string_ostream OS(S);
OS << "Unsupported expression in static initializer: ";
CE->printAsOperand(OS, /*PrintType=*/ false,
- !AP.MF ? 0 : AP.MF->getFunction()->getParent());
+ !AP.MF ? 0 : AP.MF->getFunction()->getParent());
report_fatal_error(OS.str());
}
+ case Instruction::AddrSpaceCast: {
+ // Strip any addrspace(1)->addrspace(0) addrspace casts. These will be
+ // handled by the generic() logic in the MCExpr printer
+ PointerType *DstTy = cast<PointerType>(CE->getType());
+ PointerType *SrcTy = cast<PointerType>(CE->getOperand(0)->getType());
+ if (SrcTy->getAddressSpace() == 1 && DstTy->getAddressSpace() == 0) {
+ return LowerConstant(cast<const Constant>(CE->getOperand(0)), AP);
+ }
+ std::string S;
+ raw_string_ostream OS(S);
+ OS << "Unsupported expression in static initializer: ";
+ CE->printAsOperand(OS, /*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
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);
unsigned curpos;
raw_ostream &O;
NVPTXAsmPrinter &AP;
+ bool EmitGeneric;
public:
AggBuffer(unsigned _size, raw_ostream &_O, NVPTXAsmPrinter &_AP)
size = _size;
curpos = 0;
numSymbols = 0;
+ EmitGeneric = AP.EmitGeneric;
}
~AggBuffer() { delete[] buffer; }
unsigned addBytes(unsigned char *Ptr, int Num, int Bytes) {
const Value *v = Symbols[nSym];
if (const GlobalValue *GVar = dyn_cast<GlobalValue>(v)) {
MCSymbol *Name = AP.getSymbol(GVar);
- O << *Name;
+ PointerType *PTy = dyn_cast<PointerType>(GVar->getType());
+ bool IsNonGenericPointer = false;
+ if (PTy && PTy->getAddressSpace() != 0) {
+ IsNonGenericPointer = true;
+ }
+ if (EmitGeneric && !isa<Function>(v) && !IsNonGenericPointer) {
+ O << "generic(";
+ O << *Name;
+ O << ")";
+ } else {
+ O << *Name;
+ }
} else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(v)) {
O << *nvptx::LowerConstant(Cexpr, AP);
} else
LineReader *reader;
LineReader *getReader(std::string);
+
+ // Used to control the need to emit .generic() in the initializer of
+ // module scope variables.
+ // Although ptx supports the hybrid mode like the following,
+ // .global .u32 a;
+ // .global .u32 b;
+ // .global .u32 addr[] = {a, generic(b)}
+ // we have difficulty representing the difference in the NVVM IR.
+ //
+ // Since the address value should always be generic in CUDA C and always
+ // be specific in OpenCL, we use this simple control here.
+ //
+ bool EmitGeneric;
+
public:
NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
: AsmPrinter(TM, Streamer),
nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
CurrentBankselLabelInBasicBlock = "";
reader = NULL;
+ EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
}
~NVPTXAsmPrinter() {
--- /dev/null
+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+; CHECK: .visible .global .align 4 .u32 g = 42;
+; CHECK: .visible .global .align 4 .u32 g2 = generic(g);
+; CHECK: .visible .global .align 4 .u32 g3 = g;
+
+@g = addrspace(1) global i32 42
+@g2 = addrspace(1) global i32* addrspacecast (i32 addrspace(1)* @g to i32*)
+@g3 = addrspace(1) global i32 addrspace(1)* @g