[NVPTX] Support global variables of integer type larger than i64.

[majnemer]

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

2002

Do we want the store size or the allocation size?

[jlebar]

Use AllocSize, not StoreSize.

[jlebar]

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

2002

Ah, indeed.

[majnemer]

LGTM

[jlebar]

Thank you for the review!

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

if (isManaged(*GVar)) {
     O << " .attribute(.managed)";
   }
 
   if (GVar->getAlignment() == 0)
     O << " .align " << (int)DL.getPrefTypeAlignment(ETy);
   else
     O << " .align " << GVar->getAlignment();
 
-  if (ETy->isFloatingPointTy() || ETy->isIntegerTy() || ETy->isPointerTy()) {
+  if (ETy->isFloatingPointTy() || ETy->isPointerTy() ||
+      (ETy->isIntegerTy() && ETy->getScalarSizeInBits() <= 64)) {
     O << " .";
     // Special case: ABI requires that we use .u8 for predicates
     if (ETy->isIntegerTy(1))
       O << "u8";
     else
       O << getPTXFundamentalTypeStr(ETy, false);
     O << " ";
     getSymbol(GVar)->print(O, MAI);
void NVPTXAsmPrinter::printModuleLevelGV(const GlobalVariable *GVar,
   } else {
     unsigned int ElementSize = 0;
 
     // Although PTX has direct support for struct type and array type and
     // LLVM IR is very similar to PTX, the LLVM CodeGen does not support for
     // targets that support these high level field accesses. Structs, arrays
     // and vectors are lowered into arrays of bytes.
     switch (ETy->getTypeID()) {
+    case Type::IntegerTyID: // Integers larger than 64 bits
     case Type::StructTyID:
     case Type::ArrayTyID:
     case Type::VectorTyID:
       ElementSize = DL.getTypeStoreSize(ETy);
       // Ptx allows variable initilization only for constant and
       // global state spaces.
       if (((PTy->getAddressSpace() == ADDRESS_SPACE_GLOBAL) ||
            (PTy->getAddressSpace() == ADDRESS_SPACE_CONST)) &&
void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
   }
 }
 
 void NVPTXAsmPrinter::bufferAggregateConstant(const Constant *CPV,
                                               AggBuffer *aggBuffer) {
   const DataLayout &DL = getDataLayout();
   int Bytes;
 
+  // Integers of arbitrary width
+  if (const ConstantInt *CI = dyn_cast<ConstantInt>(CPV)) {
+    APInt Val = CI->getValue();
+    for (unsigned I = 0, E = DL.getTypeStoreSize(CPV->getType()); I < E; ++I) {
+      uint8_t Byte = Val.getLoBits(8).getZExtValue();
+      aggBuffer->addBytes(&Byte, 1, 1);
+      Val = Val.lshr(8);
+    }
+    return;
+  }
+
   // Old constants
   if (isa<ConstantArray>(CPV) || isa<ConstantVector>(CPV)) {
     if (CPV->getNumOperands())
       for (unsigned i = 0, e = CPV->getNumOperands(); i != e; ++i)
         bufferLEByte(cast<Constant>(CPV->getOperand(i)), 0, aggBuffer);
     return;
   }
 

llvm/test/CodeGen/NVPTX/global-variable-big.ll

+; RUN: llc < %s | FileCheck %s
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-nvidia-cuda"
+
+; Check that we can handle global variables of large integer type.
+
+; (lsb) 0x0102'0304'0506...0F10 (msb)
+@gv = addrspace(1) externally_initialized global i128 21345817372864405881847059188222722561, align 16
+; CHECK: .visible .global .align 16 .b8 gv[16] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16};