Fix endianess issue in global variable constant initialization in NVPTX.

[sfantao]

Summary

The constant initialization for globals in NVPTX is generated as an array of bytes. The generation of this byte arrays expects the host to be little endian, which prevents big endian hosts to be used in the generation of the PTX code.

This patch fixes the problem by changing the way the bytes are extracted so that it works for either little and big endian.

I identified this issue while crosscompiling for PTX in power7 machines.

[sfantao]

Ping!

[sfantao]

Any comments on this patch?

Thanks!
Samuel

[jholewinski]

LGTM.

Sorry for the delay. I was out of town when this was posted and completely missed it.

[sfantao]

Thanks for the review! Committed in r239412.

Samuel

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
     } else {
       O << *lowerConstant(CPV);
       return;
     }
   }
   llvm_unreachable("Not scalar type found in printScalarConstant()");
 }
 
+// These utility functions assure we get the right sequence of bytes for a given
+// type even for big-endian machines
+template <typename T> static void ConvertIntToBytes(unsigned char *p, T val) {
+  int64_t vp = (int64_t)val;
+  for (unsigned i = 0; i < sizeof(T); ++i) {
+    p[i] = (unsigned char)vp;
+    vp >>= 8;
+  }
+}
+static void ConvertFloatToBytes(unsigned char *p, float val) {
+  int32_t *vp = (int32_t *)&val;
+  for (unsigned i = 0; i < sizeof(int32_t); ++i) {
+    p[i] = (unsigned char)*vp;
+    *vp >>= 8;
+  }
+}
+static void ConvertDoubleToBytes(unsigned char *p, double val) {
+  int64_t *vp = (int64_t *)&val;
+  for (unsigned i = 0; i < sizeof(int64_t); ++i) {
+    p[i] = (unsigned char)*vp;
+    *vp >>= 8;
+  }
+}
+
 void NVPTXAsmPrinter::bufferLEByte(const Constant *CPV, int Bytes,
                                    AggBuffer *aggBuffer) {
 
   const DataLayout *TD = TM.getDataLayout();
 
   if (isa<UndefValue>(CPV) || CPV->isNullValue()) {
     int s = TD->getTypeAllocSize(CPV->getType());
     if (s < Bytes)
       s = Bytes;
     aggBuffer->addZeros(s);
     return;
   }
 
-  unsigned char *ptr;
+  unsigned char ptr[8];
   switch (CPV->getType()->getTypeID()) {
 
   case Type::IntegerTyID: {
     const Type *ETy = CPV->getType();
     if (ETy == Type::getInt8Ty(CPV->getContext())) {
       unsigned char c = (unsigned char)cast<ConstantInt>(CPV)->getZExtValue();
-      ptr = &c;
+      ConvertIntToBytes<>(ptr, c);
       aggBuffer->addBytes(ptr, 1, Bytes);
     } else if (ETy == Type::getInt16Ty(CPV->getContext())) {
       short int16 = (short)cast<ConstantInt>(CPV)->getZExtValue();
-      ptr = (unsigned char *)&int16;
+      ConvertIntToBytes<>(ptr, int16);
       aggBuffer->addBytes(ptr, 2, Bytes);
     } else if (ETy == Type::getInt32Ty(CPV->getContext())) {
       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
         int int32 = (int)(constInt->getZExtValue());
-        ptr = (unsigned char *)&int32;
+        ConvertIntToBytes<>(ptr, int32);
         aggBuffer->addBytes(ptr, 4, Bytes);
         break;
       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
                 ConstantFoldConstantExpression(Cexpr, *TD))) {
           int int32 = (int)(constInt->getZExtValue());
-          ptr = (unsigned char *)&int32;
+          ConvertIntToBytes<>(ptr, int32);
           aggBuffer->addBytes(ptr, 4, Bytes);
           break;
         }
         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
           aggBuffer->addZeros(4);
           break;
         }
       }
       llvm_unreachable("unsupported integer const type");
     } else if (ETy == Type::getInt64Ty(CPV->getContext())) {
       if (const ConstantInt *constInt = dyn_cast<ConstantInt>(CPV)) {
         long long int64 = (long long)(constInt->getZExtValue());
-        ptr = (unsigned char *)&int64;
+        ConvertIntToBytes<>(ptr, int64);
         aggBuffer->addBytes(ptr, 8, Bytes);
         break;
       } else if (const ConstantExpr *Cexpr = dyn_cast<ConstantExpr>(CPV)) {
         if (const ConstantInt *constInt = dyn_cast<ConstantInt>(
                 ConstantFoldConstantExpression(Cexpr, *TD))) {
           long long int64 = (long long)(constInt->getZExtValue());
-          ptr = (unsigned char *)&int64;
+          ConvertIntToBytes<>(ptr, int64);
           aggBuffer->addBytes(ptr, 8, Bytes);
           break;
         }
         if (Cexpr->getOpcode() == Instruction::PtrToInt) {
           Value *v = Cexpr->getOperand(0)->stripPointerCasts();
           aggBuffer->addSymbol(v, Cexpr->getOperand(0));
           aggBuffer->addZeros(8);
           break;
         }
       }
       llvm_unreachable("unsupported integer const type");
     } else
       llvm_unreachable("unsupported integer const type");
     break;
   }
   case Type::FloatTyID:
   case Type::DoubleTyID: {
     const ConstantFP *CFP = dyn_cast<ConstantFP>(CPV);
     const Type *Ty = CFP->getType();
     if (Ty == Type::getFloatTy(CPV->getContext())) {
       float float32 = (float) CFP->getValueAPF().convertToFloat();
-      ptr = (unsigned char *)&float32;
+      ConvertFloatToBytes(ptr, float32);
       aggBuffer->addBytes(ptr, 4, Bytes);
     } else if (Ty == Type::getDoubleTy(CPV->getContext())) {
       double float64 = CFP->getValueAPF().convertToDouble();
-      ptr = (unsigned char *)&float64;
+      ConvertDoubleToBytes(ptr, float64);
       aggBuffer->addBytes(ptr, 8, Bytes);
     } else {
       llvm_unreachable("unsupported fp const type");
     }
     break;
   }
   case Type::PointerTyID: {
     if (const GlobalValue *GVar = dyn_cast<GlobalValue>(CPV)) {

test/CodeGen/NVPTX/globals_init.ll

+; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
+
+; Make sure the globals constant initializers are not prone to host endianess 
+; issues.
+
+; CHECK-DAG: .b8 Gbli08[2] = {171, 205};
+@Gbli08 = global [2 x i8] [i8 171, i8 205]
+
+; CHECK-DAG: .b8 Gbli16[4] = {205, 171, 1, 239};
+@Gbli16 = global [2 x i16] [i16 43981, i16 61185]
+
+; CHECK-DAG: .b8 Gbli32[8] = {1, 239, 205, 171, 137, 103, 69, 35};
+@Gbli32 = global [2 x i32] [i32 2882400001, i32 591751049]
+
+; CHECK-DAG: .b8 Gbli64[16] = {137, 103, 69, 35, 1, 239, 205, 171, 239, 205, 171, 137, 103, 69, 35, 1};
+@Gbli64 = global [2 x i64] [i64 12379813738877118345, i64 81985529216486895]
+
+; CHECK-DAG: .b8 Gblf32[8] = {192, 225, 100, 75, 0, 96, 106, 69};
+@Gblf32 = global [2 x float] [float 1.5e+7, float 3.75e+3]
+
+; CHECK-DAG: .b8 Gblf64[16] = {116, 10, 181, 48, 134, 62, 230, 58, 106, 222, 138, 98, 204, 250, 200, 75};
+@Gblf64 = global [2 x double] [double 5.75e-25, double 12.25e+56]
+