+}
+
+/// LowerBSWAP - Emit the code to lower bswap of V before the specified
+/// instruction IP.
+static Value *LowerBSWAP(Value *V, Instruction *IP) {
+ assert(V->getType()->isInteger() && "Can't bswap a non-integer type!");
+
+ unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
+
+ switch(BitSize) {
+ default: assert(0 && "Unhandled type size of value to byteswap!");
+ case 16: {
+ Value *Tmp1 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),8),"bswap.2",IP);
+ Value *Tmp2 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),8),"bswap.1",IP);
+ V = BinaryOperator::createOr(Tmp1, Tmp2, "bswap.i16", IP);
+ break;
+ }
+ case 32: {
+ Value *Tmp4 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),24),"bswap.4", IP);
+ Value *Tmp3 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),8),"bswap.3",IP);
+ Value *Tmp2 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),8),"bswap.2",IP);
+ Value *Tmp1 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),24),"bswap.1", IP);
+ Tmp3 = BinaryOperator::createAnd(Tmp3,
+ ConstantInt::get(Type::Int32Ty, 0xFF0000),
+ "bswap.and3", IP);
+ Tmp2 = BinaryOperator::createAnd(Tmp2,
+ ConstantInt::get(Type::Int32Ty, 0xFF00),
+ "bswap.and2", IP);
+ Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or1", IP);
+ Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or2", IP);
+ V = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.i32", IP);
+ break;
+ }
+ case 64: {
+ Value *Tmp8 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),56),"bswap.8", IP);
+ Value *Tmp7 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),40),"bswap.7", IP);
+ Value *Tmp6 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),24),"bswap.6", IP);
+ Value *Tmp5 = BinaryOperator::createShl(V,
+ ConstantInt::get(V->getType(),8),"bswap.5", IP);
+ Value* Tmp4 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),8),"bswap.4", IP);
+ Value* Tmp3 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),24),"bswap.3", IP);
+ Value* Tmp2 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),40),"bswap.2", IP);
+ Value* Tmp1 = BinaryOperator::createLShr(V,
+ ConstantInt::get(V->getType(),56),"bswap.1", IP);
+ Tmp7 = BinaryOperator::createAnd(Tmp7,
+ ConstantInt::get(Type::Int64Ty,
+ 0xFF000000000000ULL),
+ "bswap.and7", IP);
+ Tmp6 = BinaryOperator::createAnd(Tmp6,
+ ConstantInt::get(Type::Int64Ty, 0xFF0000000000ULL),
+ "bswap.and6", IP);
+ Tmp5 = BinaryOperator::createAnd(Tmp5,
+ ConstantInt::get(Type::Int64Ty, 0xFF00000000ULL),
+ "bswap.and5", IP);
+ Tmp4 = BinaryOperator::createAnd(Tmp4,
+ ConstantInt::get(Type::Int64Ty, 0xFF000000ULL),
+ "bswap.and4", IP);
+ Tmp3 = BinaryOperator::createAnd(Tmp3,
+ ConstantInt::get(Type::Int64Ty, 0xFF0000ULL),
+ "bswap.and3", IP);
+ Tmp2 = BinaryOperator::createAnd(Tmp2,
+ ConstantInt::get(Type::Int64Ty, 0xFF00ULL),
+ "bswap.and2", IP);
+ Tmp8 = BinaryOperator::createOr(Tmp8, Tmp7, "bswap.or1", IP);
+ Tmp6 = BinaryOperator::createOr(Tmp6, Tmp5, "bswap.or2", IP);
+ Tmp4 = BinaryOperator::createOr(Tmp4, Tmp3, "bswap.or3", IP);
+ Tmp2 = BinaryOperator::createOr(Tmp2, Tmp1, "bswap.or4", IP);
+ Tmp8 = BinaryOperator::createOr(Tmp8, Tmp6, "bswap.or5", IP);
+ Tmp4 = BinaryOperator::createOr(Tmp4, Tmp2, "bswap.or6", IP);
+ V = BinaryOperator::createOr(Tmp8, Tmp4, "bswap.i64", IP);
+ break;
+ }
+ }
+ return V;
+}
+
+/// LowerCTPOP - Emit the code to lower ctpop of V before the specified
+/// instruction IP.
+static Value *LowerCTPOP(Value *V, Instruction *IP) {
+ assert(V->getType()->isInteger() && "Can't ctpop a non-integer type!");
+
+ static const uint64_t MaskValues[6] = {
+ 0x5555555555555555ULL, 0x3333333333333333ULL,
+ 0x0F0F0F0F0F0F0F0FULL, 0x00FF00FF00FF00FFULL,
+ 0x0000FFFF0000FFFFULL, 0x00000000FFFFFFFFULL
+ };
+
+ unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
+ unsigned WordSize = (BitSize + 63) / 64;
+ Value *Count = ConstantInt::get(V->getType(), 0);
+
+ for (unsigned n = 0; n < WordSize; ++n) {
+ Value *PartValue = V;
+ for (unsigned i = 1, ct = 0; i < (BitSize>64 ? 64 : BitSize);
+ i <<= 1, ++ct) {
+ Value *MaskCst = ConstantInt::get(V->getType(), MaskValues[ct]);
+ Value *LHS = BinaryOperator::createAnd(
+ PartValue, MaskCst, "cppop.and1", IP);
+ Value *VShift = BinaryOperator::createLShr(PartValue,
+ ConstantInt::get(V->getType(), i), "ctpop.sh", IP);
+ Value *RHS = BinaryOperator::createAnd(VShift, MaskCst, "cppop.and2", IP);
+ PartValue = BinaryOperator::createAdd(LHS, RHS, "ctpop.step", IP);
+ }
+ Count = BinaryOperator::createAdd(PartValue, Count, "ctpop.part", IP);
+ if (BitSize > 64) {
+ V = BinaryOperator::createLShr(V, ConstantInt::get(V->getType(), 64),
+ "ctpop.part.sh", IP);
+ BitSize -= 64;
+ }
+ }
+
+ return Count;
+}
+
+/// LowerCTLZ - Emit the code to lower ctlz of V before the specified
+/// instruction IP.
+static Value *LowerCTLZ(Value *V, Instruction *IP) {
+
+ unsigned BitSize = V->getType()->getPrimitiveSizeInBits();
+ for (unsigned i = 1; i < BitSize; i <<= 1) {
+ Value *ShVal = ConstantInt::get(V->getType(), i);
+ ShVal = BinaryOperator::createLShr(V, ShVal, "ctlz.sh", IP);
+ V = BinaryOperator::createOr(V, ShVal, "ctlz.step", IP);
+ }
+
+ V = BinaryOperator::createNot(V, "", IP);
+ return LowerCTPOP(V, IP);
+}
+
+/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes
+/// three integer arguments. The first argument is the Value from which the
+/// bits will be selected. It may be of any bit width. The second and third
+/// arguments specify a range of bits to select with the second argument
+/// specifying the low bit and the third argument specifying the high bit. Both
+/// must be type i32. The result is the corresponding selected bits from the
+/// Value in the same width as the Value (first argument). If the low bit index
+/// is higher than the high bit index then the inverse selection is done and
+/// the bits are returned in inverse order.
+/// @brief Lowering of llvm.part.select intrinsic.
+static Instruction *LowerPartSelect(CallInst *CI) {
+ // Make sure we're dealing with a part select intrinsic here
+ Function *F = CI->getCalledFunction();
+ const FunctionType *FT = F->getFunctionType();
+ if (!F->isDeclaration() || !FT->getReturnType()->isInteger() ||
+ FT->getNumParams() != 3 || !FT->getParamType(0)->isInteger() ||
+ !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger())
+ return CI;
+
+ // Get the intrinsic implementation function by converting all the . to _
+ // in the intrinsic's function name and then reconstructing the function
+ // declaration.
+ std::string Name(F->getName());
+ for (unsigned i = 4; i < Name.length(); ++i)
+ if (Name[i] == '.')
+ Name[i] = '_';
+ Module* M = F->getParent();
+ F = cast<Function>(M->getOrInsertFunction(Name, FT));
+ F->setLinkage(GlobalValue::WeakLinkage);
+
+ // If we haven't defined the impl function yet, do so now
+ if (F->isDeclaration()) {
+
+ // Get the arguments to the function
+ Function::arg_iterator args = F->arg_begin();
+ Value* Val = args++; Val->setName("Val");
+ Value* Lo = args++; Lo->setName("Lo");
+ Value* Hi = args++; Hi->setName("High");
+
+ // We want to select a range of bits here such that [Hi, Lo] is shifted
+ // down to the low bits. However, it is quite possible that Hi is smaller
+ // than Lo in which case the bits have to be reversed.
+
+ // Create the blocks we will need for the two cases (forward, reverse)
+ BasicBlock* CurBB = new BasicBlock("entry", F);
+ BasicBlock *RevSize = new BasicBlock("revsize", CurBB->getParent());
+ BasicBlock *FwdSize = new BasicBlock("fwdsize", CurBB->getParent());
+ BasicBlock *Compute = new BasicBlock("compute", CurBB->getParent());
+ BasicBlock *Reverse = new BasicBlock("reverse", CurBB->getParent());
+ BasicBlock *RsltBlk = new BasicBlock("result", CurBB->getParent());
+
+ // Cast Hi and Lo to the size of Val so the widths are all the same
+ if (Hi->getType() != Val->getType())
+ Hi = CastInst::createIntegerCast(Hi, Val->getType(), false,
+ "tmp", CurBB);
+ if (Lo->getType() != Val->getType())
+ Lo = CastInst::createIntegerCast(Lo, Val->getType(), false,
+ "tmp", CurBB);
+
+ // Compute a few things that both cases will need, up front.
+ Constant* Zero = ConstantInt::get(Val->getType(), 0);
+ Constant* One = ConstantInt::get(Val->getType(), 1);
+ Constant* AllOnes = ConstantInt::getAllOnesValue(Val->getType());
+
+ // Compare the Hi and Lo bit positions. This is used to determine
+ // which case we have (forward or reverse)
+ ICmpInst *Cmp = new ICmpInst(ICmpInst::ICMP_ULT, Hi, Lo, "less",CurBB);
+ new BranchInst(RevSize, FwdSize, Cmp, CurBB);
+
+ // First, copmute the number of bits in the forward case.
+ Instruction* FBitSize =
+ BinaryOperator::createSub(Hi, Lo,"fbits", FwdSize);
+ new BranchInst(Compute, FwdSize);
+
+ // Second, compute the number of bits in the reverse case.
+ Instruction* RBitSize =
+ BinaryOperator::createSub(Lo, Hi, "rbits", RevSize);
+ new BranchInst(Compute, RevSize);
+
+ // Now, compute the bit range. Start by getting the bitsize and the shift
+ // amount (either Hi or Lo) from PHI nodes. Then we compute a mask for
+ // the number of bits we want in the range. We shift the bits down to the
+ // least significant bits, apply the mask to zero out unwanted high bits,
+ // and we have computed the "forward" result. It may still need to be
+ // reversed.
+
+ // Get the BitSize from one of the two subtractions
+ PHINode *BitSize = new PHINode(Val->getType(), "bits", Compute);
+ BitSize->reserveOperandSpace(2);
+ BitSize->addIncoming(FBitSize, FwdSize);
+ BitSize->addIncoming(RBitSize, RevSize);
+
+ // Get the ShiftAmount as the smaller of Hi/Lo
+ PHINode *ShiftAmt = new PHINode(Val->getType(), "shiftamt", Compute);
+ ShiftAmt->reserveOperandSpace(2);
+ ShiftAmt->addIncoming(Lo, FwdSize);
+ ShiftAmt->addIncoming(Hi, RevSize);
+
+ // Increment the bit size
+ Instruction *BitSizePlusOne =
+ BinaryOperator::createAdd(BitSize, One, "bits", Compute);
+
+ // Create a Mask to zero out the high order bits.
+ Instruction* Mask =
+ BinaryOperator::createShl(AllOnes, BitSizePlusOne, "mask", Compute);
+ Mask = BinaryOperator::createNot(Mask, "mask", Compute);
+
+ // Shift the bits down and apply the mask
+ Instruction* FRes =
+ BinaryOperator::createLShr(Val, ShiftAmt, "fres", Compute);
+ FRes = BinaryOperator::createAnd(FRes, Mask, "fres", Compute);
+ new BranchInst(Reverse, RsltBlk, Cmp, Compute);
+
+ // In the Reverse block we have the mask already in FRes but we must reverse
+ // it by shifting FRes bits right and putting them in RRes by shifting them
+ // in from left.
+
+ // First set up our loop counters
+ PHINode *Count = new PHINode(Val->getType(), "count", Reverse);
+ Count->reserveOperandSpace(2);
+ Count->addIncoming(BitSizePlusOne, Compute);
+
+ // Next, get the value that we are shifting.
+ PHINode *BitsToShift = new PHINode(Val->getType(), "val", Reverse);
+ BitsToShift->reserveOperandSpace(2);
+ BitsToShift->addIncoming(FRes, Compute);
+
+ // Finally, get the result of the last computation
+ PHINode *RRes = new PHINode(Val->getType(), "rres", Reverse);
+ RRes->reserveOperandSpace(2);
+ RRes->addIncoming(Zero, Compute);
+
+ // Decrement the counter
+ Instruction *Decr = BinaryOperator::createSub(Count, One, "decr", Reverse);
+ Count->addIncoming(Decr, Reverse);
+
+ // Compute the Bit that we want to move
+ Instruction *Bit =
+ BinaryOperator::createAnd(BitsToShift, One, "bit", Reverse);