For PR1284:

Implement the "part_set" intrinsic.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@35938 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Intrinsics.td

@@ -203,12 +203,15 @@ def int_siglongjmp : Intrinsic<[llvm_void_ty, llvm_ptr_ty, llvm_i32_ty]>;
 
 // None of these intrinsics accesses memory at all.
 let Properties = [IntrNoMem] in {
-  def int_part_select : 
-     Intrinsic<[llvm_int_ty, llvm_int_ty, llvm_i32_ty, llvm_i32_ty]>;
   def int_bswap: Intrinsic<[llvm_int_ty, llvm_int_ty]>;
   def int_ctpop: Intrinsic<[llvm_i32_ty, llvm_int_ty]>;
   def int_ctlz : Intrinsic<[llvm_i32_ty, llvm_int_ty]>;
   def int_cttz : Intrinsic<[llvm_i32_ty, llvm_int_ty]>;
+  def int_part_select : 
+     Intrinsic<[llvm_int_ty, llvm_int_ty, llvm_i32_ty, llvm_i32_ty]>;
+  def int_part_set :
+     Intrinsic<[llvm_int_ty, llvm_int_ty, llvm_int_ty, llvm_i32_ty, 
+                llvm_i32_ty]>;
 }
 
 //===------------------------ Debugger Intrinsics -------------------------===//

lib/CodeGen/IntrinsicLowering.cpp

@@ -236,14 +236,17 @@ static Value *LowerCTLZ(Value *V, Instruction *IP) {
   return LowerCTPOP(V, IP);
 }
 
-/// Convert the llvm.bit.part_select.iX.iY.iZ intrinsic. This intrinsic takes 
+/// Convert the llvm.part.select.iX.iY intrinsic. This intrinsic takes 
-/// three integer operands of arbitrary bit width. The first operand is the 
+/// three integer arguments. The first argument is the Value from which the
-/// value from which to select the bits. The second and third operands define a 
+/// bits will be selected. It may be of any bit width. The second and third
-/// range of bits to select.  The result is the bits selected and has a 
+/// arguments specify a range of bits to select with the second argument 
-/// corresponding width of Left-Right (second operand - third operand).
+/// specifying the low bit and the third argument specifying the high bit. Both
-/// @see IEEE 1666-2005, System C, Section 7.2.6, pg 175. 
+/// must be type i32. The result is the corresponding selected bits from the
-/// @brief Lowering of llvm.bit.part_select intrinsic.
+/// Value in the same width as the Value (first argument). If the low bit index
-static Instruction *LowerBitPartSelect(CallInst *CI) {
+/// 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();
@@ -268,8 +271,8 @@ static Instruction *LowerBitPartSelect(CallInst *CI) {
 
     // Get the arguments to the function
     Value* Val = F->getOperand(0);
-    Value* Left = F->getOperand(1);
+    Value* Right = F->getOperand(1);
-    Value* Right = F->getOperand(2);
+    Value* Left = F->getOperand(2);
 
     // We want to select a range of bits here such that [Left, Right] is shifted
     // down to the low bits. However, it is quite possible that Left is smaller
@@ -406,6 +409,181 @@ static Instruction *LowerBitPartSelect(CallInst *CI) {
   return new CallInst(F, Args, 3, CI->getName(), CI);
 }
 
+/// Convert the llvm.part.set.iX.iY.iZ intrinsic. This intrinsic takes 
+/// four integer arguments (iAny %Value, iAny %Replacement, i32 %Low, i32 %High)
+/// The first two arguments can be any bit width. The result is the same width
+/// as %Value. The operation replaces bits between %Low and %High with the value
+/// in %Replacement. If %Replacement is not the same width, it is truncated or
+/// zero extended as appropriate to fit the bits being replaced. If %Low is
+/// greater than %High then the inverse set of bits are replaced.
+/// @brief Lowering of llvm.bit.part.set intrinsic.
+static Instruction *LowerPartSet(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() != 4 || !FT->getParamType(0)->isInteger() ||
+      !FT->getParamType(1)->isInteger() || !FT->getParamType(2)->isInteger() ||
+      !FT->getParamType(3)->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::InternalLinkage);
+
+  // If we haven't defined the impl function yet, do so now
+  if (F->isDeclaration()) {
+    // Note: the following code is based on code generated by llvm2cpp with 
+    // the following input. This is just *one* example of a generated function.
+    // The functions vary by bit width of result and first two arguments.
+    // The generated code has been changed to deal with any bit width not just
+    // the 32/64 bitwidths used in the above sample.
+    //
+    // define i64 @part_set(i64 %Val, i32 %Rep, i32 %Lo, i32 %Hi) {
+    // entry:
+    //   %is_forward = icmp ult i32 %Lo, %Hi
+    //   %Lo.pn = select i1 %is_forward, i32 %Hi, i32 %Lo
+    //   %Hi.pn = select i1 %is_forward, i32 %Lo, i32 %Hi
+    //   %iftmp.16.0 = sub i32 %Lo.pn, %Hi.pn
+    //   icmp ult i32 %iftmp.16.0, 32
+    //   br i1 %1, label %cond_true11, label %cond_next19
+    // cond_true11:
+    //   %tmp13 = sub i32 32, %iftmp.16.0
+    //   %tmp14 = lshr i32 -1, %tmp13
+    //   %tmp16 = and i32 %tmp14, %Rep
+    //   br label %cond_next19
+    // cond_next19:
+    //   %iftmp.17.0 = phi i32 [ %tmp16, %cond_true11 ], [ %Rep, %entry ]
+    //   %tmp2021 = zext i32 %iftmp.17.0 to i64
+    //   icmp ugt i32 %Lo, %Hi
+    //   br i1 %2, label %cond_next60, label %cond_true24
+    // cond_true24:
+    //   %tmp25.cast = zext i32 %Hi to i64
+    //   %tmp26 = lshr i64 -1, %tmp25.cast
+    //   %tmp27.cast = zext i32 %Lo to i64
+    //   %tmp28 = shl i64 %tmp26, %tmp27.cast
+    //   %tmp28not = xor i64 %tmp28, -1
+    //   %tmp31 = shl i64 %tmp2021, %tmp27.cast
+    //   %tmp34 = and i64 %tmp28not, %Val
+    //   %Val_addr.064 = or i64 %tmp31, %tmp34
+    //   ret i64 %Val_addr.064
+    // cond_next60:
+    //   %tmp39.cast = zext i32 %Lo to i64
+    //   %tmp40 = shl i64 -1, %tmp39.cast
+    //   %tmp41.cast = zext i32 %Hi to i64
+    //   %tmp42 = shl i64 -1, %tmp41.cast
+    //   %tmp45.demorgan = or i64 %tmp42, %tmp40
+    //   %tmp45 = xor i64 %tmp45.demorgan, -1
+    //   %tmp47 = and i64 %tmp45, %Val
+    //   %tmp50 = shl i64 %tmp2021, %tmp39.cast
+    //   %tmp52 = sub i32 32, %Hi
+    //   %tmp52.cast = zext i32 %tmp52 to i64
+    //   %tmp54 = lshr i64 %tmp2021, %tmp52.cast
+    //   %tmp57 = or i64 %tmp50, %tmp47
+    //   %Val_addr.0 = or i64 %tmp57, %tmp54
+    //   ret i64 %Val_addr.0
+    // }
+
+    // Get the arguments for the function.
+    Function::arg_iterator args = F->arg_begin();
+    Value* Val = args++; Val->setName("Val");
+    Value* Rep = args++; Rep->setName("Rep");
+    Value* Lo  = args++; Lo->setName("Lo");
+    Value* Hi  = args++; Hi->setName("Hi");
+
+    // Get some types we need
+    const IntegerType* ValTy = cast<IntegerType>(Val->getType());
+    const IntegerType* RepTy = cast<IntegerType>(Rep->getType());
+    uint32_t ValBits = ValTy->getBitWidth();
+    uint32_t RepBits = RepTy->getBitWidth();
+
+    // Constant Definitions
+    ConstantInt* RepBitWidth = ConstantInt::get(Type::Int32Ty, RepBits);
+    ConstantInt* RepMask = ConstantInt::getAllOnesValue(RepTy);
+    ConstantInt* ValMask = ConstantInt::getAllOnesValue(ValTy);
+
+    BasicBlock* entry = new BasicBlock("entry",F,0);
+    BasicBlock* large = new BasicBlock("large",F,0);
+    BasicBlock* small = new BasicBlock("small",F,0);
+    BasicBlock* forward = new BasicBlock("cond_true24",F,0);
+    BasicBlock* reverse = new BasicBlock("cond_next60",F,0);
+
+    // Block entry (entry)
+    // First, convert Lo and Hi to ValTy bit width
+    if (ValBits > 32) {
+      Hi = new ZExtInst(Hi, ValTy, "", entry);
+      Lo = new ZExtInst(Lo, ValTy, "", entry);
+    } else if (ValBits < 32) {
+      Hi = new TruncInst(Hi, ValTy, "", entry);
+      Lo = new TruncInst(Lo, ValTy, "", entry);
+    }
+    ICmpInst* is_forward = 
+      new ICmpInst(ICmpInst::ICMP_ULT, Lo, Hi, "", entry);
+    SelectInst* Lo_pn = new SelectInst(is_forward, Hi, Lo, "", entry);
+    SelectInst* Hi_pn = new SelectInst(is_forward, Lo, Hi, "", entry);
+    BinaryOperator* NumBits = BinaryOperator::createSub(Lo_pn, Hi_pn, "",entry);
+    ICmpInst* is_large = 
+      new ICmpInst(ICmpInst::ICMP_ULT, NumBits, RepBitWidth, "", entry);
+    new BranchInst(large, small, is_large, entry);
+
+    // Block "large"
+    BinaryOperator* MaskBits = 
+      BinaryOperator::createSub(RepBitWidth, NumBits, "", large);
+    BinaryOperator* Mask1 = 
+      BinaryOperator::createLShr(RepMask, MaskBits, "", large);
+    BinaryOperator* Rep2 = BinaryOperator::createAnd(Mask1, Rep, "", large);
+    new BranchInst(small, large);
+
+    // Block "small"
+    PHINode* Rep3 = new PHINode(RepTy, "", small);
+    Rep3->reserveOperandSpace(2);
+    Rep3->addIncoming(Rep2, small);
+    Rep3->addIncoming(Rep, entry);
+    CastInst* Rep4 = new ZExtInst(Rep3, ValTy, "", small);
+    ICmpInst* is_reverse = 
+      new ICmpInst(ICmpInst::ICMP_UGT, Lo, Hi, "", small);
+    new BranchInst(reverse, forward, is_reverse, small);
+
+    // Block "forward"
+    Value* t1    = BinaryOperator::createLShr(ValMask, Hi, "", forward);
+    Value* t2    = BinaryOperator::createShl(t1, Lo, "", forward);
+    Value* nott2 = BinaryOperator::createXor(t2, ValMask, "", forward);
+    Value* t3    = BinaryOperator::createShl(Rep4, Lo, "", forward);
+    Value* t4    = BinaryOperator::createAnd(nott2, Val, "", forward);
+    Value* FRslt = BinaryOperator::createOr(t3, t4, "", forward);
+    new ReturnInst(FRslt, forward);
+
+    // Block "reverse"
+    Value* t5    = BinaryOperator::createShl(ValMask, Lo, "", reverse);
+    Value* t6    = BinaryOperator::createShl(ValMask, Hi, "", reverse);
+    Value* t7    = BinaryOperator::createOr(t6, t5, "", reverse);
+    Value* t8    = BinaryOperator::createXor(t7, ValMask, "", reverse);
+    Value* t9    = BinaryOperator::createAnd(t8, Val, "", reverse);
+    Value* t10   = BinaryOperator::createShl(Rep4, Lo, "", reverse);
+    Value* t11   = BinaryOperator::createSub(RepBitWidth, Hi, "", reverse);
+    Value* t12   = new ZExtInst(t11, ValTy, "", reverse);
+    Value* t13   = BinaryOperator::createLShr(Rep4, t12, "",reverse);
+    Value* t14   = BinaryOperator::createOr(t10, t9, "", reverse);
+    Value* RRslt = BinaryOperator::createOr(t14, t13, "", reverse);
+    new ReturnInst(RRslt, reverse);
+  }
+
+  // Return a call to the implementation function
+  Value *Args[3];
+  Args[0] = CI->getOperand(0);
+  Args[1] = CI->getOperand(1);
+  Args[2] = CI->getOperand(2);
+  Args[3] = CI->getOperand(3);
+  return new CallInst(F, Args, 4, CI->getName(), CI);
+}
+
 
 void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
   Function *Callee = CI->getCalledFunction();
@@ -476,7 +654,11 @@ void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
   }
 
   case Intrinsic::part_select:
-    CI->replaceAllUsesWith(LowerBitPartSelect(CI));
+    CI->replaceAllUsesWith(LowerPartSelect(CI));
+    break;
+
+  case Intrinsic::part_set:
+    CI->replaceAllUsesWith(LowerPartSet(CI));
     break;
 
   case Intrinsic::stacksave:

lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -2691,7 +2691,12 @@ SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
   }
   case Intrinsic::part_select: {
     // Currently not implemented: just abort
-    assert(0 && "bit_part_select intrinsic not implemented");
+    assert(0 && "part_select intrinsic not implemented");
+    abort();
+  }
+  case Intrinsic::part_set: {
+    // Currently not implemented: just abort
+    assert(0 && "part_set intrinsic not implemented");
     abort();
   }
   case Intrinsic::bswap:

lib/VMCore/Verifier.cpp

@@ -1028,13 +1028,14 @@ void Verifier::VerifyIntrinsicPrototype(Intrinsic::ID ID, Function *F, ...) {
           if (GotBits < 16 || GotBits % 16 != 0)
             CheckFailed("Intrinsic requires even byte width argument", F);
           /* FALL THROUGH */
+        case Intrinsic::part_set:
         case Intrinsic::part_select:
           if (ArgNo == 1) {
             unsigned ResultBits = 
               cast<IntegerType>(FTy->getReturnType())->getBitWidth();
             if (GotBits != ResultBits)
-              CheckFailed("Intrinsic requires parameter and result bit "
+              CheckFailed("Intrinsic requires the bit widths of the first "
-                          "widths to match", F);
+                          "parameter and the result to match", F);
           }
           break;
       }