diff --git a/debian/changelog b/debian/changelog index b9d5bcf3..63a3e515 100644 --- a/debian/changelog +++ b/debian/changelog @@ -1,3 +1,10 @@ +llvm-toolchain-3.9 (1:3.9-4) unstable; urgency=medium + + * d/p/bug-30342.diff: + Backport upstream bug 30342 to fix an infinity loop in rust testsuite + + -- Sylvestre Ledru Mon, 31 Oct 2016 10:47:52 +0100 + llvm-toolchain-3.9 (1:3.9-3) unstable; urgency=medium [ Sylvestre Ledru ] diff --git a/debian/patches/bug-30342.diff b/debian/patches/bug-30342.diff new file mode 100644 index 00000000..c3c3cd85 --- /dev/null +++ b/debian/patches/bug-30342.diff @@ -0,0 +1,4935 @@ +Index: llvm/trunk/test/Transforms/InstCombine/indexed-gep-compares.ll +=================================================================== +--- llvm/trunk/test/Transforms/InstCombine/indexed-gep-compares.ll (revision 281649) ++++ llvm/trunk/test/Transforms/InstCombine/indexed-gep-compares.ll (revision 281650) +@@ -1,170 +1,190 @@ + ; RUN: opt -instcombine -S < %s | FileCheck %s + + target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:32-f32:32:32-f64:32:32-v64:64:64-v128:128:128-a0:0:64" + + define i32 *@test1(i32* %A, i32 %Offset) { + entry: + %tmp = getelementptr inbounds i32, i32* %A, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %entry ] + %LHS = getelementptr inbounds i32, i32* %A, i32 100 + %RHS.next = getelementptr inbounds i32, i32* %RHS, i64 1 + %cond = icmp ult i32 * %LHS, %RHS + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + ; CHECK-LABEL: @test1( + ; CHECK: %[[INDEX:[0-9A-Za-z.]+]] = phi i32 [ %[[ADD:[0-9A-Za-z.]+]], %bb ], [ %Offset, %entry ] + ; CHECK: %[[ADD]] = add nsw i32 %[[INDEX]], 1 + ; CHECK: %cond = icmp sgt i32 %[[INDEX]], 100 + ; CHECK: br i1 %cond, label %bb2, label %bb + ; CHECK: %[[PTR:[0-9A-Za-z.]+]] = getelementptr inbounds i32, i32* %A, i32 %[[INDEX]] + ; CHECK: ret i32* %[[PTR]] + } + + define i32 *@test2(i32 %A, i32 %Offset) { + entry: + %A.ptr = inttoptr i32 %A to i32* + %tmp = getelementptr inbounds i32, i32* %A.ptr, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %entry ] + %LHS = getelementptr inbounds i32, i32* %A.ptr, i32 100 + %RHS.next = getelementptr inbounds i32, i32* %RHS, i64 1 + %cmp0 = ptrtoint i32 *%LHS to i32 + %cmp1 = ptrtoint i32 *%RHS to i32 + %cond = icmp ult i32 %cmp0, %cmp1 + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + ; CHECK-LABEL: @test2( + ; CHECK: %[[INDEX:[0-9A-Za-z.]+]] = phi i32 [ %[[ADD:[0-9A-Za-z.]+]], %bb ], [ %Offset, %entry ] + ; CHECK: %[[ADD]] = add nsw i32 %[[INDEX]], 1 + ; CHECK: %cond = icmp sgt i32 %[[INDEX]], 100 + ; CHECK: br i1 %cond, label %bb2, label %bb + ; CHECK: %[[TOPTR:[0-9A-Za-z.]+]] = inttoptr i32 %[[ADD:[0-9A-Za-z.]+]] to i32* + ; CHECK: %[[PTR:[0-9A-Za-z.]+]] = getelementptr inbounds i32, i32* %[[TOPTR]], i32 %[[INDEX]] + ; CHECK: ret i32* %[[PTR]] + } + + ; Perform the transformation only if we know that the GEPs used are inbounds. + define i32 *@test3(i32* %A, i32 %Offset) { + entry: + %tmp = getelementptr i32, i32* %A, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %entry ] + %LHS = getelementptr i32, i32* %A, i32 100 + %RHS.next = getelementptr i32, i32* %RHS, i64 1 + %cond = icmp ult i32 * %LHS, %RHS + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + ; CHECK-LABEL: @test3( + ; CHECK-NOT: %cond = icmp sgt i32 %{{[0-9A-Za-z.]+}}, 100 + } + + ; An inttoptr that requires an extension or truncation will be opaque when determining + ; the base pointer. In this case we can still perform the transformation by considering + ; A.ptr as being the base pointer. + define i32 *@test4(i16 %A, i32 %Offset) { + entry: + %A.ptr = inttoptr i16 %A to i32* + %tmp = getelementptr inbounds i32, i32* %A.ptr, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %entry ] + %LHS = getelementptr inbounds i32, i32* %A.ptr, i32 100 + %RHS.next = getelementptr inbounds i32, i32* %RHS, i64 1 + %cmp0 = ptrtoint i32 *%LHS to i32 + %cmp1 = ptrtoint i32 *%RHS to i32 + %cond = icmp ult i32 %cmp0, %cmp1 + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + ; CHECK-LABEL: @test4( + ; CHECK: %cond = icmp sgt i32 %{{[0-9A-Za-z.]+}}, 100 + } + + declare i32* @fun_ptr() + + define i32 *@test5(i32 %Offset) personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { + entry: + %A = invoke i32 *@fun_ptr() to label %cont unwind label %lpad + + cont: + %tmp = getelementptr inbounds i32, i32* %A, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %cont ] + %LHS = getelementptr inbounds i32, i32* %A, i32 100 + %RHS.next = getelementptr inbounds i32, i32* %RHS, i64 1 + %cond = icmp ult i32 * %LHS, %RHS + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + lpad: + %l = landingpad { i8*, i32 } cleanup + ret i32* null + + ; CHECK-LABEL: @test5( + ; CHECK: %[[INDEX:[0-9A-Za-z.]+]] = phi i32 [ %[[ADD:[0-9A-Za-z.]+]], %bb ], [ %Offset, %cont ] + ; CHECK: %[[ADD]] = add nsw i32 %[[INDEX]], 1 + ; CHECK: %cond = icmp sgt i32 %[[INDEX]], 100 + ; CHECK: br i1 %cond, label %bb2, label %bb + ; CHECK: %[[PTR:[0-9A-Za-z.]+]] = getelementptr inbounds i32, i32* %A, i32 %[[INDEX]] + ; CHECK: ret i32* %[[PTR]] + } + + declare i32 @fun_i32() + + define i32 *@test6(i32 %Offset) personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { + entry: + %A = invoke i32 @fun_i32() to label %cont unwind label %lpad + + cont: + %A.ptr = inttoptr i32 %A to i32* + %tmp = getelementptr inbounds i32, i32* %A.ptr, i32 %Offset + br label %bb + + bb: + %RHS = phi i32* [ %RHS.next, %bb ], [ %tmp, %cont ] + %LHS = getelementptr inbounds i32, i32* %A.ptr, i32 100 + %RHS.next = getelementptr inbounds i32, i32* %RHS, i64 1 + %cond = icmp ult i32 * %LHS, %RHS + br i1 %cond, label %bb2, label %bb + + bb2: + ret i32* %RHS + + lpad: + %l = landingpad { i8*, i32 } cleanup + ret i32* null + + ; CHECK-LABEL: @test6( + ; CHECK: %[[INDEX:[0-9A-Za-z.]+]] = phi i32 [ %[[ADD:[0-9A-Za-z.]+]], %bb ], [ %Offset, %cont ] + ; CHECK: %[[ADD]] = add nsw i32 %[[INDEX]], 1 + ; CHECK: %cond = icmp sgt i32 %[[INDEX]], 100 + ; CHECK: br i1 %cond, label %bb2, label %bb + ; CHECK: %[[TOPTR:[0-9A-Za-z.]+]] = inttoptr i32 %[[ADD:[0-9A-Za-z.]+]] to i32* + ; CHECK: %[[PTR:[0-9A-Za-z.]+]] = getelementptr inbounds i32, i32* %[[TOPTR]], i32 %[[INDEX]] + ; CHECK: ret i32* %[[PTR]] + } + ++ ++@pr30402 = constant i64 3 ++define i1 @test7() { ++entry: ++ br label %bb7 ++ ++bb7: ; preds = %bb10, %entry-block ++ %phi = phi i64* [ @pr30402, %entry ], [ getelementptr inbounds (i64, i64* @pr30402, i32 1), %bb7 ] ++ %cmp = icmp eq i64* %phi, getelementptr inbounds (i64, i64* @pr30402, i32 1) ++ br i1 %cmp, label %bb10, label %bb7 ++ ++bb10: ++ ret i1 %cmp ++} ++; CHECK-LABEL: @test7( ++; CHECK: %[[phi:.*]] = phi i64* [ @pr30402, %entry ], [ getelementptr inbounds (i64, i64* @pr30402, i32 1), %bb7 ] ++; CHECK: %[[cmp:.*]] = icmp eq i64* %[[phi]], getelementptr inbounds (i64, i64* @pr30402, i32 1) ++; CHECK: ret i1 %[[cmp]] ++ ++ + declare i32 @__gxx_personality_v0(...) +Index: llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp +=================================================================== +--- llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp (revision 281649) ++++ llvm/trunk/lib/Transforms/InstCombine/InstCombineCompares.cpp (revision 281650) +@@ -1,4734 +1,4734 @@ + //===- InstCombineCompares.cpp --------------------------------------------===// + // + // The LLVM Compiler Infrastructure + // + // This file is distributed under the University of Illinois Open Source + // License. See LICENSE.TXT for details. + // + //===----------------------------------------------------------------------===// + // + // This file implements the visitICmp and visitFCmp functions. + // + //===----------------------------------------------------------------------===// + + #include "InstCombineInternal.h" + #include "llvm/ADT/APSInt.h" + #include "llvm/ADT/SetVector.h" + #include "llvm/ADT/Statistic.h" + #include "llvm/Analysis/ConstantFolding.h" + #include "llvm/Analysis/InstructionSimplify.h" + #include "llvm/Analysis/MemoryBuiltins.h" + #include "llvm/Analysis/TargetLibraryInfo.h" + #include "llvm/Analysis/VectorUtils.h" + #include "llvm/IR/ConstantRange.h" + #include "llvm/IR/DataLayout.h" + #include "llvm/IR/GetElementPtrTypeIterator.h" + #include "llvm/IR/IntrinsicInst.h" + #include "llvm/IR/PatternMatch.h" + #include "llvm/Support/Debug.h" + + using namespace llvm; + using namespace PatternMatch; + + #define DEBUG_TYPE "instcombine" + + // How many times is a select replaced by one of its operands? + STATISTIC(NumSel, "Number of select opts"); + + + static ConstantInt *extractElement(Constant *V, Constant *Idx) { + return cast(ConstantExpr::getExtractElement(V, Idx)); + } + + static bool hasAddOverflow(ConstantInt *Result, + ConstantInt *In1, ConstantInt *In2, + bool IsSigned) { + if (!IsSigned) + return Result->getValue().ult(In1->getValue()); + + if (In2->isNegative()) + return Result->getValue().sgt(In1->getValue()); + return Result->getValue().slt(In1->getValue()); + } + + /// Compute Result = In1+In2, returning true if the result overflowed for this + /// type. + static bool addWithOverflow(Constant *&Result, Constant *In1, + Constant *In2, bool IsSigned = false) { + Result = ConstantExpr::getAdd(In1, In2); + + if (VectorType *VTy = dyn_cast(In1->getType())) { + for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { + Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); + if (hasAddOverflow(extractElement(Result, Idx), + extractElement(In1, Idx), + extractElement(In2, Idx), + IsSigned)) + return true; + } + return false; + } + + return hasAddOverflow(cast(Result), + cast(In1), cast(In2), + IsSigned); + } + + static bool hasSubOverflow(ConstantInt *Result, + ConstantInt *In1, ConstantInt *In2, + bool IsSigned) { + if (!IsSigned) + return Result->getValue().ugt(In1->getValue()); + + if (In2->isNegative()) + return Result->getValue().slt(In1->getValue()); + + return Result->getValue().sgt(In1->getValue()); + } + + /// Compute Result = In1-In2, returning true if the result overflowed for this + /// type. + static bool subWithOverflow(Constant *&Result, Constant *In1, + Constant *In2, bool IsSigned = false) { + Result = ConstantExpr::getSub(In1, In2); + + if (VectorType *VTy = dyn_cast(In1->getType())) { + for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) { + Constant *Idx = ConstantInt::get(Type::getInt32Ty(In1->getContext()), i); + if (hasSubOverflow(extractElement(Result, Idx), + extractElement(In1, Idx), + extractElement(In2, Idx), + IsSigned)) + return true; + } + return false; + } + + return hasSubOverflow(cast(Result), + cast(In1), cast(In2), + IsSigned); + } + + /// Given an icmp instruction, return true if any use of this comparison is a + /// branch on sign bit comparison. + static bool isBranchOnSignBitCheck(ICmpInst &I, bool isSignBit) { + for (auto *U : I.users()) + if (isa(U)) + return isSignBit; + return false; + } + + /// Given an exploded icmp instruction, return true if the comparison only + /// checks the sign bit. If it only checks the sign bit, set TrueIfSigned if the + /// result of the comparison is true when the input value is signed. + static bool isSignBitCheck(ICmpInst::Predicate Pred, const APInt &RHS, + bool &TrueIfSigned) { + switch (Pred) { + case ICmpInst::ICMP_SLT: // True if LHS s< 0 + TrueIfSigned = true; + return RHS == 0; + case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1 + TrueIfSigned = true; + return RHS.isAllOnesValue(); + case ICmpInst::ICMP_SGT: // True if LHS s> -1 + TrueIfSigned = false; + return RHS.isAllOnesValue(); + case ICmpInst::ICMP_UGT: + // True if LHS u> RHS and RHS == high-bit-mask - 1 + TrueIfSigned = true; + return RHS.isMaxSignedValue(); + case ICmpInst::ICMP_UGE: + // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc) + TrueIfSigned = true; + return RHS.isSignBit(); + default: + return false; + } + } + + /// Returns true if the exploded icmp can be expressed as a signed comparison + /// to zero and updates the predicate accordingly. + /// The signedness of the comparison is preserved. + /// TODO: Refactor with decomposeBitTestICmp()? + static bool isSignTest(ICmpInst::Predicate &Pred, const APInt &C) { + if (!ICmpInst::isSigned(Pred)) + return false; + + if (C == 0) + return ICmpInst::isRelational(Pred); + + if (C == 1) { + if (Pred == ICmpInst::ICMP_SLT) { + Pred = ICmpInst::ICMP_SLE; + return true; + } + } else if (C.isAllOnesValue()) { + if (Pred == ICmpInst::ICMP_SGT) { + Pred = ICmpInst::ICMP_SGE; + return true; + } + } + + return false; + } + + /// Given a signed integer type and a set of known zero and one bits, compute + /// the maximum and minimum values that could have the specified known zero and + /// known one bits, returning them in Min/Max. + static void computeSignedMinMaxValuesFromKnownBits(const APInt &KnownZero, + const APInt &KnownOne, + APInt &Min, APInt &Max) { + assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && + KnownZero.getBitWidth() == Min.getBitWidth() && + KnownZero.getBitWidth() == Max.getBitWidth() && + "KnownZero, KnownOne and Min, Max must have equal bitwidth."); + APInt UnknownBits = ~(KnownZero|KnownOne); + + // The minimum value is when all unknown bits are zeros, EXCEPT for the sign + // bit if it is unknown. + Min = KnownOne; + Max = KnownOne|UnknownBits; + + if (UnknownBits.isNegative()) { // Sign bit is unknown + Min.setBit(Min.getBitWidth()-1); + Max.clearBit(Max.getBitWidth()-1); + } + } + + /// Given an unsigned integer type and a set of known zero and one bits, compute + /// the maximum and minimum values that could have the specified known zero and + /// known one bits, returning them in Min/Max. + static void computeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero, + const APInt &KnownOne, + APInt &Min, APInt &Max) { + assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() && + KnownZero.getBitWidth() == Min.getBitWidth() && + KnownZero.getBitWidth() == Max.getBitWidth() && + "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth."); + APInt UnknownBits = ~(KnownZero|KnownOne); + + // The minimum value is when the unknown bits are all zeros. + Min = KnownOne; + // The maximum value is when the unknown bits are all ones. + Max = KnownOne|UnknownBits; + } + + /// This is called when we see this pattern: + /// cmp pred (load (gep GV, ...)), cmpcst + /// where GV is a global variable with a constant initializer. Try to simplify + /// this into some simple computation that does not need the load. For example + /// we can optimize "icmp eq (load (gep "foo", 0, i)), 0" into "icmp eq i, 3". + /// + /// If AndCst is non-null, then the loaded value is masked with that constant + /// before doing the comparison. This handles cases like "A[i]&4 == 0". + Instruction *InstCombiner::foldCmpLoadFromIndexedGlobal(GetElementPtrInst *GEP, + GlobalVariable *GV, + CmpInst &ICI, + ConstantInt *AndCst) { + Constant *Init = GV->getInitializer(); + if (!isa(Init) && !isa(Init)) + return nullptr; + + uint64_t ArrayElementCount = Init->getType()->getArrayNumElements(); + if (ArrayElementCount > 1024) return nullptr; // Don't blow up on huge arrays. + + // There are many forms of this optimization we can handle, for now, just do + // the simple index into a single-dimensional array. + // + // Require: GEP GV, 0, i {{, constant indices}} + if (GEP->getNumOperands() < 3 || + !isa(GEP->getOperand(1)) || + !cast(GEP->getOperand(1))->isZero() || + isa(GEP->getOperand(2))) + return nullptr; + + // Check that indices after the variable are constants and in-range for the + // type they index. Collect the indices. This is typically for arrays of + // structs. + SmallVector LaterIndices; + + Type *EltTy = Init->getType()->getArrayElementType(); + for (unsigned i = 3, e = GEP->getNumOperands(); i != e; ++i) { + ConstantInt *Idx = dyn_cast(GEP->getOperand(i)); + if (!Idx) return nullptr; // Variable index. + + uint64_t IdxVal = Idx->getZExtValue(); + if ((unsigned)IdxVal != IdxVal) return nullptr; // Too large array index. + + if (StructType *STy = dyn_cast(EltTy)) + EltTy = STy->getElementType(IdxVal); + else if (ArrayType *ATy = dyn_cast(EltTy)) { + if (IdxVal >= ATy->getNumElements()) return nullptr; + EltTy = ATy->getElementType(); + } else { + return nullptr; // Unknown type. + } + + LaterIndices.push_back(IdxVal); + } + + enum { Overdefined = -3, Undefined = -2 }; + + // Variables for our state machines. + + // FirstTrueElement/SecondTrueElement - Used to emit a comparison of the form + // "i == 47 | i == 87", where 47 is the first index the condition is true for, + // and 87 is the second (and last) index. FirstTrueElement is -2 when + // undefined, otherwise set to the first true element. SecondTrueElement is + // -2 when undefined, -3 when overdefined and >= 0 when that index is true. + int FirstTrueElement = Undefined, SecondTrueElement = Undefined; + + // FirstFalseElement/SecondFalseElement - Used to emit a comparison of the + // form "i != 47 & i != 87". Same state transitions as for true elements. + int FirstFalseElement = Undefined, SecondFalseElement = Undefined; + + /// TrueRangeEnd/FalseRangeEnd - In conjunction with First*Element, these + /// define a state machine that triggers for ranges of values that the index + /// is true or false for. This triggers on things like "abbbbc"[i] == 'b'. + /// This is -2 when undefined, -3 when overdefined, and otherwise the last + /// index in the range (inclusive). We use -2 for undefined here because we + /// use relative comparisons and don't want 0-1 to match -1. + int TrueRangeEnd = Undefined, FalseRangeEnd = Undefined; + + // MagicBitvector - This is a magic bitvector where we set a bit if the + // comparison is true for element 'i'. If there are 64 elements or less in + // the array, this will fully represent all the comparison results. + uint64_t MagicBitvector = 0; + + // Scan the array and see if one of our patterns matches. + Constant *CompareRHS = cast(ICI.getOperand(1)); + for (unsigned i = 0, e = ArrayElementCount; i != e; ++i) { + Constant *Elt = Init->getAggregateElement(i); + if (!Elt) return nullptr; + + // If this is indexing an array of structures, get the structure element. + if (!LaterIndices.empty()) + Elt = ConstantExpr::getExtractValue(Elt, LaterIndices); + + // If the element is masked, handle it. + if (AndCst) Elt = ConstantExpr::getAnd(Elt, AndCst); + + // Find out if the comparison would be true or false for the i'th element. + Constant *C = ConstantFoldCompareInstOperands(ICI.getPredicate(), Elt, + CompareRHS, DL, &TLI); + // If the result is undef for this element, ignore it. + if (isa(C)) { + // Extend range state machines to cover this element in case there is an + // undef in the middle of the range. + if (TrueRangeEnd == (int)i-1) + TrueRangeEnd = i; + if (FalseRangeEnd == (int)i-1) + FalseRangeEnd = i; + continue; + } + + // If we can't compute the result for any of the elements, we have to give + // up evaluating the entire conditional. + if (!isa(C)) return nullptr; + + // Otherwise, we know if the comparison is true or false for this element, + // update our state machines. + bool IsTrueForElt = !cast(C)->isZero(); + + // State machine for single/double/range index comparison. + if (IsTrueForElt) { + // Update the TrueElement state machine. + if (FirstTrueElement == Undefined) + FirstTrueElement = TrueRangeEnd = i; // First true element. + else { + // Update double-compare state machine. + if (SecondTrueElement == Undefined) + SecondTrueElement = i; + else + SecondTrueElement = Overdefined; + + // Update range state machine. + if (TrueRangeEnd == (int)i-1) + TrueRangeEnd = i; + else + TrueRangeEnd = Overdefined; + } + } else { + // Update the FalseElement state machine. + if (FirstFalseElement == Undefined) + FirstFalseElement = FalseRangeEnd = i; // First false element. + else { + // Update double-compare state machine. + if (SecondFalseElement == Undefined) + SecondFalseElement = i; + else + SecondFalseElement = Overdefined; + + // Update range state machine. + if (FalseRangeEnd == (int)i-1) + FalseRangeEnd = i; + else + FalseRangeEnd = Overdefined; + } + } + + // If this element is in range, update our magic bitvector. + if (i < 64 && IsTrueForElt) + MagicBitvector |= 1ULL << i; + + // If all of our states become overdefined, bail out early. Since the + // predicate is expensive, only check it every 8 elements. This is only + // really useful for really huge arrays. + if ((i & 8) == 0 && i >= 64 && SecondTrueElement == Overdefined && + SecondFalseElement == Overdefined && TrueRangeEnd == Overdefined && + FalseRangeEnd == Overdefined) + return nullptr; + } + + // Now that we've scanned the entire array, emit our new comparison(s). We + // order the state machines in complexity of the generated code. + Value *Idx = GEP->getOperand(2); + + // If the index is larger than the pointer size of the target, truncate the + // index down like the GEP would do implicitly. We don't have to do this for + // an inbounds GEP because the index can't be out of range. + if (!GEP->isInBounds()) { + Type *IntPtrTy = DL.getIntPtrType(GEP->getType()); + unsigned PtrSize = IntPtrTy->getIntegerBitWidth(); + if (Idx->getType()->getPrimitiveSizeInBits() > PtrSize) + Idx = Builder->CreateTrunc(Idx, IntPtrTy); + } + + // If the comparison is only true for one or two elements, emit direct + // comparisons. + if (SecondTrueElement != Overdefined) { + // None true -> false. + if (FirstTrueElement == Undefined) + return replaceInstUsesWith(ICI, Builder->getFalse()); + + Value *FirstTrueIdx = ConstantInt::get(Idx->getType(), FirstTrueElement); + + // True for one element -> 'i == 47'. + if (SecondTrueElement == Undefined) + return new ICmpInst(ICmpInst::ICMP_EQ, Idx, FirstTrueIdx); + + // True for two elements -> 'i == 47 | i == 72'. + Value *C1 = Builder->CreateICmpEQ(Idx, FirstTrueIdx); + Value *SecondTrueIdx = ConstantInt::get(Idx->getType(), SecondTrueElement); + Value *C2 = Builder->CreateICmpEQ(Idx, SecondTrueIdx); + return BinaryOperator::CreateOr(C1, C2); + } + + // If the comparison is only false for one or two elements, emit direct + // comparisons. + if (SecondFalseElement != Overdefined) { + // None false -> true. + if (FirstFalseElement == Undefined) + return replaceInstUsesWith(ICI, Builder->getTrue()); + + Value *FirstFalseIdx = ConstantInt::get(Idx->getType(), FirstFalseElement); + + // False for one element -> 'i != 47'. + if (SecondFalseElement == Undefined) + return new ICmpInst(ICmpInst::ICMP_NE, Idx, FirstFalseIdx); + + // False for two elements -> 'i != 47 & i != 72'. + Value *C1 = Builder->CreateICmpNE(Idx, FirstFalseIdx); + Value *SecondFalseIdx = ConstantInt::get(Idx->getType(),SecondFalseElement); + Value *C2 = Builder->CreateICmpNE(Idx, SecondFalseIdx); + return BinaryOperator::CreateAnd(C1, C2); + } + + // If the comparison can be replaced with a range comparison for the elements + // where it is true, emit the range check. + if (TrueRangeEnd != Overdefined) { + assert(TrueRangeEnd != FirstTrueElement && "Should emit single compare"); + + // Generate (i-FirstTrue) getType(), -FirstTrueElement); + Idx = Builder->CreateAdd(Idx, Offs); + } + + Value *End = ConstantInt::get(Idx->getType(), + TrueRangeEnd-FirstTrueElement+1); + return new ICmpInst(ICmpInst::ICMP_ULT, Idx, End); + } + + // False range check. + if (FalseRangeEnd != Overdefined) { + assert(FalseRangeEnd != FirstFalseElement && "Should emit single compare"); + // Generate (i-FirstFalse) >u (FalseRangeEnd-FirstFalse). + if (FirstFalseElement) { + Value *Offs = ConstantInt::get(Idx->getType(), -FirstFalseElement); + Idx = Builder->CreateAdd(Idx, Offs); + } + + Value *End = ConstantInt::get(Idx->getType(), + FalseRangeEnd-FirstFalseElement); + return new ICmpInst(ICmpInst::ICMP_UGT, Idx, End); + } + + // If a magic bitvector captures the entire comparison state + // of this load, replace it with computation that does: + // ((magic_cst >> i) & 1) != 0 + { + Type *Ty = nullptr; + + // Look for an appropriate type: + // - The type of Idx if the magic fits + // - The smallest fitting legal type if we have a DataLayout + // - Default to i32 + if (ArrayElementCount <= Idx->getType()->getIntegerBitWidth()) + Ty = Idx->getType(); + else + Ty = DL.getSmallestLegalIntType(Init->getContext(), ArrayElementCount); + + if (Ty) { + Value *V = Builder->CreateIntCast(Idx, Ty, false); + V = Builder->CreateLShr(ConstantInt::get(Ty, MagicBitvector), V); + V = Builder->CreateAnd(ConstantInt::get(Ty, 1), V); + return new ICmpInst(ICmpInst::ICMP_NE, V, ConstantInt::get(Ty, 0)); + } + } + + return nullptr; + } + + /// Return a value that can be used to compare the *offset* implied by a GEP to + /// zero. For example, if we have &A[i], we want to return 'i' for + /// "icmp ne i, 0". Note that, in general, indices can be complex, and scales + /// are involved. The above expression would also be legal to codegen as + /// "icmp ne (i*4), 0" (assuming A is a pointer to i32). + /// This latter form is less amenable to optimization though, and we are allowed + /// to generate the first by knowing that pointer arithmetic doesn't overflow. + /// + /// If we can't emit an optimized form for this expression, this returns null. + /// + static Value *evaluateGEPOffsetExpression(User *GEP, InstCombiner &IC, + const DataLayout &DL) { + gep_type_iterator GTI = gep_type_begin(GEP); + + // Check to see if this gep only has a single variable index. If so, and if + // any constant indices are a multiple of its scale, then we can compute this + // in terms of the scale of the variable index. For example, if the GEP + // implies an offset of "12 + i*4", then we can codegen this as "3 + i", + // because the expression will cross zero at the same point. + unsigned i, e = GEP->getNumOperands(); + int64_t Offset = 0; + for (i = 1; i != e; ++i, ++GTI) { + if (ConstantInt *CI = dyn_cast(GEP->getOperand(i))) { + // Compute the aggregate offset of constant indices. + if (CI->isZero()) continue; + + // Handle a struct index, which adds its field offset to the pointer. + if (StructType *STy = dyn_cast(*GTI)) { + Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); + } else { + uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); + Offset += Size*CI->getSExtValue(); + } + } else { + // Found our variable index. + break; + } + } + + // If there are no variable indices, we must have a constant offset, just + // evaluate it the general way. + if (i == e) return nullptr; + + Value *VariableIdx = GEP->getOperand(i); + // Determine the scale factor of the variable element. For example, this is + // 4 if the variable index is into an array of i32. + uint64_t VariableScale = DL.getTypeAllocSize(GTI.getIndexedType()); + + // Verify that there are no other variable indices. If so, emit the hard way. + for (++i, ++GTI; i != e; ++i, ++GTI) { + ConstantInt *CI = dyn_cast(GEP->getOperand(i)); + if (!CI) return nullptr; + + // Compute the aggregate offset of constant indices. + if (CI->isZero()) continue; + + // Handle a struct index, which adds its field offset to the pointer. + if (StructType *STy = dyn_cast(*GTI)) { + Offset += DL.getStructLayout(STy)->getElementOffset(CI->getZExtValue()); + } else { + uint64_t Size = DL.getTypeAllocSize(GTI.getIndexedType()); + Offset += Size*CI->getSExtValue(); + } + } + + // Okay, we know we have a single variable index, which must be a + // pointer/array/vector index. If there is no offset, life is simple, return + // the index. + Type *IntPtrTy = DL.getIntPtrType(GEP->getOperand(0)->getType()); + unsigned IntPtrWidth = IntPtrTy->getIntegerBitWidth(); + if (Offset == 0) { + // Cast to intptrty in case a truncation occurs. If an extension is needed, + // we don't need to bother extending: the extension won't affect where the + // computation crosses zero. + if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth) { + VariableIdx = IC.Builder->CreateTrunc(VariableIdx, IntPtrTy); + } + return VariableIdx; + } + + // Otherwise, there is an index. The computation we will do will be modulo + // the pointer size, so get it. + uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth); + + Offset &= PtrSizeMask; + VariableScale &= PtrSizeMask; + + // To do this transformation, any constant index must be a multiple of the + // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i", + // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a + // multiple of the variable scale. + int64_t NewOffs = Offset / (int64_t)VariableScale; + if (Offset != NewOffs*(int64_t)VariableScale) + return nullptr; + + // Okay, we can do this evaluation. Start by converting the index to intptr. + if (VariableIdx->getType() != IntPtrTy) + VariableIdx = IC.Builder->CreateIntCast(VariableIdx, IntPtrTy, + true /*Signed*/); + Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs); + return IC.Builder->CreateAdd(VariableIdx, OffsetVal, "offset"); + } + + /// Returns true if we can rewrite Start as a GEP with pointer Base + /// and some integer offset. The nodes that need to be re-written + /// for this transformation will be added to Explored. + static bool canRewriteGEPAsOffset(Value *Start, Value *Base, + const DataLayout &DL, + SetVector &Explored) { + SmallVector WorkList(1, Start); + Explored.insert(Base); + + // The following traversal gives us an order which can be used + // when doing the final transformation. Since in the final + // transformation we create the PHI replacement instructions first, + // we don't have to get them in any particular order. + // + // However, for other instructions we will have to traverse the + // operands of an instruction first, which means that we have to + // do a post-order traversal. + while (!WorkList.empty()) { + SetVector PHIs; + + while (!WorkList.empty()) { + if (Explored.size() >= 100) + return false; + + Value *V = WorkList.back(); + + if (Explored.count(V) != 0) { + WorkList.pop_back(); + continue; + } + + if (!isa(V) && !isa(V) && +- !isa(V) && !isa(V)) ++ !isa(V) && !isa(V)) + // We've found some value that we can't explore which is different from + // the base. Therefore we can't do this transformation. + return false; + + if (isa(V) || isa(V)) { + auto *CI = dyn_cast(V); + if (!CI->isNoopCast(DL)) + return false; + + if (Explored.count(CI->getOperand(0)) == 0) + WorkList.push_back(CI->getOperand(0)); + } + + if (auto *GEP = dyn_cast(V)) { + // We're limiting the GEP to having one index. This will preserve + // the original pointer type. We could handle more cases in the + // future. + if (GEP->getNumIndices() != 1 || !GEP->isInBounds() || + GEP->getType() != Start->getType()) + return false; + + if (Explored.count(GEP->getOperand(0)) == 0) + WorkList.push_back(GEP->getOperand(0)); + } + + if (WorkList.back() == V) { + WorkList.pop_back(); + // We've finished visiting this node, mark it as such. + Explored.insert(V); + } + + if (auto *PN = dyn_cast(V)) { + // We cannot transform PHIs on unsplittable basic blocks. + if (isa(PN->getParent()->getTerminator())) + return false; + Explored.insert(PN); + PHIs.insert(PN); + } + } + + // Explore the PHI nodes further. + for (auto *PN : PHIs) + for (Value *Op : PN->incoming_values()) + if (Explored.count(Op) == 0) + WorkList.push_back(Op); + } + + // Make sure that we can do this. Since we can't insert GEPs in a basic + // block before a PHI node, we can't easily do this transformation if + // we have PHI node users of transformed instructions. + for (Value *Val : Explored) { + for (Value *Use : Val->uses()) { + + auto *PHI = dyn_cast(Use); + auto *Inst = dyn_cast(Val); + + if (Inst == Base || Inst == PHI || !Inst || !PHI || + Explored.count(PHI) == 0) + continue; + + if (PHI->getParent() == Inst->getParent()) + return false; + } + } + return true; + } + + // Sets the appropriate insert point on Builder where we can add + // a replacement Instruction for V (if that is possible). + static void setInsertionPoint(IRBuilder<> &Builder, Value *V, + bool Before = true) { + if (auto *PHI = dyn_cast(V)) { + Builder.SetInsertPoint(&*PHI->getParent()->getFirstInsertionPt()); + return; + } + if (auto *I = dyn_cast(V)) { + if (!Before) + I = &*std::next(I->getIterator()); + Builder.SetInsertPoint(I); + return; + } + if (auto *A = dyn_cast(V)) { + // Set the insertion point in the entry block. + BasicBlock &Entry = A->getParent()->getEntryBlock(); + Builder.SetInsertPoint(&*Entry.getFirstInsertionPt()); + return; + } + // Otherwise, this is a constant and we don't need to set a new + // insertion point. + assert(isa(V) && "Setting insertion point for unknown value!"); + } + + /// Returns a re-written value of Start as an indexed GEP using Base as a + /// pointer. + static Value *rewriteGEPAsOffset(Value *Start, Value *Base, + const DataLayout &DL, + SetVector &Explored) { + // Perform all the substitutions. This is a bit tricky because we can + // have cycles in our use-def chains. + // 1. Create the PHI nodes without any incoming values. + // 2. Create all the other values. + // 3. Add the edges for the PHI nodes. + // 4. Emit GEPs to get the original pointers. + // 5. Remove the original instructions. + Type *IndexType = IntegerType::get( + Base->getContext(), DL.getPointerTypeSizeInBits(Start->getType())); + + DenseMap NewInsts; + NewInsts[Base] = ConstantInt::getNullValue(IndexType); + + // Create the new PHI nodes, without adding any incoming values. + for (Value *Val : Explored) { + if (Val == Base) + continue; + // Create empty phi nodes. This avoids cyclic dependencies when creating + // the remaining instructions. + if (auto *PHI = dyn_cast(Val)) + NewInsts[PHI] = PHINode::Create(IndexType, PHI->getNumIncomingValues(), + PHI->getName() + ".idx", PHI); + } + IRBuilder<> Builder(Base->getContext()); + + // Create all the other instructions. + for (Value *Val : Explored) { + + if (NewInsts.find(Val) != NewInsts.end()) + continue; + + if (auto *CI = dyn_cast(Val)) { + NewInsts[CI] = NewInsts[CI->getOperand(0)]; + continue; + } + if (auto *GEP = dyn_cast(Val)) { + Value *Index = NewInsts[GEP->getOperand(1)] ? NewInsts[GEP->getOperand(1)] + : GEP->getOperand(1); + setInsertionPoint(Builder, GEP); + // Indices might need to be sign extended. GEPs will magically do + // this, but we need to do it ourselves here. + if (Index->getType()->getScalarSizeInBits() != + NewInsts[GEP->getOperand(0)]->getType()->getScalarSizeInBits()) { + Index = Builder.CreateSExtOrTrunc( + Index, NewInsts[GEP->getOperand(0)]->getType(), + GEP->getOperand(0)->getName() + ".sext"); + } + + auto *Op = NewInsts[GEP->getOperand(0)]; + if (isa(Op) && dyn_cast(Op)->isZero()) + NewInsts[GEP] = Index; + else + NewInsts[GEP] = Builder.CreateNSWAdd( + Op, Index, GEP->getOperand(0)->getName() + ".add"); + continue; + } + if (isa(Val)) + continue; + + llvm_unreachable("Unexpected instruction type"); + } + + // Add the incoming values to the PHI nodes. + for (Value *Val : Explored) { + if (Val == Base) + continue; + // All the instructions have been created, we can now add edges to the + // phi nodes. + if (auto *PHI = dyn_cast(Val)) { + PHINode *NewPhi = static_cast(NewInsts[PHI]); + for (unsigned I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) { + Value *NewIncoming = PHI->getIncomingValue(I); + + if (NewInsts.find(NewIncoming) != NewInsts.end()) + NewIncoming = NewInsts[NewIncoming]; + + NewPhi->addIncoming(NewIncoming, PHI->getIncomingBlock(I)); + } + } + } + + for (Value *Val : Explored) { + if (Val == Base) + continue; + + // Depending on the type, for external users we have to emit + // a GEP or a GEP + ptrtoint. + setInsertionPoint(Builder, Val, false); + + // If required, create an inttoptr instruction for Base. + Value *NewBase = Base; + if (!Base->getType()->isPointerTy()) + NewBase = Builder.CreateBitOrPointerCast(Base, Start->getType(), + Start->getName() + "to.ptr"); + + Value *GEP = Builder.CreateInBoundsGEP( + Start->getType()->getPointerElementType(), NewBase, + makeArrayRef(NewInsts[Val]), Val->getName() + ".ptr"); + + if (!Val->getType()->isPointerTy()) { + Value *Cast = Builder.CreatePointerCast(GEP, Val->getType(), + Val->getName() + ".conv"); + GEP = Cast; + } + Val->replaceAllUsesWith(GEP); + } + + return NewInsts[Start]; + } + + /// Looks through GEPs, IntToPtrInsts and PtrToIntInsts in order to express + /// the input Value as a constant indexed GEP. Returns a pair containing + /// the GEPs Pointer and Index. + static std::pair + getAsConstantIndexedAddress(Value *V, const DataLayout &DL) { + Type *IndexType = IntegerType::get(V->getContext(), + DL.getPointerTypeSizeInBits(V->getType())); + + Constant *Index = ConstantInt::getNullValue(IndexType); + while (true) { + if (GEPOperator *GEP = dyn_cast(V)) { + // We accept only inbouds GEPs here to exclude the possibility of + // overflow. + if (!GEP->isInBounds()) + break; + if (GEP->hasAllConstantIndices() && GEP->getNumIndices() == 1 && + GEP->getType() == V->getType()) { + V = GEP->getOperand(0); + Constant *GEPIndex = static_cast(GEP->getOperand(1)); + Index = ConstantExpr::getAdd( + Index, ConstantExpr::getSExtOrBitCast(GEPIndex, IndexType)); + continue; + } + break; + } + if (auto *CI = dyn_cast(V)) { + if (!CI->isNoopCast(DL)) + break; + V = CI->getOperand(0); + continue; + } + if (auto *CI = dyn_cast(V)) { + if (!CI->isNoopCast(DL)) + break; + V = CI->getOperand(0); + continue; + } + break; + } + return {V, Index}; + } + + /// Converts (CMP GEPLHS, RHS) if this change would make RHS a constant. + /// We can look through PHIs, GEPs and casts in order to determine a common base + /// between GEPLHS and RHS. + static Instruction *transformToIndexedCompare(GEPOperator *GEPLHS, Value *RHS, + ICmpInst::Predicate Cond, + const DataLayout &DL) { + if (!GEPLHS->hasAllConstantIndices()) + return nullptr; + + Value *PtrBase, *Index; + std::tie(PtrBase, Index) = getAsConstantIndexedAddress(GEPLHS, DL); + + // The set of nodes that will take part in this transformation. + SetVector Nodes; + + if (!canRewriteGEPAsOffset(RHS, PtrBase, DL, Nodes)) + return nullptr; + + // We know we can re-write this as + // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) + // Since we've only looked through inbouds GEPs we know that we + // can't have overflow on either side. We can therefore re-write + // this as: + // OFFSET1 cmp OFFSET2 + Value *NewRHS = rewriteGEPAsOffset(RHS, PtrBase, DL, Nodes); + + // RewriteGEPAsOffset has replaced RHS and all of its uses with a re-written + // GEP having PtrBase as the pointer base, and has returned in NewRHS the + // offset. Since Index is the offset of LHS to the base pointer, we will now + // compare the offsets instead of comparing the pointers. + return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Index, NewRHS); + } + + /// Fold comparisons between a GEP instruction and something else. At this point + /// we know that the GEP is on the LHS of the comparison. + Instruction *InstCombiner::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS, + ICmpInst::Predicate Cond, + Instruction &I) { + // Don't transform signed compares of GEPs into index compares. Even if the + // GEP is inbounds, the final add of the base pointer can have signed overflow + // and would change the result of the icmp. + // e.g. "&foo[0] (RHS)) + RHS = RHS->stripPointerCasts(); + + Value *PtrBase = GEPLHS->getOperand(0); + if (PtrBase == RHS && GEPLHS->isInBounds()) { + // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0). + // This transformation (ignoring the base and scales) is valid because we + // know pointers can't overflow since the gep is inbounds. See if we can + // output an optimized form. + Value *Offset = evaluateGEPOffsetExpression(GEPLHS, *this, DL); + + // If not, synthesize the offset the hard way. + if (!Offset) + Offset = EmitGEPOffset(GEPLHS); + return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset, + Constant::getNullValue(Offset->getType())); + } else if (GEPOperator *GEPRHS = dyn_cast(RHS)) { + // If the base pointers are different, but the indices are the same, just + // compare the base pointer. + if (PtrBase != GEPRHS->getOperand(0)) { + bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands(); + IndicesTheSame &= GEPLHS->getOperand(0)->getType() == + GEPRHS->getOperand(0)->getType(); + if (IndicesTheSame) + for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i) + if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { + IndicesTheSame = false; + break; + } + + // If all indices are the same, just compare the base pointers. + if (IndicesTheSame) + return new ICmpInst(Cond, GEPLHS->getOperand(0), GEPRHS->getOperand(0)); + + // If we're comparing GEPs with two base pointers that only differ in type + // and both GEPs have only constant indices or just one use, then fold + // the compare with the adjusted indices. + if (GEPLHS->isInBounds() && GEPRHS->isInBounds() && + (GEPLHS->hasAllConstantIndices() || GEPLHS->hasOneUse()) && + (GEPRHS->hasAllConstantIndices() || GEPRHS->hasOneUse()) && + PtrBase->stripPointerCasts() == + GEPRHS->getOperand(0)->stripPointerCasts()) { + Value *LOffset = EmitGEPOffset(GEPLHS); + Value *ROffset = EmitGEPOffset(GEPRHS); + + // If we looked through an addrspacecast between different sized address + // spaces, the LHS and RHS pointers are different sized + // integers. Truncate to the smaller one. + Type *LHSIndexTy = LOffset->getType(); + Type *RHSIndexTy = ROffset->getType(); + if (LHSIndexTy != RHSIndexTy) { + if (LHSIndexTy->getPrimitiveSizeInBits() < + RHSIndexTy->getPrimitiveSizeInBits()) { + ROffset = Builder->CreateTrunc(ROffset, LHSIndexTy); + } else + LOffset = Builder->CreateTrunc(LOffset, RHSIndexTy); + } + + Value *Cmp = Builder->CreateICmp(ICmpInst::getSignedPredicate(Cond), + LOffset, ROffset); + return replaceInstUsesWith(I, Cmp); + } + + // Otherwise, the base pointers are different and the indices are + // different. Try convert this to an indexed compare by looking through + // PHIs/casts. + return transformToIndexedCompare(GEPLHS, RHS, Cond, DL); + } + + // If one of the GEPs has all zero indices, recurse. + if (GEPLHS->hasAllZeroIndices()) + return foldGEPICmp(GEPRHS, GEPLHS->getOperand(0), + ICmpInst::getSwappedPredicate(Cond), I); + + // If the other GEP has all zero indices, recurse. + if (GEPRHS->hasAllZeroIndices()) + return foldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I); + + bool GEPsInBounds = GEPLHS->isInBounds() && GEPRHS->isInBounds(); + if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) { + // If the GEPs only differ by one index, compare it. + unsigned NumDifferences = 0; // Keep track of # differences. + unsigned DiffOperand = 0; // The operand that differs. + for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i) + if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) { + if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() != + GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) { + // Irreconcilable differences. + NumDifferences = 2; + break; + } else { + if (NumDifferences++) break; + DiffOperand = i; + } + } + + if (NumDifferences == 0) // SAME GEP? + return replaceInstUsesWith(I, // No comparison is needed here. + Builder->getInt1(ICmpInst::isTrueWhenEqual(Cond))); + + else if (NumDifferences == 1 && GEPsInBounds) { + Value *LHSV = GEPLHS->getOperand(DiffOperand); + Value *RHSV = GEPRHS->getOperand(DiffOperand); + // Make sure we do a signed comparison here. + return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV); + } + } + + // Only lower this if the icmp is the only user of the GEP or if we expect + // the result to fold to a constant! + if (GEPsInBounds && (isa(GEPLHS) || GEPLHS->hasOneUse()) && + (isa(GEPRHS) || GEPRHS->hasOneUse())) { + // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2) + Value *L = EmitGEPOffset(GEPLHS); + Value *R = EmitGEPOffset(GEPRHS); + return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R); + } + } + + // Try convert this to an indexed compare by looking through PHIs/casts as a + // last resort. + return transformToIndexedCompare(GEPLHS, RHS, Cond, DL); + } + + Instruction *InstCombiner::foldAllocaCmp(ICmpInst &ICI, + const AllocaInst *Alloca, + const Value *Other) { + assert(ICI.isEquality() && "Cannot fold non-equality comparison."); + + // It would be tempting to fold away comparisons between allocas and any + // pointer not based on that alloca (e.g. an argument). However, even + // though such pointers cannot alias, they can still compare equal. + // + // But LLVM doesn't specify where allocas get their memory, so if the alloca + // doesn't escape we can argue that it's impossible to guess its value, and we + // can therefore act as if any such guesses are wrong. + // + // The code below checks that the alloca doesn't escape, and that it's only + // used in a comparison once (the current instruction). The + // single-comparison-use condition ensures that we're trivially folding all + // comparisons against the alloca consistently, and avoids the risk of + // erroneously folding a comparison of the pointer with itself. + + unsigned MaxIter = 32; // Break cycles and bound to constant-time. + + SmallVector Worklist; + for (const Use &U : Alloca->uses()) { + if (Worklist.size() >= MaxIter) + return nullptr; + Worklist.push_back(&U); + } + + unsigned NumCmps = 0; + while (!Worklist.empty()) { + assert(Worklist.size() <= MaxIter); + const Use *U = Worklist.pop_back_val(); + const Value *V = U->getUser(); + --MaxIter; + + if (isa(V) || isa(V) || isa(V) || + isa(V)) { + // Track the uses. + } else if (isa(V)) { + // Loading from the pointer doesn't escape it. + continue; + } else if (const auto *SI = dyn_cast(V)) { + // Storing *to* the pointer is fine, but storing the pointer escapes it. + if (SI->getValueOperand() == U->get()) + return nullptr; + continue; + } else if (isa(V)) { + if (NumCmps++) + return nullptr; // Found more than one cmp. + continue; + } else if (const auto *Intrin = dyn_cast(V)) { + switch (Intrin->getIntrinsicID()) { + // These intrinsics don't escape or compare the pointer. Memset is safe + // because we don't allow ptrtoint. Memcpy and memmove are safe because + // we don't allow stores, so src cannot point to V. + case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: + case Intrinsic::dbg_declare: case Intrinsic::dbg_value: + case Intrinsic::memcpy: case Intrinsic::memmove: case Intrinsic::memset: + continue; + default: + return nullptr; + } + } else { + return nullptr; + } + for (const Use &U : V->uses()) { + if (Worklist.size() >= MaxIter) + return nullptr; + Worklist.push_back(&U); + } + } + + Type *CmpTy = CmpInst::makeCmpResultType(Other->getType()); + return replaceInstUsesWith( + ICI, + ConstantInt::get(CmpTy, !CmpInst::isTrueWhenEqual(ICI.getPredicate()))); + } + + /// Fold "icmp pred (X+CI), X". + Instruction *InstCombiner::foldICmpAddOpConst(Instruction &ICI, + Value *X, ConstantInt *CI, + ICmpInst::Predicate Pred) { + // From this point on, we know that (X+C <= X) --> (X+C < X) because C != 0, + // so the values can never be equal. Similarly for all other "or equals" + // operators. + + // (X+1) X >u (MAXUINT-1) --> X == 255 + // (X+2) X >u (MAXUINT-2) --> X > 253 + // (X+MAXUINT) X >u (MAXUINT-MAXUINT) --> X != 0 + if (Pred == ICmpInst::ICMP_ULT || Pred == ICmpInst::ICMP_ULE) { + Value *R = + ConstantExpr::getSub(ConstantInt::getAllOnesValue(CI->getType()), CI); + return new ICmpInst(ICmpInst::ICMP_UGT, X, R); + } + + // (X+1) >u X --> X X != 255 + // (X+2) >u X --> X X u X --> X X X == 0 + if (Pred == ICmpInst::ICMP_UGT || Pred == ICmpInst::ICMP_UGE) + return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantExpr::getNeg(CI)); + + unsigned BitWidth = CI->getType()->getPrimitiveSizeInBits(); + ConstantInt *SMax = ConstantInt::get(X->getContext(), + APInt::getSignedMaxValue(BitWidth)); + + // (X+ 1) X >s (MAXSINT-1) --> X == 127 + // (X+ 2) X >s (MAXSINT-2) --> X >s 125 + // (X+MAXSINT) X >s (MAXSINT-MAXSINT) --> X >s 0 + // (X+MINSINT) X >s (MAXSINT-MINSINT) --> X >s -1 + // (X+ -2) X >s (MAXSINT- -2) --> X >s 126 + // (X+ -1) X >s (MAXSINT- -1) --> X != 127 + if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) + return new ICmpInst(ICmpInst::ICMP_SGT, X, ConstantExpr::getSub(SMax, CI)); + + // (X+ 1) >s X --> X X != 127 + // (X+ 2) >s X --> X X s X --> X X s X --> X X s X --> X X s X --> X X == -128 + + assert(Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE); + Constant *C = Builder->getInt(CI->getValue()-1); + return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantExpr::getSub(SMax, C)); + } + + /// Handle "(icmp eq/ne (ashr/lshr const2, A), const1)" -> + /// (icmp eq/ne A, Log2(const2/const1)) -> + /// (icmp eq/ne A, Log2(const2) - Log2(const1)). + Instruction *InstCombiner::foldICmpCstShrConst(ICmpInst &I, Value *Op, Value *A, + ConstantInt *CI1, + ConstantInt *CI2) { + assert(I.isEquality() && "Cannot fold icmp gt/lt"); + + auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { + if (I.getPredicate() == I.ICMP_NE) + Pred = CmpInst::getInversePredicate(Pred); + return new ICmpInst(Pred, LHS, RHS); + }; + + const APInt &AP1 = CI1->getValue(); + const APInt &AP2 = CI2->getValue(); + + // Don't bother doing any work for cases which InstSimplify handles. + if (AP2 == 0) + return nullptr; + bool IsAShr = isa(Op); + if (IsAShr) { + if (AP2.isAllOnesValue()) + return nullptr; + if (AP2.isNegative() != AP1.isNegative()) + return nullptr; + if (AP2.sgt(AP1)) + return nullptr; + } + + if (!AP1) + // 'A' must be large enough to shift out the highest set bit. + return getICmp(I.ICMP_UGT, A, + ConstantInt::get(A->getType(), AP2.logBase2())); + + if (AP1 == AP2) + return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType())); + + int Shift; + if (IsAShr && AP1.isNegative()) + Shift = AP1.countLeadingOnes() - AP2.countLeadingOnes(); + else + Shift = AP1.countLeadingZeros() - AP2.countLeadingZeros(); + + if (Shift > 0) { + if (IsAShr && AP1 == AP2.ashr(Shift)) { + // There are multiple solutions if we are comparing against -1 and the LHS + // of the ashr is not a power of two. + if (AP1.isAllOnesValue() && !AP2.isPowerOf2()) + return getICmp(I.ICMP_UGE, A, ConstantInt::get(A->getType(), Shift)); + return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); + } else if (AP1 == AP2.lshr(Shift)) { + return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); + } + } + + // Shifting const2 will never be equal to const1. + // FIXME: This should always be handled by InstSimplify? + auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE); + return replaceInstUsesWith(I, TorF); + } + + /// Handle "(icmp eq/ne (shl const2, A), const1)" -> + /// (icmp eq/ne A, TrailingZeros(const1) - TrailingZeros(const2)). + Instruction *InstCombiner::foldICmpCstShlConst(ICmpInst &I, Value *Op, Value *A, + ConstantInt *CI1, + ConstantInt *CI2) { + assert(I.isEquality() && "Cannot fold icmp gt/lt"); + + auto getICmp = [&I](CmpInst::Predicate Pred, Value *LHS, Value *RHS) { + if (I.getPredicate() == I.ICMP_NE) + Pred = CmpInst::getInversePredicate(Pred); + return new ICmpInst(Pred, LHS, RHS); + }; + + const APInt &AP1 = CI1->getValue(); + const APInt &AP2 = CI2->getValue(); + + // Don't bother doing any work for cases which InstSimplify handles. + if (AP2 == 0) + return nullptr; + + unsigned AP2TrailingZeros = AP2.countTrailingZeros(); + + if (!AP1 && AP2TrailingZeros != 0) + return getICmp(I.ICMP_UGE, A, + ConstantInt::get(A->getType(), AP2.getBitWidth() - AP2TrailingZeros)); + + if (AP1 == AP2) + return getICmp(I.ICMP_EQ, A, ConstantInt::getNullValue(A->getType())); + + // Get the distance between the lowest bits that are set. + int Shift = AP1.countTrailingZeros() - AP2TrailingZeros; + + if (Shift > 0 && AP2.shl(Shift) == AP1) + return getICmp(I.ICMP_EQ, A, ConstantInt::get(A->getType(), Shift)); + + // Shifting const2 will never be equal to const1. + // FIXME: This should always be handled by InstSimplify? + auto *TorF = ConstantInt::get(I.getType(), I.getPredicate() == I.ICMP_NE); + return replaceInstUsesWith(I, TorF); + } + + /// The caller has matched a pattern of the form: + /// I = icmp ugt (add (add A, B), CI2), CI1 + /// If this is of the form: + /// sum = a + b + /// if (sum+128 >u 255) + /// Then replace it with llvm.sadd.with.overflow.i8. + /// + static Instruction *processUGT_ADDCST_ADD(ICmpInst &I, Value *A, Value *B, + ConstantInt *CI2, ConstantInt *CI1, + InstCombiner &IC) { + // The transformation we're trying to do here is to transform this into an + // llvm.sadd.with.overflow. To do this, we have to replace the original add + // with a narrower add, and discard the add-with-constant that is part of the + // range check (if we can't eliminate it, this isn't profitable). + + // In order to eliminate the add-with-constant, the compare can be its only + // use. + Instruction *AddWithCst = cast(I.getOperand(0)); + if (!AddWithCst->hasOneUse()) + return nullptr; + + // If CI2 is 2^7, 2^15, 2^31, then it might be an sadd.with.overflow. + if (!CI2->getValue().isPowerOf2()) + return nullptr; + unsigned NewWidth = CI2->getValue().countTrailingZeros(); + if (NewWidth != 7 && NewWidth != 15 && NewWidth != 31) + return nullptr; + + // The width of the new add formed is 1 more than the bias. + ++NewWidth; + + // Check to see that CI1 is an all-ones value with NewWidth bits. + if (CI1->getBitWidth() == NewWidth || + CI1->getValue() != APInt::getLowBitsSet(CI1->getBitWidth(), NewWidth)) + return nullptr; + + // This is only really a signed overflow check if the inputs have been + // sign-extended; check for that condition. For example, if CI2 is 2^31 and + // the operands of the add are 64 bits wide, we need at least 33 sign bits. + unsigned NeededSignBits = CI1->getBitWidth() - NewWidth + 1; + if (IC.ComputeNumSignBits(A, 0, &I) < NeededSignBits || + IC.ComputeNumSignBits(B, 0, &I) < NeededSignBits) + return nullptr; + + // In order to replace the original add with a narrower + // llvm.sadd.with.overflow, the only uses allowed are the add-with-constant + // and truncates that discard the high bits of the add. Verify that this is + // the case. + Instruction *OrigAdd = cast(AddWithCst->getOperand(0)); + for (User *U : OrigAdd->users()) { + if (U == AddWithCst) + continue; + + // Only accept truncates for now. We would really like a nice recursive + // predicate like SimplifyDemandedBits, but which goes downwards the use-def + // chain to see which bits of a value are actually demanded. If the + // original add had another add which was then immediately truncated, we + // could still do the transformation. + TruncInst *TI = dyn_cast(U); + if (!TI || TI->getType()->getPrimitiveSizeInBits() > NewWidth) + return nullptr; + } + + // If the pattern matches, truncate the inputs to the narrower type and + // use the sadd_with_overflow intrinsic to efficiently compute both the + // result and the overflow bit. + Type *NewType = IntegerType::get(OrigAdd->getContext(), NewWidth); + Value *F = Intrinsic::getDeclaration(I.getModule(), + Intrinsic::sadd_with_overflow, NewType); + + InstCombiner::BuilderTy *Builder = IC.Builder; + + // Put the new code above the original add, in case there are any uses of the + // add between the add and the compare. + Builder->SetInsertPoint(OrigAdd); + + Value *TruncA = Builder->CreateTrunc(A, NewType, A->getName() + ".trunc"); + Value *TruncB = Builder->CreateTrunc(B, NewType, B->getName() + ".trunc"); + CallInst *Call = Builder->CreateCall(F, {TruncA, TruncB}, "sadd"); + Value *Add = Builder->CreateExtractValue(Call, 0, "sadd.result"); + Value *ZExt = Builder->CreateZExt(Add, OrigAdd->getType()); + + // The inner add was the result of the narrow add, zero extended to the + // wider type. Replace it with the result computed by the intrinsic. + IC.replaceInstUsesWith(*OrigAdd, ZExt); + + // The original icmp gets replaced with the overflow value. + return ExtractValueInst::Create(Call, 1, "sadd.overflow"); + } + + // Fold icmp Pred X, C. + Instruction *InstCombiner::foldICmpWithConstant(ICmpInst &Cmp) { + CmpInst::Predicate Pred = Cmp.getPredicate(); + Value *X = Cmp.getOperand(0); + + const APInt *C; + if (!match(Cmp.getOperand(1), m_APInt(C))) + return nullptr; + + Value *A = nullptr, *B = nullptr; + + // Match the following pattern, which is a common idiom when writing + // overflow-safe integer arithmetic functions. The source performs an addition + // in wider type and explicitly checks for overflow using comparisons against + // INT_MIN and INT_MAX. Simplify by using the sadd_with_overflow intrinsic. + // + // TODO: This could probably be generalized to handle other overflow-safe + // operations if we worked out the formulas to compute the appropriate magic + // constants. + // + // sum = a + b + // if (sum+128 >u 255) ... -> llvm.sadd.with.overflow.i8 + { + ConstantInt *CI2; // I = icmp ugt (add (add A, B), CI2), CI + if (Pred == ICmpInst::ICMP_UGT && + match(X, m_Add(m_Add(m_Value(A), m_Value(B)), m_ConstantInt(CI2)))) + if (Instruction *Res = processUGT_ADDCST_ADD( + Cmp, A, B, CI2, cast(Cmp.getOperand(1)), *this)) + return Res; + } + + // (icmp sgt smin(PosA, B) 0) -> (icmp sgt B 0) + if (*C == 0 && Pred == ICmpInst::ICMP_SGT) { + SelectPatternResult SPR = matchSelectPattern(X, A, B); + if (SPR.Flavor == SPF_SMIN) { + if (isKnownPositive(A, DL)) + return new ICmpInst(Pred, B, Cmp.getOperand(1)); + if (isKnownPositive(B, DL)) + return new ICmpInst(Pred, A, Cmp.getOperand(1)); + } + } + + // FIXME: Use m_APInt to allow folds for splat constants. + ConstantInt *CI = dyn_cast(Cmp.getOperand(1)); + if (!CI) + return nullptr; + + if (Cmp.isEquality()) { + ConstantInt *CI2; + if (match(X, m_AShr(m_ConstantInt(CI2), m_Value(A))) || + match(X, m_LShr(m_ConstantInt(CI2), m_Value(A)))) { + // (icmp eq/ne (ashr/lshr const2, A), const1) + if (Instruction *Inst = foldICmpCstShrConst(Cmp, X, A, CI, CI2)) + return Inst; + } + if (match(X, m_Shl(m_ConstantInt(CI2), m_Value(A)))) { + // (icmp eq/ne (shl const2, A), const1) + if (Instruction *Inst = foldICmpCstShlConst(Cmp, X, A, CI, CI2)) + return Inst; + } + } + + // Canonicalize icmp instructions based on dominating conditions. + BasicBlock *Parent = Cmp.getParent(); + BasicBlock *Dom = Parent->getSinglePredecessor(); + auto *BI = Dom ? dyn_cast(Dom->getTerminator()) : nullptr; + ICmpInst::Predicate Pred2; + BasicBlock *TrueBB, *FalseBB; + ConstantInt *CI2; + if (BI && match(BI, m_Br(m_ICmp(Pred2, m_Specific(X), m_ConstantInt(CI2)), + TrueBB, FalseBB)) && + TrueBB != FalseBB) { + ConstantRange CR = + ConstantRange::makeAllowedICmpRegion(Pred, CI->getValue()); + ConstantRange DominatingCR = + (Parent == TrueBB) + ? ConstantRange::makeExactICmpRegion(Pred2, CI2->getValue()) + : ConstantRange::makeExactICmpRegion( + CmpInst::getInversePredicate(Pred2), CI2->getValue()); + ConstantRange Intersection = DominatingCR.intersectWith(CR); + ConstantRange Difference = DominatingCR.difference(CR); + if (Intersection.isEmptySet()) + return replaceInstUsesWith(Cmp, Builder->getFalse()); + if (Difference.isEmptySet()) + return replaceInstUsesWith(Cmp, Builder->getTrue()); + + // If this is a normal comparison, it demands all bits. If it is a sign + // bit comparison, it only demands the sign bit. + bool UnusedBit; + bool IsSignBit = isSignBitCheck(Pred, CI->getValue(), UnusedBit); + + // Canonicalizing a sign bit comparison that gets used in a branch, + // pessimizes codegen by generating branch on zero instruction instead + // of a test and branch. So we avoid canonicalizing in such situations + // because test and branch instruction has better branch displacement + // than compare and branch instruction. + if (!isBranchOnSignBitCheck(Cmp, IsSignBit) && !Cmp.isEquality()) { + if (auto *AI = Intersection.getSingleElement()) + return new ICmpInst(ICmpInst::ICMP_EQ, X, Builder->getInt(*AI)); + if (auto *AD = Difference.getSingleElement()) + return new ICmpInst(ICmpInst::ICMP_NE, X, Builder->getInt(*AD)); + } + } + + return nullptr; + } + + /// Fold icmp (trunc X, Y), C. + Instruction *InstCombiner::foldICmpTruncConstant(ICmpInst &Cmp, + Instruction *Trunc, + const APInt *C) { + ICmpInst::Predicate Pred = Cmp.getPredicate(); + Value *X = Trunc->getOperand(0); + if (*C == 1 && C->getBitWidth() > 1) { + // icmp slt trunc(signum(V)) 1 --> icmp slt V, 1 + Value *V = nullptr; + if (Pred == ICmpInst::ICMP_SLT && match(X, m_Signum(m_Value(V)))) + return new ICmpInst(ICmpInst::ICMP_SLT, V, + ConstantInt::get(V->getType(), 1)); + } + + if (Cmp.isEquality() && Trunc->hasOneUse()) { + // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all + // of the high bits truncated out of x are known. + unsigned DstBits = Trunc->getType()->getScalarSizeInBits(), + SrcBits = X->getType()->getScalarSizeInBits(); + APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0); + computeKnownBits(X, KnownZero, KnownOne, 0, &Cmp); + + // If all the high bits are known, we can do this xform. + if ((KnownZero | KnownOne).countLeadingOnes() >= SrcBits - DstBits) { + // Pull in the high bits from known-ones set. + APInt NewRHS = C->zext(SrcBits); + NewRHS |= KnownOne & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits); + return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), NewRHS)); + } + } + + return nullptr; + } + + /// Fold icmp (xor X, Y), C. + Instruction *InstCombiner::foldICmpXorConstant(ICmpInst &Cmp, + BinaryOperator *Xor, + const APInt *C) { + Value *X = Xor->getOperand(0); + Value *Y = Xor->getOperand(1); + const APInt *XorC; + if (!match(Y, m_APInt(XorC))) + return nullptr; + + // If this is a comparison that tests the signbit (X < 0) or (x > -1), + // fold the xor. + ICmpInst::Predicate Pred = Cmp.getPredicate(); + if ((Pred == ICmpInst::ICMP_SLT && *C == 0) || + (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue())) { + + // If the sign bit of the XorCst is not set, there is no change to + // the operation, just stop using the Xor. + if (!XorC->isNegative()) { + Cmp.setOperand(0, X); + Worklist.Add(Xor); + return &Cmp; + } + + // Was the old condition true if the operand is positive? + bool isTrueIfPositive = Pred == ICmpInst::ICMP_SGT; + + // If so, the new one isn't. + isTrueIfPositive ^= true; + + Constant *CmpConstant = cast(Cmp.getOperand(1)); + if (isTrueIfPositive) + return new ICmpInst(ICmpInst::ICMP_SGT, X, SubOne(CmpConstant)); + else + return new ICmpInst(ICmpInst::ICMP_SLT, X, AddOne(CmpConstant)); + } + + if (Xor->hasOneUse()) { + // (icmp u/s (xor X SignBit), C) -> (icmp s/u X, (xor C SignBit)) + if (!Cmp.isEquality() && XorC->isSignBit()) { + Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate() + : Cmp.getSignedPredicate(); + return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC)); + } + + // (icmp u/s (xor X ~SignBit), C) -> (icmp s/u X, (xor C ~SignBit)) + if (!Cmp.isEquality() && XorC->isMaxSignedValue()) { + Pred = Cmp.isSigned() ? Cmp.getUnsignedPredicate() + : Cmp.getSignedPredicate(); + Pred = Cmp.getSwappedPredicate(Pred); + return new ICmpInst(Pred, X, ConstantInt::get(X->getType(), *C ^ *XorC)); + } + } + + // (icmp ugt (xor X, C), ~C) -> (icmp ult X, C) + // iff -C is a power of 2 + if (Pred == ICmpInst::ICMP_UGT && *XorC == ~(*C) && (*C + 1).isPowerOf2()) + return new ICmpInst(ICmpInst::ICMP_ULT, X, Y); + + // (icmp ult (xor X, C), -C) -> (icmp uge X, C) + // iff -C is a power of 2 + if (Pred == ICmpInst::ICMP_ULT && *XorC == -(*C) && C->isPowerOf2()) + return new ICmpInst(ICmpInst::ICMP_UGE, X, Y); + + return nullptr; + } + + /// Fold icmp (and (sh X, Y), C2), C1. + Instruction *InstCombiner::foldICmpAndShift(ICmpInst &Cmp, BinaryOperator *And, + const APInt *C1, const APInt *C2) { + BinaryOperator *Shift = dyn_cast(And->getOperand(0)); + if (!Shift || !Shift->isShift()) + return nullptr; + + // If this is: (X >> C3) & C2 != C1 (where any shift and any compare could + // exist), turn it into (X & (C2 << C3)) != (C1 << C3). This happens a LOT in + // code produced by the clang front-end, for bitfield access. + // This seemingly simple opportunity to fold away a shift turns out to be + // rather complicated. See PR17827 for details. + unsigned ShiftOpcode = Shift->getOpcode(); + bool IsShl = ShiftOpcode == Instruction::Shl; + const APInt *C3; + if (match(Shift->getOperand(1), m_APInt(C3))) { + bool CanFold = false; + if (ShiftOpcode == Instruction::AShr) { + // There may be some constraints that make this possible, but nothing + // simple has been discovered yet. + CanFold = false; + } else if (ShiftOpcode == Instruction::Shl) { + // For a left shift, we can fold if the comparison is not signed. We can + // also fold a signed comparison if the mask value and comparison value + // are not negative. These constraints may not be obvious, but we can + // prove that they are correct using an SMT solver. + if (!Cmp.isSigned() || (!C2->isNegative() && !C1->isNegative())) + CanFold = true; + } else if (ShiftOpcode == Instruction::LShr) { + // For a logical right shift, we can fold if the comparison is not signed. + // We can also fold a signed comparison if the shifted mask value and the + // shifted comparison value are not negative. These constraints may not be + // obvious, but we can prove that they are correct using an SMT solver. + if (!Cmp.isSigned() || + (!C2->shl(*C3).isNegative() && !C1->shl(*C3).isNegative())) + CanFold = true; + } + + if (CanFold) { + APInt NewCst = IsShl ? C1->lshr(*C3) : C1->shl(*C3); + APInt SameAsC1 = IsShl ? NewCst.shl(*C3) : NewCst.lshr(*C3); + // Check to see if we are shifting out any of the bits being compared. + if (SameAsC1 != *C1) { + // If we shifted bits out, the fold is not going to work out. As a + // special case, check to see if this means that the result is always + // true or false now. + if (Cmp.getPredicate() == ICmpInst::ICMP_EQ) + return replaceInstUsesWith(Cmp, ConstantInt::getFalse(Cmp.getType())); + if (Cmp.getPredicate() == ICmpInst::ICMP_NE) + return replaceInstUsesWith(Cmp, ConstantInt::getTrue(Cmp.getType())); + } else { + Cmp.setOperand(1, ConstantInt::get(And->getType(), NewCst)); + APInt NewAndCst = IsShl ? C2->lshr(*C3) : C2->shl(*C3); + And->setOperand(1, ConstantInt::get(And->getType(), NewAndCst)); + And->setOperand(0, Shift->getOperand(0)); + Worklist.Add(Shift); // Shift is dead. + return &Cmp; + } + } + } + + // Turn ((X >> Y) & C2) == 0 into (X & (C2 << Y)) == 0. The latter is + // preferable because it allows the C2 << Y expression to be hoisted out of a + // loop if Y is invariant and X is not. + if (Shift->hasOneUse() && *C1 == 0 && Cmp.isEquality() && + !Shift->isArithmeticShift() && !isa(Shift->getOperand(0))) { + // Compute C2 << Y. + Value *NewShift = + IsShl ? Builder->CreateLShr(And->getOperand(1), Shift->getOperand(1)) + : Builder->CreateShl(And->getOperand(1), Shift->getOperand(1)); + + // Compute X & (C2 << Y). + Value *NewAnd = Builder->CreateAnd(Shift->getOperand(0), NewShift); + Cmp.setOperand(0, NewAnd); + return &Cmp; + } + + return nullptr; + } + + /// Fold icmp (and X, C2), C1. + Instruction *InstCombiner::foldICmpAndConstConst(ICmpInst &Cmp, + BinaryOperator *And, + const APInt *C1) { + const APInt *C2; + if (!match(And->getOperand(1), m_APInt(C2))) + return nullptr; + + if (!And->hasOneUse() || !And->getOperand(0)->hasOneUse()) + return nullptr; + + // If the LHS is an 'and' of a truncate and we can widen the and/compare to + // the input width without changing the value produced, eliminate the cast: + // + // icmp (and (trunc W), C2), C1 -> icmp (and W, C2'), C1' + // + // We can do this transformation if the constants do not have their sign bits + // set or if it is an equality comparison. Extending a relational comparison + // when we're checking the sign bit would not work. + Value *W; + if (match(And->getOperand(0), m_Trunc(m_Value(W))) && + (Cmp.isEquality() || (!C1->isNegative() && !C2->isNegative()))) { + // TODO: Is this a good transform for vectors? Wider types may reduce + // throughput. Should this transform be limited (even for scalars) by using + // ShouldChangeType()? + if (!Cmp.getType()->isVectorTy()) { + Type *WideType = W->getType(); + unsigned WideScalarBits = WideType->getScalarSizeInBits(); + Constant *ZextC1 = ConstantInt::get(WideType, C1->zext(WideScalarBits)); + Constant *ZextC2 = ConstantInt::get(WideType, C2->zext(WideScalarBits)); + Value *NewAnd = Builder->CreateAnd(W, ZextC2, And->getName()); + return new ICmpInst(Cmp.getPredicate(), NewAnd, ZextC1); + } + } + + if (Instruction *I = foldICmpAndShift(Cmp, And, C1, C2)) + return I; + + // (icmp pred (and (or (lshr A, B), A), 1), 0) --> + // (icmp pred (and A, (or (shl 1, B), 1), 0)) + // + // iff pred isn't signed + if (!Cmp.isSigned() && *C1 == 0 && match(And->getOperand(1), m_One())) { + Constant *One = cast(And->getOperand(1)); + Value *Or = And->getOperand(0); + Value *A, *B, *LShr; + if (match(Or, m_Or(m_Value(LShr), m_Value(A))) && + match(LShr, m_LShr(m_Specific(A), m_Value(B)))) { + unsigned UsesRemoved = 0; + if (And->hasOneUse()) + ++UsesRemoved; + if (Or->hasOneUse()) + ++UsesRemoved; + if (LShr->hasOneUse()) + ++UsesRemoved; + + // Compute A & ((1 << B) | 1) + Value *NewOr = nullptr; + if (auto *C = dyn_cast(B)) { + if (UsesRemoved >= 1) + NewOr = ConstantExpr::getOr(ConstantExpr::getNUWShl(One, C), One); + } else { + if (UsesRemoved >= 3) + NewOr = Builder->CreateOr(Builder->CreateShl(One, B, LShr->getName(), + /*HasNUW=*/true), + One, Or->getName()); + } + if (NewOr) { + Value *NewAnd = Builder->CreateAnd(A, NewOr, And->getName()); + Cmp.setOperand(0, NewAnd); + return &Cmp; + } + } + } + + // (X & C2) > C1 --> (X & C2) != 0, if any bit set in (X & C2) will produce a + // result greater than C1. + unsigned NumTZ = C2->countTrailingZeros(); + if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && NumTZ < C2->getBitWidth() && + APInt::getOneBitSet(C2->getBitWidth(), NumTZ).ugt(*C1)) { + Constant *Zero = Constant::getNullValue(And->getType()); + return new ICmpInst(ICmpInst::ICMP_NE, And, Zero); + } + + return nullptr; + } + + /// Fold icmp (and X, Y), C. + Instruction *InstCombiner::foldICmpAndConstant(ICmpInst &Cmp, + BinaryOperator *And, + const APInt *C) { + if (Instruction *I = foldICmpAndConstConst(Cmp, And, C)) + return I; + + // TODO: These all require that Y is constant too, so refactor with the above. + + // Try to optimize things like "A[i] & 42 == 0" to index computations. + Value *X = And->getOperand(0); + Value *Y = And->getOperand(1); + if (auto *LI = dyn_cast(X)) + if (auto *GEP = dyn_cast(LI->getOperand(0))) + if (auto *GV = dyn_cast(GEP->getOperand(0))) + if (GV->isConstant() && GV->hasDefinitiveInitializer() && + !LI->isVolatile() && isa(Y)) { + ConstantInt *C2 = cast(Y); + if (Instruction *Res = foldCmpLoadFromIndexedGlobal(GEP, GV, Cmp, C2)) + return Res; + } + + if (!Cmp.isEquality()) + return nullptr; + + // X & -C == -C -> X > u ~C + // X & -C != -C -> X <= u ~C + // iff C is a power of 2 + if (Cmp.getOperand(1) == Y && (-(*C)).isPowerOf2()) { + auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_UGT + : CmpInst::ICMP_ULE; + return new ICmpInst(NewPred, X, SubOne(cast(Cmp.getOperand(1)))); + } + + // (X & C2) == 0 -> (trunc X) >= 0 + // (X & C2) != 0 -> (trunc X) < 0 + // iff C2 is a power of 2 and it masks the sign bit of a legal integer type. + const APInt *C2; + if (And->hasOneUse() && *C == 0 && match(Y, m_APInt(C2))) { + int32_t ExactLogBase2 = C2->exactLogBase2(); + if (ExactLogBase2 != -1 && DL.isLegalInteger(ExactLogBase2 + 1)) { + Type *NTy = IntegerType::get(Cmp.getContext(), ExactLogBase2 + 1); + if (And->getType()->isVectorTy()) + NTy = VectorType::get(NTy, And->getType()->getVectorNumElements()); + Value *Trunc = Builder->CreateTrunc(X, NTy); + auto NewPred = Cmp.getPredicate() == CmpInst::ICMP_EQ ? CmpInst::ICMP_SGE + : CmpInst::ICMP_SLT; + return new ICmpInst(NewPred, Trunc, Constant::getNullValue(NTy)); + } + } + + return nullptr; + } + + /// Fold icmp (or X, Y), C. + Instruction *InstCombiner::foldICmpOrConstant(ICmpInst &Cmp, BinaryOperator *Or, + const APInt *C) { + ICmpInst::Predicate Pred = Cmp.getPredicate(); + if (*C == 1) { + // icmp slt signum(V) 1 --> icmp slt V, 1 + Value *V = nullptr; + if (Pred == ICmpInst::ICMP_SLT && match(Or, m_Signum(m_Value(V)))) + return new ICmpInst(ICmpInst::ICMP_SLT, V, + ConstantInt::get(V->getType(), 1)); + } + + if (!Cmp.isEquality() || *C != 0 || !Or->hasOneUse()) + return nullptr; + + Value *P, *Q; + if (match(Or, m_Or(m_PtrToInt(m_Value(P)), m_PtrToInt(m_Value(Q))))) { + // Simplify icmp eq (or (ptrtoint P), (ptrtoint Q)), 0 + // -> and (icmp eq P, null), (icmp eq Q, null). + Value *CmpP = + Builder->CreateICmp(Pred, P, ConstantInt::getNullValue(P->getType())); + Value *CmpQ = + Builder->CreateICmp(Pred, Q, ConstantInt::getNullValue(Q->getType())); + auto LogicOpc = Pred == ICmpInst::Predicate::ICMP_EQ ? Instruction::And + : Instruction::Or; + return BinaryOperator::Create(LogicOpc, CmpP, CmpQ); + } + + return nullptr; + } + + /// Fold icmp (mul X, Y), C. + Instruction *InstCombiner::foldICmpMulConstant(ICmpInst &Cmp, + BinaryOperator *Mul, + const APInt *C) { + const APInt *MulC; + if (!match(Mul->getOperand(1), m_APInt(MulC))) + return nullptr; + + // If this is a test of the sign bit and the multiply is sign-preserving with + // a constant operand, use the multiply LHS operand instead. + ICmpInst::Predicate Pred = Cmp.getPredicate(); + if (isSignTest(Pred, *C) && Mul->hasNoSignedWrap()) { + if (MulC->isNegative()) + Pred = ICmpInst::getSwappedPredicate(Pred); + return new ICmpInst(Pred, Mul->getOperand(0), + Constant::getNullValue(Mul->getType())); + } + + return nullptr; + } + + /// Fold icmp (shl 1, Y), C. + static Instruction *foldICmpShlOne(ICmpInst &Cmp, Instruction *Shl, + const APInt *C) { + Value *Y; + if (!match(Shl, m_Shl(m_One(), m_Value(Y)))) + return nullptr; + + Type *ShiftType = Shl->getType(); + uint32_t TypeBits = C->getBitWidth(); + bool CIsPowerOf2 = C->isPowerOf2(); + ICmpInst::Predicate Pred = Cmp.getPredicate(); + if (Cmp.isUnsigned()) { + // (1 << Y) pred C -> Y pred Log2(C) + if (!CIsPowerOf2) { + // (1 << Y) < 30 -> Y <= 4 + // (1 << Y) <= 30 -> Y <= 4 + // (1 << Y) >= 30 -> Y > 4 + // (1 << Y) > 30 -> Y > 4 + if (Pred == ICmpInst::ICMP_ULT) + Pred = ICmpInst::ICMP_ULE; + else if (Pred == ICmpInst::ICMP_UGE) + Pred = ICmpInst::ICMP_UGT; + } + + // (1 << Y) >= 2147483648 -> Y >= 31 -> Y == 31 + // (1 << Y) < 2147483648 -> Y < 31 -> Y != 31 + unsigned CLog2 = C->logBase2(); + if (CLog2 == TypeBits - 1) { + if (Pred == ICmpInst::ICMP_UGE) + Pred = ICmpInst::ICMP_EQ; + else if (Pred == ICmpInst::ICMP_ULT) + Pred = ICmpInst::ICMP_NE; + } + return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, CLog2)); + } else if (Cmp.isSigned()) { + Constant *BitWidthMinusOne = ConstantInt::get(ShiftType, TypeBits - 1); + if (C->isAllOnesValue()) { + // (1 << Y) <= -1 -> Y == 31 + if (Pred == ICmpInst::ICMP_SLE) + return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne); + + // (1 << Y) > -1 -> Y != 31 + if (Pred == ICmpInst::ICMP_SGT) + return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne); + } else if (!(*C)) { + // (1 << Y) < 0 -> Y == 31 + // (1 << Y) <= 0 -> Y == 31 + if (Pred == ICmpInst::ICMP_SLT || Pred == ICmpInst::ICMP_SLE) + return new ICmpInst(ICmpInst::ICMP_EQ, Y, BitWidthMinusOne); + + // (1 << Y) >= 0 -> Y != 31 + // (1 << Y) > 0 -> Y != 31 + if (Pred == ICmpInst::ICMP_SGT || Pred == ICmpInst::ICMP_SGE) + return new ICmpInst(ICmpInst::ICMP_NE, Y, BitWidthMinusOne); + } + } else if (Cmp.isEquality() && CIsPowerOf2) { + return new ICmpInst(Pred, Y, ConstantInt::get(ShiftType, C->logBase2())); + } + + return nullptr; + } + + /// Fold icmp (shl X, Y), C. + Instruction *InstCombiner::foldICmpShlConstant(ICmpInst &Cmp, + BinaryOperator *Shl, + const APInt *C) { + const APInt *ShiftAmt; + if (!match(Shl->getOperand(1), m_APInt(ShiftAmt))) + return foldICmpShlOne(Cmp, Shl, C); + + // Check that the shift amount is in range. If not, don't perform undefined + // shifts. When the shift is visited it will be simplified. + unsigned TypeBits = C->getBitWidth(); + if (ShiftAmt->uge(TypeBits)) + return nullptr; + + ICmpInst::Predicate Pred = Cmp.getPredicate(); + Value *X = Shl->getOperand(0); + if (Cmp.isEquality()) { + // If the shift is NUW, then it is just shifting out zeros, no need for an + // AND. + Constant *LShrC = ConstantInt::get(Shl->getType(), C->lshr(*ShiftAmt)); + if (Shl->hasNoUnsignedWrap()) + return new ICmpInst(Pred, X, LShrC); + + // If the shift is NSW and we compare to 0, then it is just shifting out + // sign bits, no need for an AND either. + if (Shl->hasNoSignedWrap() && *C == 0) + return new ICmpInst(Pred, X, LShrC); + + if (Shl->hasOneUse()) { + // Otherwise strength reduce the shift into an and. + Constant *Mask = ConstantInt::get(Shl->getType(), + APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt->getZExtValue())); + + Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask"); + return new ICmpInst(Pred, And, LShrC); + } + } + + // If this is a signed comparison to 0 and the shift is sign preserving, + // use the shift LHS operand instead; isSignTest may change 'Pred', so only + // do that if we're sure to not continue on in this function. + if (Shl->hasNoSignedWrap() && isSignTest(Pred, *C)) + return new ICmpInst(Pred, X, Constant::getNullValue(X->getType())); + + // Otherwise, if this is a comparison of the sign bit, simplify to and/test. + bool TrueIfSigned = false; + if (Shl->hasOneUse() && isSignBitCheck(Pred, *C, TrueIfSigned)) { + // (X << 31) (X & 1) != 0 + Constant *Mask = ConstantInt::get( + X->getType(), + APInt::getOneBitSet(TypeBits, TypeBits - ShiftAmt->getZExtValue() - 1)); + Value *And = Builder->CreateAnd(X, Mask, Shl->getName() + ".mask"); + return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ, + And, Constant::getNullValue(And->getType())); + } + + // Transform (icmp pred iM (shl iM %v, N), C) + // -> (icmp pred i(M-N) (trunc %v iM to i(M-N)), (trunc (C>>N)) + // Transform the shl to a trunc if (trunc (C>>N)) has no loss and M-N. + // This enables us to get rid of the shift in favor of a trunc which can be + // free on the target. It has the additional benefit of comparing to a + // smaller constant, which will be target friendly. + unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1); + if (Shl->hasOneUse() && Amt != 0 && C->countTrailingZeros() >= Amt) { + Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt); + if (X->getType()->isVectorTy()) + TruncTy = VectorType::get(TruncTy, X->getType()->getVectorNumElements()); + Constant *NewC = + ConstantInt::get(TruncTy, C->ashr(*ShiftAmt).trunc(TypeBits - Amt)); + return new ICmpInst(Pred, Builder->CreateTrunc(X, TruncTy), NewC); + } + + return nullptr; + } + + /// Fold icmp ({al}shr X, Y), C. + Instruction *InstCombiner::foldICmpShrConstant(ICmpInst &Cmp, + BinaryOperator *Shr, + const APInt *C) { + // An exact shr only shifts out zero bits, so: + // icmp eq/ne (shr X, Y), 0 --> icmp eq/ne X, 0 + Value *X = Shr->getOperand(0); + CmpInst::Predicate Pred = Cmp.getPredicate(); + if (Cmp.isEquality() && Shr->isExact() && Shr->hasOneUse() && *C == 0) + return new ICmpInst(Pred, X, Cmp.getOperand(1)); + + const APInt *ShiftAmt; + if (!match(Shr->getOperand(1), m_APInt(ShiftAmt))) + return nullptr; + + // Check that the shift amount is in range. If not, don't perform undefined + // shifts. When the shift is visited it will be simplified. + unsigned TypeBits = C->getBitWidth(); + unsigned ShAmtVal = ShiftAmt->getLimitedValue(TypeBits); + if (ShAmtVal >= TypeBits || ShAmtVal == 0) + return nullptr; + + bool IsAShr = Shr->getOpcode() == Instruction::AShr; + if (!Cmp.isEquality()) { + // If we have an unsigned comparison and an ashr, we can't simplify this. + // Similarly for signed comparisons with lshr. + if (Cmp.isSigned() != IsAShr) + return nullptr; + + // Otherwise, all lshr and most exact ashr's are equivalent to a udiv/sdiv + // by a power of 2. Since we already have logic to simplify these, + // transform to div and then simplify the resultant comparison. + if (IsAShr && (!Shr->isExact() || ShAmtVal == TypeBits - 1)) + return nullptr; + + // Revisit the shift (to delete it). + Worklist.Add(Shr); + + Constant *DivCst = ConstantInt::get( + Shr->getType(), APInt::getOneBitSet(TypeBits, ShAmtVal)); + + Value *Tmp = IsAShr ? Builder->CreateSDiv(X, DivCst, "", Shr->isExact()) + : Builder->CreateUDiv(X, DivCst, "", Shr->isExact()); + + Cmp.setOperand(0, Tmp); + + // If the builder folded the binop, just return it. + BinaryOperator *TheDiv = dyn_cast(Tmp); + if (!TheDiv) + return &Cmp; + + // Otherwise, fold this div/compare. + assert(TheDiv->getOpcode() == Instruction::SDiv || + TheDiv->getOpcode() == Instruction::UDiv); + + Instruction *Res = foldICmpDivConstant(Cmp, TheDiv, C); + assert(Res && "This div/cst should have folded!"); + return Res; + } + + // Handle equality comparisons of shift-by-constant. + + // If the comparison constant changes with the shift, the comparison cannot + // succeed (bits of the comparison constant cannot match the shifted value). + // This should be known by InstSimplify and already be folded to true/false. + assert(((IsAShr && C->shl(ShAmtVal).ashr(ShAmtVal) == *C) || + (!IsAShr && C->shl(ShAmtVal).lshr(ShAmtVal) == *C)) && + "Expected icmp+shr simplify did not occur."); + + // Check if the bits shifted out are known to be zero. If so, we can compare + // against the unshifted value: + // (X & 4) >> 1 == 2 --> (X & 4) == 4. + Constant *ShiftedCmpRHS = ConstantInt::get(Shr->getType(), *C << ShAmtVal); + if (Shr->hasOneUse()) { + if (Shr->isExact()) + return new ICmpInst(Pred, X, ShiftedCmpRHS); + + // Otherwise strength reduce the shift into an 'and'. + APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal)); + Constant *Mask = ConstantInt::get(Shr->getType(), Val); + Value *And = Builder->CreateAnd(X, Mask, Shr->getName() + ".mask"); + return new ICmpInst(Pred, And, ShiftedCmpRHS); + } + + return nullptr; + } + + /// Fold icmp (udiv X, Y), C. + Instruction *InstCombiner::foldICmpUDivConstant(ICmpInst &Cmp, + BinaryOperator *UDiv, + const APInt *C) { + const APInt *C2; + if (!match(UDiv->getOperand(0), m_APInt(C2))) + return nullptr; + + assert(C2 != 0 && "udiv 0, X should have been simplified already."); + + // (icmp ugt (udiv C2, Y), C) -> (icmp ule Y, C2/(C+1)) + Value *Y = UDiv->getOperand(1); + if (Cmp.getPredicate() == ICmpInst::ICMP_UGT) { + assert(!C->isMaxValue() && + "icmp ugt X, UINT_MAX should have been simplified already."); + return new ICmpInst(ICmpInst::ICMP_ULE, Y, + ConstantInt::get(Y->getType(), C2->udiv(*C + 1))); + } + + // (icmp ult (udiv C2, Y), C) -> (icmp ugt Y, C2/C) + if (Cmp.getPredicate() == ICmpInst::ICMP_ULT) { + assert(C != 0 && "icmp ult X, 0 should have been simplified already."); + return new ICmpInst(ICmpInst::ICMP_UGT, Y, + ConstantInt::get(Y->getType(), C2->udiv(*C))); + } + + return nullptr; + } + + /// Fold icmp ({su}div X, Y), C. + Instruction *InstCombiner::foldICmpDivConstant(ICmpInst &Cmp, + BinaryOperator *Div, + const APInt *C) { + // Fold: icmp pred ([us]div X, C2), C -> range test + // Fold this div into the comparison, producing a range check. + // Determine, based on the divide type, what the range is being + // checked. If there is an overflow on the low or high side, remember + // it, otherwise compute the range [low, hi) bounding the new value. + // See: InsertRangeTest above for the kinds of replacements possible. + const APInt *C2; + if (!match(Div->getOperand(1), m_APInt(C2))) + return nullptr; + + // FIXME: If the operand types don't match the type of the divide + // then don't attempt this transform. The code below doesn't have the + // logic to deal with a signed divide and an unsigned compare (and + // vice versa). This is because (x /s C2) getOpcode() == Instruction::SDiv; + if (!Cmp.isEquality() && DivIsSigned != Cmp.isSigned()) + return nullptr; + + // The ProdOV computation fails on divide by 0 and divide by -1. Cases with + // INT_MIN will also fail if the divisor is 1. Although folds of all these + // division-by-constant cases should be present, we can not assert that they + // have happened before we reach this icmp instruction. + if (*C2 == 0 || *C2 == 1 || (DivIsSigned && C2->isAllOnesValue())) + return nullptr; + + // TODO: We could do all of the computations below using APInt. + Constant *CmpRHS = cast(Cmp.getOperand(1)); + Constant *DivRHS = cast(Div->getOperand(1)); + + // Compute Prod = CmpRHS * DivRHS. We are essentially solving an equation of + // form X / C2 = C. We solve for X by multiplying C2 (DivRHS) and C (CmpRHS). + // By solving for X, we can turn this into a range check instead of computing + // a divide. + Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS); + + // Determine if the product overflows by seeing if the product is not equal to + // the divide. Make sure we do the same kind of divide as in the LHS + // instruction that we're folding. + bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) + : ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS; + + ICmpInst::Predicate Pred = Cmp.getPredicate(); + + // If the division is known to be exact, then there is no remainder from the + // divide, so the covered range size is unit, otherwise it is the divisor. + Constant *RangeSize = + Div->isExact() ? ConstantInt::get(Div->getType(), 1) : DivRHS; + + // Figure out the interval that is being checked. For example, a comparison + // like "X /u 5 == 0" is really checking that X is in the interval [0, 5). + // Compute this interval based on the constants involved and the signedness of + // the compare/divide. This computes a half-open interval, keeping track of + // whether either value in the interval overflows. After analysis each + // overflow variable is set to 0 if it's corresponding bound variable is valid + // -1 if overflowed off the bottom end, or +1 if overflowed off the top end. + int LoOverflow = 0, HiOverflow = 0; + Constant *LoBound = nullptr, *HiBound = nullptr; + + if (!DivIsSigned) { // udiv + // e.g. X/5 op 3 --> [15, 20) + LoBound = Prod; + HiOverflow = LoOverflow = ProdOV; + if (!HiOverflow) { + // If this is not an exact divide, then many values in the range collapse + // to the same result value. + HiOverflow = addWithOverflow(HiBound, LoBound, RangeSize, false); + } + } else if (C2->isStrictlyPositive()) { // Divisor is > 0. + if (*C == 0) { // (X / pos) op 0 + // Can't overflow. e.g. X/2 op 0 --> [-1, 2) + LoBound = ConstantExpr::getNeg(SubOne(RangeSize)); + HiBound = RangeSize; + } else if (C->isStrictlyPositive()) { // (X / pos) op pos + LoBound = Prod; // e.g. X/5 op 3 --> [15, 20) + HiOverflow = LoOverflow = ProdOV; + if (!HiOverflow) + HiOverflow = addWithOverflow(HiBound, Prod, RangeSize, true); + } else { // (X / pos) op neg + // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14) + HiBound = AddOne(Prod); + LoOverflow = HiOverflow = ProdOV ? -1 : 0; + if (!LoOverflow) { + Constant *DivNeg = ConstantExpr::getNeg(RangeSize); + LoOverflow = addWithOverflow(LoBound, HiBound, DivNeg, true) ? -1 : 0; + } + } + } else if (C2->isNegative()) { // Divisor is < 0. + if (Div->isExact()) + RangeSize = ConstantExpr::getNeg(RangeSize); + if (*C == 0) { // (X / neg) op 0 + // e.g. X/-5 op 0 --> [-4, 5) + LoBound = AddOne(RangeSize); + HiBound = ConstantExpr::getNeg(RangeSize); + if (HiBound == DivRHS) { // -INTMIN = INTMIN + HiOverflow = 1; // [INTMIN+1, overflow) + HiBound = nullptr; // e.g. X/INTMIN = 0 --> X > INTMIN + } + } else if (C->isStrictlyPositive()) { // (X / neg) op pos + // e.g. X/-5 op 3 --> [-19, -14) + HiBound = AddOne(Prod); + HiOverflow = LoOverflow = ProdOV ? -1 : 0; + if (!LoOverflow) + LoOverflow = addWithOverflow(LoBound, HiBound, RangeSize, true) ? -1:0; + } else { // (X / neg) op neg + LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20) + LoOverflow = HiOverflow = ProdOV; + if (!HiOverflow) + HiOverflow = subWithOverflow(HiBound, Prod, RangeSize, true); + } + + // Dividing by a negative swaps the condition. LT <-> GT + Pred = ICmpInst::getSwappedPredicate(Pred); + } + + Value *X = Div->getOperand(0); + switch (Pred) { + default: llvm_unreachable("Unhandled icmp opcode!"); + case ICmpInst::ICMP_EQ: + if (LoOverflow && HiOverflow) + return replaceInstUsesWith(Cmp, Builder->getFalse()); + if (HiOverflow) + return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : + ICmpInst::ICMP_UGE, X, LoBound); + if (LoOverflow) + return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : + ICmpInst::ICMP_ULT, X, HiBound); + return replaceInstUsesWith( + Cmp, insertRangeTest(X, LoBound->getUniqueInteger(), + HiBound->getUniqueInteger(), DivIsSigned, true)); + case ICmpInst::ICMP_NE: + if (LoOverflow && HiOverflow) + return replaceInstUsesWith(Cmp, Builder->getTrue()); + if (HiOverflow) + return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT : + ICmpInst::ICMP_ULT, X, LoBound); + if (LoOverflow) + return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE : + ICmpInst::ICMP_UGE, X, HiBound); + return replaceInstUsesWith(Cmp, + insertRangeTest(X, LoBound->getUniqueInteger(), + HiBound->getUniqueInteger(), + DivIsSigned, false)); + case ICmpInst::ICMP_ULT: + case ICmpInst::ICMP_SLT: + if (LoOverflow == +1) // Low bound is greater than input range. + return replaceInstUsesWith(Cmp, Builder->getTrue()); + if (LoOverflow == -1) // Low bound is less than input range. + return replaceInstUsesWith(Cmp, Builder->getFalse()); + return new ICmpInst(Pred, X, LoBound); + case ICmpInst::ICMP_UGT: + case ICmpInst::ICMP_SGT: + if (HiOverflow == +1) // High bound greater than input range. + return replaceInstUsesWith(Cmp, Builder->getFalse()); + if (HiOverflow == -1) // High bound less than input range. + return replaceInstUsesWith(Cmp, Builder->getTrue()); + if (Pred == ICmpInst::ICMP_UGT) + return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound); + return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound); + } + + return nullptr; + } + + /// Fold icmp (sub X, Y), C. + Instruction *InstCombiner::foldICmpSubConstant(ICmpInst &Cmp, + BinaryOperator *Sub, + const APInt *C) { + Value *X = Sub->getOperand(0), *Y = Sub->getOperand(1); + ICmpInst::Predicate Pred = Cmp.getPredicate(); + + // The following transforms are only worth it if the only user of the subtract + // is the icmp. + if (!Sub->hasOneUse()) + return nullptr; + + if (Sub->hasNoSignedWrap()) { + // (icmp sgt (sub nsw X, Y), -1) -> (icmp sge X, Y) + if (Pred == ICmpInst::ICMP_SGT && C->isAllOnesValue()) + return new ICmpInst(ICmpInst::ICMP_SGE, X, Y); + + // (icmp sgt (sub nsw X, Y), 0) -> (icmp sgt X, Y) + if (Pred == ICmpInst::ICMP_SGT && *C == 0) + return new ICmpInst(ICmpInst::ICMP_SGT, X, Y); + + // (icmp slt (sub nsw X, Y), 0) -> (icmp slt X, Y) + if (Pred == ICmpInst::ICMP_SLT && *C == 0) + return new ICmpInst(ICmpInst::ICMP_SLT, X, Y); + + // (icmp slt (sub nsw X, Y), 1) -> (icmp sle X, Y) + if (Pred == ICmpInst::ICMP_SLT && *C == 1) + return new ICmpInst(ICmpInst::ICMP_SLE, X, Y); + } + + const APInt *C2; + if (!match(X, m_APInt(C2))) + return nullptr; + + // C2 - Y (Y | (C - 1)) == C2 + // iff (C2 & (C - 1)) == C - 1 and C is a power of 2 + if (Pred == ICmpInst::ICMP_ULT && C->isPowerOf2() && + (*C2 & (*C - 1)) == (*C - 1)) + return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateOr(Y, *C - 1), X); + + // C2 - Y >u C -> (Y | C) != C2 + // iff C2 & C == C and C + 1 is a power of 2 + if (Pred == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && (*C2 & *C) == *C) + return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateOr(Y, *C), X); + + return nullptr; + } + + /// Fold icmp (add X, Y), C. + Instruction *InstCombiner::foldICmpAddConstant(ICmpInst &Cmp, + BinaryOperator *Add, + const APInt *C) { + Value *Y = Add->getOperand(1); + const APInt *C2; + if (Cmp.isEquality() || !match(Y, m_APInt(C2))) + return nullptr; + + // Fold icmp pred (add X, C2), C. + Value *X = Add->getOperand(0); + Type *Ty = Add->getType(); + auto CR = Cmp.makeConstantRange(Cmp.getPredicate(), *C).subtract(*C2); + const APInt &Upper = CR.getUpper(); + const APInt &Lower = CR.getLower(); + if (Cmp.isSigned()) { + if (Lower.isSignBit()) + return new ICmpInst(ICmpInst::ICMP_SLT, X, ConstantInt::get(Ty, Upper)); + if (Upper.isSignBit()) + return new ICmpInst(ICmpInst::ICMP_SGE, X, ConstantInt::get(Ty, Lower)); + } else { + if (Lower.isMinValue()) + return new ICmpInst(ICmpInst::ICMP_ULT, X, ConstantInt::get(Ty, Upper)); + if (Upper.isMinValue()) + return new ICmpInst(ICmpInst::ICMP_UGE, X, ConstantInt::get(Ty, Lower)); + } + + if (!Add->hasOneUse()) + return nullptr; + + // X+C (X & -C2) == C + // iff C & (C2-1) == 0 + // C2 is a power of 2 + if (Cmp.getPredicate() == ICmpInst::ICMP_ULT && C->isPowerOf2() && + (*C2 & (*C - 1)) == 0) + return new ICmpInst(ICmpInst::ICMP_EQ, Builder->CreateAnd(X, -(*C)), + ConstantExpr::getNeg(cast(Y))); + + // X+C >u C2 -> (X & ~C2) != C + // iff C & C2 == 0 + // C2+1 is a power of 2 + if (Cmp.getPredicate() == ICmpInst::ICMP_UGT && (*C + 1).isPowerOf2() && + (*C2 & *C) == 0) + return new ICmpInst(ICmpInst::ICMP_NE, Builder->CreateAnd(X, ~(*C)), + ConstantExpr::getNeg(cast(Y))); + + return nullptr; + } + + /// Try to fold integer comparisons with a constant operand: icmp Pred X, C + /// where X is some kind of instruction. + Instruction *InstCombiner::foldICmpInstWithConstant(ICmpInst &Cmp) { + const APInt *C; + if (!match(Cmp.getOperand(1), m_APInt(C))) + return nullptr; + + BinaryOperator *BO; + if (match(Cmp.getOperand(0), m_BinOp(BO))) { + switch (BO->getOpcode()) { + case Instruction::Xor: + if (Instruction *I = foldICmpXorConstant(Cmp, BO, C)) + return I; + break; + case Instruction::And: + if (Instruction *I = foldICmpAndConstant(Cmp, BO, C)) + return I; + break; + case Instruction::Or: + if (Instruction *I = foldICmpOrConstant(Cmp, BO, C)) + return I; + break; + case Instruction::Mul: + if (Instruction *I = foldICmpMulConstant(Cmp, BO, C)) + return I; + break; + case Instruction::Shl: + if (Instruction *I = foldICmpShlConstant(Cmp, BO, C)) + return I; + break; + case Instruction::LShr: + case Instruction::AShr: + if (Instruction *I = foldICmpShrConstant(Cmp, BO, C)) + return I; + break; + case Instruction::UDiv: + if (Instruction *I = foldICmpUDivConstant(Cmp, BO, C)) + return I; + LLVM_FALLTHROUGH; + case Instruction::SDiv: + if (Instruction *I = foldICmpDivConstant(Cmp, BO, C)) + return I; + break; + case Instruction::Sub: + if (Instruction *I = foldICmpSubConstant(Cmp, BO, C)) + return I; + break; + case Instruction::Add: + if (Instruction *I = foldICmpAddConstant(Cmp, BO, C)) + return I; + break; + default: + break; + } + // TODO: These folds could be refactored to be part of the above calls. + if (Instruction *I = foldICmpBinOpEqualityWithConstant(Cmp, BO, C)) + return I; + } + + Instruction *LHSI; + if (match(Cmp.getOperand(0), m_Instruction(LHSI)) && + LHSI->getOpcode() == Instruction::Trunc) + if (Instruction *I = foldICmpTruncConstant(Cmp, LHSI, C)) + return I; + + if (Instruction *I = foldICmpIntrinsicWithConstant(Cmp, C)) + return I; + + return nullptr; + } + + /// Fold an icmp equality instruction with binary operator LHS and constant RHS: + /// icmp eq/ne BO, C. + Instruction *InstCombiner::foldICmpBinOpEqualityWithConstant(ICmpInst &Cmp, + BinaryOperator *BO, + const APInt *C) { + // TODO: Some of these folds could work with arbitrary constants, but this + // function is limited to scalar and vector splat constants. + if (!Cmp.isEquality()) + return nullptr; + + ICmpInst::Predicate Pred = Cmp.getPredicate(); + bool isICMP_NE = Pred == ICmpInst::ICMP_NE; + Constant *RHS = cast(Cmp.getOperand(1)); + Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1); + + switch (BO->getOpcode()) { + case Instruction::SRem: + // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one. + if (*C == 0 && BO->hasOneUse()) { + const APInt *BOC; + if (match(BOp1, m_APInt(BOC)) && BOC->sgt(1) && BOC->isPowerOf2()) { + Value *NewRem = Builder->CreateURem(BOp0, BOp1, BO->getName()); + return new ICmpInst(Pred, NewRem, + Constant::getNullValue(BO->getType())); + } + } + break; + case Instruction::Add: { + // Replace ((add A, B) != C) with (A != C-B) if B & C are constants. + const APInt *BOC; + if (match(BOp1, m_APInt(BOC))) { + if (BO->hasOneUse()) { + Constant *SubC = ConstantExpr::getSub(RHS, cast(BOp1)); + return new ICmpInst(Pred, BOp0, SubC); + } + } else if (*C == 0) { + // Replace ((add A, B) != 0) with (A != -B) if A or B is + // efficiently invertible, or if the add has just this one use. + if (Value *NegVal = dyn_castNegVal(BOp1)) + return new ICmpInst(Pred, BOp0, NegVal); + if (Value *NegVal = dyn_castNegVal(BOp0)) + return new ICmpInst(Pred, NegVal, BOp1); + if (BO->hasOneUse()) { + Value *Neg = Builder->CreateNeg(BOp1); + Neg->takeName(BO); + return new ICmpInst(Pred, BOp0, Neg); + } + } + break; + } + case Instruction::Xor: + if (BO->hasOneUse()) { + if (Constant *BOC = dyn_cast(BOp1)) { + // For the xor case, we can xor two constants together, eliminating + // the explicit xor. + return new ICmpInst(Pred, BOp0, ConstantExpr::getXor(RHS, BOC)); + } else if (*C == 0) { + // Replace ((xor A, B) != 0) with (A != B) + return new ICmpInst(Pred, BOp0, BOp1); + } + } + break; + case Instruction::Sub: + if (BO->hasOneUse()) { + const APInt *BOC; + if (match(BOp0, m_APInt(BOC))) { + // Replace ((sub BOC, B) != C) with (B != BOC-C). + Constant *SubC = ConstantExpr::getSub(cast(BOp0), RHS); + return new ICmpInst(Pred, BOp1, SubC); + } else if (*C == 0) { + // Replace ((sub A, B) != 0) with (A != B). + return new ICmpInst(Pred, BOp0, BOp1); + } + } + break; + case Instruction::Or: { + const APInt *BOC; + if (match(BOp1, m_APInt(BOC)) && BO->hasOneUse() && RHS->isAllOnesValue()) { + // Comparing if all bits outside of a constant mask are set? + // Replace (X | C) == -1 with (X & ~C) == ~C. + // This removes the -1 constant. + Constant *NotBOC = ConstantExpr::getNot(cast(BOp1)); + Value *And = Builder->CreateAnd(BOp0, NotBOC); + return new ICmpInst(Pred, And, NotBOC); + } + break; + } + case Instruction::And: { + const APInt *BOC; + if (match(BOp1, m_APInt(BOC))) { + // If we have ((X & C) == C), turn it into ((X & C) != 0). + if (C == BOC && C->isPowerOf2()) + return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE, + BO, Constant::getNullValue(RHS->getType())); + + // Don't perform the following transforms if the AND has multiple uses + if (!BO->hasOneUse()) + break; + + // Replace (and X, (1 << size(X)-1) != 0) with x s< 0 + if (BOC->isSignBit()) { + Constant *Zero = Constant::getNullValue(BOp0->getType()); + auto NewPred = isICMP_NE ? ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE; + return new ICmpInst(NewPred, BOp0, Zero); + } + + // ((X & ~7) == 0) --> X < 8 + if (*C == 0 && (~(*BOC) + 1).isPowerOf2()) { + Constant *NegBOC = ConstantExpr::getNeg(cast(BOp1)); + auto NewPred = isICMP_NE ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT; + return new ICmpInst(NewPred, BOp0, NegBOC); + } + } + break; + } + case Instruction::Mul: + if (*C == 0 && BO->hasNoSignedWrap()) { + const APInt *BOC; + if (match(BOp1, m_APInt(BOC)) && *BOC != 0) { + // The trivial case (mul X, 0) is handled by InstSimplify. + // General case : (mul X, C) != 0 iff X != 0 + // (mul X, C) == 0 iff X == 0 + return new ICmpInst(Pred, BOp0, Constant::getNullValue(RHS->getType())); + } + } + break; + case Instruction::UDiv: + if (*C == 0) { + // (icmp eq/ne (udiv A, B), 0) -> (icmp ugt/ule i32 B, A) + auto NewPred = isICMP_NE ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_UGT; + return new ICmpInst(NewPred, BOp1, BOp0); + } + break; + default: + break; + } + return nullptr; + } + + /// Fold an icmp with LLVM intrinsic and constant operand: icmp Pred II, C. + Instruction *InstCombiner::foldICmpIntrinsicWithConstant(ICmpInst &Cmp, + const APInt *C) { + IntrinsicInst *II = dyn_cast(Cmp.getOperand(0)); + if (!II || !Cmp.isEquality()) + return nullptr; + + // Handle icmp {eq|ne} , intcst. + switch (II->getIntrinsicID()) { + case Intrinsic::bswap: + Worklist.Add(II); + Cmp.setOperand(0, II->getArgOperand(0)); + Cmp.setOperand(1, Builder->getInt(C->byteSwap())); + return &Cmp; + case Intrinsic::ctlz: + case Intrinsic::cttz: + // ctz(A) == bitwidth(A) -> A == 0 and likewise for != + if (*C == C->getBitWidth()) { + Worklist.Add(II); + Cmp.setOperand(0, II->getArgOperand(0)); + Cmp.setOperand(1, ConstantInt::getNullValue(II->getType())); + return &Cmp; + } + break; + case Intrinsic::ctpop: { + // popcount(A) == 0 -> A == 0 and likewise for != + // popcount(A) == bitwidth(A) -> A == -1 and likewise for != + bool IsZero = *C == 0; + if (IsZero || *C == C->getBitWidth()) { + Worklist.Add(II); + Cmp.setOperand(0, II->getArgOperand(0)); + auto *NewOp = IsZero ? Constant::getNullValue(II->getType()) + : Constant::getAllOnesValue(II->getType()); + Cmp.setOperand(1, NewOp); + return &Cmp; + } + break; + } + default: + break; + } + return nullptr; + } + + /// Handle icmp (cast x to y), (cast/cst). We only handle extending casts so + /// far. + Instruction *InstCombiner::foldICmpWithCastAndCast(ICmpInst &ICmp) { + const CastInst *LHSCI = cast(ICmp.getOperand(0)); + Value *LHSCIOp = LHSCI->getOperand(0); + Type *SrcTy = LHSCIOp->getType(); + Type *DestTy = LHSCI->getType(); + Value *RHSCIOp; + + // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the + // integer type is the same size as the pointer type. + if (LHSCI->getOpcode() == Instruction::PtrToInt && + DL.getPointerTypeSizeInBits(SrcTy) == DestTy->getIntegerBitWidth()) { + Value *RHSOp = nullptr; + if (auto *RHSC = dyn_cast(ICmp.getOperand(1))) { + Value *RHSCIOp = RHSC->getOperand(0); + if (RHSCIOp->getType()->getPointerAddressSpace() == + LHSCIOp->getType()->getPointerAddressSpace()) { + RHSOp = RHSC->getOperand(0); + // If the pointer types don't match, insert a bitcast. + if (LHSCIOp->getType() != RHSOp->getType()) + RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType()); + } + } else if (auto *RHSC = dyn_cast(ICmp.getOperand(1))) { + RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy); + } + + if (RHSOp) + return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSOp); + } + + // The code below only handles extension cast instructions, so far. + // Enforce this. + if (LHSCI->getOpcode() != Instruction::ZExt && + LHSCI->getOpcode() != Instruction::SExt) + return nullptr; + + bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt; + bool isSignedCmp = ICmp.isSigned(); + + if (auto *CI = dyn_cast(ICmp.getOperand(1))) { + // Not an extension from the same type? + RHSCIOp = CI->getOperand(0); + if (RHSCIOp->getType() != LHSCIOp->getType()) + return nullptr; + + // If the signedness of the two casts doesn't agree (i.e. one is a sext + // and the other is a zext), then we can't handle this. + if (CI->getOpcode() != LHSCI->getOpcode()) + return nullptr; + + // Deal with equality cases early. + if (ICmp.isEquality()) + return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp); + + // A signed comparison of sign extended values simplifies into a + // signed comparison. + if (isSignedCmp && isSignedExt) + return new ICmpInst(ICmp.getPredicate(), LHSCIOp, RHSCIOp); + + // The other three cases all fold into an unsigned comparison. + return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, RHSCIOp); + } + + // If we aren't dealing with a constant on the RHS, exit early. + auto *C = dyn_cast(ICmp.getOperand(1)); + if (!C) + return nullptr; + + // Compute the constant that would happen if we truncated to SrcTy then + // re-extended to DestTy. + Constant *Res1 = ConstantExpr::getTrunc(C, SrcTy); + Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy); + + // If the re-extended constant didn't change... + if (Res2 == C) { + // Deal with equality cases early. + if (ICmp.isEquality()) + return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1); + + // A signed comparison of sign extended values simplifies into a + // signed comparison. + if (isSignedExt && isSignedCmp) + return new ICmpInst(ICmp.getPredicate(), LHSCIOp, Res1); + + // The other three cases all fold into an unsigned comparison. + return new ICmpInst(ICmp.getUnsignedPredicate(), LHSCIOp, Res1); + } + + // The re-extended constant changed, partly changed (in the case of a vector), + // or could not be determined to be equal (in the case of a constant + // expression), so the constant cannot be represented in the shorter type. + // Consequently, we cannot emit a simple comparison. + // All the cases that fold to true or false will have already been handled + // by SimplifyICmpInst, so only deal with the tricky case. + + if (isSignedCmp || !isSignedExt || !isa(C)) + return nullptr; + + // Evaluate the comparison for LT (we invert for GT below). LE and GE cases + // should have been folded away previously and not enter in here. + + // We're performing an unsigned comp with a sign extended value. + // This is true if the input is >= 0. [aka >s -1] + Constant *NegOne = Constant::getAllOnesValue(SrcTy); + Value *Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICmp.getName()); + + // Finally, return the value computed. + if (ICmp.getPredicate() == ICmpInst::ICMP_ULT) + return replaceInstUsesWith(ICmp, Result); + + assert(ICmp.getPredicate() == ICmpInst::ICMP_UGT && "ICmp should be folded!"); + return BinaryOperator::CreateNot(Result); + } + + bool InstCombiner::OptimizeOverflowCheck(OverflowCheckFlavor OCF, Value *LHS, + Value *RHS, Instruction &OrigI, + Value *&Result, Constant *&Overflow) { + if (OrigI.isCommutative() && isa(LHS) && !isa(RHS)) + std::swap(LHS, RHS); + + auto SetResult = [&](Value *OpResult, Constant *OverflowVal, bool ReuseName) { + Result = OpResult; + Overflow = OverflowVal; + if (ReuseName) + Result->takeName(&OrigI); + return true; + }; + + // If the overflow check was an add followed by a compare, the insertion point + // may be pointing to the compare. We want to insert the new instructions + // before the add in case there are uses of the add between the add and the + // compare. + Builder->SetInsertPoint(&OrigI); + + switch (OCF) { + case OCF_INVALID: + llvm_unreachable("bad overflow check kind!"); + + case OCF_UNSIGNED_ADD: { + OverflowResult OR = computeOverflowForUnsignedAdd(LHS, RHS, &OrigI); + if (OR == OverflowResult::NeverOverflows) + return SetResult(Builder->CreateNUWAdd(LHS, RHS), Builder->getFalse(), + true); + + if (OR == OverflowResult::AlwaysOverflows) + return SetResult(Builder->CreateAdd(LHS, RHS), Builder->getTrue(), true); + + // Fall through uadd into sadd + LLVM_FALLTHROUGH; + } + case OCF_SIGNED_ADD: { + // X + 0 -> {X, false} + if (match(RHS, m_Zero())) + return SetResult(LHS, Builder->getFalse(), false); + + // We can strength reduce this signed add into a regular add if we can prove + // that it will never overflow. + if (OCF == OCF_SIGNED_ADD) + if (WillNotOverflowSignedAdd(LHS, RHS, OrigI)) + return SetResult(Builder->CreateNSWAdd(LHS, RHS), Builder->getFalse(), + true); + break; + } + + case OCF_UNSIGNED_SUB: + case OCF_SIGNED_SUB: { + // X - 0 -> {X, false} + if (match(RHS, m_Zero())) + return SetResult(LHS, Builder->getFalse(), false); + + if (OCF == OCF_SIGNED_SUB) { + if (WillNotOverflowSignedSub(LHS, RHS, OrigI)) + return SetResult(Builder->CreateNSWSub(LHS, RHS), Builder->getFalse(), + true); + } else { + if (WillNotOverflowUnsignedSub(LHS, RHS, OrigI)) + return SetResult(Builder->CreateNUWSub(LHS, RHS), Builder->getFalse(), + true); + } + break; + } + + case OCF_UNSIGNED_MUL: { + OverflowResult OR = computeOverflowForUnsignedMul(LHS, RHS, &OrigI); + if (OR == OverflowResult::NeverOverflows) + return SetResult(Builder->CreateNUWMul(LHS, RHS), Builder->getFalse(), + true); + if (OR == OverflowResult::AlwaysOverflows) + return SetResult(Builder->CreateMul(LHS, RHS), Builder->getTrue(), true); + LLVM_FALLTHROUGH; + } + case OCF_SIGNED_MUL: + // X * undef -> undef + if (isa(RHS)) + return SetResult(RHS, UndefValue::get(Builder->getInt1Ty()), false); + + // X * 0 -> {0, false} + if (match(RHS, m_Zero())) + return SetResult(RHS, Builder->getFalse(), false); + + // X * 1 -> {X, false} + if (match(RHS, m_One())) + return SetResult(LHS, Builder->getFalse(), false); + + if (OCF == OCF_SIGNED_MUL) + if (WillNotOverflowSignedMul(LHS, RHS, OrigI)) + return SetResult(Builder->CreateNSWMul(LHS, RHS), Builder->getFalse(), + true); + break; + } + + return false; + } + + /// \brief Recognize and process idiom involving test for multiplication + /// overflow. + /// + /// The caller has matched a pattern of the form: + /// I = cmp u (mul(zext A, zext B), V + /// The function checks if this is a test for overflow and if so replaces + /// multiplication with call to 'mul.with.overflow' intrinsic. + /// + /// \param I Compare instruction. + /// \param MulVal Result of 'mult' instruction. It is one of the arguments of + /// the compare instruction. Must be of integer type. + /// \param OtherVal The other argument of compare instruction. + /// \returns Instruction which must replace the compare instruction, NULL if no + /// replacement required. + static Instruction *processUMulZExtIdiom(ICmpInst &I, Value *MulVal, + Value *OtherVal, InstCombiner &IC) { + // Don't bother doing this transformation for pointers, don't do it for + // vectors. + if (!isa(MulVal->getType())) + return nullptr; + + assert(I.getOperand(0) == MulVal || I.getOperand(1) == MulVal); + assert(I.getOperand(0) == OtherVal || I.getOperand(1) == OtherVal); + auto *MulInstr = dyn_cast(MulVal); + if (!MulInstr) + return nullptr; + assert(MulInstr->getOpcode() == Instruction::Mul); + + auto *LHS = cast(MulInstr->getOperand(0)), + *RHS = cast(MulInstr->getOperand(1)); + assert(LHS->getOpcode() == Instruction::ZExt); + assert(RHS->getOpcode() == Instruction::ZExt); + Value *A = LHS->getOperand(0), *B = RHS->getOperand(0); + + // Calculate type and width of the result produced by mul.with.overflow. + Type *TyA = A->getType(), *TyB = B->getType(); + unsigned WidthA = TyA->getPrimitiveSizeInBits(), + WidthB = TyB->getPrimitiveSizeInBits(); + unsigned MulWidth; + Type *MulType; + if (WidthB > WidthA) { + MulWidth = WidthB; + MulType = TyB; + } else { + MulWidth = WidthA; + MulType = TyA; + } + + // In order to replace the original mul with a narrower mul.with.overflow, + // all uses must ignore upper bits of the product. The number of used low + // bits must be not greater than the width of mul.with.overflow. + if (MulVal->hasNUsesOrMore(2)) + for (User *U : MulVal->users()) { + if (U == &I) + continue; + if (TruncInst *TI = dyn_cast(U)) { + // Check if truncation ignores bits above MulWidth. + unsigned TruncWidth = TI->getType()->getPrimitiveSizeInBits(); + if (TruncWidth > MulWidth) + return nullptr; + } else if (BinaryOperator *BO = dyn_cast(U)) { + // Check if AND ignores bits above MulWidth. + if (BO->getOpcode() != Instruction::And) + return nullptr; + if (ConstantInt *CI = dyn_cast(BO->getOperand(1))) { + const APInt &CVal = CI->getValue(); + if (CVal.getBitWidth() - CVal.countLeadingZeros() > MulWidth) + return nullptr; + } + } else { + // Other uses prohibit this transformation. + return nullptr; + } + } + + // Recognize patterns + switch (I.getPredicate()) { + case ICmpInst::ICMP_EQ: + case ICmpInst::ICMP_NE: + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp eq/neq mulval, zext trunc mulval + if (ZExtInst *Zext = dyn_cast(OtherVal)) + if (Zext->hasOneUse()) { + Value *ZextArg = Zext->getOperand(0); + if (TruncInst *Trunc = dyn_cast(ZextArg)) + if (Trunc->getType()->getPrimitiveSizeInBits() == MulWidth) + break; //Recognized + } + + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp eq/neq mulval, and(mulval, mask), mask selects low MulWidth bits. + ConstantInt *CI; + Value *ValToMask; + if (match(OtherVal, m_And(m_Value(ValToMask), m_ConstantInt(CI)))) { + if (ValToMask != MulVal) + return nullptr; + const APInt &CVal = CI->getValue() + 1; + if (CVal.isPowerOf2()) { + unsigned MaskWidth = CVal.logBase2(); + if (MaskWidth == MulWidth) + break; // Recognized + } + } + return nullptr; + + case ICmpInst::ICMP_UGT: + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp ugt mulval, max + if (ConstantInt *CI = dyn_cast(OtherVal)) { + APInt MaxVal = APInt::getMaxValue(MulWidth); + MaxVal = MaxVal.zext(CI->getBitWidth()); + if (MaxVal.eq(CI->getValue())) + break; // Recognized + } + return nullptr; + + case ICmpInst::ICMP_UGE: + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp uge mulval, max+1 + if (ConstantInt *CI = dyn_cast(OtherVal)) { + APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth); + if (MaxVal.eq(CI->getValue())) + break; // Recognized + } + return nullptr; + + case ICmpInst::ICMP_ULE: + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp ule mulval, max + if (ConstantInt *CI = dyn_cast(OtherVal)) { + APInt MaxVal = APInt::getMaxValue(MulWidth); + MaxVal = MaxVal.zext(CI->getBitWidth()); + if (MaxVal.eq(CI->getValue())) + break; // Recognized + } + return nullptr; + + case ICmpInst::ICMP_ULT: + // Recognize pattern: + // mulval = mul(zext A, zext B) + // cmp ule mulval, max + 1 + if (ConstantInt *CI = dyn_cast(OtherVal)) { + APInt MaxVal = APInt::getOneBitSet(CI->getBitWidth(), MulWidth); + if (MaxVal.eq(CI->getValue())) + break; // Recognized + } + return nullptr; + + default: + return nullptr; + } + + InstCombiner::BuilderTy *Builder = IC.Builder; + Builder->SetInsertPoint(MulInstr); + + // Replace: mul(zext A, zext B) --> mul.with.overflow(A, B) + Value *MulA = A, *MulB = B; + if (WidthA < MulWidth) + MulA = Builder->CreateZExt(A, MulType); + if (WidthB < MulWidth) + MulB = Builder->CreateZExt(B, MulType); + Value *F = Intrinsic::getDeclaration(I.getModule(), + Intrinsic::umul_with_overflow, MulType); + CallInst *Call = Builder->CreateCall(F, {MulA, MulB}, "umul"); + IC.Worklist.Add(MulInstr); + + // If there are uses of mul result other than the comparison, we know that + // they are truncation or binary AND. Change them to use result of + // mul.with.overflow and adjust properly mask/size. + if (MulVal->hasNUsesOrMore(2)) { + Value *Mul = Builder->CreateExtractValue(Call, 0, "umul.value"); + for (User *U : MulVal->users()) { + if (U == &I || U == OtherVal) + continue; + if (TruncInst *TI = dyn_cast(U)) { + if (TI->getType()->getPrimitiveSizeInBits() == MulWidth) + IC.replaceInstUsesWith(*TI, Mul); + else + TI->setOperand(0, Mul); + } else if (BinaryOperator *BO = dyn_cast(U)) { + assert(BO->getOpcode() == Instruction::And); + // Replace (mul & mask) --> zext (mul.with.overflow & short_mask) + ConstantInt *CI = cast(BO->getOperand(1)); + APInt ShortMask = CI->getValue().trunc(MulWidth); + Value *ShortAnd = Builder->CreateAnd(Mul, ShortMask); + Instruction *Zext = + cast(Builder->CreateZExt(ShortAnd, BO->getType())); + IC.Worklist.Add(Zext); + IC.replaceInstUsesWith(*BO, Zext); + } else { + llvm_unreachable("Unexpected Binary operation"); + } + IC.Worklist.Add(cast(U)); + } + } + if (isa(OtherVal)) + IC.Worklist.Add(cast(OtherVal)); + + // The original icmp gets replaced with the overflow value, maybe inverted + // depending on predicate. + bool Inverse = false; + switch (I.getPredicate()) { + case ICmpInst::ICMP_NE: + break; + case ICmpInst::ICMP_EQ: + Inverse = true; + break; + case ICmpInst::ICMP_UGT: + case ICmpInst::ICMP_UGE: + if (I.getOperand(0) == MulVal) + break; + Inverse = true; + break; + case ICmpInst::ICMP_ULT: + case ICmpInst::ICMP_ULE: + if (I.getOperand(1) == MulVal) + break; + Inverse = true; + break; + default: + llvm_unreachable("Unexpected predicate"); + } + if (Inverse) { + Value *Res = Builder->CreateExtractValue(Call, 1); + return BinaryOperator::CreateNot(Res); + } + + return ExtractValueInst::Create(Call, 1); + } + + /// When performing a comparison against a constant, it is possible that not all + /// the bits in the LHS are demanded. This helper method computes the mask that + /// IS demanded. + static APInt getDemandedBitsLHSMask(ICmpInst &I, unsigned BitWidth, + bool isSignCheck) { + if (isSignCheck) + return APInt::getSignBit(BitWidth); + + ConstantInt *CI = dyn_cast(I.getOperand(1)); + if (!CI) return APInt::getAllOnesValue(BitWidth); + const APInt &RHS = CI->getValue(); + + switch (I.getPredicate()) { + // For a UGT comparison, we don't care about any bits that + // correspond to the trailing ones of the comparand. The value of these + // bits doesn't impact the outcome of the comparison, because any value + // greater than the RHS must differ in a bit higher than these due to carry. + case ICmpInst::ICMP_UGT: { + unsigned trailingOnes = RHS.countTrailingOnes(); + APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingOnes); + return ~lowBitsSet; + } + + // Similarly, for a ULT comparison, we don't care about the trailing zeros. + // Any value less than the RHS must differ in a higher bit because of carries. + case ICmpInst::ICMP_ULT: { + unsigned trailingZeros = RHS.countTrailingZeros(); + APInt lowBitsSet = APInt::getLowBitsSet(BitWidth, trailingZeros); + return ~lowBitsSet; + } + + default: + return APInt::getAllOnesValue(BitWidth); + } + } + + /// \brief Check if the order of \p Op0 and \p Op1 as operand in an ICmpInst + /// should be swapped. + /// The decision is based on how many times these two operands are reused + /// as subtract operands and their positions in those instructions. + /// The rational is that several architectures use the same instruction for + /// both subtract and cmp, thus it is better if the order of those operands + /// match. + /// \return true if Op0 and Op1 should be swapped. + static bool swapMayExposeCSEOpportunities(const Value * Op0, + const Value * Op1) { + // Filter out pointer value as those cannot appears directly in subtract. + // FIXME: we may want to go through inttoptrs or bitcasts. + if (Op0->getType()->isPointerTy()) + return false; + // Count every uses of both Op0 and Op1 in a subtract. + // Each time Op0 is the first operand, count -1: swapping is bad, the + // subtract has already the same layout as the compare. + // Each time Op0 is the second operand, count +1: swapping is good, the + // subtract has a different layout as the compare. + // At the end, if the benefit is greater than 0, Op0 should come second to + // expose more CSE opportunities. + int GlobalSwapBenefits = 0; + for (const User *U : Op0->users()) { + const BinaryOperator *BinOp = dyn_cast(U); + if (!BinOp || BinOp->getOpcode() != Instruction::Sub) + continue; + // If Op0 is the first argument, this is not beneficial to swap the + // arguments. + int LocalSwapBenefits = -1; + unsigned Op1Idx = 1; + if (BinOp->getOperand(Op1Idx) == Op0) { + Op1Idx = 0; + LocalSwapBenefits = 1; + } + if (BinOp->getOperand(Op1Idx) != Op1) + continue; + GlobalSwapBenefits += LocalSwapBenefits; + } + return GlobalSwapBenefits > 0; + } + + /// \brief Check that one use is in the same block as the definition and all + /// other uses are in blocks dominated by a given block. + /// + /// \param DI Definition + /// \param UI Use + /// \param DB Block that must dominate all uses of \p DI outside + /// the parent block + /// \return true when \p UI is the only use of \p DI in the parent block + /// and all other uses of \p DI are in blocks dominated by \p DB. + /// + bool InstCombiner::dominatesAllUses(const Instruction *DI, + const Instruction *UI, + const BasicBlock *DB) const { + assert(DI && UI && "Instruction not defined\n"); + // Ignore incomplete definitions. + if (!DI->getParent()) + return false; + // DI and UI must be in the same block. + if (DI->getParent() != UI->getParent()) + return false; + // Protect from self-referencing blocks. + if (DI->getParent() == DB) + return false; + for (const User *U : DI->users()) { + auto *Usr = cast(U); + if (Usr != UI && !DT.dominates(DB, Usr->getParent())) + return false; + } + return true; + } + + /// Return true when the instruction sequence within a block is select-cmp-br. + static bool isChainSelectCmpBranch(const SelectInst *SI) { + const BasicBlock *BB = SI->getParent(); + if (!BB) + return false; + auto *BI = dyn_cast_or_null(BB->getTerminator()); + if (!BI || BI->getNumSuccessors() != 2) + return false; + auto *IC = dyn_cast(BI->getCondition()); + if (!IC || (IC->getOperand(0) != SI && IC->getOperand(1) != SI)) + return false; + return true; + } + + /// \brief True when a select result is replaced by one of its operands + /// in select-icmp sequence. This will eventually result in the elimination + /// of the select. + /// + /// \param SI Select instruction + /// \param Icmp Compare instruction + /// \param SIOpd Operand that replaces the select + /// + /// Notes: + /// - The replacement is global and requires dominator information + /// - The caller is responsible for the actual replacement + /// + /// Example: + /// + /// entry: + /// %4 = select i1 %3, %C* %0, %C* null + /// %5 = icmp eq %C* %4, null + /// br i1 %5, label %9, label %7 + /// ... + /// ;