[experiment] JIT: Introduce KnownBits

src/coreclr/jit/CMakeLists.txt

@@ -152,6 +152,7 @@ set( JIT_SOURCES
   jiteh.cpp
   jithashtable.cpp
   jitmetadata.cpp
+  knownbits.cpp
   layout.cpp
   lclmorph.cpp
   lclvars.cpp
@@ -378,6 +379,7 @@ set( JIT_HEADERS
   jitmetadatalist.h
   jitpch.h
   jitstd.h
+  knownbits.h
   lir.h
   loopcloning.h
   loopcloningopts.h

src/coreclr/jit/assertionprop.cpp

@@ -12,6 +12,7 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
 
 #include "jitpch.h"
 #include "rangecheck.h"
+#include "knownbits.h"
 #ifdef _MSC_VER
 #pragma hdrstop
 #endif
@@ -4083,6 +4084,22 @@ void Compiler::optAssertionProp_RangeProperties(ASSERT_VALARG_TP assertions,
             *isKnownNonZero = true;
         }
     }
+
+    // Known bits can also establish non-negativity (sign bit known 0) and non-zeroness (some bit
+    // known 1). Covers TYP_LONG and bit patterns an interval cannot express (e.g. "x & 7").
+    if (!*isKnownNonZero || !*isKnownNonNegative)
+    {
+        const uint64_t  signBit = 1ull << ((genTypeSize(genActualType(tree)) * BITS_PER_BYTE) - 1);
+        const KnownBits kb      = KnownBits::Compute(this, treeVN, assertions);
+        if ((kb.knownZero & signBit) != 0)
+        {
+            *isKnownNonNegative = true;
+        }
+        if (kb.knownOne != 0)
+        {
+            *isKnownNonZero = true;
+        }
+    }
 }
 
 //------------------------------------------------------------------------
@@ -4514,6 +4531,23 @@ GenTree* Compiler::optAssertionPropGlobal_RelOp(ASSERT_VALARG_TP assertions,
         }
     }
 
+    // See if we can fold the relop based on known bits. This complements the range-based folding
+    // above (which is limited to TYP_INT) by reasoning about individual bits and TYP_LONG values.
+    if (varTypeIsIntegral(op1) && (op1VN != ValueNumStore::NoVN) && (op2VN != ValueNumStore::NoVN))
+    {
+        const unsigned  width = genTypeSize(genActualType(op1)) * BITS_PER_BYTE;
+        const KnownBits kb1   = KnownBits::Compute(this, op1VN, assertions);
+        const KnownBits kb2   = KnownBits::Compute(this, op2VN, assertions);
+
+        const int relopResult = KnownBitsOps::EvalRelop(tree->OperGet(), tree->IsUnsigned(), kb1, kb2, width);
+        if (relopResult >= 0)
+        {
+            JITDUMP("Folding relop [%06u] based on known bits.\n", dspTreeID(tree));
+            newTree = gtWrapWithSideEffects(relopResult == 1 ? gtNewTrue() : gtNewFalse(), tree, GTF_ALL_EFFECT);
+            return optAssertionProp_Update(newTree, tree, stmt);
+        }
+    }
+
     // Else check if we have an equality check involving a local or an indir
     if (!tree->OperIs(GT_EQ, GT_NE))
     {
@@ -5537,6 +5571,20 @@ GenTree* Compiler::optAssertionProp_BndsChk(ASSERT_VALARG_TP assertions, GenTree
         return optAssertionProp_Update(newTree, arrBndsChk, stmt);
     };
 
+    // Known-bits elimination: redundant if (uint)index is provably < (uint)length. Catches masked
+    // indices and bit patterns the range-based paths cannot express. On 64-bit targets the index
+    // and length can both be TYP_I_IMPL (see fgMorphIndexAddr), so derive the width from the actual
+    // operand type instead of hardcoding 32 -- otherwise we'd discard the high 32 bits of a native-int
+    // index and could prove a (uint)idx < (uint)len fact that doesn't hold for the full value.
+    assert(genActualType(arrBndsChk->GetIndex()) == genActualType(arrBndsChk->GetArrayLength()));
+    const unsigned  width = genTypeSize(genActualType(arrBndsChk->GetIndex())) * BITS_PER_BYTE;
+    const KnownBits kbIdx = KnownBits::Compute(this, vnCurIdx, assertions);
+    const KnownBits kbLen = KnownBits::Compute(this, vnCurLen, assertions);
+    if (KnownBitsOps::EvalRelop(GT_LT, /* isUnsigned */ true, kbIdx, kbLen, width) == 1)
+    {
+        return dropBoundsCheck(INDEBUG("known bits prove (uint)index < (uint)length"));
+    }
+
     // First, check if we have arr[arr.Length - cns] when we know arr.Length is >= cns.
     ValueNum add0, add1;
     if (vnStore->IsVNBinFunc(vnCurIdx, VNF_ADD, &add0, &add1))

src/coreclr/jit/knownbits.cpp

@@ -0,0 +1,282 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+
+#include "jitpch.h"
+#ifdef _MSC_VER
+#pragma hdrstop
+#endif
+
+#include "knownbits.h"
+
+//------------------------------------------------------------------------
+// MergeKnownBitsAssertions: Refine "*pBits" using whatever the live assertions tell us about "num".
+//
+// Arguments:
+//    comp       - the compiler context
+//    num        - the value number being analyzed
+//    assertions - the assertion set live at the consumer
+//    width      - bit width (32 or 64) of "num"
+//    budget     - recursive search budget (currently unused here, kept for symmetry with Compute)
+//    pBits      - in/out: the lattice for "num" so far; refined in place by intersecting with each
+//                 fact this routine can extract from the assertion table
+//
+static void MergeKnownBitsAssertions(
+    Compiler* comp, ValueNum num, ASSERT_VALARG_TP assertions, unsigned width, int /*budget*/, KnownBits* pBits)
+{
+    if (BitVecOps::MayBeUninit(assertions) || BitVecOps::IsEmpty(comp->apTraits, assertions) ||
+        !comp->optAssertionHasAssertionsForVN(num))
+    {
+        return;
+    }
+
+    const uint64_t  signBit = 1ull << (width - 1);
+    const KnownBits signBitZero(signBit, 0);
+
+    // Tightest signed upper bound "num <= signedUpperBound" gathered from signed "num < C" / "num <= C"
+    // assertions with a non-negative bound. On its own a signed upper bound says nothing about the high
+    // bits (num could be negative), so we only apply it after the loop, and only once we also know num
+    // is non-negative (sign bit 0) -- then num is in [0, signedUpperBound] and its upper bits are 0.
+    bool     haveSignedUpperBound = false;
+    uint64_t signedUpperBound     = 0;
+
+    BitVecOps::Iter iter(comp->apTraits, assertions);
+    unsigned        index = 0;
+    while (iter.NextElem(&index))
+    {
+        const Compiler::AssertionDsc& cur = comp->optGetAssertion(GetAssertionIndex(index));
+        if (cur.GetOp1().GetVN() != num)
+        {
+            continue;
+        }
+
+        // "num == const": fully determines the bits.
+        if (cur.KindIs(Compiler::OAK_EQUAL))
+        {
+            int64_t eqCns;
+            if (comp->vnStore->IsVNIntegralConstant<int64_t>(cur.GetOp2().GetVN(), &eqCns))
+            {
+                *pBits = KnownBits::Intersect(*pBits, KnownBits::FromConstant((uint64_t)eqCns, width));
+            }
+            continue;
+        }
+
+        // Relops of the form "num <relop> const".
+        if (cur.IsRelop() && cur.GetOp2().KindIs(Compiler::O2K_CONST_INT))
+        {
+            const int64_t relCns = cur.GetOp2().GetIntConstant();
+
+            if (cur.KindIs(Compiler::OAK_LT_UN) && (relCns > 0))
+            {
+                // (uint)num < C  =>  num u<= C-1  =>  upper bits are 0.
+                *pBits = KnownBits::Intersect(*pBits, KnownBits::FromUnsignedUpperBound((uint64_t)(relCns - 1), width));
+            }
+            else if (cur.KindIs(Compiler::OAK_LE_UN) && (relCns >= 0))
+            {
+                // (uint)num <= C  =>  upper bits are 0.
+                *pBits = KnownBits::Intersect(*pBits, KnownBits::FromUnsignedUpperBound((uint64_t)relCns, width));
+            }
+            else if (cur.KindIs(Compiler::OAK_GE) && (relCns >= 0))
+            {
+                // num >= 0 (signed)  =>  sign bit is 0.
+                *pBits = KnownBits::Intersect(*pBits, signBitZero);
+            }
+            else if (cur.KindIs(Compiler::OAK_GT) && (relCns >= -1))
+            {
+                // num > -1 (signed)  =>  num >= 0  =>  sign bit is 0.
+                *pBits = KnownBits::Intersect(*pBits, signBitZero);
+            }
+            else if (cur.KindIs(Compiler::OAK_LT) && (relCns >= 1))
+            {
+                // num < C (signed), C >= 1. If num is also non-negative (handled after the loop),
+                // num is in [0, C-1], so record C-1 as a candidate upper bound.
+                const uint64_t ub = (uint64_t)(relCns - 1);
+                if (!haveSignedUpperBound || (ub < signedUpperBound))
+                {
+                    haveSignedUpperBound = true;
+                    signedUpperBound     = ub;
+                }
+            }
+            else if (cur.KindIs(Compiler::OAK_LE) && (relCns >= 0))
+            {
+                // num <= C (signed), C >= 0. If num is also non-negative, num is in [0, C].
+                const uint64_t ub = (uint64_t)relCns;
+                if (!haveSignedUpperBound || (ub < signedUpperBound))
+                {
+                    haveSignedUpperBound = true;
+                    signedUpperBound     = ub;
+                }
+            }
+            continue;
+        }
+
+        // "(uint)num </<= (vn + cns)" where (vn + cns) is non-negative => num is non-negative.
+        //
+        // IsVNNeverNegative on an O2K_VN_ADD_CNS asserts only that the "vn" part is non-negative.
+        // The full expression "vn + cns" can only be guaranteed non-negative when cns == 0, so we
+        // require it explicitly here -- otherwise a negative cns could make the bound itself
+        // negative and we'd derive a false non-negativity fact for num. Same shape as rangecheck.cpp.
+        //
+        if (cur.KindIs(Compiler::OAK_LT_UN, Compiler::OAK_LE_UN) && cur.GetOp2().KindIs(Compiler::O2K_VN_ADD_CNS) &&
+            cur.GetOp2().IsVNNeverNegative() && (cur.GetOp2().GetCns() == 0))
+        {
+            *pBits = KnownBits::Intersect(*pBits, signBitZero);
+        }
+    }
+
+    // If we gathered a signed upper bound and num is now known non-negative (from any of the facts
+    // above or from its value-number structure), num is in [0, signedUpperBound]: its upper bits are 0.
+    // Example: "a > 10 && a < 1000" => sign bit 0 (from a > 10) plus upper bits 0 (from a < 1000),
+    // proving a fits in a smaller type (e.g. making "checked((int)a)" non-overflowing).
+    if (haveSignedUpperBound && ((pBits->knownZero & signBit) != 0))
+    {
+        *pBits = KnownBits::Intersect(*pBits, KnownBits::FromUnsignedUpperBound(signedUpperBound, width));
+    }
+}
+
+//------------------------------------------------------------------------
+// ComputeWorker: Recursive worker for KnownBits::Compute.
+//
+// Arguments:
+//    comp       - the compiler context
+//    num        - the value number to analyze
+//    assertions - the assertion set live at the consumer
+//    budget     - recursive search budget; decremented at every recursive step. Returns the
+//                 fully-unknown lattice when the budget is exhausted.
+//    visited    - set of phi VNs we have already entered, used to guard against infinite recursion
+//                 on loop-carried phis
+//
+// Returns:
+//    KnownBits for "num" within its natural width (32 or 64). Always truncated to that width on
+//    return so the "bits above width are 0/0" invariant holds.
+//
+static KnownBits ComputeWorker(
+    Compiler* comp, ValueNum num, ASSERT_VALARG_TP assertions, int budget, ValueNumStore::SmallValueNumSet* visited)
+{
+    KnownBits result;
+    if ((num == ValueNumStore::NoVN) || (budget <= 0))
+    {
+        return result;
+    }
+
+    const var_types vnType = comp->vnStore->TypeOfVN(num);
+    if (!varTypeIsIntegral(vnType) || varTypeIsGC(vnType))
+    {
+        // We only reason about (non-GC) integral values.
+        return result;
+    }
+
+    const unsigned width = (genActualType(vnType) == TYP_LONG) ? 64 : 32;
+
+    // Constants are fully known.
+    int64_t cnsVal;
+    if (comp->vnStore->IsVNIntegralConstant<int64_t>(num, &cnsVal))
+    {
+        return KnownBits::FromConstant((uint64_t)cnsVal, width);
+    }
+
+    VNFuncApp f;
+    if (comp->vnStore->GetVNFunc(num, &f))
+    {
+        switch (f.GetFunc())
+        {
+            case VNF_AND:
+            case VNF_OR:
+            case VNF_UDIV:
+            {
+                const KnownBits a = ComputeWorker(comp, f.GetArg(0), assertions, --budget, visited);
+                const KnownBits b = ComputeWorker(comp, f.GetArg(1), assertions, --budget, visited);
+
+                if (f.FuncIs(VNF_UDIV))
+                    result = KnownBitsOps::UDiv(a, b, width);
+                else if (f.FuncIs(VNF_AND))
+                    result = KnownBitsOps::And(a, b);
+                else if (f.FuncIs(VNF_OR))
+                    result = KnownBitsOps::Or(a, b);
+                else
+                    unreached();
+                break;
+            }
+
+            case VNF_Cast:
+            case VNF_CastOvf:
+            {
+                var_types castToType;
+                bool      srcIsUnsigned;
+                comp->vnStore->GetCastOperFromVN(f.GetArg(1), &castToType, &srcIsUnsigned);
+
+                const ValueNum  srcVN   = f.GetArg(0);
+                const var_types srcType = comp->vnStore->TypeOfVN(srcVN);
+                if (varTypeIsIntegral(srcType) && !varTypeIsGC(srcType) && varTypeIsIntegral(castToType))
+                {
+                    const unsigned  srcWidth = genTypeSize(genActualType(srcType)) * BITS_PER_BYTE;
+                    const KnownBits bits     = ComputeWorker(comp, srcVN, assertions, --budget, visited);
+                    result                   = KnownBitsOps::Cast(bits, srcWidth, castToType, srcIsUnsigned);
+                }
+                break;
+            }
+
+            case VNF_EQ:
+            case VNF_NE:
+            case VNF_LT:
+            case VNF_LE:
+            case VNF_GT:
+            case VNF_GE:
+            case VNF_LT_UN:
+            case VNF_LE_UN:
+            case VNF_GT_UN:
+            case VNF_GE_UN:
+                // A relop always produces 0 or 1; we don't try to fold the comparison here, just
+                // record the [0, 1] range so a consumer reading this VN sees a single low bit.
+                result = KnownBits::FromUnsignedUpperBound(1, width);
+                break;
+
+            case VNF_MDARR_LENGTH:
+            case VNF_ARR_LENGTH:
+                // Array length is in [0, CORINFO_Array_MaxLength], so its upper bits are 0.
+                result = KnownBits::FromUnsignedUpperBound(CORINFO_Array_MaxLength, width);
+                break;
+
+            default:
+                break;
+        }
+    }
+
+    result = result.Truncate(width);
+
+    // Phi: a bit is known in the phi result only if it is known and equal along every reaching
+    // edge. We Union (LLVM's intersectWith) the per-edge KnownBits to compute that.
+    if (!result.IsConstant(width) && comp->vnStore->IsPhiDef(num) && visited->Add(comp, num))
+    {
+        KnownBits phiBits;
+        bool      first   = true;
+        auto      visitor = [comp, &phiBits, &first, &budget, visited](ValueNum vn, ASSERT_TP reachAss) {
+            const KnownBits edge = ComputeWorker(comp, vn, reachAss, --budget, visited);
+            phiBits              = first ? edge : KnownBits::Union(phiBits, edge);
+            first                = false;
+
+            // Once nothing is known, merging more edges cannot recover any information.
+            return phiBits.IsUnknown() ? Compiler::AssertVisit::Abort : Compiler::AssertVisit::Continue;
+        };
+        if ((comp->optVisitReachingAssertions(num, visitor) == Compiler::AssertVisit::Continue) && !first)
+        {
+            result = KnownBits::Intersect(result, phiBits);
+        }
+    }
+
+    MergeKnownBitsAssertions(comp, num, assertions, width, budget, &result);
+    return result.Truncate(width);
+}
+
+//------------------------------------------------------------------------
+// KnownBits::Compute: Entry point for the bit-level analog of
+//    RangeCheck::GetRangeFromAssertions. Returns which bits of "num" are known 0/1, derived from
+//    its value-number structure and the incoming assertions. Supports 32- and 64-bit integral VNs;
+//    on unsupported types returns the fully-unknown lattice.
+//
+// See KnownBits::Compute in knownbits.h for the parameter documentation.
+//
+KnownBits KnownBits::Compute(Compiler* comp, ValueNum num, ASSERT_VALARG_TP assertions, int budget)
+{
+    ValueNumStore::SmallValueNumSet visited;
+    return ComputeWorker(comp, num, assertions, budget, &visited);
+}