[KnownBits] Add knownbits analysis for mulhs/mulu 'multiply high' instructions

[RKSimon]

Summary

Split off from D98857

[RKSimon]

Fix typo in comment noticed by @foad on D98857

[foad]

LGTM.

[nikic]

llvm/lib/Support/KnownBits.cpp

498

Side note, we should rename computeForMul -> mul.

[foad]

llvm/lib/Support/KnownBits.cpp

498

I would vote for operator*.

[lebedev.ri]

LG, thank you.

[RKSimon]

Cheers - there's a sort of convergence going on that KnownBits function names match APInt (although APInt doesn't have any mulh methods) - not sure about operators though. I've been burnt in the past with the & operator as I thought it would be to determine common bits......

[RKSimon]

Closed by rGa96897219daf43a1b90e1e0c9dbf20167c0c16af

llvm/include/llvm/Support/KnownBits.h

public:
 
   /// Compute known bits resulting from adding LHS and RHS.
   static KnownBits computeForAddSub(bool Add, bool NSW, const KnownBits &LHS,
                                     KnownBits RHS);
 
   /// Compute known bits resulting from multiplying LHS and RHS.
   static KnownBits computeForMul(const KnownBits &LHS, const KnownBits &RHS);
 
+  /// Compute known bits from sign-extended multiply-hi.
+  static KnownBits mulhs(const KnownBits &LHS, const KnownBits &RHS);
+
+  /// Compute known bits from zero-extended multiply-hi.
+  static KnownBits mulhu(const KnownBits &LHS, const KnownBits &RHS);
+
   /// Compute known bits for udiv(LHS, RHS).
   static KnownBits udiv(const KnownBits &LHS, const KnownBits &RHS);
 
   /// Compute known bits for urem(LHS, RHS).
   static KnownBits urem(const KnownBits &LHS, const KnownBits &RHS);
 
   /// Compute known bits for srem(LHS, RHS).
   static KnownBits srem(const KnownBits &LHS, const KnownBits &RHS);

llvm/lib/Support/KnownBits.cpp

KnownBits KnownBits::computeForMul(const KnownBits &LHS, const KnownBits &RHS) {
 
   KnownBits Res(BitWidth);
   Res.Zero.setHighBits(LeadZ);
   Res.Zero |= (~BottomKnown).getLoBits(ResultBitsKnown);
   Res.One = BottomKnown.getLoBits(ResultBitsKnown);
   return Res;
 }
 
+KnownBits KnownBits::mulhs(const KnownBits &LHS, const KnownBits &RHS) {
+  unsigned BitWidth = LHS.getBitWidth();
+  assert(BitWidth == RHS.getBitWidth() && !LHS.hasConflict() &&
+         !RHS.hasConflict() && "Operand mismatch");
+  KnownBits WideLHS = LHS.sext(2 * BitWidth);
+  KnownBits WideRHS = RHS.sext(2 * BitWidth);
+  return computeForMul(WideLHS, WideRHS).extractBits(BitWidth, BitWidth);
+}
+
+KnownBits KnownBits::mulhu(const KnownBits &LHS, const KnownBits &RHS) {
+  unsigned BitWidth = LHS.getBitWidth();
+  assert(BitWidth == RHS.getBitWidth() && !LHS.hasConflict() &&
+         !RHS.hasConflict() && "Operand mismatch");
+  KnownBits WideLHS = LHS.zext(2 * BitWidth);
+  KnownBits WideRHS = RHS.zext(2 * BitWidth);
+  return computeForMul(WideLHS, WideRHS).extractBits(BitWidth, BitWidth);
+}
+
 KnownBits KnownBits::udiv(const KnownBits &LHS, const KnownBits &RHS) {
   unsigned BitWidth = LHS.getBitWidth();
   assert(!LHS.hasConflict() && !RHS.hasConflict());
   KnownBits Known(BitWidth);
 
   // For the purposes of computing leading zeros we can conservatively
   // treat a udiv as a logical right shift by the power of 2 known to
   // be less than the denominator.

llvm/unittests/Support/KnownBitsTest.cpp

ForeachKnownBits(Bits, [&](const KnownBits &Known2) {
       KnownAnd.One.setAllBits();
       KnownBits KnownOr(KnownAnd);
       KnownBits KnownXor(KnownAnd);
       KnownBits KnownUMax(KnownAnd);
       KnownBits KnownUMin(KnownAnd);
       KnownBits KnownSMax(KnownAnd);
       KnownBits KnownSMin(KnownAnd);
       KnownBits KnownMul(KnownAnd);
+      KnownBits KnownMulHS(KnownAnd);
+      KnownBits KnownMulHU(KnownAnd);
       KnownBits KnownUDiv(KnownAnd);
       KnownBits KnownURem(KnownAnd);
       KnownBits KnownSRem(KnownAnd);
       KnownBits KnownShl(KnownAnd);
       KnownBits KnownLShr(KnownAnd);
       KnownBits KnownAShr(KnownAnd);
 
       ForeachNumInKnownBits(Known1, [&](const APInt &N1) {
ForeachKnownBits(Bits, [&](const KnownBits &Known2) {
           Res = APIntOps::smin(N1, N2);
           KnownSMin.One &= Res;
           KnownSMin.Zero &= ~Res;
 
           Res = N1 * N2;
           KnownMul.One &= Res;
           KnownMul.Zero &= ~Res;
 
+          Res = (N1.sext(2 * Bits) * N2.sext(2 * Bits)).extractBits(Bits, Bits);
+          KnownMulHS.One &= Res;
+          KnownMulHS.Zero &= ~Res;
+
+          Res = (N1.zext(2 * Bits) * N2.zext(2 * Bits)).extractBits(Bits, Bits);
+          KnownMulHU.One &= Res;
+          KnownMulHU.Zero &= ~Res;
+
           if (!N2.isNullValue()) {
             Res = N1.udiv(N2);
             KnownUDiv.One &= Res;
             KnownUDiv.Zero &= ~Res;
 
             Res = N1.urem(N2);
             KnownURem.One &= Res;
             KnownURem.Zero &= ~Res;
ForeachKnownBits(Bits, [&](const KnownBits &Known2) {
       KnownBits ComputedSMax = KnownBits::smax(Known1, Known2);
       EXPECT_EQ(KnownSMax.Zero, ComputedSMax.Zero);
       EXPECT_EQ(KnownSMax.One, ComputedSMax.One);
 
       KnownBits ComputedSMin = KnownBits::smin(Known1, Known2);
       EXPECT_EQ(KnownSMin.Zero, ComputedSMin.Zero);
       EXPECT_EQ(KnownSMin.One, ComputedSMin.One);
 
-      // ComputedMul is conservatively correct, but not guaranteed to be
+      // The following are conservatively correct, but not guaranteed to be
       // precise.
       KnownBits ComputedMul = KnownBits::computeForMul(Known1, Known2);
       EXPECT_TRUE(ComputedMul.Zero.isSubsetOf(KnownMul.Zero));
       EXPECT_TRUE(ComputedMul.One.isSubsetOf(KnownMul.One));
 
+      KnownBits ComputedMulHS = KnownBits::mulhs(Known1, Known2);
+      EXPECT_TRUE(ComputedMulHS.Zero.isSubsetOf(KnownMulHS.Zero));
+      EXPECT_TRUE(ComputedMulHS.One.isSubsetOf(KnownMulHS.One));
+
+      KnownBits ComputedMulHU = KnownBits::mulhu(Known1, Known2);
+      EXPECT_TRUE(ComputedMulHU.Zero.isSubsetOf(KnownMulHU.Zero));
+      EXPECT_TRUE(ComputedMulHU.One.isSubsetOf(KnownMulHU.One));
+
       KnownBits ComputedUDiv = KnownBits::udiv(Known1, Known2);
       EXPECT_TRUE(ComputedUDiv.Zero.isSubsetOf(KnownUDiv.Zero));
       EXPECT_TRUE(ComputedUDiv.One.isSubsetOf(KnownUDiv.One));
 
       KnownBits ComputedURem = KnownBits::urem(Known1, Known2);
       EXPECT_TRUE(ComputedURem.Zero.isSubsetOf(KnownURem.Zero));
       EXPECT_TRUE(ComputedURem.One.isSubsetOf(KnownURem.One));