[InstCombine] Combine or-disjoint (and->mul), (and->mul) to and->mul (#136013)

The canonical pattern for bitmasked mul is currently

```
%val = and %x, %bitMask // where %bitMask is some constant
%cmp = icmp eq %val, 0
%sel = select %cmp, 0, %C // where %C is some constant = C' * %bitMask
```

In certain cases, where we are combining multiple of these bitmasked
muls with common factors, we are able to optimize into and->mul (see
#135274 )

This optimization lends itself to further optimizations. This PR
addresses one of such optimizations.

In cases where we have

`or-disjoint ( mul(and (X, C1), D) , mul (and (X, C2), D))`

we can combine into

`mul( and (X, (C1 + C2)), D) `

provided C1 and C2 are disjoint.

Generalized proof: https://alive2.llvm.org/ce/z/MQYMui

Author: jrbyrnes

llvm/lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

@@ -3592,6 +3592,73 @@ static Value *foldOrOfInversions(BinaryOperator &I,
   return nullptr;
 }
 
+// A decomposition of ((X & Mask) * Factor). The NUW / NSW bools
+// track these properities for preservation. Note that we can decompose
+// equivalent select form of this expression (e.g. (!(X & Mask) ? 0 : Mask *
+// Factor))
+struct DecomposedBitMaskMul {
+  Value *X;
+  APInt Factor;
+  APInt Mask;
+  bool NUW;
+  bool NSW;
+};
+
+static std::optional<DecomposedBitMaskMul> matchBitmaskMul(Value *V) {
+  Instruction *Op = dyn_cast<Instruction>(V);
+  if (!Op)
+    return std::nullopt;
+
+  // Decompose (A & N) * C) into BitMaskMul
+  Value *Original = nullptr;
+  const APInt *Mask = nullptr;
+  const APInt *MulConst = nullptr;
+  if (match(Op, m_Mul(m_And(m_Value(Original), m_APInt(Mask)),
+                      m_APInt(MulConst)))) {
+    if (MulConst->isZero() || Mask->isZero())
+      return std::nullopt;
+
+    return std::optional<DecomposedBitMaskMul>(
+        {Original, *MulConst, *Mask,
+         cast<BinaryOperator>(Op)->hasNoUnsignedWrap(),
+         cast<BinaryOperator>(Op)->hasNoSignedWrap()});
+  }
+
+  Value *Cond = nullptr;
+  const APInt *EqZero = nullptr, *NeZero = nullptr;
+
+  // Decompose ((A & N) ? 0 : N * C) into BitMaskMul
+  if (match(Op, m_Select(m_Value(Cond), m_APInt(EqZero), m_APInt(NeZero)))) {
+    auto ICmpDecompose =
+        decomposeBitTest(Cond, /*LookThruTrunc=*/true,
+                         /*AllowNonZeroC=*/false, /*DecomposeBitMask=*/true);
+    if (!ICmpDecompose.has_value())
+      return std::nullopt;
+
+    assert(ICmpInst::isEquality(ICmpDecompose->Pred) &&
+           ICmpDecompose->C.isZero());
+
+    if (ICmpDecompose->Pred == ICmpInst::ICMP_NE)
+      std::swap(EqZero, NeZero);
+
+    if (!EqZero->isZero() || NeZero->isZero())
+      return std::nullopt;
+
+    if (!ICmpDecompose->Mask.isPowerOf2() || ICmpDecompose->Mask.isZero() ||
+        NeZero->getBitWidth() != ICmpDecompose->Mask.getBitWidth())
+      return std::nullopt;
+
+    if (!NeZero->urem(ICmpDecompose->Mask).isZero())
+      return std::nullopt;
+
+    return std::optional<DecomposedBitMaskMul>(
+        {ICmpDecompose->X, NeZero->udiv(ICmpDecompose->Mask),
+         ICmpDecompose->Mask, /*NUW=*/false, /*NSW=*/false});
+  }
+
+  return std::nullopt;
+}
+
 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
 // here. We should standardize that construct where it is needed or choose some
 // other way to ensure that commutated variants of patterns are not missed.
@@ -3674,49 +3741,26 @@ Instruction *InstCombinerImpl::visitOr(BinaryOperator &I) {
                                    /*NSW=*/true, /*NUW=*/true))
       return R;
 
-    Value *Cond0 = nullptr, *Cond1 = nullptr;
-    const APInt *Op0Eq = nullptr, *Op0Ne = nullptr;
-    const APInt *Op1Eq = nullptr, *Op1Ne = nullptr;
-
-    //  (!(A & N) ? 0 : N * C) + (!(A & M) ? 0 : M * C) -> A & (N + M) * C
-    if (match(I.getOperand(0),
-              m_Select(m_Value(Cond0), m_APInt(Op0Eq), m_APInt(Op0Ne))) &&
-        match(I.getOperand(1),
-              m_Select(m_Value(Cond1), m_APInt(Op1Eq), m_APInt(Op1Ne)))) {
-
-      auto LHSDecompose =
-          decomposeBitTest(Cond0, /*LookThruTrunc=*/true,
-                           /*AllowNonZeroC=*/false, /*DecomposeAnd=*/true);
-      auto RHSDecompose =
-          decomposeBitTest(Cond1, /*LookThruTrunc=*/true,
-                           /*AllowNonZeroC=*/false, /*DecomposeAnd=*/true);
-
-      if (LHSDecompose && RHSDecompose && LHSDecompose->X == RHSDecompose->X &&
-          RHSDecompose->Mask.isPowerOf2() && LHSDecompose->Mask.isPowerOf2() &&
-          LHSDecompose->Mask != RHSDecompose->Mask &&
-          LHSDecompose->Mask.getBitWidth() == Op0Ne->getBitWidth() &&
-          RHSDecompose->Mask.getBitWidth() == Op1Ne->getBitWidth()) {
-        assert(Op0Ne->getBitWidth() == Op1Ne->getBitWidth());
-        assert(ICmpInst::isEquality(LHSDecompose->Pred));
-        if (LHSDecompose->Pred == ICmpInst::ICMP_NE)
-          std::swap(Op0Eq, Op0Ne);
-        if (RHSDecompose->Pred == ICmpInst::ICMP_NE)
-          std::swap(Op1Eq, Op1Ne);
-
-        if (!Op0Ne->isZero() && !Op1Ne->isZero() && Op0Eq->isZero() &&
-            Op1Eq->isZero() && Op0Ne->urem(LHSDecompose->Mask).isZero() &&
-            Op1Ne->urem(RHSDecompose->Mask).isZero() &&
-            Op0Ne->udiv(LHSDecompose->Mask) ==
-                Op1Ne->udiv(RHSDecompose->Mask)) {
-          auto NewAnd = Builder.CreateAnd(
-              LHSDecompose->X,
-              ConstantInt::get(LHSDecompose->X->getType(),
-                               (LHSDecompose->Mask + RHSDecompose->Mask)));
-
-          return BinaryOperator::CreateMul(
-              NewAnd, ConstantInt::get(NewAnd->getType(),
-                                       Op0Ne->udiv(LHSDecompose->Mask)));
-        }
+    // (A & N) * C + (A & M) * C -> (A & (N + M)) & C
+    // This also accepts the equivalent select form of (A & N) * C
+    // expressions i.e. !(A & N) ? 0 : N * C)
+    auto Decomp1 = matchBitmaskMul(I.getOperand(1));
+    if (Decomp1) {
+      auto Decomp0 = matchBitmaskMul(I.getOperand(0));
+      if (Decomp0 && Decomp0->X == Decomp1->X &&
+          (Decomp0->Mask & Decomp1->Mask).isZero() &&
+          Decomp0->Factor == Decomp1->Factor) {
+
+        Value *NewAnd = Builder.CreateAnd(
+            Decomp0->X, ConstantInt::get(Decomp0->X->getType(),
+                                         (Decomp0->Mask + Decomp1->Mask)));
+
+        auto *Combined = BinaryOperator::CreateMul(
+            NewAnd, ConstantInt::get(NewAnd->getType(), Decomp1->Factor));
+
+        Combined->setHasNoUnsignedWrap(Decomp0->NUW && Decomp1->NUW);
+        Combined->setHasNoSignedWrap(Decomp0->NSW && Decomp1->NSW);
+        return Combined;
       }
     }
   }

llvm/test/Transforms/InstCombine/or-bitmask.ll

@@ -36,13 +36,9 @@ define i32 @add_select_cmp_and2(i32 %in) {
 
 define i32 @add_select_cmp_and3(i32 %in) {
 ; CHECK-LABEL: @add_select_cmp_and3(
-; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 3
-; CHECK-NEXT:    [[TEMP:%.*]] = mul nuw nsw i32 [[TMP1]], 72
-; CHECK-NEXT:    [[BITOP2:%.*]] = and i32 [[IN]], 4
-; CHECK-NEXT:    [[CMP2:%.*]] = icmp eq i32 [[BITOP2]], 0
-; CHECK-NEXT:    [[SEL2:%.*]] = select i1 [[CMP2]], i32 0, i32 288
-; CHECK-NEXT:    [[OUT:%.*]] = or disjoint i32 [[TEMP]], [[SEL2]]
-; CHECK-NEXT:    ret i32 [[OUT]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 7
+; CHECK-NEXT:    [[TEMP1:%.*]] = mul nuw nsw i32 [[TMP1]], 72
+; CHECK-NEXT:    ret i32 [[TEMP1]]
 ;
   %bitop0 = and i32 %in, 1
   %cmp0 = icmp eq i32 %bitop0, 0
@@ -60,12 +56,9 @@ define i32 @add_select_cmp_and3(i32 %in) {
 
 define i32 @add_select_cmp_and4(i32 %in) {
 ; CHECK-LABEL: @add_select_cmp_and4(
-; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 3
-; CHECK-NEXT:    [[OUT:%.*]] = mul nuw nsw i32 [[TMP1]], 72
-; CHECK-NEXT:    [[TMP2:%.*]] = and i32 [[IN]], 12
-; CHECK-NEXT:    [[TEMP3:%.*]] = mul nuw nsw i32 [[TMP2]], 72
-; CHECK-NEXT:    [[OUT1:%.*]] = or disjoint i32 [[OUT]], [[TEMP3]]
-; CHECK-NEXT:    ret i32 [[OUT1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = and i32 [[IN:%.*]], 15
+; CHECK-NEXT:    [[TEMP2:%.*]] = mul nuw nsw i32 [[TMP2]], 72
+; CHECK-NEXT:    ret i32 [[TEMP2]]
 ;
   %bitop0 = and i32 %in, 1
   %cmp0 = icmp eq i32 %bitop0, 0
@@ -361,6 +354,103 @@ define i64 @mask_select_types_1(i64 %in) {
   ret i64 %out
 }
 
+define i32 @add_select_cmp_mixed1(i32 %in) {
+; CHECK-LABEL: @add_select_cmp_mixed1(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 3
+; CHECK-NEXT:    [[OUT:%.*]] = mul nuw nsw i32 [[TMP1]], 72
+; CHECK-NEXT:    ret i32 [[OUT]]
+;
+  %mask = and i32 %in, 1
+  %sel0 = mul i32 %mask, 72
+  %bitop1 = and i32 %in, 2
+  %cmp1 = icmp eq i32 %bitop1, 0
+  %sel1 = select i1 %cmp1, i32 0, i32 144
+  %out = or disjoint i32 %sel0, %sel1
+  ret i32 %out
+}
+
+define i32 @add_select_cmp_mixed2(i32 %in) {
+; CHECK-LABEL: @add_select_cmp_mixed2(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 3
+; CHECK-NEXT:    [[OUT:%.*]] = mul nuw nsw i32 [[TMP1]], 72
+; CHECK-NEXT:    ret i32 [[OUT]]
+;
+  %bitop0 = and i32 %in, 1
+  %cmp0 = icmp eq i32 %bitop0, 0
+  %mask = and i32 %in, 2
+  %sel0 = select i1 %cmp0, i32 0, i32 72
+  %sel1 = mul i32 %mask, 72
+  %out = or disjoint i32 %sel0, %sel1
+  ret i32 %out
+}
+
+define i32 @add_select_cmp_and_mul(i32 %in) {
+; CHECK-LABEL: @add_select_cmp_and_mul(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 3
+; CHECK-NEXT:    [[OUT:%.*]] = mul nuw nsw i32 [[TMP1]], 72
+; CHECK-NEXT:    ret i32 [[OUT]]
+;
+  %mask0 = and i32 %in, 1
+  %sel0 = mul i32 %mask0, 72
+  %mask1 = and i32 %in, 2
+  %sel1 = mul i32 %mask1, 72
+  %out = or disjoint i32 %sel0, %sel1
+  ret i32 %out
+}
+
+define i32 @add_select_cmp_mixed2_mismatch(i32 %in) {
+; CHECK-LABEL: @add_select_cmp_mixed2_mismatch(
+; CHECK-NEXT:    [[BITOP0:%.*]] = and i32 [[IN:%.*]], 1
+; CHECK-NEXT:    [[CMP0:%.*]] = icmp eq i32 [[BITOP0]], 0
+; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[IN]], 2
+; CHECK-NEXT:    [[SEL0:%.*]] = select i1 [[CMP0]], i32 0, i32 73
+; CHECK-NEXT:    [[SEL1:%.*]] = mul nuw nsw i32 [[MASK]], 72
+; CHECK-NEXT:    [[OUT:%.*]] = or disjoint i32 [[SEL0]], [[SEL1]]
+; CHECK-NEXT:    ret i32 [[OUT]]
+;
+  %bitop0 = and i32 %in, 1
+  %cmp0 = icmp eq i32 %bitop0, 0
+  %mask = and i32 %in, 2
+  %sel0 = select i1 %cmp0, i32 0, i32 73
+  %sel1 = mul i32 %mask, 72
+  %out = or disjoint i32 %sel0, %sel1
+  ret i32 %out
+}
+
+define i32 @add_select_cmp_and_mul_mismatch(i32 %in) {
+; CHECK-LABEL: @add_select_cmp_and_mul_mismatch(
+; CHECK-NEXT:    [[TMP1:%.*]] = trunc i32 [[IN:%.*]] to i1
+; CHECK-NEXT:    [[SEL0:%.*]] = select i1 [[TMP1]], i32 73, i32 0
+; CHECK-NEXT:    [[MASK1:%.*]] = and i32 [[IN]], 2
+; CHECK-NEXT:    [[SEL1:%.*]] = mul nuw nsw i32 [[MASK1]], 72
+; CHECK-NEXT:    [[OUT:%.*]] = or disjoint i32 [[SEL0]], [[SEL1]]
+; CHECK-NEXT:    ret i32 [[OUT]]
+;
+  %mask0 = and i32 %in, 1
+  %sel0 = mul i32 %mask0, 73
+  %mask1 = and i32 %in, 2
+  %sel1 = mul i32 %mask1, 72
+  %out = or disjoint i32 %sel0, %sel1
+  ret i32 %out
+}
+
+define i32 @and_mul_non_disjoint(i32 %in) {
+; CHECK-LABEL: @and_mul_non_disjoint(
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[IN:%.*]], 2
+; CHECK-NEXT:    [[OUT:%.*]] = mul nuw nsw i32 [[TMP1]], 72
+; CHECK-NEXT:    [[MASK1:%.*]] = and i32 [[IN]], 4
+; CHECK-NEXT:    [[SEL1:%.*]] = mul nuw nsw i32 [[MASK1]], 72
+; CHECK-NEXT:    [[OUT1:%.*]] = or i32 [[OUT]], [[SEL1]]
+; CHECK-NEXT:    ret i32 [[OUT1]]
+;
+  %mask0 = and i32 %in, 2
+  %sel0 = mul i32 %mask0, 72
+  %mask1 = and i32 %in, 4
+  %sel1 = mul i32 %mask1, 72
+  %out = or i32 %sel0, %sel1
+  ret i32 %out
+}
+
 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
 ; CONSTSPLAT: {{.*}}
 ; CONSTVEC: {{.*}}