[InstCombine] Fold adds + shifts with nsw and nuw flags

I also added mul nsw/nuw 3, div 2 since this was the canonical version of ((x << 1) + x) / 2, which is a specific expression which canonicalization causes the InstCombine to miss it.

llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp

@@ -1267,6 +1267,19 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
       match(Op1, m_SpecificIntAllowUndef(BitWidth - 1)))
     return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
 
+  // Special Case:
+  // if both the add and the shift are nuw, we can omit the AND entirely
+  // ((X << Y) nuw + Z nuw) >>u Z --> (X + (Y >>u Z))
+  Value *Y;
+  if (match(Op0, m_OneUse(m_c_NUWAdd((m_NUWShl(m_Value(X), m_Specific(Op1))),
+                                     m_Value(Y))))) {
+    Value *NewLshr = Builder.CreateLShr(Y, Op1, "", I.isExact());
+    auto *newAdd = BinaryOperator::CreateNUWAdd(NewLshr, X);
+    if (auto *Op0Bin = cast<BinaryOperator>(Op0))
+      newAdd->setHasNoSignedWrap(Op0Bin->hasNoSignedWrap());
+    return newAdd;
+  }
+
   if (match(Op1, m_APInt(C))) {
     unsigned ShAmtC = C->getZExtValue();
     auto *II = dyn_cast<IntrinsicInst>(Op0);
@@ -1283,7 +1296,6 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
       return new ZExtInst(Cmp, Ty);
     }
 
-    Value *X;
     const APInt *C1;
     if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) {
       if (C1->ult(ShAmtC)) {
@@ -1328,7 +1340,7 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
     // ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C)
     // TODO: Consolidate with the more general transform that starts from shl
     //       (the shifts are in the opposite order).
-    Value *Y;
+
     if (match(Op0,
               m_OneUse(m_c_Add(m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))),
                                m_Value(Y))))) {
@@ -1450,9 +1462,25 @@ Instruction *InstCombinerImpl::visitLShr(BinaryOperator &I) {
           NewMul->setHasNoSignedWrap(true);
           return NewMul;
         }
+
+        // Special case:
+        // lshr nuw (mul (X, 3), 1) -> add nuw nsw (X, lshr(X, 1)
+        if (ShAmtC == 1 && MulC->getZExtValue() == 3) {
+          auto *NewAdd = BinaryOperator::CreateNUWAdd(
+              X,
+              Builder.CreateLShr(X, ConstantInt::get(Ty, 1), "", I.isExact()));
+          NewAdd->setHasNoSignedWrap(true);
+          return NewAdd;
+        }
       }
     }
 
+    // // lshr nsw (mul (X, 3), 1) -> add nsw (X, lshr(X, 1)
+    if (match(Op0, m_OneUse(m_NSWMul(m_Value(X), m_SpecificInt(3)))) &&
+        ShAmtC == 1)
+      return BinaryOperator::CreateNSWAdd(
+          X, Builder.CreateLShr(X, ConstantInt::get(Ty, 1), "", I.isExact()));
+
     // Try to narrow bswap.
     // In the case where the shift amount equals the bitwidth difference, the
     // shift is eliminated.
@@ -1656,6 +1684,26 @@ Instruction *InstCombinerImpl::visitAShr(BinaryOperator &I) {
       if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
         return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty);
     }
+
+    // Special case: ashr nuw (mul (X, 3), 1) -> add nuw nsw (X, lshr(X, 1)
+    if (match(Op0, m_OneUse(m_NSWMul(m_Value(X), m_SpecificInt(3)))) &&
+        ShAmt == 1) {
+      Value *Shift;
+      if (auto *Op0Bin = cast<BinaryOperator>(Op0)) {
+        if (Op0Bin->hasNoUnsignedWrap())
+          // We can use lshr if the mul is nuw and nsw
+          Shift =
+              Builder.CreateLShr(X, ConstantInt::get(Ty, 1), "", I.isExact());
+        else
+          Shift =
+              Builder.CreateAShr(X, ConstantInt::get(Ty, 1), "", I.isExact());
+
+        auto *NewAdd = BinaryOperator::CreateNSWAdd(X, Shift);
+        NewAdd->setHasNoUnsignedWrap(Op0Bin->hasNoUnsignedWrap());
+
+        return NewAdd;
+      }
+    }
   }
 
   const SimplifyQuery Q = SQ.getWithInstruction(&I);

llvm/test/Transforms/InstCombine/lshr.ll

@@ -362,9 +362,9 @@ define <3 x i14> @mul_splat_fold_vec(<3 x i14> %x) {
 
 define i32 @mul_times_3_div_2 (i32 %x) {
 ; CHECK-LABEL: @mul_times_3_div_2(
-; CHECK-NEXT:    [[TMP1:%.*]] = mul nuw nsw i32 [[X:%.*]], 3
-; CHECK-NEXT:    [[TMP2:%.*]] = lshr i32 [[TMP1]], 1
-; CHECK-NEXT:    ret i32 [[TMP2]]
+; CHECK-NEXT:    [[TMP2:%.*]] = lshr i32 [[TMP1:%.*]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP2]], [[TMP1]]
+; CHECK-NEXT:    ret i32 [[TMP3]]
 ;
   %2 = mul nsw nuw i32 %x, 3
   %3 = lshr i32 %2, 1
@@ -373,21 +373,20 @@ define i32 @mul_times_3_div_2 (i32 %x) {
 
   define i32 @shl_add_lshr (i32 %x, i32 %c, i32 %y) {
 ; CHECK-LABEL: @shl_add_lshr(
-; CHECK-NEXT:    [[TMP1:%.*]] = shl nuw i32 [[X:%.*]], [[C:%.*]]
-; CHECK-NEXT:    [[TMP2:%.*]] = add nuw nsw i32 [[TMP1]], [[Y:%.*]]
-; CHECK-NEXT:    [[TMP3:%.*]] = lshr exact i32 [[TMP2]], [[C]]
-; CHECK-NEXT:    ret i32 [[TMP3]]
+; CHECK-NEXT:    [[TMP3:%.*]] = lshr exact i32 [[TMP2:%.*]], [[C:%.*]]
+; CHECK-NEXT:    [[TMP4:%.*]] = add nuw nsw i32 [[TMP3]], [[X:%.*]]
+; CHECK-NEXT:    ret i32 [[TMP4]]
 ;
   %2 = shl nuw i32 %x, %c
-  %3 = add nsw nuw i32 %2, %y
+  %3 = add nuw nsw i32 %2, %y
   %4 = lshr exact i32 %3, %c
   ret i32 %4
 }
 
   define i32 @ashr_mul_times_3_div_2 (i32 %0) {
 ; CHECK-LABEL: @ashr_mul_times_3_div_2(
-; CHECK-NEXT:    [[TMP2:%.*]] = mul nuw nsw i32 [[TMP0:%.*]], 3
-; CHECK-NEXT:    [[TMP3:%.*]] = ashr i32 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP2:%.*]] = lshr i32 [[TMP0:%.*]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = add nuw nsw i32 [[TMP2]], [[TMP0]]
 ; CHECK-NEXT:    ret i32 [[TMP3]]
 ;
   %2 = mul nsw nuw i32 %0, 3
@@ -397,9 +396,9 @@ define i32 @mul_times_3_div_2 (i32 %x) {
 
 define i32 @ashr_mul_times_3_div_2_exact (i32 %0) {
 ; CHECK-LABEL: @ashr_mul_times_3_div_2_exact(
-; CHECK-NEXT:    [[TMP2:%.*]] = mul nsw i32 [[TMP0:%.*]], 3
-; CHECK-NEXT:    [[TMP3:%.*]] = ashr exact i32 [[TMP2]], 1
-; CHECK-NEXT:    ret i32 [[TMP3]]
+; CHECK-NEXT:    [[TMP3:%.*]] = ashr exact i32 [[TMP2:%.*]], 1
+; CHECK-NEXT:    [[TMP4:%.*]] = add nsw i32 [[TMP3]], [[TMP2]]
+; CHECK-NEXT:    ret i32 [[TMP4]]
 ;
   %2 = mul nsw i32 %0, 3
   %3 = ashr exact i32 %2, 1