[InstCombine] Fold mul (shr exact (X, N)), 2^N + 1 -> add (X , shr exact (X, N)) (#112407)

Alive2 Proofs:
https://alive2.llvm.org/ce/z/aJnxyp
https://alive2.llvm.org/ce/z/dyeGEv

Author: AZero13

llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

@@ -255,6 +255,33 @@ Instruction *InstCombinerImpl::visitMul(BinaryOperator &I) {
     }
   }
 
+  // mul (shr exact X, N), (2^N + 1) -> add (X, shr exact (X, N))
+  {
+    Value *NewOp;
+    const APInt *ShiftC;
+    const APInt *MulAP;
+    if (BitWidth > 2 &&
+        match(&I, m_Mul(m_Exact(m_Shr(m_Value(NewOp), m_APInt(ShiftC))),
+                        m_APInt(MulAP))) &&
+        (*MulAP - 1).isPowerOf2() && *ShiftC == MulAP->logBase2()) {
+      Value *BinOp = Op0;
+      BinaryOperator *OpBO = cast<BinaryOperator>(Op0);
+
+      // mul nuw (ashr exact X, N) -> add nuw (X, lshr exact (X, N))
+      if (HasNUW && OpBO->getOpcode() == Instruction::AShr && OpBO->hasOneUse())
+        BinOp = Builder.CreateLShr(NewOp, ConstantInt::get(Ty, *ShiftC), "",
+                                   /*isExact=*/true);
+
+      auto *NewAdd = BinaryOperator::CreateAdd(NewOp, BinOp);
+      if (HasNSW && (HasNUW || OpBO->getOpcode() == Instruction::LShr ||
+                     ShiftC->getZExtValue() < BitWidth - 1))
+        NewAdd->setHasNoSignedWrap(true);
+
+      NewAdd->setHasNoUnsignedWrap(HasNUW);
+      return NewAdd;
+    }
+  }
+
   if (Op0->hasOneUse() && match(Op1, m_NegatedPower2())) {
     // Interpret  X * (-1<<C)  as  (-X) * (1<<C)  and try to sink the negation.
     // The "* (1<<C)" thus becomes a potential shifting opportunity.

llvm/test/Transforms/InstCombine/ashr-lshr.ll

@@ -1077,4 +1077,148 @@ entry:
   ret i32 %shr
 }
 
+define i32 @ashr_shift_mul(i32 %x) {
+; CHECK-LABEL: @ashr_shift_mul(
+; CHECK-NEXT:    [[A:%.*]] = ashr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr exact i32 %x, 3
+  %res = mul i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @ashr_shift_mul_nuw(i32 %x) {
+; CHECK-LABEL: @ashr_shift_mul_nuw(
+; CHECK-NEXT:    [[TMP1:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add nuw i32 [[X]], [[TMP1]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr exact i32 %x, 3
+  %res = mul nuw i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @ashr_shift_mul_nsw(i32 %x) {
+; CHECK-LABEL: @ashr_shift_mul_nsw(
+; CHECK-NEXT:    [[A:%.*]] = ashr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add nsw i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr exact i32 %x, 3
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @lshr_shift_mul_nuw(i32 %x) {
+; CHECK-LABEL: @lshr_shift_mul_nuw(
+; CHECK-NEXT:    [[A:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add nuw i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr exact i32 %x, 3
+  %res = mul nuw i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @lshr_shift_mul(i32 %x) {
+; CHECK-LABEL: @lshr_shift_mul(
+; CHECK-NEXT:    [[A:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr exact i32 %x, 3
+  %res = mul i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @lshr_shift_mul_nsw(i32 %x) {
+; CHECK-LABEL: @lshr_shift_mul_nsw(
+; CHECK-NEXT:    [[A:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = add nsw i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr exact i32 %x, 3
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
+; Negative test
+
+define i32 @lshr_no_exact(i32 %x) {
+; CHECK-LABEL: @lshr_no_exact(
+; CHECK-NEXT:    [[A:%.*]] = lshr i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = mul nuw nsw i32 [[A]], 9
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr i32 %x, 3
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
+; Negative test
+
+define i32 @ashr_no_exact(i32 %x) {
+; CHECK-LABEL: @ashr_no_exact(
+; CHECK-NEXT:    [[A:%.*]] = ashr i32 [[X:%.*]], 3
+; CHECK-NEXT:    [[RES:%.*]] = mul nsw i32 [[A]], 9
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr i32 %x, 3
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @lshr_multiuse(i32 %x) {
+; CHECK-LABEL: @lshr_multiuse(
+; CHECK-NEXT:    [[A:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    call void @use(i32 [[A]])
+; CHECK-NEXT:    [[RES:%.*]] = add nsw i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr exact i32 %x, 3
+  call void @use(i32 %a)
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @lshr_multiuse_no_flags(i32 %x) {
+; CHECK-LABEL: @lshr_multiuse_no_flags(
+; CHECK-NEXT:    [[A:%.*]] = lshr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    call void @use(i32 [[A]])
+; CHECK-NEXT:    [[RES:%.*]] = add i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = lshr exact i32 %x, 3
+  call void @use(i32 %a)
+  %res = mul i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @ashr_multiuse_no_flags(i32 %x) {
+; CHECK-LABEL: @ashr_multiuse_no_flags(
+; CHECK-NEXT:    [[A:%.*]] = ashr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    call void @use(i32 [[A]])
+; CHECK-NEXT:    [[RES:%.*]] = add i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr exact i32 %x, 3
+  call void @use(i32 %a)
+  %res = mul i32 %a, 9
+  ret i32 %res
+}
+
+define i32 @ashr_multiuse(i32 %x) {
+; CHECK-LABEL: @ashr_multiuse(
+; CHECK-NEXT:    [[A:%.*]] = ashr exact i32 [[X:%.*]], 3
+; CHECK-NEXT:    call void @use(i32 [[A]])
+; CHECK-NEXT:    [[RES:%.*]] = add nsw i32 [[X]], [[A]]
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %a = ashr exact i32 %x, 3
+  call void @use(i32 %a)
+  %res = mul nsw i32 %a, 9
+  ret i32 %res
+}
+
 declare void @use(i32)