[InstCombine] Canonicalize `(X^(X-1)) u{ge,lt} X` as pow2 test

[goldstein.w.n]

Summary

Also hoist the current canoniclizations from foldICmpEquality to
keep all the logic in the same place.
https://alive2.llvm.org/ce/z/T8osF6

[nikic]

Could you please split off the extraction of foldICmpPow2Test into an NFC patch (feel free to just commit it)?

[goldstein.w.n]

In D152673#4413211, @nikic wrote:

Could you please split off the extraction of foldICmpPow2Test into an NFC patch (feel free to just commit it)?

Done.

[goldstein.w.n]

Split off the factoring out

[nikic]

LGTM

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

5000

format

[goldstein.w.n]

Fix fmt

[goldstein.w.n]

Rebase

[goldstein.w.n]

Rebase

[goldstein.w.n]

Rebase

[goldstein.w.n]

Rebase

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

 }
 
 // Canonicalize checking for a power-of-2-or-zero value:
 static Instruction *foldICmpPow2Test(ICmpInst &I,
                                      InstCombiner::BuilderTy &Builder) {
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
   const CmpInst::Predicate Pred = I.getPredicate();
   Value *A = nullptr;
+  bool CheckIs;
   if (I.isEquality()) {
     // (A & (A-1)) == 0 --> ctpop(A) < 2 (two commuted variants)
     // ((A-1) & A) != 0 --> ctpop(A) > 1 (two commuted variants)
     if (!match(Op0, m_OneUse(m_c_And(m_Add(m_Value(A), m_AllOnes()),
                                      m_Deferred(A)))) ||
         !match(Op1, m_ZeroInt()))
       A = nullptr;
 
     // (A & -A) == A --> ctpop(A) < 2 (four commuted variants)
     // (-A & A) != A --> ctpop(A) > 1 (four commuted variants)
     if (match(Op0, m_OneUse(m_c_And(m_Neg(m_Specific(Op1)), m_Specific(Op1)))))
       A = Op1;
     else if (match(Op1,
                    m_OneUse(m_c_And(m_Neg(m_Specific(Op0)), m_Specific(Op0)))))
       A = Op0;
+
+    CheckIs = Pred == ICmpInst::ICMP_EQ;
+  } else if (Pred == ICmpInst::ICMP_UGE || Pred == ICmpInst::ICMP_ULT) {
+    // (A ^ (A-1)) u>= A --> ctpop(A) < 2 (two commuted variants)
+    // ((A-1) ^ A) u< A --> ctpop(A) > 1 (two commuted variants)
+    if (match(Op0, m_OneUse(m_c_Xor(m_Add(m_Specific(Op1), m_AllOnes()),
+                                    m_Specific(Op1)))))
+      A = Op1;
+    else if (match(Op1, m_OneUse(m_c_Xor(m_Add(m_Specific(Op0), m_AllOnes()),
+                                         m_Specific(Op0)))))
+      A = Op0;
+
+    CheckIs = Pred == ICmpInst::ICMP_UGE;
   }
+
   if (A) {
     Type *Ty = A->getType();
     CallInst *CtPop = Builder.CreateUnaryIntrinsic(Intrinsic::ctpop, A);
-    return Pred == ICmpInst::ICMP_EQ ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop,
-                                                    ConstantInt::get(Ty, 2))
-                                     : new ICmpInst(ICmpInst::ICMP_UGT, CtPop,
-                                                    ConstantInt::get(Ty, 1));
+    return CheckIs ? new ICmpInst(ICmpInst::ICMP_ULT, CtPop,
   }
 
   return nullptr;
 }
 
 Instruction *InstCombinerImpl::foldICmpEquality(ICmpInst &I) {
   if (!I.isEquality())
     return nullptr;

llvm/test/Transforms/InstCombine/ispow2.ll

;
   %x_and_nx = and i32 %x, %nx
   %r = icmp sge i32 %x_and_nx, %x
   ret i1 %r
 }
 
 define i1 @blsmsk_is_p2_or_z(i32 %xx, i32 %yy) {
 ; CHECK-LABEL: @blsmsk_is_p2_or_z(
 ; CHECK-NEXT:    [[X:%.*]] = or i32 [[XX:%.*]], [[YY:%.*]]
-; CHECK-NEXT:    [[XM1:%.*]] = add i32 [[X]], -1
-; CHECK-NEXT:    [[Y:%.*]] = xor i32 [[X]], [[XM1]]
-; CHECK-NEXT:    [[R:%.*]] = icmp uge i32 [[X]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X]]), !range [[RNG0]]
+; CHECK-NEXT:    [[R:%.*]] = icmp ult i32 [[TMP1]], 2
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %x = or i32 %xx, %yy
   %xm1 = add i32 %x, -1
   %y = xor i32 %x, %xm1
   %r = icmp uge i32 %x, %y
   ret i1 %r
 }
 
 define i1 @blsmsk_isnt_p2_or_z(i32 %x) {
 ; CHECK-LABEL: @blsmsk_isnt_p2_or_z(
-; CHECK-NEXT:    [[XM1:%.*]] = add i32 [[X:%.*]], -1
-; CHECK-NEXT:    [[Y:%.*]] = xor i32 [[XM1]], [[X]]
-; CHECK-NEXT:    [[R:%.*]] = icmp ult i32 [[Y]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
+; CHECK-NEXT:    [[R:%.*]] = icmp ugt i32 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   %xm1 = add i32 %x, -1
   %y = xor i32 %x, %xm1
   %r = icmp ult i32 %y, %x
   ret i1 %r
 }