[ValueTracking] Recognize that and(x, add (x, -1)) clears the low bit

[reames]

Summary

This is a cleaned up version of a patch by John Regehr with permission. Originally found via the souper tool.

If we add an odd number to x, then bitwise-and the result with x, we know that the low bit of the result must be zero. Either it was zero in x originally, or the add cleared it in the temporary value. As a result, one of the two values anded together must have the bit cleared.

[sanjoy]

Minor optional to fix nit inline

lib/Analysis/ValueTracking.cpp

1092 ↗

(On Diff #39148)

This is minor:

I think (but perhaps @majnemer can confirm this) instcombine canonicalizes operands (for commutative operations) so that the LHS is more complex than the RHS, so maybe there is not much value in checking for both "x & (x - 1)" and "(x - 1) & x" -- instcombine will transform the former to the latter.

[reames]

lib/Analysis/ValueTracking.cpp

1092 ↗

(On Diff #39148)

I don't think it will in this case. I believe the definition of complexity is on the type of the operand (constant, non-constant, etc..), not the form of the operand.

llvm/trunk/lib/Analysis/ValueTracking.cpp

case Instruction::And: {
     // If either the LHS or the RHS are Zero, the result is zero.
     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
 
     // Output known-1 bits are only known if set in both the LHS & RHS.
     KnownOne &= KnownOne2;
     // Output known-0 are known to be clear if zero in either the LHS | RHS.
     KnownZero |= KnownZero2;
+
+    // and(x, add (x, -1)) is a common idiom that always clears the low bit;
+    // here we handle the more general case of adding any odd number by
+    // matching the form add(x, add(x, y)) where y is odd.
+    // TODO: This could be generalized to clearing any bit set in y where the
+    // following bit is known to be unset in y.
+    Value *Y = nullptr;
+    if (match(I->getOperand(0), m_Add(m_Specific(I->getOperand(1)),
+                                      m_Value(Y))) ||
+        match(I->getOperand(1), m_Add(m_Specific(I->getOperand(0)),
+                                      m_Value(Y)))) {
+      APInt KnownZero3(BitWidth, 0), KnownOne3(BitWidth, 0);
+      computeKnownBits(Y, KnownZero3, KnownOne3, DL, Depth + 1, Q);
+      if (KnownOne3.countTrailingOnes() > 0)
+        KnownZero |= APInt::getLowBitsSet(BitWidth, 1);
+    }
     break;
   }
   case Instruction::Or: {
     computeKnownBits(I->getOperand(1), KnownZero, KnownOne, DL, Depth + 1, Q);
     computeKnownBits(I->getOperand(0), KnownZero2, KnownOne2, DL, Depth + 1, Q);
 
     // Output known-0 bits are only known if clear in both the LHS & RHS.
     KnownZero &= KnownZero2;

llvm/trunk/test/Transforms/InstSimplify/add-mask.ll

+; RUN: opt -S -instsimplify < %s | FileCheck %s
+
+define i1 @test(i32 %a) {
+; CHECK-LABEL: @test
+; CHECK: ret i1 false
+  %rhs = add i32 %a, -1
+  %and = and i32 %a, %rhs
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}
+
+define i1 @test2(i32 %a) {
+; CHECK-LABEL: @test2
+; CHECK: ret i1 false
+  %rhs = add i32 %a, 1
+  %and = and i32 %a, %rhs
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}
+
+define i1 @test3(i32 %a) {
+; CHECK-LABEL: @test3
+; CHECK: ret i1 false
+  %rhs = add i32 %a, 7
+  %and = and i32 %a, %rhs
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}
+
+@B = external global i32
+declare void @llvm.assume(i1)
+
+; Known bits without a constant
+define i1 @test4(i32 %a) {
+; CHECK-LABEL: @test4
+; CHECK: ret i1 false
+  %b = load i32, i32* @B
+  %b.and = and i32 %b, 1
+  %b.cnd = icmp eq i32 %b.and, 1
+  call void @llvm.assume(i1 %b.cnd)
+
+  %rhs = add i32 %a, %b
+  %and = and i32 %a, %rhs
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}
+
+; Negative test - even number
+define i1 @test5(i32 %a) {
+; CHECK-LABEL: @test5
+; CHECK: ret i1 %res
+  %rhs = add i32 %a, 2
+  %and = and i32 %a, %rhs
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}
+
+define i1 @test6(i32 %a) {
+; CHECK-LABEL: @test6
+; CHECK: ret i1 false
+  %lhs = add i32 %a, -1
+  %and = and i32 %lhs, %a
+  %res = icmp eq i32 %and, 1
+  ret i1 %res
+}