[InstCombine] SimplifyDemandedBits - mul(x,x) is odd iff x is odd

[spatel]

Summary

This was mentioned in D118539.

With proof adapted from a comment there:
https://alive2.llvm.org/ce/z/AXPr3k

[RKSimon]

LGTM - cheers

[lebedev.ri]

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

550–556

Can we generalize this to something along the lines of:
if we only demand 2 low bits, then replace with and %x, 0b00001 ?

[lebedev.ri]

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

550–556

https://alive2.llvm.org/ce/z/C2ihC2

[spatel]

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

550–556

Yes, that does seem better. I'll go ahead and commit this patch as a minimum improvement, then update with more tests and fix up the code to reduce risk.

[spatel]

79b3fe80707b

llvm/lib/Transforms/InstCombine/InstCombineSimplifyDemanded.cpp

case Instruction::Sub: {
 
     // Otherwise just compute the known bits of the result.
     bool NSW = cast<OverflowingBinaryOperator>(I)->hasNoSignedWrap();
     Known = KnownBits::computeForAddSub(I->getOpcode() == Instruction::Add,
                                         NSW, LHSKnown, RHSKnown);
     break;
   }
   case Instruction::Mul: {
+    if (DemandedMask.isPowerOf2()) {
+      if (I->getOperand(0) == I->getOperand(1)) {
+        // X * X is odd iff X is odd.
+        if (DemandedMask == 1)
+          return I->getOperand(0);
+
+        // 'Quadratic Reciprocity': mul(x,x) -> 0 if we're only demanding bit[1]
+        if (DemandedMask == 2)
+          return ConstantInt::getNullValue(VTy);
+      }
+
       // The LSB of X*Y is set only if (X & 1) == 1 and (Y & 1) == 1.
       // If we demand exactly one bit N and we have "X * (C' << N)" where C' is
       // odd (has LSB set), then the left-shifted low bit of X is the answer.
-    if (DemandedMask.isPowerOf2()) {
       unsigned CTZ = DemandedMask.countTrailingZeros();
       const APInt *C;
       if (match(I->getOperand(1), m_APInt(C)) &&
           C->countTrailingZeros() == CTZ) {
         Constant *ShiftC = ConstantInt::get(I->getType(), CTZ);
         Instruction *Shl = BinaryOperator::CreateShl(I->getOperand(0), ShiftC);
         return InsertNewInstWith(Shl, *I);
       }
-      // 'Quadratic Reciprocity': mul(x,x) -> 0 if we're only demanding bit[1]
-      if (DemandedMask == 2 && I->getOperand(0) == I->getOperand(1))
-        return ConstantInt::getNullValue(VTy);
     }
     computeKnownBits(I, Known, Depth, CxtI);
     break;
   }
   case Instruction::Shl: {
     const APInt *SA;
     if (match(I->getOperand(1), m_APInt(SA))) {
       const APInt *ShrAmt;

llvm/test/Transforms/InstCombine/mul-masked-bits.ll

 ;
   %m = mul i67 %x, -63 ; any odd number will do
   %r = and i67 %m, 1
   ret i67 %r
 }
 
 define i33 @squared_one_demanded_low_bit(i33 %x) {
 ; CHECK-LABEL: @squared_one_demanded_low_bit(
-; CHECK-NEXT:    [[MUL:%.*]] = mul i33 [[X:%.*]], [[X]]
-; CHECK-NEXT:    [[AND:%.*]] = and i33 [[MUL]], 1
-; CHECK-NEXT:    ret i33 [[AND]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i33 [[X:%.*]], 1
+; CHECK-NEXT:    ret i33 [[TMP1]]
 ;
   %mul = mul i33 %x, %x
   %and = and i33 %mul, 1
   ret i33 %and
 }
 
 define <2 x i8> @squared_one_demanded_low_bit_splat(<2 x i8> %x) {
 ; CHECK-LABEL: @squared_one_demanded_low_bit_splat(
-; CHECK-NEXT:    [[MUL:%.*]] = mul <2 x i8> [[X:%.*]], [[X]]
-; CHECK-NEXT:    [[AND:%.*]] = or <2 x i8> [[MUL]], <i8 -2, i8 -2>
+; CHECK-NEXT:    [[AND:%.*]] = or <2 x i8> [[X:%.*]], <i8 -2, i8 -2>
 ; CHECK-NEXT:    ret <2 x i8> [[AND]]
 ;
   %mul = mul <2 x i8> %x, %x
   %and = or <2 x i8> %mul, <i8 254, i8 254>
   ret <2 x i8> %and
 }