[InstCombine] Canonicalize variable mask in masked merge

[lebedev.ri]

Summary

Masked merge has a pattern of: ((x ^ y) & M) ^ y.
But, there is no difference between ((x ^ y) & M) ^ y and ((x ^ y) & ~M) ^ x,
We should canonicalize it and drop that xor.

https://rise4fun.com/Alive/Yol

[lebedev.ri]

Slightly cleanup the matcher by using m_CombineAnd(),
which allows to use m_c_And() when looking for A part.

I'm not sure whether it actually matter because i was not able
to come up with a testcase, but it looks a bit cleaner anyway..

[lebedev.ri]

Rebased

[spatel]

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

2493–2495

Make sure I haven't missed anything, but it seems easier to match this inline here more simply as:

// Eliminate a 'not' from xor form of masked-merge (aka, bitwise select).
// The 'not' must be operand 1 of the 'and' (complexity-based op ordering).
// ((Op1 ^ X) & ~Y) ^ Op1 --> ((Op1 ^ X) & Y) ^ X
if (match(Op0, m_OneUse(m_And(m_c_Xor(m_Specific(Op1), m_Value(X)),
                              m_Not(m_Value(Y)))))) {
  Value *Op00 = cast<BinaryOperator>(Op0)->getOperand(0);
  return BinaryOperator::CreateXor(Builder.CreateAnd(Op00, Y), X);
}
// Op0 ^ ((Op0 ^ X) & ~Y) --> ((Op0 ^ X) & Y) ^ X
if (match(Op1, m_OneUse(m_And(m_c_Xor(m_Specific(Op0), m_Value(X)),
                              m_Not(m_Value(Y)))))) {
  Value *Op10 = cast<BinaryOperator>(Op1)->getOperand(0);
  return BinaryOperator::CreateXor(Builder.CreateAnd(Op10, Y), X);
}

[lebedev.ri]

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

2493–2495

I'm sorry, i don't really like this, especially after seeing the DAGCombiner :)
I'm going to work on fixing the commutability problem of BinaryOp_match and m_Specific().
(if that fix won't be accepted, will revisit this.)

[lebedev.ri]

Rebased ontop of D45828

[lebedev.ri]

Streamlined the matcher, rebased ontop of revisited multiuse tests.

[lebedev.ri]

• Rebased.
• Ping. This differential is the base, and has no unmerged dependencies, so this is the one to review :)

[spatel]

LGTM - see inline for some nits.

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

2421–2427

The comment seems inverted from our usual description of:
Starting pattern --> Ending pattern

2439

Call this 'NotM' to comply with LLVM capitalization rules?

[lebedev.ri]

@spatel thank you for the review!

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

2421–2427

It is because this very function will be dealing with unfolding the pattern in case of constant mask (D45867), then the description will be appended.
But i tried to reword it a little.

lib/Transforms/InstCombine/InstCombineAndOrXor.cpp

if (Value *AndICmp = SimplifyBinOp(Instruction::And, LHS, RHS, SQ)) {
         return Builder.CreateAnd(LHS, RHS);
       }
     }
   }
 
   return nullptr;
 }
 
+/// If we have a masked merge, in the form of:
+///   |        A  |  |B|
+///   ((x ^ y) & M) ^ y
+///    |  D  |
+/// If A has one use, and M is inverted, we can convert it into:
+///    |  D  |
+///   ((x ^ y) & ~M) ^ x
+static Instruction *invertMaskInMaskedMerge(BinaryOperator &I,
+                                            InstCombiner::BuilderTy &Builder) {
+  Value *B, *X, *D, *M;
+  if (!match(&I, m_c_Xor(m_Value(B),
+                         m_OneUse(m_c_And(
+                             m_CombineAnd(m_c_Xor(m_c_Specific(&B), m_Value(X)),
+                                          m_Value(D)),
+                             m_Not(m_Value(M)))))))
+    return nullptr;
+
+  // De-invert the mask and swap the value in B part.
+  Value *NewA = Builder.CreateAnd(D, M);
+  return BinaryOperator::CreateXor(NewA, X);
+}
+
 // FIXME: We use commutative matchers (m_c_*) for some, but not all, matches
 // here. We should standardize that construct where it is needed or choose some
 // other way to ensure that commutated variants of patterns are not missed.
 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
   bool Changed = SimplifyAssociativeOrCommutative(I);
   Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
 
   if (Value *V = SimplifyVectorOp(I))
Instruction *InstCombiner::visitXor(BinaryOperator &I) {
   }
   // ~(~X | Y) --> (X & ~Y)
   // ~(Y | ~X) --> (X & ~Y)
   if (match(&I, m_Not(m_OneUse(m_c_Or(m_Not(m_Value(X)), m_Value(Y)))))) {
     Value *NotY = Builder.CreateNot(Y, Y->getName() + ".not");
     return BinaryOperator::CreateAnd(X, NotY);
   }
 
+  if (Instruction *Xor = invertMaskInMaskedMerge(I, Builder))
+    return Xor;
+
   // Is this a 'not' (~) fed by a binary operator?
   BinaryOperator *NotVal;
   if (match(&I, m_Not(m_BinOp(NotVal)))) {
     if (NotVal->getOpcode() == Instruction::And ||
         NotVal->getOpcode() == Instruction::Or) {
       // Apply DeMorgan's Law when inverts are free:
       // ~(X & Y) --> (~X | ~Y)
       // ~(X | Y) --> (~X & ~Y)

test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-scalar.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 ; If we have a masked merge, in the form of: (M is not constant)
 ;   ((x ^ y) & ~M) ^ y
 ; We can drop de-invert the M:
 ;   ((x ^ y) & M) ^ x
 
 define i4 @scalar (i4 %x, i4 %y, i4 %m) {
 ; CHECK-LABEL: @scalar(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %n0 = xor i4 %x, %y
   %n1 = and i4 %n0, %im
   %r  = xor i4 %n1, %y
   ret i4 %r
 }
 
 ; ============================================================================ ;
 ; Commutativity
 ; ============================================================================ ;
 
 ; Used to make sure that the IR complexity sorting does not interfere.
 declare i4 @gen4()
 
 ; FIXME: should the  %n1 = and i4 %im, %n0  swapped order pattern be tested?
 
 define i4 @c_1_0_0 (i4 %x, i4 %y, i4 %m) {
 ; CHECK-LABEL: @c_1_0_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %n0 = xor i4 %y, %x ; swapped order
   %n1 = and i4 %n0, %im
   %r  = xor i4 %n1, %y
   ret i4 %r
 }
 
 define i4 @c_0_1_0 (i4 %x, i4 %y, i4 %m) {
 ; CHECK-LABEL: @c_0_1_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %n0 = xor i4 %x, %y
   %n1 = and i4 %n0, %im
   %r  = xor i4 %n1, %x ; %x instead of %y
   ret i4 %r
 }
 
 define i4 @c_0_0_1 (i4 %m) {
 ; CHECK-LABEL: @c_0_0_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[X:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[Y:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X]], [[Y]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[Y]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %x  = call i4 @gen4()
   %y  = call i4 @gen4()
   %n0 = xor i4 %x, %y
   %n1 = and i4 %n0, %im
   %r  = xor i4 %y, %n1 ; swapped order
   ret i4 %r
 }
 
 define i4 @c_1_1_0 (i4 %x, i4 %y, i4 %m) {
 ; CHECK-LABEL: @c_1_1_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %n0 = xor i4 %y, %x ; swapped order
   %n1 = and i4 %n0, %im
   %r  = xor i4 %n1, %x ; %x instead of %y
   ret i4 %r
 }
 
 define i4 @c_1_0_1 (i4 %x, i4 %m) {
 ; CHECK-LABEL: @c_1_0_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[Y:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[Y]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %y  = call i4 @gen4()
   %n0 = xor i4 %y, %x ; swapped order
   %n1 = and i4 %n0, %im
   %r  = xor i4 %y, %n1 ; swapped order
   ret i4 %r
 }
 
 define i4 @c_0_1_1 (i4 %y, i4 %m) {
 ; CHECK-LABEL: @c_0_1_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[X:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[X]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %x  = call i4 @gen4()
   %n0 = xor i4 %x, %y
   %n1 = and i4 %n0, %im
   %r  = xor i4 %x, %n1 ; swapped order, %x instead of %y
   ret i4 %r
 }
 
 define i4 @c_1_1_1 (i4 %m) {
 ; CHECK-LABEL: @c_1_1_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor i4 [[M:%.*]], -1
 ; CHECK-NEXT:    [[X:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[Y:%.*]] = call i4 @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y]], [[X]]
-; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i4 [[X]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret i4 [[R]]
 ;
   %im = xor i4 %m, -1
   %x  = call i4 @gen4()
   %y  = call i4 @gen4()
   %n0 = xor i4 %y, %x ; swapped order
   %n1 = and i4 %n0, %im
   %r  = xor i4 %x, %n1 ; swapped order, %x instead of %y

test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-vector.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 ; If we have a masked merge, in the form of: (M is not constant)
 ;   ((x ^ y) & ~M) ^ y
 ; We can drop de-invert the M:
 ;   ((x ^ y) & M) ^ x
 
 define <2 x i4> @vector (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
 ; CHECK-LABEL: @vector(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %n0 = xor <2 x i4> %x, %y
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %n1, %y
   ret <2 x i4> %r
 }
 
 define <3 x i4> @vector_undef (<3 x i4> %x, <3 x i4> %y, <3 x i4> %m) {
 ; CHECK-LABEL: @vector_undef(
-; CHECK-NEXT:    [[IM:%.*]] = xor <3 x i4> [[M:%.*]], <i4 -1, i4 undef, i4 -1>
 ; CHECK-NEXT:    [[N0:%.*]] = xor <3 x i4> [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <3 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <3 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret <3 x i4> [[R]]
 ;
   %im = xor <3 x i4> %m, <i4 -1, i4 undef, i4 -1>
   %n0 = xor <3 x i4> %x, %y
   %n1 = and <3 x i4> %n0, %im
   %r  = xor <3 x i4> %n1, %y
   ret <3 x i4> %r
 }
 
 ; ============================================================================ ;
 ; Commutativity
 ; ============================================================================ ;
 
 ; Used to make sure that the IR complexity sorting does not interfere.
 declare <2 x i4> @gen4()
 
 ; FIXME: should  %n1 = and <2 x i4> %im, %n0  swapped order pattern be tested?
 
 define <2 x i4> @c_1_0_0 (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
 ; CHECK-LABEL: @c_1_0_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %n0 = xor <2 x i4> %y, %x ; swapped order
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %n1, %y
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_0_1_0 (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
 ; CHECK-LABEL: @c_0_1_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %n0 = xor <2 x i4> %x, %y
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %n1, %x ; %x instead of %y
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_0_0_1 (<2 x i4> %m) {
 ; CHECK-LABEL: @c_0_0_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[X:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[Y:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X]], [[Y]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[Y]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %x  = call <2 x i4> @gen4()
   %y  = call <2 x i4> @gen4()
   %n0 = xor <2 x i4> %x, %y
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %y, %n1 ; swapped order
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_1_1_0 (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
 ; CHECK-LABEL: @c_1_1_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[X]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %n0 = xor <2 x i4> %y, %x ; swapped order
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %n1, %x ; %x instead of %y
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_1_0_1 (<2 x i4> %x, <2 x i4> %m) {
 ; CHECK-LABEL: @c_1_0_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[Y:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[Y]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %y  = call <2 x i4> @gen4()
   %n0 = xor <2 x i4> %y, %x ; swapped order
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %y, %n1 ; swapped order
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_0_1_1 (<2 x i4> %y, <2 x i4> %m) {
 ; CHECK-LABEL: @c_0_1_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[X:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[X]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %x  = call <2 x i4> @gen4()
   %n0 = xor <2 x i4> %x, %y
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %x, %n1 ; swapped order, %x instead of %y
   ret <2 x i4> %r
 }
 
 define <2 x i4> @c_1_1_1 (<2 x i4> %m) {
 ; CHECK-LABEL: @c_1_1_1(
-; CHECK-NEXT:    [[IM:%.*]] = xor <2 x i4> [[M:%.*]], <i4 -1, i4 -1>
 ; CHECK-NEXT:    [[X:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[Y:%.*]] = call <2 x i4> @gen4()
 ; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y]], [[X]]
-; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[X]], [[N1]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[TMP1]], [[Y]]
 ; CHECK-NEXT:    ret <2 x i4> [[R]]
 ;
   %im = xor <2 x i4> %m, <i4 -1, i4 -1>
   %x  = call <2 x i4> @gen4()
   %y  = call <2 x i4> @gen4()
   %n0 = xor <2 x i4> %y, %x ; swapped order
   %n1 = and <2 x i4> %n0, %im
   %r  = xor <2 x i4> %x, %n1 ; swapped order, %x instead of %y

test/Transforms/InstCombine/masked-merge-with-xors.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -instcombine -S | FileCheck %s
 
 ; Masked merge-*like* pattern, but with some operands inverted.
 
 define i32 @f_0(i32 %x, i32 %y, i32 %m) {
 ; CHECK-LABEL: @f_0(
-; CHECK-NEXT:    [[IM:%.*]] = xor i32 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i32 [[Y:%.*]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i32 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i32 [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i32 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %im = xor i32 %m, -1
   %n0 = xor i32 %y, %x
   %n1 = and i32 %n0, %im ; could de-invert the mask
   %r = xor i32 %n1, %y
   ret i32 %r
 }
 
 define i32 @f_1(i32 %x, i32 %y, i32 %m) {
 ; CHECK-LABEL: @f_1(
 ; CHECK-NEXT:    [[IX:%.*]] = xor i32 [[X:%.*]], -1
-; CHECK-NEXT:    [[IM:%.*]] = xor i32 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i32 [[IX]], [[Y:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i32 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i32 [[N1]], [[Y]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i32 [[TMP1]], [[IX]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %ix = xor i32 %x, -1
   %im = xor i32 %m, -1
   %n0 = xor i32 %ix, %y
   %n1 = and i32 %n0, %im ; could de-invert the mask
   %r = xor i32 %n1, %y
   ret i32 %r
 }
 
 ; Should be
 ; %iy = xor i32 %y, -1
 ; %b0 = and i32 %iy, %m
 ; %b1 = and i32 %x, %im
 ; %r = or i32 %b0, %b1
 define i32 @f_2(i32 %x, i32 %y, i32 %m) {
 ; CHECK-LABEL: @f_2(
 ; CHECK-NEXT:    [[IY:%.*]] = xor i32 [[Y:%.*]], -1
-; CHECK-NEXT:    [[IM:%.*]] = xor i32 [[M:%.*]], -1
 ; CHECK-NEXT:    [[N0:%.*]] = xor i32 [[IY]], [[X:%.*]]
-; CHECK-NEXT:    [[N1:%.*]] = and i32 [[N0]], [[IM]]
-; CHECK-NEXT:    [[R:%.*]] = xor i32 [[N1]], [[IY]]
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i32 [[TMP1]], [[X]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
   %iy = xor i32 %y, -1
   %im = xor i32 %m, -1
   %n0 = xor i32 %iy, %x
   %n1 = and i32 %n0, %im
   %r = xor i32 %n1, %iy
   ret i32 %r