Index: test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-scalar.ll
===================================================================
--- /dev/null
+++ test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-scalar.ll
@@ -0,0 +1,205 @@
+; 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:    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:    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:    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:    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:    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:    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:    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:    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
+  ret i4 %r
+}
+
+; ============================================================================ ;
+; Negative tests. Should not be folded.
+; ============================================================================ ;
+
+; One use only.
+
+declare void @use4(i4)
+
+define i4 @n_oneuse (i4 %x, i4 %y, i4 %m) {
+; CHECK-LABEL: @n_oneuse(
+; 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:    call void @use4(i4 [[N1]])
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %im = xor i4 %m, -1
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, %im ; two uses of %n1, which is going to be replaced
+  %r  = xor i4 %n1, %y
+  call void @use4(i4 %n1)
+  ret i4 %r
+}
+
+; Some third variable is used
+
+define i4 @n_third_var (i4 %x, i4 %y, i4 %z, i4 %m) {
+; CHECK-LABEL: @n_third_var(
+; 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]], [[Z:%.*]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %im = xor i4 %m, -1
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, %im
+  %r  = xor i4 %n1, %z ; not %x or %y
+  ret i4 %r
+}
+
+define i4 @n_badxor (i4 %x, i4 %y, i4 %m) {
+; CHECK-LABEL: @n_badxor(
+; 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:    ret i4 [[R]]
+;
+  %im = xor i4 %m, 1 ; not -1
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, %im
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
Index: test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-vector.ll
===================================================================
--- /dev/null
+++ test/Transforms/InstCombine/invert-variable-mask-in-masked-merge-vector.ll
@@ -0,0 +1,237 @@
+; 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:    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:    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:    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:    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:    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:    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:    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:    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:    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
+  ret <2 x i4> %r
+}
+
+; ============================================================================ ;
+; Negative tests. Should not be folded.
+; ============================================================================ ;
+
+; One use only.
+
+declare void @use4(<2 x i4>)
+
+define <2 x i4> @n_oneuse (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
+; CHECK-LABEL: @n_oneuse(
+; 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:    call void @use4(<2 x i4> [[N1]])
+; 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 ; two uses of %n1, which is going to be replaced
+  %r  = xor <2 x i4> %n1, %y
+  call void @use4(<2 x i4> %n1)
+  ret <2 x i4> %r
+}
+
+; Some third variable is used
+
+define <2 x i4> @n_third_var (<2 x i4> %x, <2 x i4> %y, <2 x i4> %z, <2 x i4> %m) {
+; CHECK-LABEL: @n_third_var(
+; 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]], [[Z:%.*]]
+; 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, %z ; not %x or %y
+  ret <2 x i4> %r
+}
+
+; Bad xor
+
+define <2 x i4> @n_badxor_splat (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
+; CHECK-LABEL: @n_badxor_splat(
+; 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:    ret <2 x i4> [[R]]
+;
+  %im = xor <2 x i4> %m, <i4 1, i4 1> ; not -1
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, %im ; two uses of %n1, which is going to be replaced
+  %r  = xor <2 x i4> %n1, %y
+  ret <2 x i4> %r
+}
+
+define <2 x i4> @n_badxor (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
+; CHECK-LABEL: @n_badxor(
+; 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:    ret <2 x i4> [[R]]
+;
+  %im = xor <2 x i4> %m, <i4 -1, i4 1> ; not -1
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, %im ; two uses of %n1, which is going to be replaced
+  %r  = xor <2 x i4> %n1, %y
+  ret <2 x i4> %r
+}