Index: test/Transforms/InstCombine/invert-const-mask-in-masked-merge-scalar.ll
===================================================================
--- /dev/null
+++ test/Transforms/InstCombine/invert-const-mask-in-masked-merge-scalar.ll
@@ -0,0 +1,436 @@
+; 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 constant)
+;   ((x ^ y) & M) ^ y
+; Invert it, if the inverted mask looks better
+;   ((x ^ y) & ~M) ^ x
+
+; This file, unlike vector.ll is kinda big. This allows to keep that file small.
+; And to have a good test coverage, for all the combinations that could result
+; in cyclic canonicalization, if the canon form is not well-defined.
+
+; Naming scheme: @t_<bitwidth>_<mask>
+
+define i3 @t_3_1 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_1(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], 1
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, 1
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i3 @t_3_2 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_2(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], 2
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, 2
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i3 @t_3_3 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_3(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], 3
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, 3
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i3 @t_3_4 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_4(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], -4
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, -4
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i3 @t_3_5 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_5(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], -3
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, -3
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i3 @t_3_6 (i3 %x, i3 %y) {
+; CHECK-LABEL: @t_3_6(
+; CHECK-NEXT:    [[N0:%.*]] = xor i3 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i3 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i3 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i3 [[R]]
+;
+  %n0 = xor i3 %x, %y
+  %n1 = and i3 %n0, -2
+  %r  = xor i3 %n1, %y
+  ret i3 %r
+}
+
+define i4 @t_4_1 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_1(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 1
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 1
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_2 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_2(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 2
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_3 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_3(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 3
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 3
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_4 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_4(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 4
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 4
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_5 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_5(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 5
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 5
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_6 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_6(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 6
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 6
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_7 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_7(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], 7
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, 7
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_8 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_8(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -8
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -8
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_9 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_9(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -7
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -7
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_10 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_10(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -6
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -6
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_11 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_11(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -5
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -5
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_12 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_12(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -4
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -4
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_13 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_13(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -3
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -3
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @t_4_14 (i4 %x, i4 %y) {
+; CHECK-LABEL: @t_4_14(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+; ============================================================================ ;
+; Commutativity
+; ============================================================================ ;
+
+; Used to make sure that the IR complexity sorting does not interfere.
+declare i4 @gen4()
+
+define i4 @c_1_0_0 (i4 %x, i4 %y) {
+; CHECK-LABEL: @c_1_0_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y:%.*]], [[X:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %y, %x ; swapped order
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+define i4 @c_0_1_0 (i4 %x, i4 %y) {
+; CHECK-LABEL: @c_0_1_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[X]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %x ; %x instead of %y
+  ret i4 %r
+}
+
+define i4 @c_0_0_1 () {
+; CHECK-LABEL: @c_0_0_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]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[Y]], [[N1]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %x  = call i4 @gen4()
+  %y  = call i4 @gen4()
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %y, %n1 ; swapped order
+  ret i4 %r
+}
+
+define i4 @c_1_1_0 (i4 %x, i4 %y) {
+; CHECK-LABEL: @c_1_1_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y:%.*]], [[X:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[X]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %y, %x ; swapped order
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %x ; %x instead of %y
+  ret i4 %r
+}
+
+define i4 @c_1_0_1 (i4 %x) {
+; CHECK-LABEL: @c_1_0_1(
+; CHECK-NEXT:    [[Y:%.*]] = call i4 @gen4()
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[Y]], [[X:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[Y]], [[N1]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %y  = call i4 @gen4()
+  %n0 = xor i4 %y, %x ; swapped order
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %y, %n1 ; swapped order
+  ret i4 %r
+}
+
+define i4 @c_0_1_1 (i4 %y) {
+; CHECK-LABEL: @c_0_1_1(
+; CHECK-NEXT:    [[X:%.*]] = call i4 @gen4()
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[X]], [[N1]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %x  = call i4 @gen4()
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %x, %n1 ; swapped order, %x instead of %y
+  ret i4 %r
+}
+
+define i4 @c_1_1_1 () {
+; CHECK-LABEL: @c_1_1_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]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[X]], [[N1]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %x  = call i4 @gen4()
+  %y  = call i4 @gen4()
+  %n0 = xor i4 %y, %x ; swapped order
+  %n1 = and i4 %n0, -2
+  %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) {
+; CHECK-LABEL: @n_oneuse(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    call void @use4(i4 [[N1]])
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2 ; two uses of %n1, which is going to be replaced
+  %r  = xor i4 %n1, %y
+  call void @use4(i4 %n1)
+  ret i4 %r
+}
+
+; Mask is not constant
+
+define i4 @n_var_mask (i4 %x, i4 %y, i4 %m) {
+; CHECK-LABEL: @n_var_mask(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Y]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, %m
+  %r  = xor i4 %n1, %y
+  ret i4 %r
+}
+
+; Some third variable is used
+
+define i4 @n_third_var (i4 %x, i4 %y, i4 %z) {
+; CHECK-LABEL: @n_third_var(
+; CHECK-NEXT:    [[N0:%.*]] = xor i4 [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and i4 [[N0]], -2
+; CHECK-NEXT:    [[R:%.*]] = xor i4 [[N1]], [[Z:%.*]]
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %z ; not %x or %y
+  ret i4 %r
+}
Index: test/Transforms/InstCombine/invert-const-mask-in-masked-merge-vector.ll
===================================================================
--- /dev/null
+++ test/Transforms/InstCombine/invert-const-mask-in-masked-merge-vector.ll
@@ -0,0 +1,258 @@
+; 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 constant)
+;   ((x ^ y) & M) ^ y
+; Invert it, if the inverted mask looks better
+;   ((x ^ y) & ~M) ^ x
+
+; In vector case, do inversion if there are more mask elements that are
+; non-canonical, than canonical elements. Undef is considered canonical.
+
+; All the heavy lifting has already been done in scalar.ll testfile,
+; so here we can simply test just a few simple cases.
+
+define <2 x i4> @splat (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @splat(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %r  = xor <2 x i4> %n1, %y
+  ret <2 x i4> %r
+}
+
+define <3 x i4> @splat_undef (<3 x i4> %x, <3 x i4> %y) {
+; CHECK-LABEL: @splat_undef(
+; CHECK-NEXT:    [[N0:%.*]] = xor <3 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <3 x i4> [[N0]], <i4 -2, i4 undef, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <3 x i4> [[R]]
+;
+  %n0 = xor <3 x i4> %x, %y
+  %n1 = and <3 x i4> %n0, <i4 -2, i4 undef, i4 -2>
+  %r  = xor <3 x i4> %n1, %y
+  ret <3 x i4> %r
+}
+
+define <3 x i4> @nonsplat (<3 x i4> %x, <3 x i4> %y) {
+; CHECK-LABEL: @nonsplat(
+; CHECK-NEXT:    [[N0:%.*]] = xor <3 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <3 x i4> [[N0]], <i4 -2, i4 1, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <3 x i4> [[R]]
+;
+  %n0 = xor <3 x i4> %x, %y
+  %n1 = and <3 x i4> %n0, <i4 -2, i4 1, i4 -2> ; 2 of 3 -> do invert
+  %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()
+
+define <2 x i4> @c_1_0_0 (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @c_1_0_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y:%.*]], [[X:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %y, %x ; swapped order
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %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) {
+; CHECK-LABEL: @c_0_1_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[X]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %r  = xor <2 x i4> %n1, %x ; %x instead of %y
+  ret <2 x i4> %r
+}
+
+define <2 x i4> @c_0_0_1 () {
+; CHECK-LABEL: @c_0_0_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]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[Y]], [[N1]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %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, <i4 -2, i4 -2>
+  %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) {
+; CHECK-LABEL: @c_1_1_0(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[Y:%.*]], [[X:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[X]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %y, %x ; swapped order
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %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) {
+; CHECK-LABEL: @c_1_0_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]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[Y]], [[N1]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %y  = call <2 x i4> @gen4()
+  %n0 = xor <2 x i4> %y, %x ; swapped order
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %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) {
+; CHECK-LABEL: @c_0_1_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]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[X]], [[N1]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %x  = call <2 x i4> @gen4()
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %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 () {
+; CHECK-LABEL: @c_1_1_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]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[X]], [[N1]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %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, <i4 -2, i4 -2>
+  %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) {
+; CHECK-LABEL: @n_oneuse(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; 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]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2> ; 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
+}
+
+; Mask is not constant
+
+define <2 x i4> @n_var_mask (<2 x i4> %x, <2 x i4> %y, <2 x i4> %m) {
+; CHECK-LABEL: @n_var_mask(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], [[M:%.*]]
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, %m
+  %r  = xor <2 x i4> %n1, %y
+  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) {
+; CHECK-LABEL: @n_third_var(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 -2>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Z:%.*]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %r  = xor <2 x i4> %n1, %z ; not %x or %y
+  ret <2 x i4> %r
+}
+
+; ============================================================================ ;
+; Negative tests with non-splat vectors. Should not be folded.
+; ============================================================================ ;
+
+define <2 x i4> @n_vec (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @n_vec(
+; CHECK-NEXT:    [[N0:%.*]] = xor <2 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <2 x i4> [[N0]], <i4 -2, i4 1>
+; CHECK-NEXT:    [[R:%.*]] = xor <2 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 1> ; 1 of 2 -> don't invert
+  %r  = xor <2 x i4> %n1, %y
+  ret <2 x i4> %r
+}
+
+define <3 x i4> @n_vec_undef (<3 x i4> %x, <3 x i4> %y) {
+; CHECK-LABEL: @n_vec_undef(
+; CHECK-NEXT:    [[N0:%.*]] = xor <3 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <3 x i4> [[N0]], <i4 -2, i4 undef, i4 1>
+; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <3 x i4> [[R]]
+;
+  %n0 = xor <3 x i4> %x, %y
+  %n1 = and <3 x i4> %n0, <i4 -2, i4 undef, i4 1> ; 1 of 3 -> don't invert
+  %r  = xor <3 x i4> %n1, %y
+  ret <3 x i4> %r
+}
+
+define <3 x i4> @n_vec_good (<3 x i4> %x, <3 x i4> %y) {
+; CHECK-LABEL: @n_vec_good(
+; CHECK-NEXT:    [[N0:%.*]] = xor <3 x i4> [[X:%.*]], [[Y:%.*]]
+; CHECK-NEXT:    [[N1:%.*]] = and <3 x i4> [[N0]], <i4 1, i4 -2, i4 1>
+; CHECK-NEXT:    [[R:%.*]] = xor <3 x i4> [[N1]], [[Y]]
+; CHECK-NEXT:    ret <3 x i4> [[R]]
+;
+  %n0 = xor <3 x i4> %x, %y
+  %n1 = and <3 x i4> %n0, <i4 1, i4 -2, i4 1> ; 1 of 3 -> don't invert
+  %r  = xor <3 x i4> %n1, %y
+  ret <3 x i4> %r
+}