[InstCombine][NFC] Add tests for unfolding masked merge with constant mask

[lebedev.ri]

Summary

As discussed in D45733, we want to do this in InstCombine.

[lebedev.ri]

Self-accepting since this is just NFC new tests, and dependent differential is accepted.

test/Transforms/InstCombine/unfold-masked-merge-with-const-mask-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 constant)
+;   ((x ^ y) & M) ^ y
+; Unfold it to
+;   (x & M) | (y & ~M)
+
+define i4 @scalar0 (i4 %x, i4 %y) {
+; CHECK-LABEL: @scalar0(
+; 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 @scalar1 (i4 %x, i4 %y) {
+; CHECK-LABEL: @scalar1(
+; 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_D (i4 %x, i4 %y) {
+; CHECK-LABEL: @n_oneuse_D(
+; 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 [[N0]])
+; CHECK-NEXT:    ret i4 [[R]]
+;
+  %n0 = xor i4 %x, %y ; two uses of %n0, which is going to be replaced
+  %n1 = and i4 %n0, -2
+  %r  = xor i4 %n1, %y
+  call void @use4(i4 %n0)
+  ret i4 %r
+}
+
+define i4 @n_oneuse_A (i4 %x, i4 %y) {
+; CHECK-LABEL: @n_oneuse_A(
+; 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
+}
+
+define i4 @n_oneuse_AD (i4 %x, i4 %y) {
+; CHECK-LABEL: @n_oneuse_AD(
+; 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 [[N0]])
+; 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 %n0)
+  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
+}

test/Transforms/InstCombine/unfold-masked-merge-with-const-mask-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 constant)
+;   ((x ^ y) & M) ^ y
+; Unfold it to
+;   (x & M) | (y & ~M)
+
+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 <2 x i4> @nonsplat (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @nonsplat(
+; 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>
+  %r  = xor <2 x i4> %n1, %y
+  ret <2 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_D (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @n_oneuse_D(
+; 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> [[N0]])
+; CHECK-NEXT:    ret <2 x i4> [[R]]
+;
+  %n0 = xor <2 x i4> %x, %y ; two uses of %n0, which is going to be replaced
+  %n1 = and <2 x i4> %n0, <i4 -2, i4 -2>
+  %r  = xor <2 x i4> %n1, %y
+  call void @use4(<2 x i4> %n0)
+  ret <2 x i4> %r
+}
+
+define <2 x i4> @n_oneuse_A (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @n_oneuse_A(
+; 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
+}
+
+define <2 x i4> @n_oneuse_AD (<2 x i4> %x, <2 x i4> %y) {
+; CHECK-LABEL: @n_oneuse_AD(
+; 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> [[N0]])
+; 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> %n0)
+  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
+}