[llvm] [PatternMatch] Do not accept undef elements in m_AllOnes() and friends (PR #88217)

[Nikita Popov via llvm-commits]

================
@@ -306,11 +306,22 @@ define <2 x i7> @all_ones_left_right_splat(<2 x i7> %x) {
 
 ; Poison could propagate, but undef must not.
 
-define <3 x i7> @all_ones_left_right_splat_poison_undef_elt(<3 x i7> %x) {
-; CHECK-LABEL: @all_ones_left_right_splat_poison_undef_elt(
+define <3 x i7> @all_ones_left_right_splat_undef_elt(<3 x i7> %x) {
+; CHECK-LABEL: @all_ones_left_right_splat_undef_elt(
+; CHECK-NEXT:    [[LEFT:%.*]] = shl <3 x i7> <i7 undef, i7 -1, i7 undef>, [[X:%.*]]
+; CHECK-NEXT:    [[RIGHT:%.*]] = ashr <3 x i7> [[LEFT]], [[X]]
+; CHECK-NEXT:    ret <3 x i7> [[RIGHT]]
+;
+  %left = shl <3 x i7> <i7 undef, i7 -1, i7 undef>, %x
+  %right = ashr <3 x i7> %left, %x
+  ret <3 x i7> %right
+}
+
+define <3 x i7> @all_ones_left_right_splat_poison__elt(<3 x i7> %x) {
+; CHECK-LABEL: @all_ones_left_right_splat_poison__elt(
 ; CHECK-NEXT:    ret <3 x i7> <i7 -1, i7 -1, i7 -1>
 ;
-  %left = shl <3 x i7> <i7 poison, i7 -1, i7 undef>, %x
+  %left = shl <3 x i7> <i7 poison, i7 -1, i7 poison>, %x
----------------
nikic wrote:

No, the transform was fine as it did not propagate the undef. This test corresponds to the `// We could return the original -1 constant to preserve poison elements.` comment change in InstructionSimplify.cpp: Now that we only handle poison elements, we *could* propagate the poison elements. I didn't make the actual change for that in this PR.

https://github.com/llvm/llvm-project/pull/88217