[llvm] [VectorCombine] Scalarize extracts of ZExt if profitable. (PR #142976)

[Florian Hahn via llvm-commits]

================
@@ -9,15 +9,25 @@ define void @zext_v4i8_all_lanes_used(<4 x i8> %src) {
 ; CHECK-LABEL: define void @zext_v4i8_all_lanes_used(
 ; CHECK-SAME: <4 x i8> [[SRC:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP0:%.*]] = freeze <4 x i8> [[SRC]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <4 x i8> [[TMP0]] to i32
+; CHECK-NEXT:    [[TMP2:%.*]] = lshr i32 [[TMP1]], 24
+; CHECK-NEXT:    [[TMP3:%.*]] = and i32 [[TMP2]], 255
+; CHECK-NEXT:    [[TMP4:%.*]] = lshr i32 [[TMP1]], 16
+; CHECK-NEXT:    [[TMP5:%.*]] = and i32 [[TMP4]], 255
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP1]], 8
+; CHECK-NEXT:    [[TMP7:%.*]] = and i32 [[TMP6]], 255
+; CHECK-NEXT:    [[TMP8:%.*]] = lshr i32 [[TMP1]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = and i32 [[TMP8]], 255
 ; CHECK-NEXT:    [[EXT9:%.*]] = zext nneg <4 x i8> [[SRC]] to <4 x i32>
 ; CHECK-NEXT:    [[EXT_0:%.*]] = extractelement <4 x i32> [[EXT9]], i64 0
 ; CHECK-NEXT:    [[EXT_1:%.*]] = extractelement <4 x i32> [[EXT9]], i64 1
 ; CHECK-NEXT:    [[EXT_2:%.*]] = extractelement <4 x i32> [[EXT9]], i64 2
 ; CHECK-NEXT:    [[EXT_3:%.*]] = extractelement <4 x i32> [[EXT9]], i64 3
-; CHECK-NEXT:    call void @use.i32(i32 [[EXT_0]])
-; CHECK-NEXT:    call void @use.i32(i32 [[EXT_1]])
-; CHECK-NEXT:    call void @use.i32(i32 [[EXT_2]])
-; CHECK-NEXT:    call void @use.i32(i32 [[EXT_3]])
+; CHECK-NEXT:    call void @use.i32(i32 [[TMP9]])
+; CHECK-NEXT:    call void @use.i32(i32 [[TMP7]])
+; CHECK-NEXT:    call void @use.i32(i32 [[TMP5]])
+; CHECK-NEXT:    call void @use.i32(i32 [[TMP3]])
 ; CHECK-NEXT:    ret void
----------------
fhahn wrote:

The issue currently is that the extends cannot be removed currently when we run the transform, as they come after the extend, and removing them will invalidate the iterator over the BB unfortunately.

https://github.com/llvm/llvm-project/pull/142976