[llvm] [LV] Also clamp MaxVF by trip count when maximizing vector bandwidth. (PR #149794)

[Florian Hahn via llvm-commits]

================
@@ -6,96 +6,41 @@ target triple = "aarch64-linux-gnu"
 define i64 @main_vector_loop_fixed_with_no_remaining_iterations(ptr %src, ptr noalias %dst, i32 %x) #0 {
 ; CHECK-LABEL: define i64 @main_vector_loop_fixed_with_no_remaining_iterations(
 ; CHECK-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]], i32 [[X:%.*]]) #[[ATTR0:[0-9]+]] {
-; CHECK-NEXT:  [[ITER_CHECK:.*]]:
-; CHECK-NEXT:    [[TMP2:%.*]] = call i64 @llvm.vscale.i64()
-; CHECK-NEXT:    [[TMP3:%.*]] = mul nuw i64 [[TMP2]], 2
-; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 16, [[TMP3]]
-; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[VEC_EPILOG_SCALAR_PH:.*]], label %[[VECTOR_MAIN_LOOP_ITER_CHECK:.*]]
-; CHECK:       [[VECTOR_MAIN_LOOP_ITER_CHECK]]:
-; CHECK-NEXT:    br i1 true, label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK-NEXT:  [[ENTRY:.*]]:
+; CHECK-NEXT:    br i1 false, label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
 ; CHECK:       [[VECTOR_PH]]:
-; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i32> poison, i32 [[X]], i64 0
-; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i32> [[BROADCAST_SPLATINSERT]], <16 x i32> poison, <16 x i32> zeroinitializer
-; CHECK-NEXT:    [[TMP0:%.*]] = call <16 x i32> @llvm.abs.v16i32(<16 x i32> [[BROADCAST_SPLAT]], i1 false)
-; CHECK-NEXT:    [[TMP1:%.*]] = call <16 x i32> @llvm.abs.v16i32(<16 x i32> [[BROADCAST_SPLAT]], i1 false)
+; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> poison, i32 [[X]], i64 0
----------------
fhahn wrote:

The loop has an interleave group with gaps, so the scalar epilogue is required and we can only execute up to 15 iterations in the vector loop (note that the main loop with VF =16 never executed ( `; CHECK-NEXT:    br i1 true, label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]]`) in the removed check lines).

I added some comments and renamed the test file to `interleave-groups-with-gaps.ll`, as none of the tests should have their epilogue vectorized, if there are no remaining iterations.

https://github.com/llvm/llvm-project/pull/149794