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

[David Sherwood 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
----------------
david-arm wrote:

I'm not sure why this VF has been clamped to 8, given the trip count is 16 (at least it seems to be from looking at the scalar loop)? I'd expect the absolute best choice to be VF=16,IC=1 so that we delete the loop.

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