[llvm] [LV] Don't vectorize epilogue with scalable VF if no iterations remain. (PR #149789)

[David Sherwood via llvm-commits]

================
@@ -0,0 +1,399 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals none --version 5
+; RUN: opt -passes=loop-vectorize -S %s | FileCheck %s
+
+target triple = "aarch64-linux-gnu"
+
+; Original loop has trip count 16, but contains interleave groups with gaps, so
+; the last iteration must execute in the scalar loop. Thus the vector loop can
+; only execute up to 15 iterations.
+define i64 @vector_loop_with_remaining_iterations(ptr %src, ptr noalias %dst, i32 %x) #0 {
+; CHECK-LABEL: define i64 @vector_loop_with_remaining_iterations(
+; CHECK-SAME: ptr [[SRC:%.*]], ptr noalias [[DST:%.*]], i32 [[X:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  [[ITER_CHECK:.*]]:
+; CHECK-NEXT:    [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
+; CHECK-NEXT:    [[TMP1:%.*]] = mul nuw i64 [[TMP0]], 2
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ule i64 17, [[TMP1]]
+; 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 false, label %[[VEC_EPILOG_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:    [[TMP2:%.*]] = call <16 x i32> @llvm.abs.v16i32(<16 x i32> [[BROADCAST_SPLAT]], i1 false)
+; CHECK-NEXT:    [[TMP3:%.*]] = call <16 x i32> @llvm.abs.v16i32(<16 x i32> [[BROADCAST_SPLAT]], i1 false)
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <16 x i64> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP11:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr { [4 x i8] }, ptr [[SRC]], i64 [[INDEX]], i32 0, i64 3
+; CHECK-NEXT:    [[WIDE_VEC:%.*]] = load <64 x i8>, ptr [[TMP4]], align 1
+; CHECK-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <64 x i8> [[WIDE_VEC]], <64 x i8> poison, <16 x i32> <i32 0, i32 4, i32 8, i32 12, i32 16, i32 20, i32 24, i32 28, i32 32, i32 36, i32 40, i32 44, i32 48, i32 52, i32 56, i32 60>
+; CHECK-NEXT:    [[TMP5:%.*]] = zext <16 x i8> [[STRIDED_VEC]] to <16 x i32>
+; CHECK-NEXT:    [[TMP6:%.*]] = call <16 x i32> @llvm.umin.v16i32(<16 x i32> [[TMP2]], <16 x i32> [[TMP5]])
+; CHECK-NEXT:    [[TMP7:%.*]] = call <16 x i32> @llvm.umin.v16i32(<16 x i32> [[TMP3]], <16 x i32> [[TMP6]])
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 [[INDEX]]
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i8, ptr [[TMP8]], i32 0
+; CHECK-NEXT:    store <16 x i8> zeroinitializer, ptr [[TMP9]], align 1
+; CHECK-NEXT:    [[TMP10:%.*]] = zext <16 x i32> [[TMP7]] to <16 x i64>
+; CHECK-NEXT:    [[TMP11]] = or <16 x i64> [[VEC_PHI]], [[TMP10]]
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 16
+; CHECK-NEXT:    [[TMP12:%.*]] = icmp eq i64 [[INDEX_NEXT]], 16
----------------
david-arm wrote:

Perhaps not required for this patch, but we also shouldn't be generating a vector epilogue here because we know the last remaining iteration has to be done in the scalar epilogue. I suspect that `RemainingIterations` needs to check isOne if a scalar epilogue is required.

https://github.com/llvm/llvm-project/pull/149789