[llvm] [LV][NFC] Add branch weight test showing incorrect behaviour (PR #144682)

[David Sherwood via llvm-commits]

================
@@ -1,53 +1,81 @@
-; RUN: opt < %s -S -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=4  -enable-epilogue-vectorization -epilogue-vectorization-force-VF=4 | FileCheck %s
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --filter "br" --filter "^.*:" --version 5
+; RUN: opt < %s -S -passes=loop-vectorize -force-vector-interleave=1 -force-vector-width=4  -enable-epilogue-vectorization \
+; RUN:   -epilogue-vectorization-force-VF=4 | FileCheck %s --check-prefix=MAINVF4IC1_EPI4
+; RUN: opt < %s -S -passes=loop-vectorize -force-vector-interleave=2 -force-vector-width=4  -enable-epilogue-vectorization \
+; RUN:   -epilogue-vectorization-force-VF=4 | FileCheck %s --check-prefix=MAINVF4IC2_EPI4
 
-; CHECK-LABEL: @f0(
-;
-; CHECK: entry:
-; CHECK:   br i1 %cmp.entry, label %iter.check, label %exit, !prof [[PROF_F0_ENTRY:![0-9]+]]
-;
-; CHECK: iter.check:
-; CHECK:   br i1 %min.iters.check, label %vec.epilog.scalar.ph, label %vector.scevcheck, !prof [[PROF_F0_UNLIKELY:![0-9]+]]
-;
-; CHECK: vector.scevcheck:
-; CHECK:   br i1 %4, label %vec.epilog.scalar.ph, label %vector.main.loop.iter.check, !prof [[PROF_F0_UNLIKELY]]
-;
-; CHECK: vector.main.loop.iter.check:
-; CHECK:   br i1 %min.iters.check1, label %vec.epilog.ph, label %vector.ph, !prof [[PROF_F0_UNLIKELY]]
-;
-; CHECK: vector.ph:
-; CHECK:   br label %vector.body
-;
-; CHECK: vector.body:
-; CHECK:   br i1 {{.+}}, label %middle.block, label %vector.body, !prof [[PROF_F0_VECTOR_BODY:![0-9]+]]
-;
-; CHECK: middle.block:
-; CHECK:   br i1 %cmp.n, label %exit.loopexit, label %vec.epilog.iter.check, !prof [[PROF_F0_MIDDLE_BLOCKS:![0-9]+]]
-;
-; CHECK: vec.epilog.iter.check:
-; CHECK:   br i1 %min.epilog.iters.check, label %vec.epilog.scalar.ph, label %vec.epilog.ph, !prof [[PROF_F0_VEC_EPILOGUE_SKIP:![0-9]+]]
-;
-; CHECK: vec.epilog.ph:
-; CHECK:   br label %vec.epilog.vector.body
-;
-; CHECK: vec.epilog.vector.body:
-; CHECK:   br i1 {{.+}}, label %vec.epilog.middle.block, label %vec.epilog.vector.body, !prof [[PROF_F0_VEC_EPILOG_VECTOR_BODY:![0-9]+]]
-;
-; CHECK: vec.epilog.middle.block:
-; CHECK:   br i1 %cmp.n{{.+}}, label %exit.loopexit, label %vec.epilog.scalar.ph, !prof [[PROF_F0_MIDDLE_BLOCKS:![0-9]+]]
-;
-; CHECK: vec.epilog.scalar.ph:
-; CHECK:   br label %loop
-;
-; CHECK: loop:
-; CHECK:   br i1 %cmp.loop, label %loop, label %exit.loopexit, !prof [[PROF_F0_LOOP:![0-9]+]]
+; FIXME: For MAINVF4IC2_EPI4 the branch weights in the terminator of
+; the VEC_EPILOG_ITER_CHECK block should be [4,4] since we process 8
+; scalar iterations in the main loop, leaving the remaining count to
+; be in the range [0,7]. That gives a 4:4 chance of skipping the
+; vector epilogue. I believe the problem lies in
+; EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck
+; where the main loop VF is set to the same value as the epilogue VF.
+define void @f0(i8 %n, i32 %len, ptr %p) !prof !0 {
+; MAINVF4IC1_EPI4-LABEL: define void @f0(
+; MAINVF4IC1_EPI4-SAME: i8 [[N:%.*]], i32 [[LEN:%.*]], ptr [[P:%.*]]) !prof [[PROF0:![0-9]+]] {
+; MAINVF4IC1_EPI4:  [[ENTRY:.*:]]
+; MAINVF4IC1_EPI4:    br i1 [[CMP_ENTRY:%.*]], label %[[ITER_CHECK:.*]], label %[[EXIT:.*]], !prof [[PROF1:![0-9]+]]
+; MAINVF4IC1_EPI4:  [[ITER_CHECK]]:
+; MAINVF4IC1_EPI4:    br i1 [[MIN_ITERS_CHECK:%.*]], label %[[VEC_EPILOG_SCALAR_PH:.*]], label %[[VECTOR_SCEVCHECK:.*]], !prof [[PROF2:![0-9]+]]
+; MAINVF4IC1_EPI4:  [[VECTOR_SCEVCHECK]]:
+; MAINVF4IC1_EPI4:    br i1 [[TMP4:%.*]], label %[[VEC_EPILOG_SCALAR_PH]], label %[[VECTOR_MAIN_LOOP_ITER_CHECK:.*]], !prof [[PROF2]]
+; MAINVF4IC1_EPI4:  [[VECTOR_MAIN_LOOP_ITER_CHECK]]:
+; MAINVF4IC1_EPI4:    br i1 [[MIN_ITERS_CHECK1:%.*]], label %[[VEC_EPILOG_PH:.*]], label %[[VECTOR_PH:.*]], !prof [[PROF2]]
+; MAINVF4IC1_EPI4:  [[VECTOR_PH]]:
+; MAINVF4IC1_EPI4:    br label %[[VECTOR_BODY:.*]]
+; MAINVF4IC1_EPI4:  [[VECTOR_BODY]]:
+; MAINVF4IC1_EPI4:    br i1 [[TMP8:%.*]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !prof [[PROF3:![0-9]+]], !llvm.loop [[LOOP4:![0-9]+]]
+; MAINVF4IC1_EPI4:  [[MIDDLE_BLOCK]]:
+; MAINVF4IC1_EPI4:    br i1 [[CMP_N:%.*]], label %[[EXIT_LOOPEXIT:.*]], label %[[VEC_EPILOG_ITER_CHECK:.*]], !prof [[PROF7:![0-9]+]]
+; MAINVF4IC1_EPI4:  [[VEC_EPILOG_ITER_CHECK]]:
+; MAINVF4IC1_EPI4:    br i1 [[MIN_EPILOG_ITERS_CHECK:%.*]], label %[[VEC_EPILOG_SCALAR_PH]], label %[[VEC_EPILOG_PH]], !prof [[PROF8:![0-9]+]]
+; MAINVF4IC1_EPI4:  [[VEC_EPILOG_PH]]:
+; MAINVF4IC1_EPI4:    br label %[[VEC_EPILOG_VECTOR_BODY:.*]]
+; MAINVF4IC1_EPI4:  [[VEC_EPILOG_VECTOR_BODY]]:
+; MAINVF4IC1_EPI4:    br i1 [[TMP12:%.*]], label %[[VEC_EPILOG_MIDDLE_BLOCK:.*]], label %[[VEC_EPILOG_VECTOR_BODY]], !prof [[PROF9:![0-9]+]], !llvm.loop [[LOOP10:![0-9]+]]
----------------
david-arm wrote:

It's easy enough to include the exit condition by adding a filter for icmp.

https://github.com/llvm/llvm-project/pull/144682