[llvm] [LV] Mask off possibly aliasing vector lanes (PR #100579)

[Sander de Smalen via llvm-commits]

================
@@ -0,0 +1,350 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --filter-out-after "^middle.block:" --filter-out-after "^scalar.ph:" --version 4
+; RUN: opt -S -mtriple=aarch64-unknown-linux-gnu -mattr=+sve2 -passes=loop-vectorize -prefer-predicate-over-epilogue=predicate-dont-vectorize -force-vector-interleave=1 %s | FileCheck %s
+
+define dso_local void @alias_mask(ptr noalias %a, ptr %b, ptr %c, i64 %n) {
+; CHECK-LABEL: define dso_local void @alias_mask(
+; CHECK-SAME: ptr noalias [[A:%.*]], ptr [[B:%.*]], ptr [[C:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[B4:%.*]] = ptrtoint ptr [[B]] to i64
+; CHECK-NEXT:    [[C3:%.*]] = ptrtoint ptr [[C]] to i64
+; CHECK-NEXT:    [[B2:%.*]] = ptrtoint ptr [[B]] to i64
+; CHECK-NEXT:    [[C1:%.*]] = ptrtoint ptr [[C]] to i64
+; CHECK-NEXT:    [[CMP11:%.*]] = icmp sgt i64 [[N]], 0
+; CHECK-NEXT:    br i1 [[CMP11]], label [[FOR_BODY_PREHEADER:%.*]], label [[FOR_COND_CLEANUP:%.*]]
+; CHECK:       for.body.preheader:
+; CHECK-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_MEMCHECK:%.*]]
+; CHECK:       vector.memcheck:
+; CHECK-NEXT:    [[ALIAS_LANE_MASK:%.*]] = call <vscale x 16 x i1> @llvm.loop.dependence.war.mask.nxv16i1(ptr [[B]], ptr [[C]], i64 1)
+; CHECK-NEXT:    [[TMP0:%.*]] = zext <vscale x 16 x i1> [[ALIAS_LANE_MASK]] to <vscale x 16 x i8>
+; CHECK-NEXT:    [[TMP1:%.*]] = call i8 @llvm.vector.reduce.add.nxv16i8(<vscale x 16 x i8> [[TMP0]])
+; CHECK-NEXT:    [[TMP2:%.*]] = zext i8 [[TMP1]] to i64
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ugt i64 [[TMP2]], 0
+; CHECK-NEXT:    br i1 [[TMP3]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]
----------------
sdesmalen-arm wrote:

If the popcount is higher than 0, it branches to the scalar preheader, isn't it supposed to branch to the vector preheader instead?

https://github.com/llvm/llvm-project/pull/100579