[llvm] [LV] Fix for miscompile with disjoint or (PR #81922)

[Florian Hahn via llvm-commits]

================
@@ -0,0 +1,102 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S -passes=loop-vectorize < %s | FileCheck %s
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at global = external global ptr addrspace(1), align 8
+
+; PR 81872 explains the issue.
+
+; If we vectorize, we have a miscompile where array IV and thereby value stored in (arr[99],
+; arr[98]) is calculated incorrectly since disjoint or was only disjoint because
+; of dominating conditions. Dropping the disjoint to avoid poison still changes
+; the behaviour since now the or is not longer equivalent to the add.
+; Function Attrs: uwtable
+define void @test(ptr addrspace(1) noundef align 8 dereferenceable_or_null(16) %arg1) #0 {
+; CHECK-LABEL: define void @test(
+; CHECK-SAME: ptr addrspace(1) noundef align 8 dereferenceable_or_null(16) [[ARG1:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  bb5:
+; CHECK-NEXT:    [[GETELEMENTPTR:%.*]] = getelementptr inbounds i8, ptr addrspace(1) [[ARG1]], i64 16
+; CHECK-NEXT:    br label [[BB8:%.*]]
+; CHECK:       bb6:
+; CHECK-NEXT:    [[ADD:%.*]] = add nsw i32 [[PHI9:%.*]], 1
+; CHECK-NEXT:    [[ICMP7:%.*]] = icmp sgt i32 [[PHI9]], -2
+; CHECK-NEXT:    br i1 [[ICMP7]], label [[BB10:%.*]], label [[BB8]]
+; CHECK:       bb8:
+; CHECK-NEXT:    [[PHI9]] = phi i32 [ -60516, [[BB5:%.*]] ], [ [[ADD]], [[BB6:%.*]] ]
+; CHECK-NEXT:    br label [[BB15:%.*]]
+; CHECK:       bb10:
+; CHECK-NEXT:    [[GETELEMENTPTR11:%.*]] = getelementptr inbounds i8, ptr addrspace(1) [[ARG1]], i64 96
+; CHECK-NEXT:    [[LOAD12:%.*]] = load i64, ptr addrspace(1) [[GETELEMENTPTR11]], align 8, !noundef [[META0:![0-9]+]]
+; CHECK-NEXT:    [[LOAD13:%.*]] = load ptr addrspace(1), ptr @global, align 8, !invariant.load [[META0]], !nonnull [[META0]], !dereferenceable_or_null !1, !align [[META2:![0-9]+]]
+; CHECK-NEXT:    [[GETELEMENTPTR14:%.*]] = getelementptr inbounds i8, ptr addrspace(1) [[LOAD13]], i64 848
+; CHECK-NEXT:    store i64 [[LOAD12]], ptr addrspace(1) [[GETELEMENTPTR14]], align 8
+; CHECK-NEXT:    ret void
+; CHECK:       bb15:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 99, [[BB8]] ], [ [[IV_NEXT:%.*]], [[BB20:%.*]] ]
+; CHECK-NEXT:    [[AND:%.*]] = and i64 [[IV]], 1
+; CHECK-NEXT:    [[ICMP17:%.*]] = icmp eq i64 [[AND]], 0
+; CHECK-NEXT:    br i1 [[ICMP17]], label [[BB18:%.*]], label [[BB20]], !prof [[PROF3:![0-9]+]]
+; CHECK:       bb18:
+; CHECK-NEXT:    [[OR:%.*]] = or disjoint i64 [[IV]], 1
+; CHECK-NEXT:    [[GETELEMENTPTR19:%.*]] = getelementptr inbounds i64, ptr addrspace(1) [[GETELEMENTPTR]], i64 [[OR]]
+; CHECK-NEXT:    store i64 1, ptr addrspace(1) [[GETELEMENTPTR19]], align 8
+; CHECK-NEXT:    br label [[BB20]]
+; CHECK:       bb20:
+; CHECK-NEXT:    [[IV_NEXT]] = add nsw i64 [[IV]], -1
+; CHECK-NEXT:    [[ICMP22:%.*]] = icmp eq i64 [[IV_NEXT]], 90
+; CHECK-NEXT:    br i1 [[ICMP22]], label [[BB6]], label [[BB15]], !prof [[PROF4:![0-9]+]]
+;
+bb5:
+  %getelementptr = getelementptr inbounds i8, ptr addrspace(1) %arg1, i64 16
+  br label %bb8
+
+bb6:                                              ; preds = %bb20
+  %add = add nsw i32 %phi9, 1
----------------
fhahn wrote:

Is this bb6, bb8 & bb10 needed to show the regression?

https://github.com/llvm/llvm-project/pull/81922