[AIEX] OuterLoopPipeliner Bug fix 2

llvm/test/CodeGen/AIE/aie2p/schedule/outer-loop-pipeliner/outer-loop-pipelining-shared-exit.ll

@@ -0,0 +1,63 @@
+; This file is licensed under the Apache License v2.0 with LLVM Exceptions.
+; See https://llvm.org/LICENSE.txt for license information.
+; SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+;
+; (c) Copyright 2026 Advanced Micro Devices, Inc. or its affiliates
+;
+; REQUIRES: asserts
+; RUN: not --crash llc -mtriple=aie2p -O2 -aie-enable-outer-loop-pipelining \
+; RUN:     -stop-after=aie-outer-loop-pipeliner \
+; RUN:     -o - %s 2>&1 | FileCheck %s
+
+; BASELINE (pre-fix): the outer loop's exit block is shared between the
+; preheader-bypass edge (entry -> exit) and the loop back-edge, so the live-out
+; is an LCSSA phi with an incoming value from entry and no dedicated
+; exit.loopexit block. swapInClonedLS/wireLastIterIntoCFG retarget the latch
+; incoming edge, but removeFromCFG later deletes the original latch and asserts
+; because the exit phi no longer lists it as a predecessor.
+;
+; This test pins the current crash; the follow-up fix replaces these CHECK
+; lines with the expected pipelined output.
+
+; CHECK: Invalid basic block argument to remove!
+
+define i32 @shared_exit(ptr noalias %a, ptr noalias %c, i32 %N, i32 %M) {
+entry:
+  %cmp.outer = icmp sgt i32 %N, 1
+  br i1 %cmp.outer, label %outer.header, label %exit
+
+outer.header:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %outer.latch ]
+  %a.ptr = phi ptr [ %a, %entry ], [ %a.ptr.next, %outer.latch ]
+  %c.ptr = phi ptr [ %c, %entry ], [ %c.ptr.next, %outer.latch ]
+  %v0 = load i32, ptr %a.ptr, align 4
+  call void @llvm.set.loop.iterations.i32(i32 %M)
+  br label %inner.header
+
+inner.header:
+  %acc = phi i32 [ 0, %outer.header ], [ %acc.next, %inner.header ]
+  %acc.next = add i32 %acc, %v0
+  %inner.cond = call i1 @llvm.loop.decrement.i32(i32 1)
+  br i1 %inner.cond, label %inner.header, label %outer.latch, !llvm.loop !1
+
+outer.latch:
+  store i32 %acc.next, ptr %c.ptr, align 4
+  %a.ptr.next = getelementptr inbounds i32, ptr %a.ptr, i32 1
+  %c.ptr.next = getelementptr inbounds i32, ptr %c.ptr, i32 1
+  %i.next = add i32 %i, 1
+  %outer.cond = icmp slt i32 %i.next, %N
+  br i1 %outer.cond, label %outer.header, label %exit, !llvm.loop !0
+
+exit:
+  ; Live-out shared with the entry-bypass edge: LCSSA phi, no exit.loopexit.
+  %acc.lcssa = phi i32 [ 0, %entry ], [ %acc.next, %outer.latch ]
+  ret i32 %acc.lcssa
+}
+
+declare void @llvm.set.loop.iterations.i32(i32)
+declare i1 @llvm.loop.decrement.i32(i32)
+
+!0 = distinct !{!0, !2, !3}
+!1 = distinct !{!1, !2}
+!2 = !{!"llvm.loop.mustprogress"}
+!3 = !{!"llvm.loop.itercount.range", i32 2}