[llvm] [LoopVectorize] Add regression test for issue #182646 (PR #184356)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Aviral Garg (aviralgarg05)

<details>
<summary>Changes</summary>

## Summary
This PR adds a regression test for [#<!-- -->182646](https://github.com/llvm/llvm-project/issues/182646), where LoopVectorize could hit an assertion about a mismatch between the VPlan and legacy cost models.

The original report included a reduced C++ reproducer with `vectorize_predicate(enable)` and a symbolic induction step. On current `main`, I could not reproduce the crash anymore, but this pattern is now covered with an explicit test so we don’t regress.

## What changed
- Added `llvm/test/Transforms/LoopVectorize/pr182646-vplan-legacy-cost-model-mismatch.ll`
- The test exercises the reduced loop shape from the report and checks that running loop-vectorize does not assert.

No functional code changes are included in this PR.

## Why this is useful
Even though the crash appears fixed by other upstream changes, this keeps the issue closed for good by making the scenario part of the test suite.

## Validation
I validated this in an assertions-enabled build:

1. Confirmed the reported reduced case does not crash on current `main`.
2. Ran the new regression test with `llvm-lit`.
3. Re-ran verification multiple times:
   - `opt -passes=loop-vectorize -disable-output`
   - `opt -passes='default<O3>' -disable-output`

All runs completed successfully.


---
Full diff: https://github.com/llvm/llvm-project/pull/184356.diff


1 Files Affected:

- (added) llvm/test/Transforms/LoopVectorize/pr182646-vplan-legacy-cost-model-mismatch.ll (+45) 


``````````diff
diff --git a/llvm/test/Transforms/LoopVectorize/pr182646-vplan-legacy-cost-model-mismatch.ll b/llvm/test/Transforms/LoopVectorize/pr182646-vplan-legacy-cost-model-mismatch.ll
new file mode 100644
index 0000000000000..e0da9672096f5
--- /dev/null
+++ b/llvm/test/Transforms/LoopVectorize/pr182646-vplan-legacy-cost-model-mismatch.ll
@@ -0,0 +1,45 @@
+; RUN: opt < %s -passes=loop-vectorize -disable-output
+
+; Regression test for issue #182646:
+; https://github.com/llvm/llvm-project/issues/182646
+;
+; This reduced test case has a symbolic induction step and loop vectorization
+; hints. It previously hit the "VPlan cost model and legacy cost model
+; disagreed" assertion in computeBestVF().
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+ at a = external global [0 x [3 x [3 x i8]]], align 1
+ at b = external global [0 x [3 x i8]], align 1
+
+define void @test(i32 %d, i1 %e, ptr %f) {
+entry:
+  %assume.cmp = icmp eq i32 %d, -1430399074
+  call void @llvm.assume(i1 %assume.cmp)
+  %step = add nsw i32 %d, 1430399077
+  br label %for.body
+
+for.body:
+  %i = phi i32 [ 1, %entry ], [ %next, %for.body ]
+  %idx = sext i32 %i to i64
+  %a.gep = getelementptr inbounds [3 x i8], ptr @a, i64 0, i64 %idx
+  store i8 0, ptr %a.gep, align 1
+  %f.gep = getelementptr inbounds [3 x [3 x i8]], ptr %f, i64 %idx, i64 %idx
+  %f.val = load i8, ptr %f.gep, align 1
+  %f.bool = trunc i8 %f.val to i1
+  %stored.bool = or i1 %f.bool, %e
+  %stored.val = zext i1 %stored.bool to i8
+  store i8 %stored.val, ptr @b, align 1
+  %sum = add i32 %step, %i
+  %cmp = icmp slt i32 %sum, 10
+  %next = select i1 %cmp, i32 %sum, i32 1
+  br label %for.body, !llvm.loop !0
+}
+
+declare void @llvm.assume(i1)
+
+!0 = distinct !{!0, !1, !2, !3}
+!1 = !{!"llvm.loop.mustprogress"}
+!2 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
+!3 = !{!"llvm.loop.vectorize.enable", i1 true}

``````````

</details>


https://github.com/llvm/llvm-project/pull/184356