[llvm] [SLP] Fix crash of shuffle poison (PR #106857)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: tcwzxx (tcwzxx)

<details>
<summary>Changes</summary>

When the shuffle masks are `PoisonMaskElem`, there is not need to check the cost of `SK_ExtractSubvector`. It is free. Otherwise, it will cause the compiler to crash.

 Assertion `(Idx + EltsPerVector) <= alignTo(NumElts, EltsPerVector) && "SK_ExtractSubvector index out of range"' failed.

---
Full diff: https://github.com/llvm/llvm-project/pull/106857.diff


2 Files Affected:

- (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+2-1) 
- (added) llvm/test/Transforms/SLPVectorizer/crash_extractelement_from_null.ll (+34) 


``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 3d41c978281351..fa16f4632174de 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -8703,7 +8703,8 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
     auto CheckPerRegistersShuffle = [&](MutableArrayRef<int> Mask,
                                         SmallVectorImpl<unsigned> &Indices)
         -> std::optional<TTI::ShuffleKind> {
-      if (NumElts <= EltsPerVector)
+      if (NumElts <= EltsPerVector ||
+          all_of(Mask, [](int I) { return I == PoisonMaskElem; }))
         return std::nullopt;
       int OffsetReg0 =
           alignDown(std::accumulate(Mask.begin(), Mask.end(), INT_MAX,
diff --git a/llvm/test/Transforms/SLPVectorizer/crash_extractelement_from_null.ll b/llvm/test/Transforms/SLPVectorizer/crash_extractelement_from_null.ll
new file mode 100644
index 00000000000000..87d92706bf36a4
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/crash_extractelement_from_null.ll
@@ -0,0 +1,34 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+
+; RUN: opt -S --passes=slp-vectorizer < %s | FileCheck %s
+
+define void @test(i8 %0, i8 %1) {
+; CHECK-LABEL: define void @test(
+; CHECK-SAME: i8 [[TMP0:%.*]], i8 [[TMP1:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[L:%.*]] = load <4 x i8>, ptr getelementptr (i8, ptr null, i32 8), align 1
+; CHECK-NEXT:    [[LI15:%.*]] = extractelement <4 x i8> [[L]], i64 15
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ne i8 [[TMP0]], 0
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp ne i8 [[TMP1]], 0
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp ne i8 [[TMP0]], 0
+; CHECK-NEXT:    [[DOTI15:%.*]] = icmp ne i8 [[LI15]], 0
+; CHECK-NEXT:    [[I0244:%.*]] = insertelement <4 x i1> zeroinitializer, i1 [[TMP2]], i64 0
+; CHECK-NEXT:    [[I1245:%.*]] = insertelement <4 x i1> [[I0244]], i1 [[TMP3]], i64 1
+; CHECK-NEXT:    [[I2246:%.*]] = insertelement <4 x i1> [[I1245]], i1 [[TMP4]], i64 2
+; CHECK-NEXT:    [[TMP5:%.*]] = insertelement <4 x i1> [[I2246]], i1 [[DOTI15]], i64 3
+; CHECK-NEXT:    ret void
+;
+entry:
+  %l = load <4 x i8>, ptr getelementptr (i8, ptr null, i32 8), align 1
+  %li15 = extractelement <4 x i8> %l, i64 15
+  %2 = icmp ne i8 %0, 0
+  %3 = icmp ne i8 %1, 0
+  %4 = icmp ne i8 %0, 0
+  %.i15 = icmp ne i8 %li15, 0
+
+  %i0244 = insertelement <4 x i1> zeroinitializer, i1 %2, i64 0
+  %i1245 = insertelement <4 x i1> %i0244, i1 %3, i64 1
+  %i2246 = insertelement <4 x i1> %i1245, i1 %4, i64 2
+  %14 = insertelement <4 x i1> %i2246, i1 %.i15, i64 3
+  ret void
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/106857