[DAGCombiner] avoid crash on out-of-bounds insert index (PR44139)

We already have this simplification at node-creation-time, but
the test from:
https://bugs.llvm.org/show_bug.cgi?id=44139
...shows that we can combine our way to an assert/crash too.

3 files changed, 33 insertions, 4 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index d56e737226e..793352c16d3 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -16756,6 +16756,11 @@ SDValue DAGCombiner::visitINSERT_VECTOR_ELT(SDNode *N) {
   EVT VT = InVec.getValueType();
   unsigned NumElts = VT.getVectorNumElements();
 
+  // Insert into out-of-bounds element is undefined.
+  if (auto *IndexC = dyn_cast<ConstantSDNode>(EltNo))
+    if (IndexC->getZExtValue() >= VT.getVectorNumElements())
+      return DAG.getUNDEF(VT);
+
   // Remove redundant insertions:
   // (insert_vector_elt x (extract_vector_elt x idx) idx) -> x
   if (InVal.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
diff --git a/llvm/test/CodeGen/X86/insertelement-var-index.ll b/llvm/test/CodeGen/X86/insertelement-var-index.ll
index c6ab2cd7fa9..a37fe63944d 100644
--- a/llvm/test/CodeGen/X86/insertelement-var-index.ll
+++ b/llvm/test/CodeGen/X86/insertelement-var-index.ll
@@ -623,3 +623,31 @@ define <4 x double> @load_f64_v4f64(double* %p, i32 %y) nounwind {
   ret <4 x double> %ins
 }
 
+; Don't die trying to insert to an invalid index.
+
+define i32 @PR44139(<16 x i64>* %p) {
+; ALL-LABEL: PR44139:
+; ALL:       # %bb.0:
+; ALL-NEXT:    movl (%rdi), %eax
+; ALL-NEXT:    leal 2147483647(%rax), %ecx
+; ALL-NEXT:    testl %eax, %eax
+; ALL-NEXT:    cmovnsl %eax, %ecx
+; ALL-NEXT:    andl $-2147483648, %ecx # imm = 0x80000000
+; ALL-NEXT:    addl %eax, %ecx
+; ALL-NEXT:    # kill: def $eax killed $eax killed $rax
+; ALL-NEXT:    xorl %edx, %edx
+; ALL-NEXT:    divl %ecx
+; ALL-NEXT:    retq
+  %L = load <16 x i64>, <16 x i64>* %p
+  %E1 = extractelement <16 x i64> %L, i64 0
+  %tempvector = insertelement <16 x i64> undef, i64 %E1, i32 0
+  %vector = shufflevector <16 x i64> %tempvector, <16 x i64> undef, <16 x i32> zeroinitializer
+  %C3 = icmp sgt i64 9223372036854775807, -9223372036854775808
+  %t0 = trunc <16 x i64> %vector to <16 x i32>
+  %I4 = insertelement <16 x i64> %vector, i64 %E1, i1 %C3
+  store <16 x i64> %I4, <16 x i64>* %p
+  %elt = extractelement <16 x i32> %t0, i32 0
+  %B = srem i32 %elt, -2147483648
+  %B9 = udiv i32 %elt, %B
+  ret i32 %B9
+}
diff --git a/llvm/test/CodeGen/X86/vec_extract.ll b/llvm/test/CodeGen/X86/vec_extract.ll
index 2d52bec473a..9b347c12194 100644
--- a/llvm/test/CodeGen/X86/vec_extract.ll
+++ b/llvm/test/CodeGen/X86/vec_extract.ll
@@ -110,15 +110,11 @@ define <4 x i32> @ossfuzz15662(<4 x i32*>* %in) {
 ; X32-LABEL: ossfuzz15662:
 ; X32:       # %bb.0:
 ; X32-NEXT:    xorps %xmm0, %xmm0
-; X32-NEXT:    movaps %xmm0, (%eax)
-; X32-NEXT:    xorps %xmm0, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: ossfuzz15662:
 ; X64:       # %bb.0:
 ; X64-NEXT:    xorps %xmm0, %xmm0
-; X64-NEXT:    movaps %xmm0, (%rax)
-; X64-NEXT:    xorps %xmm0, %xmm0
 ; X64-NEXT:    retq
    %C10 = icmp ule i1 false, false
    %C3 = icmp ule i1 true, undef