From 4e1ce0cbafa9e2b4f7cb11249898a877deab48b4 Mon Sep 17 00:00:00 2001
From: Simon Pilgrim <llvm-dev@redking.me.uk>
Date: Sun, 10 Mar 2024 16:23:51 +0000
Subject: [PATCH] [X86] combineAndShuffleNot - ensure the type is legal before
 create X86ISD::ANDNP target nodes

Fixes #84660

(cherry picked from commit 862c7e0218f27b55a5b75ae59a4f73cd4610448d)
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 11 +++++++++--
 llvm/test/CodeGen/X86/combine-and.ll    | 19 +++++++++++++++++++
 2 files changed, 28 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index a071c5a3ca0326..9e64726fb6fff7 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -47878,6 +47878,7 @@ static SDValue combineAndShuffleNot(SDNode *N, SelectionDAG &DAG,
   SDValue X, Y;
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   if (SDValue Not = GetNot(N0)) {
     X = Not;
@@ -47891,9 +47892,11 @@ static SDValue combineAndShuffleNot(SDNode *N, SelectionDAG &DAG,
   X = DAG.getBitcast(VT, X);
   Y = DAG.getBitcast(VT, Y);
   SDLoc DL(N);
+
   // We do not split for SSE at all, but we need to split vectors for AVX1 and
   // AVX2.
-  if (!Subtarget.useAVX512Regs() && VT.is512BitVector()) {
+  if (!Subtarget.useAVX512Regs() && VT.is512BitVector() && 
+      TLI.isTypeLegal(VT.getHalfNumVectorElementsVT(*DAG.getContext()))) {
     SDValue LoX, HiX;
     std::tie(LoX, HiX) = splitVector(X, DAG, DL);
     SDValue LoY, HiY;
@@ -47903,7 +47906,11 @@ static SDValue combineAndShuffleNot(SDNode *N, SelectionDAG &DAG,
     SDValue HiV = DAG.getNode(X86ISD::ANDNP, DL, SplitVT, {HiX, HiY});
     return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, {LoV, HiV});
   }
-  return DAG.getNode(X86ISD::ANDNP, DL, VT, {X, Y});
+
+  if (TLI.isTypeLegal(VT))
+    return DAG.getNode(X86ISD::ANDNP, DL, VT, {X, Y});
+
+  return SDValue();
 }
 
 // Try to widen AND, OR and XOR nodes to VT in order to remove casts around
diff --git a/llvm/test/CodeGen/X86/combine-and.ll b/llvm/test/CodeGen/X86/combine-and.ll
index d223b75419ac47..294fcd6a9563eb 100644
--- a/llvm/test/CodeGen/X86/combine-and.ll
+++ b/llvm/test/CodeGen/X86/combine-and.ll
@@ -1171,6 +1171,25 @@ define <4 x i32> @neg_scalar_broadcast_two_uses(i32 %a0, <4 x i32> %a1, ptr %a2)
   ret <4 x i32> %4
 }
 
+; PR84660 - check for illegal types
+define <2 x i128> @neg_scalar_broadcast_illegaltype(i128 %arg) {
+; CHECK-LABEL: neg_scalar_broadcast_illegaltype:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movq %rdi, %rax
+; CHECK-NEXT:    notl %esi
+; CHECK-NEXT:    andl $1, %esi
+; CHECK-NEXT:    movq %rsi, 16(%rdi)
+; CHECK-NEXT:    movq %rsi, (%rdi)
+; CHECK-NEXT:    movq $0, 24(%rdi)
+; CHECK-NEXT:    movq $0, 8(%rdi)
+; CHECK-NEXT:    retq
+  %i = xor i128 %arg, 1
+  %i1 = insertelement <2 x i128> zeroinitializer, i128 %i, i64 0
+  %i2 = shufflevector <2 x i128> %i1, <2 x i128> zeroinitializer, <2 x i32> zeroinitializer
+  %i3 = and <2 x i128> <i128 1, i128 1>, %i2
+  ret <2 x i128> %i3
+}
+
 define <2 x i64> @andnp_xx(<2 x i64> %v0) nounwind {
 ; SSE-LABEL: andnp_xx:
 ; SSE:       # %bb.0: