From b06d1aa3a1826bde799eb7a48375e3ecf69369b9 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Tue, 20 Nov 2018 09:03:58 +0000
Subject: [PATCH] [X86] Add custom type legalization for v8i8->v8i32 sign
 extend pre-SSE4.1

This helps with a future patch and makes us less reliant on DAG combine merging shuffles.

llvm-svn: 347295
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 33 +++++++++++++++++++++++++
 1 file changed, 33 insertions(+)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index f60f019e7759..57bf86e051f3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -946,6 +946,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
 
     if (ExperimentalVectorWideningLegalization) {
+      setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
       setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
 
       setOperationAction(ISD::TRUNCATE,    MVT::v2i8,  Custom);
@@ -26369,6 +26370,38 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
       return;
     }
 
+    if (!Subtarget.hasSSE41() && VT == MVT::v8i32 && InVT == MVT::v8i8) {
+      // Widen the input to 128 bits.
+      In = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v16i8, In,
+                       DAG.getUNDEF(InVT));
+      // Emit a shuffle that will become punpcklbw putting the input elements
+      // in the high half of the expansion.
+      In = DAG.getVectorShuffle(MVT::v16i8, dl, In, In,
+                                {-1, 0, -1, 1, -1, 2, -1, 3,
+                                 -1, 4, -1, 5, -1, 6, -1, 7});
+      In = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, In);
+
+      // Emit a shuffle that will become punpcklwd. Shift right to fill with
+      // sign bits.
+      SDValue Lo = DAG.getVectorShuffle(MVT::v8i16, dl, In, In,
+                                        {-1, 0, -1, 1, -1, 2, -1, 3});
+      Lo = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Lo);
+      Lo = DAG.getNode(ISD::SRA, dl, MVT::v4i32, Lo,
+                       DAG.getConstant(24, dl, MVT::v4i32));
+
+      // Emit a shuffle that will become punpckhwd. Shift right to fill with
+      // sign bits.
+      SDValue Hi = DAG.getVectorShuffle(MVT::v8i16, dl, In, In,
+                                        {-1, 4, -1, 5, -1, 6, -1, 7});
+      Hi = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Hi);
+      Hi = DAG.getNode(ISD::SRA, dl, MVT::v4i32, Hi,
+                       DAG.getConstant(24, dl, MVT::v4i32));
+
+      SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
+      Results.push_back(Res);
+      return;
+    }
+
     if ((VT == MVT::v16i32 || VT == MVT::v8i64) && InVT.is128BitVector()) {
       // Perform custom splitting instead of the two stage extend we would get
       // by default.