From bc65b686618e51b9b39e57fc18a356dd6ac2ccee Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@gmail.com>
Date: Sat, 7 Mar 2020 11:33:10 -0800
Subject: [PATCH] [X86] Add a DAG combine to turn vbroadcast(vzload X) ->
 vbroadcast_load

Remove now unneeded isel patterns.
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 16 ++++++++++++++++
 llvm/lib/Target/X86/X86InstrAVX512.td   | 15 ---------------
 llvm/lib/Target/X86/X86InstrSSE.td      |  9 ---------
 3 files changed, 16 insertions(+), 24 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index da8b68c..e65a29d 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -35146,6 +35146,22 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
       return N; // Return N so it doesn't get rechecked!
     }
 
+    // vbroadcast(vzload X) -> vbroadcast_load X
+    if (Src.getOpcode() == X86ISD::VZEXT_LOAD && Src.hasOneUse()) {
+      MemSDNode *LN = cast<MemIntrinsicSDNode>(Src);
+      if (LN->getMemoryVT().getSizeInBits() == VT.getScalarSizeInBits()) {
+        SDVTList Tys = DAG.getVTList(VT, MVT::Other);
+        SDValue Ops[] = { LN->getChain(), LN->getBasePtr() };
+        SDValue BcastLd =
+            DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, DL, Tys, Ops,
+                                    LN->getMemoryVT(), LN->getMemOperand());
+        DCI.CombineTo(N.getNode(), BcastLd);
+        DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BcastLd.getValue(1));
+        DCI.recursivelyDeleteUnusedNodes(LN);
+        return N; // Return N so it doesn't get rechecked!
+      }
+    }
+
     return SDValue();
   }
   case X86ISD::BLENDI: {
diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 2ed3387..f5e06e4 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -1423,19 +1423,6 @@ multiclass avx512_subvec_broadcast_rm_dq<bits<8> opc, string OpcodeStr,
                            AVX5128IBase, EVEX;
 }
 
-let Predicates = [HasAVX512] in {
-  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
-  def : Pat<(v8i64 (X86VBroadcast (v2i64 (X86vzload64 addr:$src)))),
-            (VPBROADCASTQZrm addr:$src)>;
-}
-
-let Predicates = [HasVLX] in {
-  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
-  def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload64 addr:$src)))),
-            (VPBROADCASTQZ128rm addr:$src)>;
-  def : Pat<(v4i64 (X86VBroadcast (v2i64 (X86vzload64 addr:$src)))),
-            (VPBROADCASTQZ256rm addr:$src)>;
-}
 let Predicates = [HasVLX, HasBWI] in {
   // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
   // This means we'll encounter truncated i32 loads; match that here.
@@ -10873,8 +10860,6 @@ def : Pat<(v2f64 (X86VBroadcast f64:$src)),
           (VMOVDDUPZ128rr (v2f64 (COPY_TO_REGCLASS FR64X:$src, VR128X)))>;
 def : Pat<(v2f64 (X86VBroadcast (v2f64 (simple_load addr:$src)))),
           (VMOVDDUPZ128rm addr:$src)>;
-def : Pat<(v2f64 (X86VBroadcast (v2f64 (X86vzload64 addr:$src)))),
-          (VMOVDDUPZ128rm addr:$src)>;
 
 def : Pat<(vselect (v2i1 VK2WM:$mask), (v2f64 (X86VBroadcast f64:$src)),
                    (v2f64 VR128X:$src0)),
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 73bba72..9659145 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -7523,13 +7523,6 @@ defm VPBROADCASTD  : avx2_broadcast<0x58, "vpbroadcastd", i32mem, X86VBroadcastl
 defm VPBROADCASTQ  : avx2_broadcast<0x59, "vpbroadcastq", i64mem, X86VBroadcastld64,
                                     v2i64, v4i64, NoVLX>;
 
-let Predicates = [HasAVX2, NoVLX] in {
-  // 32-bit targets will fail to load a i64 directly but can use ZEXT_LOAD.
-  def : Pat<(v2i64 (X86VBroadcast (v2i64 (X86vzload64 addr:$src)))),
-            (VPBROADCASTQrm addr:$src)>;
-  def : Pat<(v4i64 (X86VBroadcast (v2i64 (X86vzload64 addr:$src)))),
-            (VPBROADCASTQYrm addr:$src)>;
-}
 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in {
   // loadi16 is tricky to fold, because !isTypeDesirableForOp, justifiably.
   // This means we'll encounter truncated i32 loads; match that here.
@@ -7621,8 +7614,6 @@ let Predicates = [HasAVX, NoVLX] in {
             (VMOVDDUPrr VR128:$src)>;
   def : Pat<(v2f64 (X86VBroadcast (v2f64 (simple_load addr:$src)))),
             (VMOVDDUPrm addr:$src)>;
-  def : Pat<(v2f64 (X86VBroadcast (v2f64 (X86vzload64 addr:$src)))),
-            (VMOVDDUPrm addr:$src)>;
 }
 
 let Predicates = [HasAVX1Only] in {
-- 
2.7.4