From 61f81f9637cf81ee25d941b09c41a372d9387f3e Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Wed, 8 Nov 2017 22:26:39 +0000
Subject: [PATCH] [X86] Preserve memory refs when folding loads into divides.

This is similar to what we already do for multiplies. Without this we can't unfold and hoist an invariant load.

llvm-svn: 317732
---
 llvm/lib/Target/X86/X86ISelDAGToDAG.cpp       |  6 +++++-
 llvm/test/CodeGen/X86/hoist-invariant-load.ll | 30 ++++++++++++++++++++++++++-
 2 files changed, 34 insertions(+), 2 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
index 21fee9a..09954324 100644
--- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
+++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp
@@ -2885,11 +2885,15 @@ void X86DAGToDAGISel::Select(SDNode *Node) {
     if (foldedLoad) {
       SDValue Ops[] = { Tmp0, Tmp1, Tmp2, Tmp3, Tmp4, N1.getOperand(0),
                         InFlag };
-      SDNode *CNode =
+      MachineSDNode *CNode =
         CurDAG->getMachineNode(MOpc, dl, MVT::Other, MVT::Glue, Ops);
       InFlag = SDValue(CNode, 1);
       // Update the chain.
       ReplaceUses(N1.getValue(1), SDValue(CNode, 0));
+      // Record the mem-refs
+      MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+      MemOp[0] = cast<LoadSDNode>(N1)->getMemOperand();
+      CNode->setMemRefs(MemOp, MemOp + 1);
     } else {
       InFlag =
         SDValue(CurDAG->getMachineNode(Opc, dl, MVT::Glue, N1, InFlag), 0);
diff --git a/llvm/test/CodeGen/X86/hoist-invariant-load.ll b/llvm/test/CodeGen/X86/hoist-invariant-load.ll
index e7929c9..a8b4e0a 100644
--- a/llvm/test/CodeGen/X86/hoist-invariant-load.ll
+++ b/llvm/test/CodeGen/X86/hoist-invariant-load.ll
@@ -1,10 +1,12 @@
 ; REQUIRES: asserts
-; RUN: llc -mcpu=haswell < %s -stats -O2 2>&1 | grep "4 machinelicm.*hoisted"
+; RUN: llc -mcpu=haswell < %s -stats -O2 2>&1 | grep "7 machinelicm.*hoisted"
 ; For test:
 ; 2 invariant loads, 1 for OBJC_SELECTOR_REFERENCES_
 ; and 1 for objc_msgSend from the GOT
 ; For test_multi_def:
 ; 2 invariant load (full multiply, both loads should be hoisted.)
+; For test_div_def:
+; 2 invariant load (full divide, both loads should be hoisted.) 1 additional instruction for a zeroing edx that gets hoisted and then rematerialized.
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.7.2"
@@ -60,4 +62,30 @@ exit:
   ret void
 }
 
+define void @test_div_def(i32* dereferenceable(8) %x1,
+                          i32* dereferenceable(8) %x2,
+                          i32* %y, i32 %count) nounwind {
+entry:
+  br label %for.body
+
+for.check:
+  %inc = add nsw i32 %i, 1
+  %done = icmp sge i32 %inc, %count
+  br i1 %done, label %exit, label %for.body
+
+for.body:
+  %i = phi i32 [ 0, %entry ], [ %inc, %for.check ]
+  %x1_load = load i32, i32* %x1, align 8, !invariant.load !0
+  %x2_load = load i32, i32* %x2, align 8, !invariant.load !0
+  %x_quot = udiv i32 %x1_load, %x2_load
+  %y_elem = getelementptr inbounds i32, i32* %y, i32 %i
+  %y_load = load i32, i32* %y_elem, align 8
+  %y_plus = add i32 %x_quot, %y_load
+  store i32 %y_plus, i32* %y_elem, align 8
+  br label %for.check
+
+exit:
+  ret void
+}
+
 !0 = !{}
-- 
2.7.4