From eaee28b5caec0972aa35479fecbe4196528c8fb8 Mon Sep 17 00:00:00 2001
From: Tim Northover <tnorthover@apple.com>
Date: Mon, 19 Sep 2016 09:11:09 +0000
Subject: [PATCH] ARM: check alignment before transforming ldr -> ldm (or
 similar).

ldm and stm instructions always require 4-byte alignment on the pointer, but we
weren't checking this before trying to reduce code-size by replacing a
post-indexed load/store with them. Unfortunately, we were also dropping this
incormation in DAG ISel too, but that's easy enough to fix.

llvm-svn: 281893
---
 llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp     | 32 +++++++++++++++++++++--------
 llvm/lib/Target/ARM/Thumb2SizeReduction.cpp |  4 ++++
 llvm/test/CodeGen/ARM/thumb2-size-opt.ll    | 13 ++++++++++++
 3 files changed, 41 insertions(+), 8 deletions(-)

diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index cb01b63..977a082 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -193,6 +193,8 @@ public:
 #include "ARMGenDAGISel.inc"
 
 private:
+  void transferMemOperands(SDNode *Src, SDNode *Dst);
+
   /// Indexed (pre/post inc/dec) load matching code for ARM.
   bool tryARMIndexedLoad(SDNode *N);
   bool tryT1IndexedLoad(SDNode *N);
@@ -1471,6 +1473,12 @@ static inline SDValue getAL(SelectionDAG *CurDAG, const SDLoc &dl) {
   return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, dl, MVT::i32);
 }
 
+void ARMDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) {
+  MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
+  MemOp[0] = cast<MemSDNode>(N)->getMemOperand();
+  cast<MachineSDNode>(Result)->setMemRefs(MemOp, MemOp + 1);
+}
+
 bool ARMDAGToDAGISel::tryARMIndexedLoad(SDNode *N) {
   LoadSDNode *LD = cast<LoadSDNode>(N);
   ISD::MemIndexedMode AM = LD->getAddressingMode();
@@ -1529,16 +1537,20 @@ bool ARMDAGToDAGISel::tryARMIndexedLoad(SDNode *N) {
       SDValue Base = LD->getBasePtr();
       SDValue Ops[]= { Base, AMOpc, getAL(CurDAG, SDLoc(N)),
                        CurDAG->getRegister(0, MVT::i32), Chain };
-      ReplaceNode(N, CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32,
-                                            MVT::i32, MVT::Other, Ops));
+      SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
+                                           MVT::Other, Ops);
+      transferMemOperands(N, New);
+      ReplaceNode(N, New);
       return true;
     } else {
       SDValue Chain = LD->getChain();
       SDValue Base = LD->getBasePtr();
       SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG, SDLoc(N)),
                        CurDAG->getRegister(0, MVT::i32), Chain };
-      ReplaceNode(N, CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32,
-                                            MVT::i32, MVT::Other, Ops));
+      SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
+                                           MVT::Other, Ops);
+      transferMemOperands(N, New);
+      ReplaceNode(N, New);
       return true;
     }
   }
@@ -1566,8 +1578,10 @@ bool ARMDAGToDAGISel::tryT1IndexedLoad(SDNode *N) {
   SDValue Base = LD->getBasePtr();
   SDValue Ops[]= { Base, getAL(CurDAG, SDLoc(N)),
                    CurDAG->getRegister(0, MVT::i32), Chain };
-  ReplaceNode(N, CurDAG->getMachineNode(ARM::tLDR_postidx, SDLoc(N), MVT::i32, MVT::i32,
-                                        MVT::Other, Ops));
+  SDNode *New = CurDAG->getMachineNode(ARM::tLDR_postidx, SDLoc(N), MVT::i32,
+                                       MVT::i32, MVT::Other, Ops);
+  transferMemOperands(N, New);
+  ReplaceNode(N, New);
   return true;
 }
 
@@ -1612,8 +1626,10 @@ bool ARMDAGToDAGISel::tryT2IndexedLoad(SDNode *N) {
     SDValue Base = LD->getBasePtr();
     SDValue Ops[]= { Base, Offset, getAL(CurDAG, SDLoc(N)),
                      CurDAG->getRegister(0, MVT::i32), Chain };
-    ReplaceNode(N, CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
-                                          MVT::Other, Ops));
+    SDNode *New = CurDAG->getMachineNode(Opcode, SDLoc(N), MVT::i32, MVT::i32,
+                                         MVT::Other, Ops);
+    transferMemOperands(N, New);
+    ReplaceNode(N, New);
     return true;
   }
 
diff --git a/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp b/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp
index 5a60893..07158f9 100644
--- a/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp
+++ b/llvm/lib/Target/ARM/Thumb2SizeReduction.cpp
@@ -430,6 +430,10 @@ Thumb2SizeReduce::ReduceLoadStore(MachineBasicBlock &MBB, MachineInstr *MI,
     if (!MBB.getParent()->getFunction()->optForMinSize())
       return false;
 
+    if (!MI->hasOneMemOperand() ||
+        (*MI->memoperands_begin())->getAlignment() < 4)
+      return false;
+
     // We're creating a completely different type of load/store - LDM from LDR.
     // For this reason we can't reuse the logic at the end of this function; we
     // have to implement the MI building here.
diff --git a/llvm/test/CodeGen/ARM/thumb2-size-opt.ll b/llvm/test/CodeGen/ARM/thumb2-size-opt.ll
index aba0276..f8d6489 100644
--- a/llvm/test/CodeGen/ARM/thumb2-size-opt.ll
+++ b/llvm/test/CodeGen/ARM/thumb2-size-opt.ll
@@ -98,3 +98,16 @@ false:
   store i32* %next, i32** %addr2
   ret i32 %res
 }
+
+; ldm instructions fault on misaligned accesses so we mustn't convert
+; this post-indexed ldr into one.
+define i32* @misaligned_post(i32* %src, i32* %dest) minsize {
+; CHECK-LABEL: misaligned_post:
+; CHECK: ldr [[VAL:.*]], [r0], #4
+; CHECK: str [[VAL]], [r1]
+
+  %val = load i32, i32* %src, align 1
+  store i32 %val, i32* %dest
+  %next = getelementptr i32, i32* %src, i32 1
+  ret i32* %next
+}
-- 
2.7.4