1 files changed, 27 insertions, 29 deletions

diff --git a/contrib/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp b/contrib/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp
index 6cafbbb..6fded9c 100644
--- a/contrib/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp
+++ b/contrib/llvm/lib/Target/ARM/ARMSelectionDAGInfo.cpp
@@ -160,41 +160,39 @@ ARMSelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
   unsigned VTSize = 4;
   unsigned i = 0;
   // Emit a maximum of 4 loads in Thumb1 since we have fewer registers
-  const unsigned MAX_LOADS_IN_LDM = Subtarget.isThumb1Only() ? 4 : 6;
+  const unsigned MaxLoadsInLDM = Subtarget.isThumb1Only() ? 4 : 6;
   SDValue TFOps[6];
   SDValue Loads[6];
   uint64_t SrcOff = 0, DstOff = 0;
 
-  // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the
-  // same number of stores.  The loads and stores will get combined into
-  // ldm/stm later on.
-  while (EmittedNumMemOps < NumMemOps) {
-    for (i = 0;
-         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
-      Loads[i] = DAG.getLoad(VT, dl, Chain,
-                             DAG.getNode(ISD::ADD, dl, MVT::i32, Src,
-                                         DAG.getConstant(SrcOff, dl, MVT::i32)),
-                             SrcPtrInfo.getWithOffset(SrcOff), isVolatile,
-                             false, false, 0);
-      TFOps[i] = Loads[i].getValue(1);
-      SrcOff += VTSize;
-    }
-    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
-                        makeArrayRef(TFOps, i));
+  // FIXME: We should invent a VMEMCPY pseudo-instruction that lowers to
+  // VLDM/VSTM and make this code emit it when appropriate. This would reduce
+  // pressure on the general purpose registers. However this seems harder to map
+  // onto the register allocator's view of the world.
 
-    for (i = 0;
-         i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) {
-      TFOps[i] = DAG.getStore(Chain, dl, Loads[i],
-                              DAG.getNode(ISD::ADD, dl, MVT::i32, Dst,
-                                          DAG.getConstant(DstOff, dl, MVT::i32)),
-                              DstPtrInfo.getWithOffset(DstOff),
-                              isVolatile, false, 0);
-      DstOff += VTSize;
-    }
-    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
-                        makeArrayRef(TFOps, i));
+  // The number of MEMCPY pseudo-instructions to emit. We use up to
+  // MaxLoadsInLDM registers per mcopy, which will get lowered into ldm/stm
+  // later on. This is a lower bound on the number of MEMCPY operations we must
+  // emit.
+  unsigned NumMEMCPYs = (NumMemOps + MaxLoadsInLDM - 1) / MaxLoadsInLDM;
+
+  SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other, MVT::Glue);
+
+  for (unsigned I = 0; I != NumMEMCPYs; ++I) {
+    // Evenly distribute registers among MEMCPY operations to reduce register
+    // pressure.
+    unsigned NextEmittedNumMemOps = NumMemOps * (I + 1) / NumMEMCPYs;
+    unsigned NumRegs = NextEmittedNumMemOps - EmittedNumMemOps;
+
+    Dst = DAG.getNode(ARMISD::MEMCPY, dl, VTs, Chain, Dst, Src,
+                      DAG.getConstant(NumRegs, dl, MVT::i32));
+    Src = Dst.getValue(1);
+    Chain = Dst.getValue(2);
+
+    DstPtrInfo = DstPtrInfo.getWithOffset(NumRegs * VTSize);
+    SrcPtrInfo = SrcPtrInfo.getWithOffset(NumRegs * VTSize);
 
-    EmittedNumMemOps += i;
+    EmittedNumMemOps = NextEmittedNumMemOps;
   }
 
   if (BytesLeft == 0)