X86 memcpy lowering: use "rep movs" even when esi is used as base pointer

For functions where esi is used as base pointer, we would previously fall back
from lowering memcpy with "rep movs" because that clobbers esi.

With this patch, we just store esi in another physical register, and restore
it afterwards. This adds a little bit of register preassure, but the more
efficient memcpy should be worth it.

Differential Revision: http://llvm-reviews.chandlerc.com/D2968

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204174 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 29 insertions, 13 deletions

diff --git a/lib/Target/X86/X86SelectionDAGInfo.cpp b/lib/Target/X86/X86SelectionDAGInfo.cpp
index b9c620fddc..3b3ab2a83a 100644
--- a/lib/Target/X86/X86SelectionDAGInfo.cpp
+++ b/lib/Target/X86/X86SelectionDAGInfo.cpp
@@ -200,13 +200,11 @@ X86SelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
       SrcPtrInfo.getAddrSpace() >= 256)
     return SDValue();
 
-  // ESI might be used as a base pointer, in that case we can't simply overwrite
-  // the register.  Fall back to generic code.
+  // If ESI is used as a base pointer, we must preserve it when doing rep movs.
   const X86RegisterInfo *TRI =
       static_cast<const X86RegisterInfo *>(DAG.getTarget().getRegisterInfo());
-  if (TRI->hasBasePointer(DAG.getMachineFunction()) &&
-      TRI->getBaseRegister() == X86::ESI)
-    return SDValue();
+  bool PreserveESI = TRI->hasBasePointer(DAG.getMachineFunction()) &&
+                     TRI->getBaseRegister() == X86::ESI;
 
   MVT AVT;
   if (Align & 1)
@@ -225,27 +223,45 @@ X86SelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl,
   SDValue Count = DAG.getIntPtrConstant(CountVal);
   unsigned BytesLeft = SizeVal % UBytes;
 
-  SDValue InFlag(0, 0);
+
+  if (PreserveESI) {
+    // Save ESI to a physical register. (We cannot use a virtual register
+    // because if it is spilled we wouldn't be able to reload it.)
+    // We don't glue this because the register dependencies are explicit.
+    Chain = DAG.getCopyToReg(Chain, dl, X86::EDX,
+                             DAG.getRegister(X86::ESI, MVT::i32));
+  }
+
+  SDValue InGlue(0, 0);
   Chain  = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
                                                               X86::ECX,
-                            Count, InFlag);
-  InFlag = Chain.getValue(1);
+                            Count, InGlue);
+  InGlue = Chain.getValue(1);
   Chain  = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
                                                               X86::EDI,
-                            Dst, InFlag);
-  InFlag = Chain.getValue(1);
+                            Dst, InGlue);
+  InGlue = Chain.getValue(1);
   Chain  = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
                                                               X86::ESI,
-                            Src, InFlag);
-  InFlag = Chain.getValue(1);
+                            Src, InGlue);
+  InGlue = Chain.getValue(1);
 
   SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
-  SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
+  SDValue Ops[] = { Chain, DAG.getValueType(AVT), InGlue };
+  // FIXME: Make X86rep_movs explicitly use FCX, RDI, RSI instead of glue.
   SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
                                 array_lengthof(Ops));
 
+  if (PreserveESI) {
+    InGlue = RepMovs.getValue(1);
+    RepMovs = DAG.getCopyToReg(RepMovs, dl, X86::ESI,
+                               DAG.getRegister(X86::EDX, MVT::i32), InGlue);
+  }
+
   SmallVector<SDValue, 4> Results;
   Results.push_back(RepMovs);
+
+
   if (BytesLeft) {
     // Handle the last 1 - 7 bytes.
     unsigned Offset = SizeVal - BytesLeft;