[x86] Combine x86mmx/i64 to v2i64 conversion to use scalar_to_vector

Handle the poor codegen for i64/x86xmm->v2i64 (%mm -> %xmm) moves. Instead of
using stack store/load pair to do the job, use scalar_to_vector directly, which
in the MMX case can use movq2dq. This was the current behavior prior to
improvements for vector legalization of extloads in r213897.

This commit fixes the regression and as a side-effect also remove some
unnecessary shuffles.

In the new attached testcase, we go from:

pshufw  $-18, (%rdi), %mm0
movq    %mm0, -8(%rsp)
movq    -8(%rsp), %xmm0
pshufd  $-44, %xmm0, %xmm0
movd    %xmm0, %eax
...

To:

pshufw  $-18, (%rdi), %mm0
movq2dq %mm0, %xmm0
movd    %xmm0, %eax
...

Differential Revision: http://reviews.llvm.org/D7126
rdar://problem/19413324

llvm-svn: 226953

1 files changed, 29 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e92a099753b..37026ce0f12 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -24757,6 +24757,8 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
   LoadSDNode *Ld = cast<LoadSDNode>(N);
   EVT RegVT = Ld->getValueType(0);
   EVT MemVT = Ld->getMemoryVT();
+  SDValue Ptr   = Ld->getBasePtr();
+  SDValue Chain = Ld->getChain();
   SDLoc dl(Ld);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
@@ -24795,6 +24797,33 @@ static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
     return DCI.CombineTo(N, NewVec, TF, true);
   }
 
+  // Conversion from x86mmx/i64 to v2i64 types is often done via stack
+  // store/load. Under certain conditions we can bypass the memory access and
+  // combine this load to use a scalar_to_vector instead. This leads to
+  // a reduction in the stack use, redundant emission of shuffles and create
+  // isel matching candidates for movq2dq instructions.
+  if (RegVT == MVT::v2i64 && Subtarget->hasSSE2() && Ext == ISD::EXTLOAD &&
+      !Ld->isVolatile() && ISD::isNON_TRUNCStore(Chain.getNode())) {
+
+    // If this load is directly stored, get the original source value.
+    StoreSDNode *PrevST = cast<StoreSDNode>(Chain);
+    EVT SrcTy = PrevST->getValue().getValueType();
+    if (PrevST->getBasePtr() != Ptr ||
+        !(SrcTy == MVT::i64 || SrcTy == MVT::x86mmx))
+      return SDValue();
+    SDValue SrcVal = Chain.getOperand(1);
+
+    // On 32bit systems, we can't save 64bit integers, use f64 instead.
+    bool Usef64 = TLI.isTypeLegal(MVT::f64) && !Subtarget->is64Bit();
+    if (Usef64)
+      SrcVal = DAG.getNode(ISD::BITCAST, dl, MVT::f64, SrcVal);
+    SrcVal = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, Usef64 ? MVT::v2f64 : RegVT,
+                              SrcVal);
+
+    return DCI.CombineTo(N, Usef64 ?
+        DAG.getNode(ISD::BITCAST, dl, RegVT, SrcVal) : SrcVal, Chain);
+  }
+
   return SDValue();
 }