From 1ffd8e8114bc34295380b7fed815c7cf7644e84e Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@intel.com>
Date: Fri, 22 Mar 2019 20:46:56 +0000
Subject: [X86] Use movq for i64 atomic load on 32-bit targets when sse2 is
 enable

We used a lock cmpxchg8b to do i64 atomic loads. But if we have SSE2 we can do better and use a plain movq to do the load instead.

I tried to just use an f64 atomic load and add isel patterns to MOVSD(which the domain fixing pass can turn to MOVQ), but the atomic_load SDNode in TargetSelectionDAG.td requires the type to be integer.

So I've emitted VZEXT_LOAD instead which should be selected by isel to a MOVQ. Hopefully we don't need a specific atomic flavor of this. I kept the memory operand from the original AtomicSDNode. I wasn't sure if I might need to set the MOVolatile flag?

I've left some FIXMEs for improvements we can do without SSE2.

Differential Revision: https://reviews.llvm.org/D59679

llvm-svn: 356807
---
 llvm/lib/Target/X86/X86ISelLowering.cpp | 49 +++++++++++++++++++++++++++++----
 1 file changed, 44 insertions(+), 5 deletions(-)

(limited to 'llvm/lib')

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 5de8ea9c934..dc3292eed8c 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -485,6 +485,9 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
   }
 
+  if (!Subtarget.is64Bit())
+    setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
+
   if (Subtarget.hasCmpxchg16b()) {
     setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
   }
@@ -25494,11 +25497,22 @@ bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
 }
 
 // Note: this turns large loads into lock cmpxchg8b/16b.
-// FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
+// TODO: In 32-bit mode, use MOVLPS when SSE1 is available?
+// TODO: In 32-bit mode, use FILD/FISTP when X87 is available?
 TargetLowering::AtomicExpansionKind
 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
-  return needsCmpXchgNb(LI->getType()) ? AtomicExpansionKind::CmpXChg
-                                       : AtomicExpansionKind::None;
+  Type *MemType = LI->getType();
+
+  // If this a 64 bit atomic load on a 32-bit target and SSE2 is enabled, we
+  // can use movq to do the load.
+  bool NoImplicitFloatOps =
+      LI->getFunction()->hasFnAttribute(Attribute::NoImplicitFloat);
+  if (MemType->getPrimitiveSizeInBits() == 64 && !Subtarget.is64Bit() &&
+      !Subtarget.useSoftFloat() && !NoImplicitFloatOps && Subtarget.hasSSE2())
+    return AtomicExpansionKind::None;
+
+  return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
+                                 : AtomicExpansionKind::None;
 }
 
 TargetLowering::AtomicExpansionKind
@@ -27312,6 +27326,32 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
     Results.push_back(EFLAGS.getValue(1));
     return;
   }
+  case ISD::ATOMIC_LOAD: {
+    assert(N->getValueType(0) == MVT::i64 && "Unexpected VT!");
+    bool NoImplicitFloatOps =
+        DAG.getMachineFunction().getFunction().hasFnAttribute(
+            Attribute::NoImplicitFloat);
+    if (!Subtarget.useSoftFloat() && !NoImplicitFloatOps &&
+        Subtarget.hasSSE2()) {
+      auto *Node = cast<AtomicSDNode>(N);
+      // Use a VZEXT_LOAD which will be selected as MOVQ. Then extract the lower
+      // 64-bits.
+      SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
+      SDValue Ops[] = { Node->getChain(), Node->getBasePtr() };
+      SDValue Ld = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
+                                           MVT::i64, Node->getMemOperand());
+      SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Ld,
+                                DAG.getIntPtrConstant(0, dl));
+      Results.push_back(Res);
+      Results.push_back(Ld.getValue(1));
+      return;
+    }
+    // TODO: Use MOVLPS when SSE1 is available?
+    // TODO: Use FILD/FISTP when X87 is available?
+    // Delegate to generic TypeLegalization. Situations we can really handle
+    // should have already been dealt with by AtomicExpandPass.cpp.
+    break;
+  }
   case ISD::ATOMIC_SWAP:
   case ISD::ATOMIC_LOAD_ADD:
   case ISD::ATOMIC_LOAD_SUB:
@@ -27323,11 +27363,10 @@ void X86TargetLowering::ReplaceNodeResults(SDNode *N,
   case ISD::ATOMIC_LOAD_MAX:
   case ISD::ATOMIC_LOAD_UMIN:
   case ISD::ATOMIC_LOAD_UMAX:
-  case ISD::ATOMIC_LOAD: {
     // Delegate to generic TypeLegalization. Situations we can really handle
     // should have already been dealt with by AtomicExpandPass.cpp.
     break;
-  }
+
   case ISD::BITCAST: {
     assert(Subtarget.hasSSE2() && "Requires at least SSE2!");
     EVT DstVT = N->getValueType(0);
-- 
cgit v1.2.3