[Hexagon] Make memcpy lowering thread-safe

This removes an unpleasant hack involving a global variable for special
lowering of certain memcpy calls. These are now lowered as intended in
EmitTargetCodeForMemcpy in the same way that other targets do it.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@255785 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/Hexagon/HexagonISelLowering.cpp

@@ -819,12 +819,7 @@ HexagonTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
   // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
   // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
   // node so that legalize doesn't hack it.
-  if (flag_aligned_memcpy) {
+  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
-    const char *MemcpyName =
-      "__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes";
-    Callee = DAG.getTargetExternalSymbol(MemcpyName, PtrVT);
-    flag_aligned_memcpy = false;
-  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
     Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, PtrVT);
   } else if (ExternalSymbolSDNode *S =
              dyn_cast<ExternalSymbolSDNode>(Callee)) {

lib/Target/Hexagon/HexagonSelectionDAGInfo.cpp

@@ -12,12 +12,11 @@
 //===----------------------------------------------------------------------===//
 
 #include "HexagonTargetMachine.h"
+#include "llvm/CodeGen/SelectionDAG.h"
 using namespace llvm;
 
 #define DEBUG_TYPE "hexagon-selectiondag-info"
 
-bool llvm::flag_aligned_memcpy;
-
 SDValue
 HexagonSelectionDAGInfo::
 EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain,
@@ -25,15 +24,40 @@ EmitTargetCodeForMemcpy(SelectionDAG &DAG, SDLoc dl, SDValue Chain,
                         bool isVolatile, bool AlwaysInline,
                         MachinePointerInfo DstPtrInfo,
                         MachinePointerInfo SrcPtrInfo) const {
-  flag_aligned_memcpy = false;
+  ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
-  if ((Align & 0x3) == 0) {
+  if (AlwaysInline || (Align & 0x3) != 0 || !ConstantSize)
-    ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
+    return SDValue();
-    if (ConstantSize) {
-      uint64_t SizeVal = ConstantSize->getZExtValue();
-      if ((SizeVal > 32) && ((SizeVal % 8) == 0))
-        flag_aligned_memcpy = true;
-    }
-  }
 
-  return SDValue();
+  uint64_t SizeVal = ConstantSize->getZExtValue();
+  if (SizeVal < 32 || (SizeVal % 8) != 0)
+    return SDValue();
+
+  // Special case aligned memcpys with size >= 32 bytes and a multiple of 8.
+  //
+  const TargetLowering &TLI = *DAG.getSubtarget().getTargetLowering();
+  TargetLowering::ArgListTy Args;
+  TargetLowering::ArgListEntry Entry;
+  Entry.Ty = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
+  Entry.Node = Dst;
+  Args.push_back(Entry);
+  Entry.Node = Src;
+  Args.push_back(Entry);
+  Entry.Node = Size;
+  Args.push_back(Entry);
+
+  const char *SpecialMemcpyName =
+      "__hexagon_memcpy_likely_aligned_min32bytes_mult8bytes";
+
+  TargetLowering::CallLoweringInfo CLI(DAG);
+  CLI.setDebugLoc(dl)
+      .setChain(Chain)
+      .setCallee(TLI.getLibcallCallingConv(RTLIB::MEMCPY),
+                 Type::getVoidTy(*DAG.getContext()),
+                 DAG.getTargetExternalSymbol(
+                     SpecialMemcpyName, TLI.getPointerTy(DAG.getDataLayout())),
+                 std::move(Args), 0)
+      .setDiscardResult();
+
+  std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
+  return CallResult.second;
 }

lib/Target/Hexagon/HexagonTargetMachine.h

@@ -44,8 +44,6 @@ public:
   }
 };
 
-extern bool flag_aligned_memcpy;
-
 } // end namespace llvm
 
 #endif

test/CodeGen/Hexagon/memcpy-likely-aligned.ll

@@ -0,0 +1,32 @@
+; RUN: llc -march=hexagon < %s | FileCheck %s
+; CHECK: __hexagon_memcpy_likely_aligned_min32bytes_mult8bytes
+
+target datalayout = "e-p:32:32:32-i64:64:64-i32:32:32-i16:16:16-i1:32:32-f64:64:64-f32:32:32-a0:0-n32"
+target triple = "hexagon-unknown-linux-gnu"
+
+%struct.e = type { i8, i8, [2 x i8] }
+%struct.s = type { i8* }
+%struct.o = type { %struct.n }
+%struct.n = type { [2 x %struct.l] }
+%struct.l = type { %struct.e, %struct.d, %struct.e }
+%struct.d = type <{ i8, i8, i8, i8, [2 x i8], [2 x i8] }>
+
+@y = global { <{ { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e }, { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } }> } { <{ { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e }, { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } }> <{ { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } { %struct.e { i8 3, i8 0, [2 x i8] undef }, { i8, i8, i8, [5 x i8] } { i8 -47, i8 2, i8 0, [5 x i8] undef }, %struct.e { i8 3, i8 0, [2 x i8] undef } }, { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } { %struct.e { i8 3, i8 0, [2 x i8] undef }, { i8, i8, i8, [5 x i8] } { i8 -47, i8 2, i8 0, [5 x i8] undef }, %struct.e { i8 3, i8 0, [2 x i8] undef } } }> }, align 4
+@t = common global %struct.s zeroinitializer, align 4
+@q = internal global %struct.o* null, align 4
+
+define void @foo() nounwind {
+entry:
+  %0 = load i8*, i8** getelementptr inbounds (%struct.s, %struct.s* @t, i32 0, i32 0), align 4
+  %1 = bitcast i8* %0 to %struct.o*
+  store %struct.o* %1, %struct.o** @q, align 4
+  %2 = load %struct.o*, %struct.o** @q, align 4
+  %p = getelementptr inbounds %struct.o, %struct.o* %2, i32 0, i32 0
+  %m = getelementptr inbounds %struct.n, %struct.n* %p, i32 0, i32 0
+  %arraydecay = getelementptr inbounds [2 x %struct.l], [2 x %struct.l]* %m, i32 0, i32 0
+  %3 = bitcast %struct.l* %arraydecay to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %3, i8* getelementptr inbounds ({ <{ { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e }, { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } }> }, { <{ { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e }, { %struct.e, { i8, i8, i8, [5 x i8] }, %struct.e } }> }* @y, i32 0, i32 0, i32 0, i32 0, i32 0), i32 32, i32 4, i1 false)
+  ret void
+}
+
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind