[TailCall] Allow llvm.memcpy/memset/memmove to be tail calls when parent function return the intrinsics's first argument

[wmi]

Summary

llvm.memcpy/memset/memmove return void but they will return the first argument after they are expanded as libcalls. Now if the parent function has any return value, llvm.memcpy cannot be turned into tail call after expansion.

The patch is to handle that case in SelectionDAGBuilder so when caller function return the same value as the first argument of llvm.memcpy, tail call is allowed.

[efriedma]

llvm/trunk/lib/CodeGen/Analysis.cpp

569

This is only true if the argument is specifically expanded to a libcall to the C library function "memcpy". If the target expands it to some other library call (e.g., __aeabi_memcpy, like we do on Android), it isn't a tail call, and we miscompile.

llvm/trunk/lib/CodeGen/Analysis.cpp

bool llvm::returnTypeIsEligibleForTailCall(const Function *F,
   if (isa<UndefValue>(Ret->getOperand(0))) return true;
 
   // Make sure the attributes attached to each return are compatible.
   bool AllowDifferingSizes;
   if (!attributesPermitTailCall(F, I, Ret, TLI, &AllowDifferingSizes))
     return false;
 
   const Value *RetVal = Ret->getOperand(0), *CallVal = I;
+  // Intrinsic like llvm.memcpy has no return value, but will return the
+  // first argument if it is expanded as libcall.
+  const CallInst *Call = cast<CallInst>(I);
+  if (Function *F = Call->getCalledFunction()) {
+    Intrinsic::ID IID = F->getIntrinsicID();
+    if ((IID == Intrinsic::memcpy || IID == Intrinsic::memmove ||
+         IID == Intrinsic::memset) &&
+        RetVal == Call->getArgOperand(0))
+      return true;
+  }
+
   SmallVector<unsigned, 4> RetPath, CallPath;
   SmallVector<CompositeType *, 4> RetSubTypes, CallSubTypes;
 
   bool RetEmpty = !firstRealType(RetVal->getType(), RetSubTypes, RetPath);
   bool CallEmpty = !firstRealType(CallVal->getType(), CallSubTypes, CallPath);
 
   // Nothing's actually returned, it doesn't matter what the callee put there
   // it's a valid tail call.

llvm/trunk/test/CodeGen/X86/tailcall-mem-intrinsics.ll

 ; CHECK-LABEL: tail_memset
 ; CHECK: jmp memset
 define void @tail_memset(i8* nocapture %p, i8 %c, i32 %n) #0 {
 entry:
   tail call void @llvm.memset.p0i8.i32(i8* %p, i8 %c, i32 %n, i32 1, i1 false)
   ret void
 }
 
+; CHECK-LABEL: tail_memcpy_ret
+; CHECK: jmp memcpy
+define i8* @tail_memcpy_ret(i8* nocapture %p, i8* nocapture readonly %q, i32 %n) #0 {
+entry:
+  tail call void @llvm.memcpy.p0i8.p0i8.i32(i8* %p, i8* %q, i32 %n, i32 1, i1 false)
+  ret i8* %p
+}
+
+; CHECK-LABEL: tail_memmove_ret
+; CHECK: jmp memmove
+define i8* @tail_memmove_ret(i8* nocapture %p, i8* nocapture readonly %q, i32 %n) #0 {
+entry:
+  tail call void @llvm.memmove.p0i8.p0i8.i32(i8* %p, i8* %q, i32 %n, i32 1, i1 false)
+  ret i8* %p
+}
+
+; CHECK-LABEL: tail_memset_ret
+; CHECK: jmp memset
+define i8* @tail_memset_ret(i8* nocapture %p, i8 %c, i32 %n) #0 {
+entry:
+  tail call void @llvm.memset.p0i8.i32(i8* %p, i8 %c, i32 %n, i32 1, i1 false)
+  ret i8* %p
+}
+
 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture readonly, i32, i32, i1) #0
 declare void @llvm.memmove.p0i8.p0i8.i32(i8* nocapture, i8* nocapture readonly, i32, i32, i1) #0
 declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) #0
 
 attributes #0 = { nounwind }