Recommitting r275284: add support to inline __builtin_mempcpy

Patch by Sunita Marathe

Third try, now following fixes to MSan to handle mempcy in such a way that this commit won't break the MSan buildbots. (Thanks, Evegenii!)



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@277189 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/Analysis/TargetLibraryInfo.def

@@ -734,6 +734,9 @@ TLI_DEFINE_STRING_INTERNAL("memcpy")
 /// void *memmove(void *s1, const void *s2, size_t n);
 TLI_DEFINE_ENUM_INTERNAL(memmove)
 TLI_DEFINE_STRING_INTERNAL("memmove")
+/// void *mempcpy(void *s1, const void *s2, size_t n);
+TLI_DEFINE_ENUM_INTERNAL(mempcpy)
+TLI_DEFINE_STRING_INTERNAL("mempcpy")
 // void *memrchr(const void *s, int c, size_t n);
 TLI_DEFINE_ENUM_INTERNAL(memrchr)
 TLI_DEFINE_STRING_INTERNAL("memrchr")

include/llvm/Analysis/TargetLibraryInfo.h

@@ -252,7 +252,7 @@ public:
     case LibFunc::exp2:      case LibFunc::exp2f:      case LibFunc::exp2l:
     case LibFunc::memcmp:    case LibFunc::strcmp:     case LibFunc::strcpy:
     case LibFunc::stpcpy:    case LibFunc::strlen:     case LibFunc::strnlen:
-    case LibFunc::memchr:
+    case LibFunc::memchr:    case LibFunc::mempcpy:
       return true;
     }
     return false;

lib/Analysis/TargetLibraryInfo.cpp

@@ -644,6 +644,7 @@ bool TargetLibraryInfoImpl::isValidProtoForLibFunc(const FunctionType &FTy,
       return false;
   // fallthrough
   case LibFunc::memcpy:
+  case LibFunc::mempcpy:
   case LibFunc::memmove:
     return (NumParams == 3 && FTy.getReturnType() == FTy.getParamType(0) &&
             FTy.getParamType(0)->isPointerTy() &&

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -6043,6 +6043,49 @@ bool SelectionDAGBuilder::visitMemChrCall(const CallInst &I) {
   return false;
 }
 
+///
+/// visitMemPCpyCall -- lower a mempcpy call as a memcpy followed by code to
+/// to adjust the dst pointer by the size of the copied memory.
+bool SelectionDAGBuilder::visitMemPCpyCall(const CallInst &I) {
+
+  // Verify argument count: void *mempcpy(void *, const void *, size_t)
+  if (I.getNumArgOperands() != 3)
+    return false;
+
+  SDValue Dst = getValue(I.getArgOperand(0));
+  SDValue Src = getValue(I.getArgOperand(1));
+  SDValue Size = getValue(I.getArgOperand(2));
+
+  unsigned DstAlign = DAG.InferPtrAlignment(Dst);
+  unsigned SrcAlign = DAG.InferPtrAlignment(Src);
+  unsigned Align = std::min(DstAlign, SrcAlign);
+  if (Align == 0) // Alignment of one or both could not be inferred.
+    Align = 1; // 0 and 1 both specify no alignment, but 0 is reserved.
+
+  bool isVol = false;
+  SDLoc sdl = getCurSDLoc();
+
+  // In the mempcpy context we need to pass in a false value for isTailCall
+  // because the return pointer needs to be adjusted by the size of
+  // the copied memory.
+  SDValue MC = DAG.getMemcpy(getRoot(), sdl, Dst, Src, Size, Align, isVol,
+                             false, /*isTailCall=*/false,
+                             MachinePointerInfo(I.getArgOperand(0)),
+                             MachinePointerInfo(I.getArgOperand(1)));
+  assert(MC.getNode() != nullptr &&
+         "** memcpy should not be lowered as TailCall in mempcpy context **");
+  DAG.setRoot(MC);
+
+  // Check if Size needs to be truncated or extended.
+  Size = DAG.getSExtOrTrunc(Size, sdl, Dst.getValueType());
+
+  // Adjust return pointer to point just past the last dst byte.
+  SDValue DstPlusSize = DAG.getNode(ISD::ADD, sdl, Dst.getValueType(),
+                                    Dst, Size);
+  setValue(&I, DstPlusSize);
+  return true;
+}
+
 /// visitStrCpyCall -- See if we can lower a strcpy or stpcpy call into an
 /// optimized form.  If so, return true and lower it, otherwise return false
 /// and it will be lowered like a normal call.
@@ -6333,6 +6376,10 @@ void SelectionDAGBuilder::visitCall(const CallInst &I) {
         if (visitMemCmpCall(I))
           return;
         break;
+      case LibFunc::mempcpy:
+        if (visitMemPCpyCall(I))
+          return;
+        break;
       case LibFunc::memchr:
         if (visitMemChrCall(I))
           return;

lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h

@@ -885,6 +885,7 @@ private:
   void visitPHI(const PHINode &I);
   void visitCall(const CallInst &I);
   bool visitMemCmpCall(const CallInst &I);
+  bool visitMemPCpyCall(const CallInst &I);
   bool visitMemChrCall(const CallInst &I);
   bool visitStrCpyCall(const CallInst &I, bool isStpcpy);
   bool visitStrCmpCall(const CallInst &I);

lib/Transforms/Utils/BuildLibCalls.cpp

@@ -250,6 +250,7 @@ bool llvm::inferLibFuncAttributes(Function &F, const TargetLibraryInfo &TLI) {
     Changed |= setDoesNotCapture(F, 2);
     return Changed;
   case LibFunc::memcpy:
+  case LibFunc::mempcpy:
   case LibFunc::memccpy:
   case LibFunc::memmove:
     Changed |= setDoesNotThrow(F);

test/CodeGen/X86/mempcpy.ll

@@ -0,0 +1,28 @@
+;  RUN: llc < %s -mtriple=x86_64-unknown-linux -O2 | FileCheck %s
+;  RUN: llc < %s -mtriple=i686-unknown-linux -O2 | FileCheck %s
+
+; This test checks that:
+; (1)  mempcpy is lowered as memcpy, and 
+; (2)  its return value is DST+N i.e. the dst pointer adjusted by the copy size.
+; To keep the testing of (2) independent of the exact instructions used to
+; adjust the dst pointer, DST+N is explicitly computed and stored to a global
+; variable G before the mempcpy call. This instance of DST+N causes the repeat
+; DST+N done in the context of the return value of mempcpy to be redundant, and
+; the first instance to be reused as the return value. This allows the check for
+; (2) to be expressed as verifying that the MOV to store DST+N to G and
+; the MOV to copy DST+N to %rax use the same source register.
+@G = common global i8* null, align 8
+
+; CHECK-LABEL: RET_MEMPCPY:
+; CHECK: mov{{.*}} [[REG:%[er][a-z0-9]+]], {{.*}}G
+; CHECK: call{{.*}} {{.*}}memcpy
+; CHECK: mov{{.*}} [[REG]], %{{[er]}}ax
+;
+define i8* @RET_MEMPCPY(i8* %DST, i8* %SRC, i64 %N) {
+  %add.ptr = getelementptr inbounds i8, i8* %DST, i64 %N
+  store i8* %add.ptr, i8** @G, align 8
+  %call = tail call i8* @mempcpy(i8* %DST, i8* %SRC, i64 %N) 
+  ret i8* %call
+}
+
+declare i8* @mempcpy(i8*, i8*, i64)

test/Transforms/InferFunctionAttrs/annotate.ll

@@ -499,6 +499,9 @@ declare i32 @memcmp(i8*, i8*, i64)
 ; CHECK: declare i8* @memcpy(i8*, i8* nocapture readonly, i64) [[G0]]
 declare i8* @memcpy(i8*, i8*, i64)
 
+; CHECK: declare i8* @mempcpy(i8*, i8* nocapture readonly, i64) [[G0]]
+declare i8* @mempcpy(i8*, i8*, i64)
+
 ; CHECK: declare i8* @memmove(i8*, i8* nocapture readonly, i64) [[G0]]
 declare i8* @memmove(i8*, i8*, i64)
 

test/Transforms/InferFunctionAttrs/no-proto.ll

@@ -480,6 +480,9 @@ declare void @memcmp(...)
 ; CHECK: declare void @memcpy(...)
 declare void @memcpy(...)
 
+; CHECK: declare void @mempcpy(...)
+declare void @mempcpy(...)
+
 ; CHECK: declare void @memmove(...)
 declare void @memmove(...)