JIT: Disallow implicit byref args and return buffers from aliasing (#…

…104616)

The recent work to allow more retbuf definitions uncovered cases where
we would end up with the retbuffer aliasing an implicit byref argument.
The JIT does not handle this aliasing, which would require proper
`GTF_GLOB_REF` flags on the local nodes accesses implicit byrefs. Fix
the problem by disallowing last-use copy elision when the local would
alias the return buffer.

Fix #104613

src/coreclr/jit/gentree.cpp

@@ -17974,11 +17974,14 @@ GenTreeLclVarCommon* GenTree::IsImplicitByrefParameterValuePreMorph(Compiler* co
 //    compiler - compiler instance
 //    addr     - [out] tree representing the address computation on top of the implicit byref.
 //               Will be the same as the return value if the whole implicit byref is used, for example.
+//    offset   - [out] Offset that "addr" is adding on top of the returned local.
 //
 // Return Value:
 //    Node for the local, or nullptr.
 //
-GenTreeLclVar* GenTree::IsImplicitByrefParameterValuePostMorph(Compiler* compiler, GenTree** addr)
+GenTreeLclVar* GenTree::IsImplicitByrefParameterValuePostMorph(Compiler*       compiler,
+                                                               GenTree**       addr,
+                                                               target_ssize_t* offset)
 {
 #if FEATURE_IMPLICIT_BYREFS && !defined(TARGET_LOONGARCH64) // TODO-LOONGARCH64-CQ: enable this.
 
@@ -17987,13 +17990,10 @@ GenTreeLclVar* GenTree::IsImplicitByrefParameterValuePostMorph(Compiler* compile
         return nullptr;
     }
 
-    *addr              = AsIndir()->Addr();
-    GenTree* innerAddr = *addr;
+    *addr = AsIndir()->Addr();
 
-    while (innerAddr->OperIs(GT_ADD) && innerAddr->gtGetOp2()->IsCnsIntOrI())
-    {
-        innerAddr = innerAddr->gtGetOp1();
-    }
+    GenTree* innerAddr = *addr;
+    compiler->gtPeelOffsets(&innerAddr, offset);
 
     if (innerAddr->OperIs(GT_LCL_VAR))
     {

src/coreclr/jit/gentree.h

@@ -1972,7 +1972,7 @@ struct GenTree
                       unsigned*             pSize     = nullptr);
 
     GenTreeLclVarCommon* IsImplicitByrefParameterValuePreMorph(Compiler* compiler);
-    GenTreeLclVar*       IsImplicitByrefParameterValuePostMorph(Compiler* compiler, GenTree** addr);
+    GenTreeLclVar* IsImplicitByrefParameterValuePostMorph(Compiler* compiler, GenTree** addr, target_ssize_t* offset);
 
     unsigned IsLclVarUpdateTree(GenTree** otherTree, genTreeOps* updateOper);
 

src/coreclr/jit/morph.cpp

@@ -3157,13 +3157,15 @@ void Compiler::fgMakeOutgoingStructArgCopy(GenTreeCall* call, CallArg* arg)
     if (opts.OptimizationEnabled() && arg->AbiInfo.PassedByRef)
     {
         GenTree*             implicitByRefLclAddr;
+        target_ssize_t       implicitByRefLclOffs;
         GenTreeLclVarCommon* implicitByRefLcl =
-            argx->IsImplicitByrefParameterValuePostMorph(this, &implicitByRefLclAddr);
+            argx->IsImplicitByrefParameterValuePostMorph(this, &implicitByRefLclAddr, &implicitByRefLclOffs);
 
         GenTreeLclVarCommon* lcl = implicitByRefLcl;
         if ((lcl == nullptr) && argx->OperIsLocal())
         {
-            lcl = argx->AsLclVarCommon();
+            lcl                  = argx->AsLclVarCommon();
+            implicitByRefLclOffs = lcl->GetLclOffs();
         }
 
         if (lcl != nullptr)
@@ -3191,6 +3193,28 @@ void Compiler::fgMakeOutgoingStructArgCopy(GenTreeCall* call, CallArg* arg)
                            !varDsc->lvPromoted && !varDsc->lvIsStructField && ((lcl->gtFlags & GTF_VAR_DEATH) != 0);
             }
 
+            // Disallow the argument from potentially aliasing the return
+            // buffer.
+            if (omitCopy)
+            {
+                GenTreeLclVarCommon* retBuffer = gtCallGetDefinedRetBufLclAddr(call);
+                if ((retBuffer != nullptr) && (retBuffer->GetLclNum() == varNum))
+                {
+                    unsigned       retBufferSize  = typGetObjLayout(call->gtRetClsHnd)->GetSize();
+                    target_ssize_t retBufferStart = retBuffer->GetLclOffs();
+                    target_ssize_t retBufferEnd   = retBufferStart + static_cast<target_ssize_t>(retBufferSize);
+
+                    unsigned       argSize        = arg->GetSignatureType() == TYP_STRUCT
+                                                        ? typGetObjLayout(arg->GetSignatureClassHandle())->GetSize()
+                                                        : genTypeSize(arg->GetSignatureType());
+                    target_ssize_t implByrefStart = implicitByRefLclOffs;
+                    target_ssize_t implByrefEnd   = implByrefStart + static_cast<target_ssize_t>(argSize);
+
+                    bool disjoint = (retBufferEnd <= implByrefStart) || (implByrefEnd <= retBufferStart);
+                    omitCopy      = disjoint;
+                }
+            }
+
             if (omitCopy)
             {
                 if (implicitByRefLcl != nullptr)
@@ -3218,6 +3242,7 @@ void Compiler::fgMakeOutgoingStructArgCopy(GenTreeCall* call, CallArg* arg)
 #endif
 
     JITDUMP("making an outgoing copy for struct arg\n");
+    assert(!call->IsTailCall() || !arg->AbiInfo.PassedByRef);
 
     CORINFO_CLASS_HANDLE copyBlkClass = arg->GetSignatureClassHandle();
     unsigned             tmp          = 0;