Avoid creating illegal byref pointers

Byref pointers need to point within their "host" object -- thus
the alternate name "interior pointers". If the JIT creates and
reports a pointer as a "byref", but it points outside the host
object, and a GC occurs that moves the host object, the byref
pointer will not be updated. If a subsequent calculation puts
the byref "back" into the host object, it will actually be pointing
to garbage, since the host object has moved.

This occurred on ARM with array index calculations, in particular
because ARM doesn't have a single-instruction "base + scale*index + offset"
addressing mode. Thus, we were generating, for the jaggedarr_cs_do
test case, `ProcessJagged3DArray()` function:
```
// r0 = array object, r6 = computed index offset. We mark r4 as a byref.
add r4, r0, r6

// r4 - 32 is the offset of the object we care about. Then we load the array element.
// In this case, the loaded element is a gcref, so r4 becomes a gcref.
ldr r4, [r4-32]
```
We get this math because the user code uses `a[i - 10]`, which is
essentially `a + (i - 10) * 4 + 8` for element size 4. This is optimized
to `a + i * 4 - 32`. In the above code, `r6` is `i * 4`. In this case,
after the first instruction, `r4` can point beyond the array.
If a GC happens, `r4` isn't updated, and the second instruction loads garbage.

There are several fixes:
1. Change array morphing in `fgMorphArrayIndex()` to rearrange the array index
IR node creation to only create a byref pointer that is precise; don't create
"intermediate" byref pointers that don't represent the actual array element
address being computed. The tree matching code that annotates the generated tree
with field sequences needs to be updated to match the new form.
2. Change `fgMoveOpsLeft()` to prevent the left-weighted reassociation optimization
`[byref]+ (ref, [int]+ (int, int)) => [byref]+ ([byref]+ (ref, int), int)`. This
optimization creates "incorrect" byrefs that don't necessarily point within
the host object.
3. Add an additional condition to the `Fold "((x+icon1)+icon2) to (x+(icon1+icon2))"`
morph optimization to prevent merging of constant TYP_REF nodes, which now were
being recognized due to different tree shapes. This was probably always a problem,
but the particular tree shape wasn't seen before.

These fixes are all-platform. However, to reduce risk at this point, the are
enabled for ARM only, under the `FEATURE_PREVENT_BAD_BYREFS` `#ifdef`.

Fixes #17517.

There are many, many diffs.

For ARM32 ngen-based desktop asm diffs, it is a 0.30% improvement across all
framework assemblies. A lot of the diffs seem to be because we CSE the entire
array address offset expression, not just the index expression.

Author: BruceForstall

src/jit/morph.cpp

@@ -5810,6 +5810,25 @@ void Compiler::fgMoveOpsLeft(GenTree* tree)
             break;
         }
 
+#if FEATURE_PREVENT_BAD_BYREFS
+
+        // Don't split up a byref calculation and create a new byref. E.g.,
+        // [byref]+ (ref, [int]+ (int, int)) => [byref]+ ([byref]+ (ref, int), int).
+        // Doing this transformation could create a situation where the first
+        // addition (that is, [byref]+ (ref, int) ) creates a byref pointer that
+        // no longer points within the ref object. If a GC happens, the byref won't
+        // get updated. This can happen, for instance, if one of the int components
+        // is negative. It also requires the address generation be in a fully-interruptible
+        // code region.
+        //
+        if (varTypeIsGC(op1->TypeGet()) && op2->TypeGet() == TYP_I_IMPL)
+        {
+            assert(varTypeIsGC(tree->TypeGet()) && (oper == GT_ADD));
+            break;
+        }
+
+#endif // FEATURE_PREVENT_BAD_BYREFS
+
         /* Change "(x op (y op z))" to "(x op y) op z" */
         /* ie.    "(op1 op (ad1 op ad2))" to "(op1 op ad1) op ad2" */
 
@@ -6224,6 +6243,27 @@ GenTree* Compiler::fgMorphArrayIndex(GenTree* tree)
         addr = index;
     }
 
+#if FEATURE_PREVENT_BAD_BYREFS
+
+    // Be careful to only create the byref pointer when the full index expression is added to the array reference.
+    // We don't want to create a partial byref address expression that doesn't include the full index offset:
+    // a byref must point within the containing object. It is dangerous (especially when optimizations come into
+    // play) to create a "partial" byref that doesn't point exactly to the correct object; there is risk that
+    // the partial byref will not point within the object, and thus not get updated correctly during a GC.
+    // This is mostly a risk in fully-interruptible code regions.
+
+    /* Add the first element's offset */
+
+    GenTree* cns = gtNewIconNode(elemOffs, TYP_I_IMPL);
+
+    addr = gtNewOperNode(GT_ADD, TYP_I_IMPL, addr, cns);
+
+    /* Add the object ref to the element's offset */
+
+    addr = gtNewOperNode(GT_ADD, TYP_BYREF, arrRef, addr);
+
+#else // !FEATURE_PREVENT_BAD_BYREFS
+
     /* Add the object ref to the element's offset */
 
     addr = gtNewOperNode(GT_ADD, TYP_BYREF, arrRef, addr);
@@ -6234,6 +6274,8 @@ GenTree* Compiler::fgMorphArrayIndex(GenTree* tree)
 
     addr = gtNewOperNode(GT_ADD, TYP_BYREF, addr, cns);
 
+#endif // !FEATURE_PREVENT_BAD_BYREFS
+
 #if SMALL_TREE_NODES
     assert((tree->gtDebugFlags & GTF_DEBUG_NODE_LARGE) || GenTree::s_gtNodeSizes[GT_IND] == TREE_NODE_SZ_SMALL);
 #endif
@@ -6329,6 +6371,35 @@ GenTree* Compiler::fgMorphArrayIndex(GenTree* tree)
     GenTree* cnsOff = nullptr;
     if (addr->OperGet() == GT_ADD)
     {
+
+#if FEATURE_PREVENT_BAD_BYREFS
+
+        assert(addr->TypeGet() == TYP_BYREF);
+        assert(addr->gtOp.gtOp1->TypeGet() == TYP_REF);
+
+        addr = addr->gtOp.gtOp2;
+
+        // Look for the constant [#FirstElem] node here, or as the RHS of an ADD.
+
+        if (addr->gtOper == GT_CNS_INT)
+        {
+            cnsOff = addr;
+            addr   = nullptr;
+        }
+        else
+        {
+            if ((addr->OperGet() == GT_ADD) && (addr->gtOp.gtOp2->gtOper == GT_CNS_INT))
+            {
+                cnsOff = addr->gtOp.gtOp2;
+                addr   = addr->gtOp.gtOp1;
+            }
+
+            // Label any constant array index contributions with #ConstantIndex and any LclVars with GTF_VAR_ARR_INDEX
+            addr->LabelIndex(this);
+        }
+
+#else // !FEATURE_PREVENT_BAD_BYREFS
+
         if (addr->gtOp.gtOp2->gtOper == GT_CNS_INT)
         {
             cnsOff = addr->gtOp.gtOp2;
@@ -6346,6 +6417,8 @@ GenTree* Compiler::fgMorphArrayIndex(GenTree* tree)
             addr = addr->gtOp.gtOp1;
         }
         assert(addr->TypeGet() == TYP_REF);
+
+#endif // !FEATURE_PREVENT_BAD_BYREFS
     }
     else if (addr->OperGet() == GT_CNS_INT)
     {
@@ -13477,9 +13550,25 @@ GenTree* Compiler::fgMorphSmpOp(GenTree* tree, MorphAddrContext* mac)
                 if (op2->IsCnsIntOrI() && varTypeIsIntegralOrI(typ))
                 {
                     /* Fold "((x+icon1)+icon2) to (x+(icon1+icon2))" */
+                    CLANG_FORMAT_COMMENT_ANCHOR;
+
+#if FEATURE_PREVENT_BAD_BYREFS
+
+                    if (op1->gtOper == GT_ADD &&                          //
+                        !gtIsActiveCSE_Candidate(op1) &&                  //
+                        !op1->gtOverflow() &&                             //
+                        op1->gtOp.gtOp2->IsCnsIntOrI() &&                 //
+                        (op1->gtOp.gtOp2->OperGet() == op2->OperGet()) && //
+                        (op1->gtOp.gtOp2->TypeGet() != TYP_REF) &&        // Don't fold REFs
+                        (op2->TypeGet() != TYP_REF))                      // Don't fold REFs
+
+#else // !FEATURE_PREVENT_BAD_BYREFS
 
                     if (op1->gtOper == GT_ADD && !gtIsActiveCSE_Candidate(op1) && op1->gtOp.gtOp2->IsCnsIntOrI() &&
                         !op1->gtOverflow() && op1->gtOp.gtOp2->OperGet() == op2->OperGet())
+
+#endif // !FEATURE_PREVENT_BAD_BYREFS
+
                     {
                         cns1 = op1->gtOp.gtOp2;
                         op2->gtIntConCommon.SetIconValue(cns1->gtIntConCommon.IconValue() +

src/jit/target.h

@@ -317,6 +317,21 @@ typedef unsigned short regPairNoSmall; // arm: need 12 bits
 // #define PSEUDORANDOM_NOP_INSERTION
 // #endif
 
+// TODO-Cleanup: FEATURE_PREVENT_BAD_BYREFS guards code that prevents creating byref pointers to array elements
+// that are not "complete". That is, it only allows byref pointers to the exact array element, not to a portion
+// of the address expression leading to the full addressing expression. This prevents the possibility of creating
+// an illegal byref, which is an expression that points outside of the "host" object. Such bad byrefs won't get
+// updated properly during a GC, leaving them to point to garbage. This led to a GC hole on ARM due to ARM's
+// limited addressing modes (found in GCStress). The change is applicable and possibly desirable for other platforms,
+// but was put under ifdef to avoid introducing potential destabilizing change to those platforms at the end of the
+// .NET Core 2.1 ship cycle. More detail here: https://github.com/dotnet/coreclr/pull/17524. Consider making this
+// all-platform and removing these #ifdefs.
+#if defined(_TARGET_ARM_)
+#define FEATURE_PREVENT_BAD_BYREFS 1
+#else
+#define FEATURE_PREVENT_BAD_BYREFS 0
+#endif
+
 /*****************************************************************************/
 
 // clang-format off