Merge pull request #11268 from sdmaclea/PR-ARM64-two-gc-attributes

[Arm64] Support two GC attributes in pair forms

6 files changed, 83 insertions, 67 deletions

diff --git a/src/jit/codegenarm64.cpp b/src/jit/codegenarm64.cpp
index c075c0ac18..a8c92c2fba 100644
--- a/src/jit/codegenarm64.cpp
+++ b/src/jit/codegenarm64.cpp
@@ -3487,31 +3487,29 @@ void CodeGen::genCodeForCpObj(GenTreeObj* cpObjNode)
     // If we can prove it's on the stack we don't need to use the write barrier.
     if (dstOnStack)
     {
-        for (unsigned i = 0; i < slots; ++i)
+        unsigned i = 0;
+        // Check if two or more remaining slots and use a ldp/stp sequence
+        while (i < slots - 1)
         {
-            emitAttr attr = EA_8BYTE;
-            if (gcPtrs[i] == GCT_GCREF)
-                attr = EA_GCREF;
-            else if (gcPtrs[i] == GCT_BYREF)
-                attr = EA_BYREF;
-
-            // Check if two or more remaining slots and use a ldp/stp sequence
-            // TODO-ARM64-CQ: Current limitations only allows using ldp/stp when both of the GC types match
-            if ((i + 1 < slots) && (gcPtrs[i] == gcPtrs[i + 1]))
-            {
-                emit->emitIns_R_R_R_I(INS_ldp, attr, tmpReg, tmpReg2, REG_WRITE_BARRIER_SRC_BYREF,
-                                      2 * TARGET_POINTER_SIZE, INS_OPTS_POST_INDEX);
-                emit->emitIns_R_R_R_I(INS_stp, attr, tmpReg, tmpReg2, REG_WRITE_BARRIER_DST_BYREF,
-                                      2 * TARGET_POINTER_SIZE, INS_OPTS_POST_INDEX);
-                ++i; // extra increment of i, since we are copying two items
-            }
-            else
-            {
-                emit->emitIns_R_R_I(INS_ldr, attr, tmpReg, REG_WRITE_BARRIER_SRC_BYREF, TARGET_POINTER_SIZE,
-                                    INS_OPTS_POST_INDEX);
-                emit->emitIns_R_R_I(INS_str, attr, tmpReg, REG_WRITE_BARRIER_DST_BYREF, TARGET_POINTER_SIZE,
-                                    INS_OPTS_POST_INDEX);
-            }
+            emitAttr attr0 = emitTypeSize(compiler->getJitGCType(gcPtrs[i + 0]));
+            emitAttr attr1 = emitTypeSize(compiler->getJitGCType(gcPtrs[i + 1]));
+
+            emit->emitIns_R_R_R_I(INS_ldp, attr0, tmpReg, tmpReg2, REG_WRITE_BARRIER_SRC_BYREF, 2 * TARGET_POINTER_SIZE,
+                                  INS_OPTS_POST_INDEX, attr1);
+            emit->emitIns_R_R_R_I(INS_stp, attr0, tmpReg, tmpReg2, REG_WRITE_BARRIER_DST_BYREF, 2 * TARGET_POINTER_SIZE,
+                                  INS_OPTS_POST_INDEX, attr1);
+            i += 2;
+        }
+
+        // Use a ldr/str sequence for the last remainder
+        if (i < slots)
+        {
+            emitAttr attr0 = emitTypeSize(compiler->getJitGCType(gcPtrs[i + 0]));
+
+            emit->emitIns_R_R_I(INS_ldr, attr0, tmpReg, REG_WRITE_BARRIER_SRC_BYREF, TARGET_POINTER_SIZE,
+                                INS_OPTS_POST_INDEX);
+            emit->emitIns_R_R_I(INS_str, attr0, tmpReg, REG_WRITE_BARRIER_DST_BYREF, TARGET_POINTER_SIZE,
+                                INS_OPTS_POST_INDEX);
         }
     }
     else
diff --git a/src/jit/codegenarmarch.cpp b/src/jit/codegenarmarch.cpp
index 4a79ad1bba..0888e83963 100644
--- a/src/jit/codegenarmarch.cpp
+++ b/src/jit/codegenarmarch.cpp
@@ -356,20 +356,9 @@ void CodeGen::genPutArgStk(GenTreePutArgStk* treeNode)
                     assert(loReg != addrReg);
                     noway_assert((remainingSize == 2 * TARGET_POINTER_SIZE) || (hiReg != addrReg));
 
-                    // Use a ldp instruction if types match
-                    // TODO-ARM64-CQ: Current limitations only allows using ldp/stp when both of the GC types match
-                    if (type0 == type1)
-                    {
-                        // Load from our address expression source
-                        emit->emitIns_R_R_R_I(INS_ldp, emitTypeSize(type0), loReg, hiReg, addrReg, structOffset);
-                    }
-                    else
-                    {
-                        // Load from our address expression source
-                        emit->emitIns_R_R_I(ins_Load(type0), emitTypeSize(type0), loReg, addrReg, structOffset);
-                        emit->emitIns_R_R_I(ins_Load(type1), emitTypeSize(type1), hiReg, addrReg,
-                                            structOffset + TARGET_POINTER_SIZE);
-                    }
+                    // Load from our address expression source
+                    emit->emitIns_R_R_R_I(INS_ldp, emitTypeSize(type0), loReg, hiReg, addrReg, structOffset,
+                                          INS_OPTS_NONE, emitTypeSize(type0));
                 }
 
                 // Emit two store instructions to store the two registers into the outgoing argument area
diff --git a/src/jit/emit.cpp b/src/jit/emit.cpp
index d2aa29fd7a..ff2475a4f0 100644
--- a/src/jit/emit.cpp
+++ b/src/jit/emit.cpp
@@ -1381,6 +1381,10 @@ void* emitter::emitAllocInstr(size_t sz, emitAttr opsz)
         id->idOpSize(EA_SIZE(opsz));
     }
 
+#ifdef _TARGET_ARM64_
+    id->idGCrefReg2(GCT_NONE);
+#endif
+
     // Amd64: ip-relative addressing is supported even when not generating relocatable ngen code
     if (EA_IS_DSP_RELOC(opsz)
 #ifndef _TARGET_AMD64_
diff --git a/src/jit/emit.h b/src/jit/emit.h
index e1c924f467..a63242f0b0 100644
--- a/src/jit/emit.h
+++ b/src/jit/emit.h
@@ -655,6 +655,9 @@ protected:
         // unnecessarily since the GC-ness of the second register is only needed for call instructions.
         // The instrDescCGCA struct's member keeping the GC-ness of the first return register is _idcSecondRetRegGCType.
         GCtype _idGCref : 2; // GCref operand? (value is a "GCtype")
+#ifdef _TARGET_ARM64_
+        GCtype _idGCref2 : 2;
+#endif
 
         // Note that we use the _idReg1 and _idReg2 fields to hold
         // the live gcrefReg mask for the call instructions on x86/x64
@@ -668,7 +671,7 @@ protected:
         // x86:   30 bits
         // amd64: 38 bits
         // arm:   32 bits
-        // arm64: 30 bits
+        // arm64: 32 bits
         CLANG_FORMAT_COMMENT_ANCHOR;
 
 #if HAS_TINY_DESC
@@ -718,8 +721,8 @@ protected:
 #define ID_EXTRA_BITFIELD_BITS (16)
 
 #elif defined(_TARGET_ARM64_)
-// For Arm64, we have used 15 bits from the second DWORD.
-#define ID_EXTRA_BITFIELD_BITS (16)
+// For Arm64, we have used 18 bits from the second DWORD.
+#define ID_EXTRA_BITFIELD_BITS (18)
 #elif defined(_TARGET_XARCH_) && !defined(LEGACY_BACKEND)
 // For xarch !LEGACY_BACKEND, we have used 14 bits from the second DWORD.
 #define ID_EXTRA_BITFIELD_BITS (14)
@@ -735,7 +738,7 @@ protected:
         // x86:   38 bits  // if HAS_TINY_DESC is not defined (which it is)
         // amd64: 46 bits
         // arm:   48 bits
-        // arm64: 48 bits
+        // arm64: 50 bits
         CLANG_FORMAT_COMMENT_ANCHOR;
 
         unsigned _idCnsReloc : 1; // LargeCns is an RVA and needs reloc tag
@@ -748,7 +751,7 @@ protected:
         // x86:   40 bits
         // amd64: 48 bits
         // arm:   50 bits
-        // arm64: 50 bits
+        // arm64: 52 bits
         CLANG_FORMAT_COMMENT_ANCHOR;
 
 #define ID_EXTRA_BITS (ID_EXTRA_RELOC_BITS + ID_EXTRA_BITFIELD_BITS)
@@ -764,7 +767,7 @@ protected:
         // x86:   24 bits
         // amd64: 16 bits
         // arm:   14 bits
-        // arm64: 14 bits
+        // arm64: 12 bits
 
         unsigned _idSmallCns : ID_BIT_SMALL_CNS;
 
@@ -1072,6 +1075,17 @@ protected:
             assert(reg == _idReg1);
         }
 
+#ifdef _TARGET_ARM64_
+        GCtype idGCrefReg2() const
+        {
+            return (GCtype)_idGCref2;
+        }
+        void idGCrefReg2(GCtype gctype)
+        {
+            _idGCref2 = gctype;
+        }
+#endif // _TARGET_ARM64_
+
         regNumber idReg2() const
         {
             return _idReg2;
diff --git a/src/jit/emitarm64.cpp b/src/jit/emitarm64.cpp
index 1a56e863e5..e2a1ae14ec 100644
--- a/src/jit/emitarm64.cpp
+++ b/src/jit/emitarm64.cpp
@@ -5072,7 +5072,8 @@ void emitter::emitIns_R_R_R_I(instruction ins,
                               regNumber   reg2,
                               regNumber   reg3,
                               ssize_t     imm,
-                              insOpts     opt /* = INS_OPTS_NONE */)
+                              insOpts     opt /* = INS_OPTS_NONE */,
+                              emitAttr    attrReg2 /* = EA_UNKNOWN */)
 {
     emitAttr  size     = EA_SIZE(attr);
     emitAttr  elemsize = EA_UNKNOWN;
@@ -5347,6 +5348,23 @@ void emitter::emitIns_R_R_R_I(instruction ins,
     id->idReg2(reg2);
     id->idReg3(reg3);
 
+    if (attrReg2 != EA_UNKNOWN)
+    {
+        assert((fmt == IF_LS_3B) || (fmt == IF_LS_3C));
+        if (EA_IS_GCREF(attrReg2))
+        {
+            /* A special value indicates a GCref pointer value */
+
+            id->idGCrefReg2(GCT_GCREF);
+        }
+        else if (EA_IS_BYREF(attrReg2))
+        {
+            /* A special value indicates a Byref pointer value */
+
+            id->idGCrefReg2(GCT_BYREF);
+        }
+    }
+
     dispIns(id);
     appendToCurIG(id);
 }
@@ -9324,35 +9342,27 @@ size_t emitter::emitOutputInstr(insGroup* ig, instrDesc* id, BYTE** dp)
     // for stores, but we ignore those cases here.)
     if (emitInsMayWriteToGCReg(id)) // True if "id->idIns()" writes to a register than can hold GC ref.
     {
-        // If we ever generate instructions that write to multiple registers,
-        // then we'd need to more work here to ensure that changes in the status of GC refs are
-        // tracked properly.
-        if (emitInsMayWriteMultipleRegs(id))
+        // We assume that "idReg1" is the primary destination register for all instructions
+        if (id->idGCref() != GCT_NONE)
         {
-            // INS_ldp etc...
-            // We assume that "idReg1" and "idReg2" are the destination register for all instructions
-            // TODO-ARM64-CQ: Current limitations only allows using ldp/stp when both of the GC types match
-            if (id->idGCref() != GCT_NONE)
-            {
-                emitGCregLiveUpd(id->idGCref(), id->idReg1(), dst);
-                emitGCregLiveUpd(id->idGCref(), id->idReg2(), dst);
-            }
-            else
-            {
-                emitGCregDeadUpd(id->idReg1(), dst);
-                emitGCregDeadUpd(id->idReg2(), dst);
-            }
+            emitGCregLiveUpd(id->idGCref(), id->idReg1(), dst);
         }
         else
         {
-            // We assume that "idReg1" is the destination register for all instructions
-            if (id->idGCref() != GCT_NONE)
+            emitGCregDeadUpd(id->idReg1(), dst);
+        }
+
+        if (emitInsMayWriteMultipleRegs(id))
+        {
+            // INS_ldp etc...
+            // "idReg2" is the secondary destination register
+            if (id->idGCrefReg2() != GCT_NONE)
             {
-                emitGCregLiveUpd(id->idGCref(), id->idReg1(), dst);
+                emitGCregLiveUpd(id->idGCrefReg2(), id->idReg2(), dst);
             }
             else
             {
-                emitGCregDeadUpd(id->idReg1(), dst);
+                emitGCregDeadUpd(id->idReg2(), dst);
             }
         }
     }
diff --git a/src/jit/emitarm64.h b/src/jit/emitarm64.h
index 6a8e42b86f..8dcb3286bb 100644
--- a/src/jit/emitarm64.h
+++ b/src/jit/emitarm64.h
@@ -724,7 +724,8 @@ void emitIns_R_R_R_I(instruction ins,
                      regNumber   reg2,
                      regNumber   reg3,
                      ssize_t     imm,
-                     insOpts     opt = INS_OPTS_NONE);
+                     insOpts     opt      = INS_OPTS_NONE,
+                     emitAttr    attrReg2 = EA_UNKNOWN);
 
 void emitIns_R_R_R_Ext(instruction ins,
                        emitAttr    attr,