Removing GT_STORE_DYN_BLK, part 1

src/coreclr/jit/importer.cpp

@@ -10285,9 +10285,19 @@ void Compiler::impImportBlockCode(BasicBlock* block)
             case CEE_CPBLK:
             {
                 GenTreeFlags indirFlags = impPrefixFlagsToIndirFlags(prefixFlags);
-                op3                     = impPopStack().val; // Size
-                op2                     = impPopStack().val; // Value / Src addr
-                op1                     = impPopStack().val; // Dst addr
+                const bool   isVolatile = (indirFlags & GTF_IND_VOLATILE) != 0;
+#ifndef TARGET_X86
+                if (isVolatile && !impStackTop(0).val->IsCnsIntOrI())
+                {
+                    // We're going to emit a helper call surrounded by memory barriers, so we need to spill any side
+                    // effects.
+                    impSpillSideEffects(true, CHECK_SPILL_ALL DEBUGARG("spilling side-effects"));
+                }
+#endif
+
+                op3 = impPopStack().val; // Size
+                op2 = impPopStack().val; // Value / Src addr
+                op1 = impPopStack().val; // Dst addr
 
                 if (op3->IsCnsIntOrI())
                 {
@@ -10324,6 +10334,29 @@ void Compiler::impImportBlockCode(BasicBlock* block)
                 }
                 else
                 {
+                    if (TARGET_POINTER_SIZE == 8)
+                    {
+                        // Cast size to TYP_LONG on 64-bit targets
+                        op3 = gtNewCastNode(TYP_LONG, op3, /* fromUnsigned */ true, TYP_LONG);
+                    }
+
+// TODO: enable for X86 as well, it currently doesn't support memset/memcpy helpers
+// Then, get rid of GT_STORE_DYN_BLK entirely.
+#ifndef TARGET_X86
+                    const unsigned helper = opcode == CEE_INITBLK ? CORINFO_HELP_MEMSET : CORINFO_HELP_MEMCPY;
+                    if (isVolatile)
+                    {
+                        // Wrap with memory barriers: full-barrier + call + load-barrier
+                        impAppendTree(gtNewMemoryBarrier(), CHECK_SPILL_ALL, impCurStmtDI);
+                        impAppendTree(gtNewHelperCallNode(helper, TYP_VOID, op1, op2, op3), CHECK_SPILL_ALL,
+                                      impCurStmtDI);
+                        op1 = gtNewMemoryBarrier(true);
+                    }
+                    else
+                    {
+                        op1 = gtNewHelperCallNode(helper, TYP_VOID, op1, op2, op3);
+                    }
+#else
                     if (opcode == CEE_INITBLK)
                     {
                         if (!op2->IsIntegralConst(0))
@@ -10335,13 +10368,8 @@ void Compiler::impImportBlockCode(BasicBlock* block)
                     {
                         op2 = gtNewIndir(TYP_STRUCT, op2);
                     }
-
-#ifdef TARGET_64BIT
-                    // STORE_DYN_BLK takes a native uint size as it turns into call to memcpy.
-                    op3 = gtNewCastNode(TYP_I_IMPL, op3, /* fromUnsigned */ true, TYP_I_IMPL);
-#endif
-
                     op1 = gtNewStoreDynBlkNode(op1, op2, op3, indirFlags);
+#endif
                 }
                 goto SPILL_APPEND;
             }

src/coreclr/jit/lower.cpp

@@ -1844,6 +1844,7 @@ GenTree* Lowering::AddrGen(void* addr)
 
 //------------------------------------------------------------------------
 // LowerCallMemmove: Replace Buffer.Memmove(DST, SRC, CNS_SIZE) with a GT_STORE_BLK:
+//    Do the same for CORINFO_HELP_MEMCPY(DST, SRC, CNS_SIZE)
 //
 //    *  STORE_BLK struct<CNS_SIZE> (copy) (Unroll)
 //    +--*  LCL_VAR   byref  dst
@@ -1860,7 +1861,8 @@ GenTree* Lowering::AddrGen(void* addr)
 bool Lowering::LowerCallMemmove(GenTreeCall* call, GenTree** next)
 {
     JITDUMP("Considering Memmove [%06d] for unrolling.. ", comp->dspTreeID(call))
-    assert(comp->lookupNamedIntrinsic(call->gtCallMethHnd) == NI_System_Buffer_Memmove);
+    assert(call->IsHelperCall(comp, CORINFO_HELP_MEMCPY) ||
+           (comp->lookupNamedIntrinsic(call->gtCallMethHnd) == NI_System_Buffer_Memmove));
 
     assert(call->gtArgs.CountUserArgs() == 3);
 
@@ -1894,7 +1896,8 @@ bool Lowering::LowerCallMemmove(GenTreeCall* call, GenTree** next)
 
             // TODO-CQ: Use GenTreeBlk::BlkOpKindUnroll here if srcAddr and dstAddr don't overlap, thus, we can
             // unroll this memmove as memcpy - it doesn't require lots of temp registers
-            storeBlk->gtBlkOpKind = GenTreeBlk::BlkOpKindUnrollMemmove;
+            storeBlk->gtBlkOpKind = call->IsHelperCall(comp, CORINFO_HELP_MEMCPY) ? GenTreeBlk::BlkOpKindUnroll
+                                                                                  : GenTreeBlk::BlkOpKindUnrollMemmove;
 
             BlockRange().InsertBefore(call, srcBlk);
             BlockRange().InsertBefore(call, storeBlk);
@@ -2215,16 +2218,22 @@ GenTree* Lowering::LowerCall(GenTree* node)
     }
 
 #if defined(TARGET_AMD64) || defined(TARGET_ARM64)
+    GenTree* nextNode = nullptr;
     if (call->gtCallMoreFlags & GTF_CALL_M_SPECIAL_INTRINSIC)
     {
-        GenTree*       nextNode = nullptr;
-        NamedIntrinsic ni       = comp->lookupNamedIntrinsic(call->gtCallMethHnd);
+        NamedIntrinsic ni = comp->lookupNamedIntrinsic(call->gtCallMethHnd);
         if (((ni == NI_System_Buffer_Memmove) && LowerCallMemmove(call, &nextNode)) ||
             ((ni == NI_System_SpanHelpers_SequenceEqual) && LowerCallMemcmp(call, &nextNode)))
         {
             return nextNode;
         }
     }
+
+    // Try to lower CORINFO_HELP_MEMCPY to unrollable STORE_BLK
+    if (call->IsHelperCall(comp, CORINFO_HELP_MEMCPY) && LowerCallMemmove(call, &nextNode))
+    {
+        return nextNode;
+    }
 #endif
 
     call->ClearOtherRegs();