Merge pull request #13897 from hseok-oh/ryujit/ftc_arg_count

Argument register & stack size calculation for fast tail call

Author: Jarret Shook

src/jit/compiler.h

@@ -8344,10 +8344,8 @@ XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
         unsigned  compArgsCount;     // Number of arguments (incl. implicit and     hidden)
 
 #if FEATURE_FASTTAILCALL
-        unsigned compArgRegCount;      // Number of incoming integer argument registers used for incoming arguments
-        unsigned compFloatArgRegCount; // Number of incoming floating argument registers used for incoming arguments
-        size_t   compArgStackSize;     // Incoming argument stack size in bytes
-#endif                                 // FEATURE_FASTTAILCALL
+        size_t compArgStackSize; // Incoming argument stack size in bytes
+#endif                           // FEATURE_FASTTAILCALL
 
         unsigned compRetBuffArg; // position of hidden return param var (0, 1) (BAD_VAR_NUM means not present);
         int compTypeCtxtArg; // position of hidden param for type context for generic code (CORINFO_CALLCONV_PARAMTYPE)

src/jit/lclvars.cpp

@@ -235,60 +235,6 @@ void Compiler::lvaInitTypeRef()
 
     lvaInitArgs(&varDscInfo);
 
-#if FEATURE_FASTTAILCALL
-
-    //-------------------------------------------------------------------------
-    // Calculate the argument register usage.
-    //
-    // This will later be used for fastTailCall determination
-    //-------------------------------------------------------------------------
-
-    unsigned argRegCount      = 0;
-    unsigned floatingRegCount = 0;
-    size_t   stackSize        = 0;
-
-    auto incrementRegCount = [&floatingRegCount, &argRegCount](LclVarDsc* varDsc) {
-        if (varDsc->lvIsHfa())
-        {
-            floatingRegCount += varDsc->lvHfaSlots();
-        }
-        else
-        {
-            varDsc->IsFloatRegType() ? ++floatingRegCount : ++argRegCount;
-        }
-    };
-
-    unsigned   argNum;
-    LclVarDsc* curDsc;
-
-    for (curDsc = lvaTable, argNum = 0; argNum < varDscInfo.varNum; argNum++, curDsc++)
-    {
-        if (curDsc->lvIsRegArg)
-        {
-            incrementRegCount(curDsc);
-#if FEATURE_MULTIREG_ARGS
-            if (curDsc->lvOtherArgReg != REG_NA)
-            {
-                incrementRegCount(curDsc);
-            }
-#endif // FEATURE_MULTIREG_ARGS
-        }
-        else
-        {
-            stackSize += curDsc->lvArgStackSize();
-        }
-    }
-
-    //-------------------------------------------------------------------------
-    // Save the register usage information and stack size.
-    //-------------------------------------------------------------------------
-
-    info.compArgRegCount      = argRegCount;
-    info.compFloatArgRegCount = floatingRegCount;
-    info.compArgStackSize     = stackSize;
-
-#endif // FEATURE_FASTTAILCALL
-
     //-------------------------------------------------------------------------
     // Finally the local variables
     //-------------------------------------------------------------------------
@@ -410,6 +356,13 @@ void Compiler::lvaInitArgs(InitVarDscInfo* varDscInfo)
     codeGen->floatRegState.rsCalleeRegArgCount = varDscInfo->floatRegArgNum;
 #endif // FEATURE_STACK_FP_X87
 
+#if FEATURE_FASTTAILCALL
+    // Save the stack usage information
+    // We can get register usage information using codeGen->intRegState and
+    // codeGen->floatRegState
+    info.compArgStackSize = varDscInfo->stackArgSize;
+#endif // FEATURE_FASTTAILCALL
+
     // The total argument size must be aligned.
     noway_assert((compArgSize % sizeof(void*)) == 0);
 
@@ -978,7 +931,6 @@ void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo)
         else
         {
 #if defined(_TARGET_ARM_)
-
             varDscInfo->setAllRegArgUsed(argType);
             if (varTypeIsFloating(argType))
             {
@@ -994,6 +946,10 @@ void Compiler::lvaInitUserArgs(InitVarDscInfo* varDscInfo)
             varDscInfo->setAllRegArgUsed(argType);
 
 #endif // _TARGET_XXX_
+
+#if FEATURE_FASTTAILCALL
+            varDscInfo->stackArgSize += (unsigned)roundUp(argSize, TARGET_POINTER_SIZE);
+#endif // FEATURE_FASTTAILCALL
         }
 
 #ifdef FEATURE_UNIX_AMD64_STRUCT_PASSING
@@ -1090,6 +1046,9 @@ void Compiler::lvaInitGenericsCtxt(InitVarDscInfo* varDscInfo)
             // For the RyuJIT backend, we need to mark these as being on the stack,
             // as this is not done elsewhere in the case that canEnreg returns false.
             varDsc->lvOnFrame = true;
+#if FEATURE_FASTTAILCALL
+            varDscInfo->stackArgSize += TARGET_POINTER_SIZE;
+#endif // FEATURE_FASTTAILCALL
         }
 #endif // !LEGACY_BACKEND
 
@@ -1163,6 +1122,9 @@ void Compiler::lvaInitVarArgsHandle(InitVarDscInfo* varDscInfo)
             // For the RyuJIT backend, we need to mark these as being on the stack,
             // as this is not done elsewhere in the case that canEnreg returns false.
             varDsc->lvOnFrame = true;
+#if FEATURE_FASTTAILCALL
+            varDscInfo->stackArgSize += TARGET_POINTER_SIZE;
+#endif // FEATURE_FASTTAILCALL
         }
 #endif // !LEGACY_BACKEND
 

src/jit/morph.cpp

@@ -7253,8 +7253,8 @@ bool Compiler::fgCanFastTailCall(GenTreeCall* callee)
     // Note that callee being a vararg method is not a problem since we can account the params being passed.
     unsigned nCallerArgs = info.compArgsCount;
 
-    size_t callerArgRegCount      = info.compArgRegCount;
-    size_t callerFloatArgRegCount = info.compFloatArgRegCount;
+    size_t callerArgRegCount      = codeGen->intRegState.rsCalleeRegArgCount;
+    size_t callerFloatArgRegCount = codeGen->floatRegState.rsCalleeRegArgCount;
 
     // Count the callee args including implicit and hidden.
     // Note that GenericContext and VarargCookie are added by importer while

src/jit/register_arg_convention.h

@@ -26,6 +26,11 @@ struct InitVarDscInfo
     bool      anyFloatStackArgs;
 #endif // _TARGET_ARM_
 
+#if FEATURE_FASTTAILCALL
+    // It is used to calculate argument stack size information in byte
+    unsigned stackArgSize;
+#endif // FEATURE_FASTTAILCALL
+
 public:
     // set to initial values
     void Init(LclVarDsc* lvaTable, bool _hasRetBufArg)
@@ -42,6 +47,10 @@ struct InitVarDscInfo
         fltArgSkippedRegMask = RBM_NONE;
         anyFloatStackArgs    = false;
 #endif // _TARGET_ARM_
+
+#if FEATURE_FASTTAILCALL
+        stackArgSize = 0;
+#endif // FEATURE_FASTTAILCALL
     }
 
     // return ref to current register arg for this type