JIT: Make stack size computation in fgCanFastTailCall more precise (#…

…103538)

Two changes:
1. Skip applying alignment when computing a call's arg stack size, which does
not make sense/is not correct
2. Round up the incoming parameter stack space in terms of number of slots

Without (1) we overestimate the stack size usage for some calls. Without (2) we
underestimate the incoming stack size for some methods. Both of these just
result in fewer tailcalls than possible, so cause no correctness issues.
However, I hit some diffs in #103537 because of them.

src/coreclr/jit/morph.cpp

@@ -5351,7 +5351,7 @@ bool Compiler::fgCanFastTailCall(GenTreeCall* callee, const char** failReason)
     callee->gtArgs.AddFinalArgsAndDetermineABIInfo(this, callee);
 
     unsigned calleeArgStackSize = 0;
-    unsigned callerArgStackSize = info.compArgStackSize;
+    unsigned callerArgStackSize = roundUp(info.compArgStackSize, TARGET_POINTER_SIZE);
 
     auto reportFastTailCallDecision = [&](const char* thisFailReason) {
         if (failReason != nullptr)
@@ -5405,8 +5405,12 @@ bool Compiler::fgCanFastTailCall(GenTreeCall* callee, const char** failReason)
 
     for (CallArg& arg : callee->gtArgs.Args())
     {
-        calleeArgStackSize = roundUp(calleeArgStackSize, arg.AbiInfo.ByteAlignment);
-        calleeArgStackSize += arg.AbiInfo.GetStackByteSize();
+        unsigned stackSize = arg.AbiInfo.GetStackByteSize();
+        if (stackSize > 0)
+        {
+            calleeArgStackSize = roundUp(calleeArgStackSize, arg.AbiInfo.ByteAlignment);
+            calleeArgStackSize += stackSize;
+        }
 
 #if defined(TARGET_ARM) || defined(TARGET_RISCV64)
         if (arg.AbiInfo.IsSplit())