Don't force vzeroupper for helpers written in managed

Author: tannergooding

src/coreclr/jit/gentree.cpp

@@ -2040,8 +2040,13 @@ bool GenTreeCall::NeedsVzeroupper(Compiler* comp)
 
             // Most other helpers are well known to not use any floating-point or SIMD logic internally, but
             // a few do exist so we need to ensure they are handled. They are identified by taking or
-            // returning a floating-point or SIMD type, regardless of how it is actually passed/returned.
-            checkSignature = !needsVzeroupper;
+            // returning a floating-point or SIMD type, regardless of how it is actually passed/returned but
+            // are excluded if we know they are implemented in managed.
+            checkSignature = !needsVzeroupper &&
+                             !IsHelperCall(comp, CORINFO_HELP_DBL2INT_OVF) &&
+                             !IsHelperCall(comp, CORINFO_HELP_DBL2LNG_OVF) &&
+                             !IsHelperCall(comp, CORINFO_HELP_DBL2UINT_OVF) &&
+                             !IsHelperCall(comp, CORINFO_HELP_DBL2ULNG_OVF);
             break;
         }
 

src/coreclr/jit/importercalls.cpp

@@ -9549,9 +9549,15 @@ GenTree* Compiler::impMathIntrinsic(CORINFO_METHOD_HANDLE method,
 
     if (isIntrinsicImplementedByUserCall)
     {
+#if defined(TARGET_XARCH)
         // We want to track math intrinsics implemented as user calls as special
         // to ensure we don't lose track of the fact it will call into native code
+        //
+        // This is used on xarch to track that it may need vzeroupper inserted to
+        // avoid the perf penalty on some hardware.
+
         *isSpecial = true;
+#endif // TARGET_XARCH
     }
 
 #if !defined(TARGET_X86)

src/coreclr/jit/rationalize.cpp

@@ -44,8 +44,14 @@ void Rationalizer::RewriteNodeAsCall(GenTree**             use,
 
     if (isSpecialIntrinsic)
     {
+#if defined(TARGET_XARCH)
         // Mark this as having been a special intrinsic node
+        //
+        // This is used on xarch to track that it may need vzeroupper inserted to
+        // avoid the perf penalty on some hardware.
+
         call->gtCallMoreFlags |= GTF_CALL_M_SPECIAL_INTRINSIC;
+#endif // TARGET_XARCH
     }
 
     assert(sig != nullptr);