JIT: emit vzeroupper before calls to the bulk write barrier helper (#…

…106908)

The helper uses SSE2, so we need to take care to avoid AVX-SSE
transition penalties.

Fixes #106679.

src/coreclr/jit/gentree.cpp

@@ -1997,9 +1997,14 @@ void CallArgs::Remove(CallArg* arg)
 //
 bool GenTreeCall::NeedsVzeroupper(Compiler* comp)
 {
+    if (!comp->canUseVexEncoding())
+    {
+        return false;
+    }
+
     bool needsVzeroupper = false;
 
-    if (IsPInvoke() && comp->canUseVexEncoding())
+    if (IsPInvoke())
     {
         // The Intel optimization manual guidance in `3.11.5.3 Fixing Instruction Slowdowns` states:
         //   Insert a VZEROUPPER to tell the hardware that the state of the higher registers is clean
@@ -2028,6 +2033,7 @@ bool GenTreeCall::NeedsVzeroupper(Compiler* comp)
                 if (varTypeUsesFloatReg(this))
                 {
                     needsVzeroupper = true;
+                    break;
                 }
                 else
                 {
@@ -2050,6 +2056,13 @@ bool GenTreeCall::NeedsVzeroupper(Compiler* comp)
         }
     }
 
+    // Other special cases
+    //
+    if (!needsVzeroupper && IsHelperCall(comp, CORINFO_HELP_BULK_WRITEBARRIER))
+    {
+        needsVzeroupper = true;
+    }
+
     return needsVzeroupper;
 }
 #endif // TARGET_XARCH