block_of_code: Add vzeroupper instructions where AVX-SSE transitions may occur

src/backend_x64/block_of_code.cpp

@@ -86,6 +86,12 @@ void BlockOfCode::GenRunCode() {
     align();
     run_code = getCurr<RunCodeFuncType>();
 
+    // As we currently do not emit AVX instructions, AVX-SSE transition may occur.
+    // We avoid the transition penalty by calling vzeroupper.
+    if (DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
+        vzeroupper();
+    }
+
     // This serves two purposes:
     // 1. It saves all the registers we as a callee need to save.
     // 2. It aligns the stack so that the code the JIT emits can assume

src/backend_x64/block_of_code.h

@@ -49,6 +49,12 @@ public:
         const u64 address  = reinterpret_cast<u64>(fn);
         const u64 distance = address - (getCurr<u64>() + 5);
 
+        // As we do not know if user-code is AVX or SSE, an AVX-SSE transition may occur.
+        // We avoid the transition penalty by calling vzeroupper.
+        if (DoesCpuSupport(Xbyak::util::Cpu::tAVX)) {
+            vzeroupper();
+        }
+
         if (distance >= 0x0000000080000000ULL && distance < 0xFFFFFFFF80000000ULL) {
             // Far call
             mov(rax, address);