diff --git a/platform/pc/service.c b/platform/pc/service.c
index c24bbf9ed..6000db0bb 100644
--- a/platform/pc/service.c
+++ b/platform/pc/service.c
@@ -260,6 +260,9 @@ static void __attribute__((noinline)) init_service_new_stack()
     init_debug("init_hwrand");
     init_hwrand();
 
+    init_debug("init cpu features");
+    init_cpu_features();
+
     init_debug("calling kernel_runtime_init");
     kernel_runtime_init(kh);
     while(1);
diff --git a/src/x86_64/crt0.s b/src/x86_64/crt0.s
index 424ecd371..e9a99be61 100644
--- a/src/x86_64/crt0.s
+++ b/src/x86_64/crt0.s
@@ -45,19 +45,32 @@ extern  init_service
         wrmsr
 %endmacro
 
+extern use_xsave
 
-;; XXX - stick with fx until we can choose according to capabilities -
-;; otherwise existing stuff breaks with ops, etc.
 %macro load_extended_registers 1
-;        mov edx, 0xffffffff
-;        mov eax, edx
+        mov al, [use_xsave]
+        test al, al
+        jnz %%xs
         fxrstor [%1+FRAME_EXTENDED_SAVE*8]
+        jmp %%out
+%%xs:
+        mov edx, 0xffffffff
+        mov eax, edx
+        xrstor [%1+FRAME_EXTENDED_SAVE*8]
+%%out:
 %endmacro
         
 %macro save_extended_registers 1
-;        mov edx, 0xffffffff
-;        mov eax, edx
+        mov al, [use_xsave]
+        test al, al
+        jnz %%xs
         fxsave [%1+FRAME_EXTENDED_SAVE*8]  ; we wouldn't have to do this if we could guarantee no other user thread ran before us
+        jmp %%out
+%%xs:
+        mov edx, 0xffffffff
+        mov eax, edx
+        xsave [%1+FRAME_EXTENDED_SAVE*8]
+%%out:
 %endmacro
 
         
diff --git a/src/x86_64/kernel_machine.c b/src/x86_64/kernel_machine.c
index 628dbfc1a..9fd400bf1 100644
--- a/src/x86_64/kernel_machine.c
+++ b/src/x86_64/kernel_machine.c
@@ -36,7 +36,7 @@ heap allocate_tagged_region(kernel_heaps kh, u64 tag)
 void clone_frame_pstate(context dest, context src)
 {
     runtime_memcpy(dest, src, sizeof(u64) * (FRAME_N_PSTATE + 1));
-    runtime_memcpy(dest + FRAME_EXTENDED_SAVE, src + FRAME_EXTENDED_SAVE, extended_frame_size());
+    runtime_memcpy(dest + FRAME_EXTENDED_SAVE, src + FRAME_EXTENDED_SAVE, extended_frame_size);
 }
 
 void init_cpuinfo_machine(cpuinfo ci, heap backed)
diff --git a/src/x86_64/kernel_machine.h b/src/x86_64/kernel_machine.h
index d33254164..d35bb0ec1 100644
--- a/src/x86_64/kernel_machine.h
+++ b/src/x86_64/kernel_machine.h
@@ -110,6 +110,7 @@ void start_cpu(int index);
 void allocate_apboot(heap stackheap, void (*ap_entry)());
 void deallocate_apboot(heap stackheap);
 void install_idt(void);
+void init_cpu_features();
 
 #define IST_EXCEPTION 1
 #define IST_INTERRUPT 2
@@ -165,6 +166,16 @@ static inline void cpuid(u32 fn, u32 ecx, u32 * v)
     asm volatile("cpuid" : "=a" (v[0]), "=b" (v[1]), "=c" (v[2]), "=d" (v[3]) : "0" (fn), "2" (ecx));
 }
 
+static inline void xsetbv(u32 ecx, u32 eax, u32 edx)
+{
+    asm volatile("xsetbv" : : "a" (eax), "d" (edx), "c" (ecx));
+}
+
+static inline void xgetbv(u32 ecx, u32 *eax, u32 *edx)
+{
+    asm volatile("xgetbv" : "=a" (*eax), "=d" (*edx) : "c" (ecx));
+}
+
 /* syscall entry */
 
 static inline void set_syscall_handler(void *syscall_entry)
@@ -239,20 +250,10 @@ static inline cpuinfo current_cpu(void)
     return (cpuinfo)pointer_from_u64(addr);
 }
 
-static inline u64 extended_frame_size(void)
-{
-#if 0
-    u32 v[4];
-    cpuid(0xd, 0, v);
-    return v[1];
-#else
-    return 512;                 /* XXX fx only right now */
-#endif
-}
-
+extern u64 extended_frame_size;
 static inline u64 total_frame_size(void)
 {
-    return FRAME_EXTENDED_SAVE * sizeof(u64) + extended_frame_size();
+    return FRAME_EXTENDED_SAVE * sizeof(u64) + extended_frame_size;
 }
 
 static inline void frame_enable_interrupts(context f)
diff --git a/src/x86_64/mp.c b/src/x86_64/mp.c
index d51414818..e84f1a23c 100644
--- a/src/x86_64/mp.c
+++ b/src/x86_64/mp.c
@@ -22,16 +22,45 @@ static void (*start_callback)();
 #define mp_debug_u64(x)
 #endif
 
-void cpu_init(int cpu)
+#define CPUID_XSAVE (1<<26)
+#define CPUID_AVX (1<<28)
+
+#define XCR0_SSE (1<<1)
+#define XCR0_AVX (1<<2)
+u8 use_xsave;
+u64 extended_frame_size = 512;
+
+void init_cpu_features()
 {
     u64 cr;
+    u32 v[4];
+
+    cpuid(1, 0, v);
+    if (v[2] & CPUID_XSAVE)
+        use_xsave = 1;
     mov_from_cr("cr4", cr);
     cr |= CR4_PGE | CR4_OSFXSR | CR4_OSXMMEXCPT;
+    if (use_xsave)
+        cr |= CR4_OSXSAVE;
     mov_to_cr("cr4", cr);
     mov_from_cr("cr0", cr);
     cr |= C0_MP | C0_WP;
     cr &= ~C0_EM;
     mov_to_cr("cr0", cr);
+    if (use_xsave) {
+        if (v[2] & CPUID_AVX) {
+            xgetbv(0, &v[0], &v[1]);
+            v[0] |= XCR0_SSE | XCR0_AVX;
+            xsetbv(0, v[0], v[1]);
+        }
+        cpuid(0xd, 0, v);
+        extended_frame_size = v[1];
+    }
+}
+
+void cpu_init(int cpu)
+{
+    init_cpu_features();
     u64 addr = u64_from_pointer(cpuinfo_from_id(cpu));
     write_msr(KERNEL_GS_MSR, 0); /* clear user GS */
     write_msr(GS_MSR, addr);