Update x86cpuid.pl to correctly detect shared cache and to support new

[openssl.git] / crypto / x86cpuid.pl

diff --git a/crypto/x86cpuid.pl b/crypto/x86cpuid.pl
index 4a58651f0f9d4326971d7ffaf468786b2647c36a..7d924a60b77d78c9352c50bbe89795332a78ee56 100644 (file)
--- a/crypto/x86cpuid.pl
+++ b/crypto/x86cpuid.pl
@@ -20,12 +20,36 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
        &xor    ("ecx","eax");
        &bt     ("ecx",21);
        &jnc    (&label("nocpuid"));
+       &xor    ("eax","eax");
+       &cpuid  ();
+       &xor    ("eax","eax");
+       &cmp    ("ebx",0x756e6547);     # "Genu"
+       &setne  (&LB("eax"));
+       &mov    ("ebp","eax");
+       &cmp    ("edx",0x49656e69);     # "ineI"
+       &setne  (&LB("eax"));
+       &or     ("ebp","eax");
+       &cmp    ("ecx",0x6c65746e);     # "ntel"
+       &setne  (&LB("eax"));
+       &or     ("ebp","eax");
        &mov    ("eax",1);
        &cpuid  ();
+       &bt     ("edx",28);             # test hyper-threading bit
+       &jnc    (&label("nocpuid"));
+       &cmp    ("ebp",0);
+       &jne    (&label("notintel"));
+       &or     ("edx",1<<20);          # use reserved bit to engage RC4_CHAR
+&set_label("notintel");
+       &shr    ("ebx",16);
+       &cmp    (&LB("ebx"),1);         # see if cache is shared(*)
+       &ja     (&label("nocpuid"));
+       &and    ("edx",~(1<<28));       # clear hyper-threading bit if not
 &set_label("nocpuid");
        &mov    ("eax","edx");
        &mov    ("edx","ecx");
 &function_end("OPENSSL_ia32_cpuid");
+# (*)  on Core2 this value is set to 2 denoting the fact that L2
+#      cache is shared between cores.
 
 &external_label("OPENSSL_ia32cap_P");
 
@@ -153,6 +177,45 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
        &ret    ();
 &function_end_B("OPENSSL_atomic_add");
 
+# This function can become handy under Win32 in situations when
+# we don't know which calling convention, __stdcall or __cdecl(*),
+# indirect callee is using. In C it can be deployed as
+#
+#ifdef OPENSSL_CPUID_OBJ
+#      type OPENSSL_indirect_call(void *f,...);
+#      ...
+#      OPENSSL_indirect_call(func,[up to $max arguments]);
+#endif
+#
+# (*)  it's designed to work even for __fastcall if number of
+#      arguments is 1 or 2!
+&function_begin_B("OPENSSL_indirect_call");
+       {
+       my $i,$max=7;           # $max has to be chosen as 4*n-1
+                               # in order to preserve eventual
+                               # stack alignment
+       &push   ("ebp");
+       &mov    ("ebp","esp");
+       &sub    ("esp",$max*4);
+       &mov    ("ecx",&DWP(12,"ebp"));
+       &mov    (&DWP(0,"esp"),"ecx");
+       &mov    ("edx",&DWP(16,"ebp"));
+       &mov    (&DWP(4,"esp"),"edx");
+       for($i=2;$i<$max;$i++)
+               {
+               # Some copies will be redundant/bogus...
+               &mov    ("eax",&DWP(12+$i*4,"ebp"));
+               &mov    (&DWP(0+$i*4,"esp"),"eax");
+               }
+       &call_ptr       (&DWP(8,"ebp"));# make the call...
+       &mov    ("esp","ebp");  # ... and just restore the stack pointer
+                               # without paying attention to what we called,
+                               # (__cdecl *func) or (__stdcall *one).
+       &pop    ("ebp");
+       &ret    ();
+       }
+&function_end_B("OPENSSL_indirect_call");
+
 &initseg("OPENSSL_cpuid_setup");
 
 &asm_finish();