blob: 151493ab9aa875f164f1ce23c30ca7baa7cf3976 [file] [log] [blame]
#!/usr/bin/env perl
$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
push(@INC, "${dir}perlasm", "perlasm");
require "";
for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
&xor ("edx","edx");
&pushf ();
&pop ("eax");
&mov ("ecx","eax");
&xor ("eax",1<<21);
&push ("eax");
&popf ();
&pushf ();
&pop ("eax");
&xor ("ecx","eax");
&xor ("eax","eax");
&bt ("ecx",21);
&jnc (&label("nocpuid"));
&mov ("esi",&wparam(0));
&mov (&DWP(8,"esi"),"eax"); # clear 3rd word
&cpuid ();
&mov ("edi","eax"); # max value for standard query level
&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"); # 0 indicates Intel CPU
&jz (&label("intel"));
&cmp ("ebx",0x68747541); # "Auth"
&setne (&LB("eax"));
&mov ("esi","eax");
&cmp ("edx",0x69746E65); # "enti"
&setne (&LB("eax"));
&or ("esi","eax");
&cmp ("ecx",0x444D4163); # "cAMD"
&setne (&LB("eax"));
&or ("esi","eax"); # 0 indicates AMD CPU
&jnz (&label("intel"));
# AMD specific
&mov ("eax",0x80000000);
&cpuid ();
&cmp ("eax",0x80000001);
&jb (&label("intel"));
&mov ("esi","eax");
&mov ("eax",0x80000001);
&cpuid ();
&or ("ebp","ecx");
&and ("ebp",1<<11|1); # isolate XOP bit
&cmp ("esi",0x80000008);
&jb (&label("intel"));
&mov ("eax",0x80000008);
&cpuid ();
&movz ("esi",&LB("ecx")); # number of cores - 1
&inc ("esi"); # number of cores
&mov ("eax",1);
&xor ("ecx","ecx");
&cpuid ();
&bt ("edx",28);
&jnc (&label("generic"));
&shr ("ebx",16);
&and ("ebx",0xff);
&cmp ("ebx","esi");
&ja (&label("generic"));
&and ("edx",0xefffffff); # clear hyper-threading bit
&jmp (&label("generic"));
&cmp ("edi",7);
&jb (&label("cacheinfo"));
&mov ("esi",&wparam(0));
&mov ("eax",7);
&xor ("ecx","ecx");
&cpuid ();
&mov (&DWP(8,"esi"),"ebx");
&cmp ("edi",4);
&mov ("edi",-1);
&jb (&label("nocacheinfo"));
&mov ("eax",4);
&mov ("ecx",0); # query L1D
&cpuid ();
&mov ("edi","eax");
&shr ("edi",14);
&and ("edi",0xfff); # number of cores -1 per L1D
&mov ("eax",1);
&xor ("ecx","ecx");
&cpuid ();
&and ("edx",0xbfefffff); # force reserved bits #20, #30 to 0
&cmp ("ebp",0);
&jne (&label("notintel"));
&or ("edx",1<<30); # set reserved bit#30 on Intel CPUs
&and (&HB("eax"),15); # familiy ID
&cmp (&HB("eax"),15); # P4?
&jne (&label("notintel"));
&or ("edx",1<<20); # set reserved bit#20 to engage RC4_CHAR
&bt ("edx",28); # test hyper-threading bit
&jnc (&label("generic"));
&and ("edx",0xefffffff);
&cmp ("edi",0);
&je (&label("generic"));
&or ("edx",0x10000000);
&shr ("ebx",16);
&cmp (&LB("ebx"),1);
&ja (&label("generic"));
&and ("edx",0xefffffff); # clear hyper-threading bit if not
&and ("ebp",1<<11); # isolate AMD XOP flag
&and ("ecx",0xfffff7ff); # force 11th bit to 0
&mov ("esi","edx");
&or ("ebp","ecx"); # merge AMD XOP flag
&bt ("ecx",27); # check OSXSAVE bit
&jnc (&label("clear_avx"));
&xor ("ecx","ecx");
&data_byte(0x0f,0x01,0xd0); # xgetbv
&and ("eax",6);
&cmp ("eax",6);
&je (&label("done"));
&cmp ("eax",2);
&je (&label("clear_avx"));
&and ("ebp",0xfdfffffd); # clear AESNI and PCLMULQDQ bits
&and ("esi",0xfeffffff); # clear FXSR
&and ("ebp",0xefffe7ff); # clear AVX, FMA and AMD XOP bits
&mov ("edi",&wparam(0));
&and (&DWP(8,"edi"),0xffffffdf); # clear AVX2
&mov ("eax","esi");
&mov ("edx","ebp");
&xor ("eax","eax");
&xor ("edx","edx");
&bt (&DWP(0,"ecx"),4);
&jnc (&label("notsc"));
&rdtsc ();
&ret ();
# This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
# but it's safe to call it on any [supported] 32-bit platform...
# Just check for [non-]zero return value...
&bt (&DWP(0,"ecx"),4);
&jnc (&label("nohalt")); # no TSC
&data_word(0x9058900e); # push %cs; pop %eax
&and ("eax",3);
&jnz (&label("nohalt")); # not enough privileges
&pushf ();
&pop ("eax");
&bt ("eax",9);
&jnc (&label("nohalt")); # interrupts are disabled
&rdtsc ();
&push ("edx");
&push ("eax");
&halt ();
&rdtsc ();
&sub ("eax",&DWP(0,"esp"));
&sbb ("edx",&DWP(4,"esp"));
&add ("esp",8);
&ret ();
&xor ("eax","eax");
&xor ("edx","edx");
&ret ();
# Essentially there is only one use for this function. Under DJGPP:
# #include <go32.h>
# ...
# i=OPENSSL_far_spin(_dos_ds,0x46c);
# ...
# to obtain the number of spins till closest timer interrupt.
&pushf ();
&pop ("eax");
&bt ("eax",9);
&jnc (&label("nospin")); # interrupts are disabled
&mov ("eax",&DWP(4,"esp"));
&mov ("ecx",&DWP(8,"esp"));
&data_word (0x90d88e1e); # push %ds, mov %eax,%ds
&xor ("eax","eax");
&mov ("edx",&DWP(0,"ecx"));
&jmp (&label("spin"));
&align (16);
&inc ("eax");
&cmp ("edx",&DWP(0,"ecx"));
&je (&label("spin"));
&data_word (0x1f909090); # pop %ds
&ret ();
&xor ("eax","eax");
&xor ("edx","edx");
&ret ();
&xor ("eax","eax");
&xor ("edx","edx");
&mov ("ecx",&DWP(0,"ecx"));
&bt (&DWP(0,"ecx"),1);
&jnc (&label("no_x87"));
if ($sse2) {
&and ("ecx",1<<26|1<<24); # check SSE2 and FXSR bits
&cmp ("ecx",1<<26|1<<24);
&jne (&label("no_sse2"));
&pxor ("xmm0","xmm0");
&pxor ("xmm1","xmm1");
&pxor ("xmm2","xmm2");
&pxor ("xmm3","xmm3");
&pxor ("xmm4","xmm4");
&pxor ("xmm5","xmm5");
&pxor ("xmm6","xmm6");
&pxor ("xmm7","xmm7");
# just a bunch of fldz to zap the fp/mm bank followed by finit...
&lea ("eax",&DWP(4,"esp"));
&ret ();
&mov ("edx",&DWP(4,"esp")); # fetch the pointer, 1st arg
&mov ("ecx",&DWP(8,"esp")); # fetch the increment, 2nd arg
&push ("ebx");
&nop ();
&mov ("eax",&DWP(0,"edx"));
&lea ("ebx",&DWP(0,"eax","ecx"));
&nop ();
&data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx) # %eax is envolved and is always reloaded
&jne (&label("spin"));
&mov ("eax","ebx"); # OpenSSL expects the new value
&pop ("ebx");
&ret ();
# 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
# type OPENSSL_indirect_call(void *f,...);
# ...
# OPENSSL_indirect_call(func,[up to $max arguments]);
# (*) it's designed to work even for __fastcall if number of
# arguments is 1 or 2!
my ($max,$i)=(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");
# 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 ();
&mov ("edx",&wparam(0));
&mov ("ecx",&wparam(1));
&xor ("eax","eax");
&cmp ("ecx",7);
&jae (&label("lot"));
&cmp ("ecx",0);
&je (&label("ret"));
&mov (&BP(0,"edx"),"al");
&sub ("ecx",1);
&lea ("edx",&DWP(1,"edx"));
&jnz (&label("little"));
&ret ();
&test ("edx",3);
&jz (&label("aligned"));
&mov (&BP(0,"edx"),"al");
&lea ("ecx",&DWP(-1,"ecx"));
&lea ("edx",&DWP(1,"edx"));
&jmp (&label("lot"));
&mov (&DWP(0,"edx"),"eax");
&lea ("ecx",&DWP(-4,"ecx"));
&test ("ecx",-4);
&lea ("edx",&DWP(4,"edx"));
&jnz (&label("aligned"));
&cmp ("ecx",0);
&jne (&label("little"));
&ret ();
&mov ("ecx",8);
&rdrand ("eax");
&jc (&label("break"));
&loop (&label("loop"));
&cmp ("eax",0);
&cmove ("eax","ecx");
&ret ();