[thirdparty/openssl.git] / crypto / x86cpuid.pl

#!/usr/bin/env perl

$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
push(@INC, "${dir}perlasm", "perlasm");
require "x86asm.pl";

&asm_init($ARGV[0],"x86cpuid");

for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }

&function_begin("OPENSSL_ia32_cpuid");
	&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"));
	&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);
	&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"));
	
&set_label("intel");
	&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

&set_label("nocacheinfo");
	&mov	("eax",1);
	&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
&set_label("notintel");
	&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

&set_label("generic");
	&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"));
&set_label("clear_xmm");
	&and	("ebp",0xfdfffffd);	# clear AESNI and PCLMULQDQ bits
	&and	("esi",0xfeffffff);	# clear FXSR
&set_label("clear_avx");
	&and	("ebp",0xefffe7ff);	# clear AVX, FMA and AMD XOP bits
&set_label("done");
	&mov	("eax","esi");
	&mov	("edx","ebp");
&set_label("nocpuid");
&function_end("OPENSSL_ia32_cpuid");

&external_label("OPENSSL_ia32cap_P");

&function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	&xor	("eax","eax");
	&xor	("edx","edx");
	&picmeup("ecx","OPENSSL_ia32cap_P");
	&bt	(&DWP(0,"ecx"),4);
	&jnc	(&label("notsc"));
	&rdtsc	();
&set_label("notsc");
	&ret	();
&function_end_B("OPENSSL_rdtsc");

# 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...
&function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	&picmeup("ecx","OPENSSL_ia32cap_P");
	&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	();

&set_label("nohalt");
	&xor	("eax","eax");
	&xor	("edx","edx");
	&ret	();
&function_end_B("OPENSSL_instrument_halt");

# 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.

&function_begin_B("OPENSSL_far_spin");
	&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);
&set_label("spin");
	&inc	("eax");
	&cmp	("edx",&DWP(0,"ecx"));
	&je	(&label("spin"));

	&data_word (0x1f909090);	# pop	%ds
	&ret	();

&set_label("nospin");
	&xor	("eax","eax");
	&xor	("edx","edx");
	&ret	();
&function_end_B("OPENSSL_far_spin");

&function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	&xor	("eax","eax");
	&xor	("edx","edx");
	&picmeup("ecx","OPENSSL_ia32cap_P");
	&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");
	&set_label("no_sse2");
	}
	# just a bunch of fldz to zap the fp/mm bank followed by finit...
	&data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
&set_label("no_x87");
	&lea	("eax",&DWP(4,"esp"));
	&ret	();
&function_end_B("OPENSSL_wipe_cpu");

&function_begin_B("OPENSSL_atomic_add");
	&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"));
&set_label("spin");
	&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	();
&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 $max arguments]);
#endif
#
# (*)	it's designed to work even for __fastcall if number of
#	arguments is 1 or 2!
&function_begin_B("OPENSSL_indirect_call");
	{
	my ($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");
	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");

&function_begin_B("OPENSSL_cleanse");
	&mov	("edx",&wparam(0));
	&mov	("ecx",&wparam(1));
	&xor	("eax","eax");
	&cmp	("ecx",7);
	&jae	(&label("lot"));
	&cmp	("ecx",0);
	&je	(&label("ret"));
&set_label("little");
	&mov	(&BP(0,"edx"),"al");
	&sub	("ecx",1);
	&lea	("edx",&DWP(1,"edx"));
	&jnz	(&label("little"));
&set_label("ret");
	&ret	();

&set_label("lot",16);
	&test	("edx",3);
	&jz	(&label("aligned"));
	&mov	(&BP(0,"edx"),"al");
	&lea	("ecx",&DWP(-1,"ecx"));
	&lea	("edx",&DWP(1,"edx"));
	&jmp	(&label("lot"));
&set_label("aligned");
	&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	();
&function_end_B("OPENSSL_cleanse");

{
my $lasttick = "esi";
my $lastdiff = "ebx";
my $out = "edi";
my $cnt = "ecx";
my $max = "ebp";

&function_begin("OPENSSL_instrument_bus");
    &mov	("eax",0);
    if ($sse2) {
	&picmeup("edx","OPENSSL_ia32cap_P");
	&bt	(&DWP(0,"edx"),4);
	&jnc	(&label("nogo"));	# no TSC
	&bt	(&DWP(0,"edx"),19);
	&jnc	(&label("nogo"));	# no CLFLUSH

	&mov	($out,&wparam(0));	# load arguments
	&mov	($cnt,&wparam(1));

	# collect 1st tick
	&rdtsc	();
	&mov	($lasttick,"eax");	# lasttick = tick
	&mov	($lastdiff,0);		# lastdiff = 0
	&clflush(&DWP(0,$out));
	&data_byte(0xf0);		# lock
	&add	(&DWP(0,$out),$lastdiff);
	&jmp	(&label("loop"));

&set_label("loop",16);
	&rdtsc	();
	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
	&sub	("eax",$lasttick);	# diff
	&mov	($lasttick,"edx");	# lasttick = tick
	&mov	($lastdiff,"eax");	# lastdiff = diff
	&clflush(&DWP(0,$out));
	&data_byte(0xf0);		# lock
	&add	(&DWP(0,$out),"eax");	# accumulate diff
	&lea	($out,&DWP(4,$out));	# ++$out
	&sub	($cnt,1);		# --$cnt
	&jnz	(&label("loop"));

	&mov	("eax",&wparam(1));
&set_label("nogo");
    }
&function_end("OPENSSL_instrument_bus");

&function_begin("OPENSSL_instrument_bus2");
    &mov	("eax",0);
    if ($sse2) {
	&picmeup("edx","OPENSSL_ia32cap_P");
	&bt	(&DWP(0,"edx"),4);
	&jnc	(&label("nogo"));	# no TSC
	&bt	(&DWP(0,"edx"),19);
	&jnc	(&label("nogo"));	# no CLFLUSH

	&mov	($out,&wparam(0));	# load arguments
	&mov	($cnt,&wparam(1));
	&mov	($max,&wparam(2));

	&rdtsc	();			# collect 1st tick
	&mov	($lasttick,"eax");	# lasttick = tick
	&mov	($lastdiff,0);		# lastdiff = 0

	&clflush(&DWP(0,$out));
	&data_byte(0xf0);		# lock
	&add	(&DWP(0,$out),$lastdiff);

	&rdtsc	();			# collect 1st diff
	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
	&sub	("eax",$lasttick);	# diff
	&mov	($lasttick,"edx");	# lasttick = tick
	&mov	($lastdiff,"eax");	# lastdiff = diff
	&jmp	(&label("loop2"));

&set_label("loop2",16);
	&clflush(&DWP(0,$out));
	&data_byte(0xf0);		# lock
	&add	(&DWP(0,$out),"eax");	# accumulate diff

	&sub	($max,1);
	&jz	(&label("done2"));

	&rdtsc	();
	&mov	("edx","eax");		# put aside tick (yes, I neglect edx)
	&sub	("eax",$lasttick);	# diff
	&mov	($lasttick,"edx");	# lasttick = tick
	&cmp	("eax",$lastdiff);
	&mov	($lastdiff,"eax");	# lastdiff = diff
	&mov	("edx",0);
	&setne	("dl");
	&sub	($cnt,"edx");		# conditional --$cnt
	&lea	($out,&DWP(0,$out,"edx",4));	# conditional ++$out
	&jnz	(&label("loop2"));

&set_label("done2");
	&mov	("eax",&wparam(1));
	&sub	("eax",$cnt);
&set_label("nogo");
    }
&function_end("OPENSSL_instrument_bus2");
}

&function_begin_B("OPENSSL_ia32_rdrand");
	&mov	("ecx",8);
&set_label("loop");
	&rdrand	("eax");
	&jc	(&label("break"));
	&loop	(&label("loop"));
&set_label("break");
	&cmp	("eax",0);
	&cmove	("eax","ecx");
	&ret	();
&function_end_B("OPENSSL_ia32_rdrand");

&initseg("OPENSSL_cpuid_setup");

&asm_finish();
Commit	Line	Data
14e21f86 AP	1	#!/usr/bin/env perl
14e21f86 AP	2
4d1f3f7a DSH	3	$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
4d1f3f7a DSH	4	push(@INC, "${dir}perlasm", "perlasm");
14e21f86 AP	5	require "x86asm.pl";
	6
	7	&asm_init($ARGV[0],"x86cpuid");
	8
c5022659 AP	9	for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
c5022659 AP	10
14e21f86 AP	11	&function_begin("OPENSSL_ia32_cpuid");
	12	&xor ("edx","edx");
	13	&pushf ();
	14	&pop ("eax");
	15	&mov ("ecx","eax");
	16	&xor ("eax",1<<21);
	17	&push ("eax");
	18	&popf ();
	19	&pushf ();
	20	&pop ("eax");
	21	&xor ("ecx","eax");
162f677d	22	&xor ("eax","eax");
0f2ece87 AP	23	&bt ("ecx",21);
0f2ece87 AP	24	&jnc (&label("nocpuid"));
162f677d	25	&cpuid ();
761393bb AP	26	&mov ("edi","eax"); # max value for standard query level
761393bb AP	27
162f677d AP	28	&xor ("eax","eax");
	29	&cmp ("ebx",0x756e6547); # "Genu"
	30	&setne (&LB("eax"));
	31	&mov ("ebp","eax");
	32	&cmp ("edx",0x49656e69); # "ineI"
	33	&setne (&LB("eax"));
	34	&or ("ebp","eax");
	35	&cmp ("ecx",0x6c65746e); # "ntel"
	36	&setne (&LB("eax"));
58672313	37	&or ("ebp","eax"); # 0 indicates Intel CPU
58672313 AP	38	&jz (&label("intel"));
	39
	40	&cmp ("ebx",0x68747541); # "Auth"
	41	&setne (&LB("eax"));
	42	&mov ("esi","eax");
	43	&cmp ("edx",0x69746E65); # "enti"
	44	&setne (&LB("eax"));
	45	&or ("esi","eax");
	46	&cmp ("ecx",0x444D4163); # "cAMD"
	47	&setne (&LB("eax"));
	48	&or ("esi","eax"); # 0 indicates AMD CPU
	49	&jnz (&label("intel"));
	50
761393bb	51	# AMD specific
58672313 AP	52	&mov ("eax",0x80000000);
58672313 AP	53	&cpuid ();
b9064221 AP	54	&cmp ("eax",0x80000001);
	55	&jb (&label("intel"));
	56	&mov ("esi","eax");
	57	&mov ("eax",0x80000001);
	58	&cpuid ();
	59	&or ("ebp","ecx");
	60	&and ("ebp",1<<11\|1); # isolate XOP bit
	61	&cmp ("esi",0x80000008);
58672313 AP	62	&jb (&label("intel"));
	63
	64	&mov ("eax",0x80000008);
	65	&cpuid ();
	66	&movz ("esi",&LB("ecx")); # number of cores - 1
	67	&inc ("esi"); # number of cores
	68
761393bb AP	69	&mov ("eax",1);
	70	&cpuid ();
	71	&bt ("edx",28);
b9064221	72	&jnc (&label("generic"));
761393bb AP	73	&shr ("ebx",16);
	74	&and ("ebx",0xff);
	75	&cmp ("ebx","esi");
b9064221	76	&ja (&label("generic"));
761393bb	77	&and ("edx",0xefffffff); # clear hyper-threading bit
b9064221	78	&jmp (&label("generic"));
761393bb	79
58672313	80	&set_label("intel");
761393bb AP	81	&cmp ("edi",4);
	82	&mov ("edi",-1);
	83	&jb (&label("nocacheinfo"));
	84
	85	&mov ("eax",4);
	86	&mov ("ecx",0); # query L1D
	87	&cpuid ();
	88	&mov ("edi","eax");
	89	&shr ("edi",14);
	90	&and ("edi",0xfff); # number of cores -1 per L1D
	91
	92	&set_label("nocacheinfo");
14e21f86 AP	93	&mov ("eax",1);
14e21f86 AP	94	&cpuid ();
18f5603c	95	&and ("edx",0xbfefffff); # force reserved bits #20, #30 to 0
162f677d	96	&cmp ("ebp",0);
4bb90087 AP	97	&jne (&label("notintel"));
4bb90087 AP	98	&or ("edx",1<<30); # set reserved bit#30 on Intel CPUs
3df2eff4 AP	99	&and (&HB("eax"),15); # familiy ID
3df2eff4 AP	100	&cmp (&HB("eax"),15); # P4?
4bb90087 AP	101	&jne (&label("notintel"));
	102	&or ("edx",1<<20); # set reserved bit#20 to engage RC4_CHAR
	103	&set_label("notintel");
3df2eff4	104	&bt ("edx",28); # test hyper-threading bit
b9064221	105	&jnc (&label("generic"));
761393bb AP	106	&and ("edx",0xefffffff);
761393bb AP	107	&cmp ("edi",0);
b9064221	108	&je (&label("generic"));
761393bb AP	109
761393bb AP	110	&or ("edx",0x10000000);
162f677d	111	&shr ("ebx",16);
761393bb	112	&cmp (&LB("ebx"),1);
b9064221	113	&ja (&label("generic"));
b1e8b4e6	114	&and ("edx",0xefffffff); # clear hyper-threading bit if not
b9064221 AP	115
	116	&set_label("generic");
	117	&and ("ebp",1<<11); # isolate AMD XOP flag
18f5603c	118	&and ("ecx",0xfffff7ff); # force 11th bit to 0
b9064221 AP	119	&mov ("esi","edx");
	120	&or ("ebp","ecx"); # merge AMD XOP flag
	121
b9064221	122	&bt ("ecx",27); # check OSXSAVE bit
e879dd43	123	&jnc (&label("clear_avx"));
b9064221 AP	124	&xor ("ecx","ecx");
	125	&data_byte(0x0f,0x01,0xd0); # xgetbv
	126	&and ("eax",6);
	127	&cmp ("eax",6);
	128	&je (&label("done"));
	129	&cmp ("eax",2);
	130	&je (&label("clear_avx"));
	131	&set_label("clear_xmm");
18f5603c AP	132	&and ("ebp",0xfdfffffd); # clear AESNI and PCLMULQDQ bits
18f5603c AP	133	&and ("esi",0xfeffffff); # clear FXSR
b9064221	134	&set_label("clear_avx");
18f5603c	135	&and ("ebp",0xefffe7ff); # clear AVX, FMA and AMD XOP bits
3df2eff4	136	&set_label("done");
b9064221 AP	137	&mov ("eax","esi");
b9064221 AP	138	&mov ("edx","ebp");
0f2ece87	139	&set_label("nocpuid");
14e21f86 AP	140	&function_end("OPENSSL_ia32_cpuid");
14e21f86 AP	141
2b247cf8	142	&external_label("OPENSSL_ia32cap_P");
14e21f86	143
2b247cf8	144	&function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
14e21f86 AP	145	&xor ("eax","eax");
14e21f86 AP	146	&xor ("edx","edx");
2b247cf8	147	&picmeup("ecx","OPENSSL_ia32cap_P");
14e21f86 AP	148	&bt (&DWP(0,"ecx"),4);
	149	&jnc (&label("notsc"));
	150	&rdtsc ();
	151	&set_label("notsc");
	152	&ret ();
	153	&function_end_B("OPENSSL_rdtsc");
	154
c85c5c40 AP	155	# This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
	156	# but it's safe to call it on any [supported] 32-bit platform...
	157	# Just check for [non-]zero return value...
	158	&function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	159	&picmeup("ecx","OPENSSL_ia32cap_P");
	160	&bt (&DWP(0,"ecx"),4);
	161	&jnc (&label("nohalt")); # no TSC
	162
	163	&data_word(0x9058900e); # push %cs; pop %eax
	164	&and ("eax",3);
	165	&jnz (&label("nohalt")); # not enough privileges
	166
	167	&pushf ();
f9c5e5d9	168	&pop ("eax");
c85c5c40 AP	169	&bt ("eax",9);
	170	&jnc (&label("nohalt")); # interrupts are disabled
	171
	172	&rdtsc ();
	173	&push ("edx");
	174	&push ("eax");
	175	&halt ();
	176	&rdtsc ();
	177
	178	&sub ("eax",&DWP(0,"esp"));
	179	&sbb ("edx",&DWP(4,"esp"));
	180	&add ("esp",8);
	181	&ret ();
	182
	183	&set_label("nohalt");
	184	&xor ("eax","eax");
	185	&xor ("edx","edx");
	186	&ret ();
	187	&function_end_B("OPENSSL_instrument_halt");
	188
cee73df3 AP	189	# Essentially there is only one use for this function. Under DJGPP:
	190	#
	191	# #include <go32.h>
	192	# ...
	193	# i=OPENSSL_far_spin(_dos_ds,0x46c);
	194	# ...
	195	# to obtain the number of spins till closest timer interrupt.
	196
	197	&function_begin_B("OPENSSL_far_spin");
	198	&pushf ();
8d1b199d	199	&pop ("eax");
cee73df3 AP	200	&bt ("eax",9);
	201	&jnc (&label("nospin")); # interrupts are disabled
	202
	203	&mov ("eax",&DWP(4,"esp"));
	204	&mov ("ecx",&DWP(8,"esp"));
	205	&data_word (0x90d88e1e); # push %ds, mov %eax,%ds
	206	&xor ("eax","eax");
	207	&mov ("edx",&DWP(0,"ecx"));
	208	&jmp (&label("spin"));
	209
	210	&align (16);
	211	&set_label("spin");
	212	&inc ("eax");
	213	&cmp ("edx",&DWP(0,"ecx"));
	214	&je (&label("spin"));
	215
	216	&data_word (0x1f909090); # pop %ds
	217	&ret ();
	218
	219	&set_label("nospin");
	220	&xor ("eax","eax");
	221	&xor ("edx","edx");
	222	&ret ();
	223	&function_end_B("OPENSSL_far_spin");
	224
	225	&function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	226	&xor ("eax","eax");
	227	&xor ("edx","edx");
	228	&picmeup("ecx","OPENSSL_ia32cap_P");
	229	&mov ("ecx",&DWP(0,"ecx"));
	230	&bt (&DWP(0,"ecx"),1);
	231	&jnc (&label("no_x87"));
c5022659	232	if ($sse2) {
b9064221 AP	233	&and ("ecx",1<<26\|1<<24); # check SSE2 and FXSR bits
	234	&cmp ("ecx",1<<26\|1<<24);
	235	&jne (&label("no_sse2"));
c5022659 AP	236	&pxor ("xmm0","xmm0");
	237	&pxor ("xmm1","xmm1");
	238	&pxor ("xmm2","xmm2");
	239	&pxor ("xmm3","xmm3");
	240	&pxor ("xmm4","xmm4");
	241	&pxor ("xmm5","xmm5");
	242	&pxor ("xmm6","xmm6");
	243	&pxor ("xmm7","xmm7");
	244	&set_label("no_sse2");
	245	}
c49a0aa0 AP	246	# just a bunch of fldz to zap the fp/mm bank followed by finit...
c49a0aa0 AP	247	&data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
cee73df3 AP	248	&set_label("no_x87");
	249	&lea ("eax",&DWP(4,"esp"));
	250	&ret ();
	251	&function_end_B("OPENSSL_wipe_cpu");
	252
	253	&function_begin_B("OPENSSL_atomic_add");
	254	&mov ("edx",&DWP(4,"esp")); # fetch the pointer, 1st arg
	255	&mov ("ecx",&DWP(8,"esp")); # fetch the increment, 2nd arg
	256	&push ("ebx");
	257	&nop ();
	258	&mov ("eax",&DWP(0,"edx"));
	259	&set_label("spin");
	260	&lea ("ebx",&DWP(0,"eax","ecx"));
	261	&nop ();
	262	&data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx) # %eax is envolved and is always reloaded
	263	&jne (&label("spin"));
	264	&mov ("eax","ebx"); # OpenSSL expects the new value
	265	&pop ("ebx");
	266	&ret ();
	267	&function_end_B("OPENSSL_atomic_add");
	268
061bebc0 AP	269	# This function can become handy under Win32 in situations when
	270	# we don't know which calling convention, __stdcall or __cdecl(*),
	271	# indirect callee is using. In C it can be deployed as
	272	#
	273	#ifdef OPENSSL_CPUID_OBJ
	274	# type OPENSSL_indirect_call(void *f,...);
	275	# ...
	276	# OPENSSL_indirect_call(func,[up to $max arguments]);
	277	#endif
	278	#
	279	# (*) it's designed to work even for __fastcall if number of
	280	# arguments is 1 or 2!
	281	&function_begin_B("OPENSSL_indirect_call");
	282	{
f9c5e5d9	283	my ($max,$i)=(7,); # $max has to be chosen as 4*n-1
061bebc0 AP	284	# in order to preserve eventual
	285	# stack alignment
	286	&push ("ebp");
	287	&mov ("ebp","esp");
	288	&sub ("esp",$max*4);
	289	&mov ("ecx",&DWP(12,"ebp"));
	290	&mov (&DWP(0,"esp"),"ecx");
	291	&mov ("edx",&DWP(16,"ebp"));
	292	&mov (&DWP(4,"esp"),"edx");
	293	for($i=2;$i<$max;$i++)
	294	{
	295	# Some copies will be redundant/bogus...
	296	&mov ("eax",&DWP(12+$i*4,"ebp"));
	297	&mov (&DWP(0+$i*4,"esp"),"eax");
	298	}
3ebf898e	299	&call_ptr (&DWP(8,"ebp"));# make the call...
061bebc0 AP	300	&mov ("esp","ebp"); # ... and just restore the stack pointer
	301	# without paying attention to what we called,
	302	# (__cdecl func) or (__stdcall one).
	303	&pop ("ebp");
	304	&ret ();
	305	}
	306	&function_end_B("OPENSSL_indirect_call");
	307
b2dba9bf AP	308	&function_begin_B("OPENSSL_cleanse");
	309	&mov ("edx",&wparam(0));
	310	&mov ("ecx",&wparam(1));
	311	&xor ("eax","eax");
	312	&cmp ("ecx",7);
	313	&jae (&label("lot"));
7676eebf AP	314	&cmp ("ecx",0);
7676eebf AP	315	&je (&label("ret"));
b2dba9bf AP	316	&set_label("little");
	317	&mov (&BP(0,"edx"),"al");
	318	&sub ("ecx",1);
	319	&lea ("edx",&DWP(1,"edx"));
	320	&jnz (&label("little"));
7676eebf	321	&set_label("ret");
b2dba9bf AP	322	&ret ();
	323
	324	&set_label("lot",16);
	325	&test ("edx",3);
	326	&jz (&label("aligned"));
	327	&mov (&BP(0,"edx"),"al");
	328	&lea ("ecx",&DWP(-1,"ecx"));
	329	&lea ("edx",&DWP(1,"edx"));
	330	&jmp (&label("lot"));
	331	&set_label("aligned");
	332	&mov (&DWP(0,"edx"),"eax");
	333	&lea ("ecx",&DWP(-4,"ecx"));
	334	&test ("ecx",-4);
	335	&lea ("edx",&DWP(4,"edx"));
	336	&jnz (&label("aligned"));
	337	&cmp ("ecx",0);
	338	&jne (&label("little"));
	339	&ret ();
	340	&function_end_B("OPENSSL_cleanse");
	341
5fabb88a AP	342	{
	343	my $lasttick = "esi";
	344	my $lastdiff = "ebx";
	345	my $out = "edi";
	346	my $cnt = "ecx";
	347	my $max = "ebp";
	348
	349	&function_begin("OPENSSL_instrument_bus");
	350	&mov ("eax",0);
	351	if ($sse2) {
	352	&picmeup("edx","OPENSSL_ia32cap_P");
	353	&bt (&DWP(0,"edx"),4);
	354	&jnc (&label("nogo")); # no TSC
	355	&bt (&DWP(0,"edx"),19);
	356	&jnc (&label("nogo")); # no CLFLUSH
	357
	358	&mov ($out,&wparam(0)); # load arguments
	359	&mov ($cnt,&wparam(1));
	360
	361	# collect 1st tick
	362	&rdtsc ();
	363	&mov ($lasttick,"eax"); # lasttick = tick
	364	&mov ($lastdiff,0); # lastdiff = 0
	365	&clflush(&DWP(0,$out));
b9064221	366	&data_byte(0xf0); # lock
5fabb88a AP	367	&add (&DWP(0,$out),$lastdiff);
	368	&jmp (&label("loop"));
	369
	370	&set_label("loop",16);
	371	&rdtsc ();
	372	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	373	&sub ("eax",$lasttick); # diff
	374	&mov ($lasttick,"edx"); # lasttick = tick
	375	&mov ($lastdiff,"eax"); # lastdiff = diff
	376	&clflush(&DWP(0,$out));
b9064221	377	&data_byte(0xf0); # lock
5fabb88a AP	378	&add (&DWP(0,$out),"eax"); # accumulate diff
	379	&lea ($out,&DWP(4,$out)); # ++$out
	380	&sub ($cnt,1); # --$cnt
	381	&jnz (&label("loop"));
	382
	383	&mov ("eax",&wparam(1));
	384	&set_label("nogo");
	385	}
	386	&function_end("OPENSSL_instrument_bus");
	387
	388	&function_begin("OPENSSL_instrument_bus2");
	389	&mov ("eax",0);
	390	if ($sse2) {
	391	&picmeup("edx","OPENSSL_ia32cap_P");
	392	&bt (&DWP(0,"edx"),4);
	393	&jnc (&label("nogo")); # no TSC
	394	&bt (&DWP(0,"edx"),19);
	395	&jnc (&label("nogo")); # no CLFLUSH
	396
	397	&mov ($out,&wparam(0)); # load arguments
	398	&mov ($cnt,&wparam(1));
	399	&mov ($max,&wparam(2));
	400
	401	&rdtsc (); # collect 1st tick
	402	&mov ($lasttick,"eax"); # lasttick = tick
	403	&mov ($lastdiff,0); # lastdiff = 0
	404
	405	&clflush(&DWP(0,$out));
b9064221	406	&data_byte(0xf0); # lock
5fabb88a AP	407	&add (&DWP(0,$out),$lastdiff);
	408
	409	&rdtsc (); # collect 1st diff
	410	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	411	&sub ("eax",$lasttick); # diff
	412	&mov ($lasttick,"edx"); # lasttick = tick
	413	&mov ($lastdiff,"eax"); # lastdiff = diff
	414	&jmp (&label("loop2"));
	415
	416	&set_label("loop2",16);
	417	&clflush(&DWP(0,$out));
b9064221	418	&data_byte(0xf0); # lock
5fabb88a AP	419	&add (&DWP(0,$out),"eax"); # accumulate diff
	420
	421	&sub ($max,1);
	422	&jz (&label("done2"));
	423
	424	&rdtsc ();
	425	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	426	&sub ("eax",$lasttick); # diff
	427	&mov ($lasttick,"edx"); # lasttick = tick
	428	&cmp ("eax",$lastdiff);
	429	&mov ($lastdiff,"eax"); # lastdiff = diff
	430	&mov ("edx",0);
	431	&setne ("dl");
	432	&sub ($cnt,"edx"); # conditional --$cnt
	433	&lea ($out,&DWP(0,$out,"edx",4)); # conditional ++$out
	434	&jnz (&label("loop2"));
	435
	436	&set_label("done2");
	437	&mov ("eax",&wparam(1));
	438	&sub ("eax",$cnt);
	439	&set_label("nogo");
	440	}
	441	&function_end("OPENSSL_instrument_bus2");
	442	}
	443
301799b8 AP	444	&function_begin_B("OPENSSL_ia32_rdrand");
	445	&mov ("ecx",8);
	446	&set_label("loop");
	447	&rdrand ("eax");
	448	&jc (&label("break"));
	449	&loop (&label("loop"));
	450	&set_label("break");
	451	&cmp ("eax",0);
	452	&cmove ("eax","ecx");
	453	&ret ();
	454	&function_end_B("OPENSSL_ia32_rdrand");
	455
2b247cf8	456	&initseg("OPENSSL_cpuid_setup");
14e21f86 AP	457
14e21f86 AP	458	&asm_finish();