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

#! /usr/bin/env perl
# Copyright 2004-2020 The OpenSSL Project Authors. All Rights Reserved.
#
# Licensed under the Apache License 2.0 (the "License").  You may not use
# this file except in compliance with the License.  You can obtain a copy
# in the file LICENSE in the source distribution or at
# https://www.openssl.org/source/license.html

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

$output = pop and open STDOUT,">$output";

&asm_init($ARGV[0]);

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");
	&mov	("esi",&wparam(0));
	&mov	(&DWP(8,"esi"),"eax");	# clear extended feature flags
	&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);
	&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"));

&set_label("intel");
	&cmp	("edi",4);
	&mov	("esi",-1);
	&jb	(&label("nocacheinfo"));

	&mov	("eax",4);
	&mov	("ecx",0);		# query L1D
	&cpuid	();
	&mov	("esi","eax");
	&shr	("esi",14);
	&and	("esi",0xfff);		# number of cores -1 per L1D

&set_label("nocacheinfo");
	&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);	# family 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	("esi",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");		# %ebp:%esi is copy of %ecx:%edx
	&or	("ebp","ecx");		# merge AMD XOP flag

	&cmp	("edi",7);
	&mov	("edi",&wparam(0));
	&jb	(&label("no_extended_info"));
	&mov	("eax",7);
	&xor	("ecx","ecx");
	&cpuid	();
	&mov	(&DWP(8,"edi"),"ebx");	# save extended feature flag
&set_label("no_extended_info");

	&bt	("ebp",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
	&and	(&DWP(8,"edi"),0xffffffdf);	# clear AVX2
&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 involved and is always reloaded
	&jne	(&label("spin"));
	&mov	("eax","ebx");	# OpenSSL expects the new value
	&pop	("ebx");
	&ret	();
&function_end_B("OPENSSL_atomic_add");

&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");

&function_begin_B("CRYPTO_memcmp");
	&push	("esi");
	&push	("edi");
	&mov	("esi",&wparam(0));
	&mov	("edi",&wparam(1));
	&mov	("ecx",&wparam(2));
	&xor	("eax","eax");
	&xor	("edx","edx");
	&cmp	("ecx",0);
	&je	(&label("no_data"));
&set_label("loop");
	&mov	("dl",&BP(0,"esi"));
	&lea	("esi",&DWP(1,"esi"));
	&xor	("dl",&BP(0,"edi"));
	&lea	("edi",&DWP(1,"edi"));
	&or	("al","dl");
	&dec	("ecx");
	&jnz	(&label("loop"));
	&neg	("eax");
	&shr	("eax",31);
&set_label("no_data");
	&pop	("edi");
	&pop	("esi");
	&ret	();
&function_end_B("CRYPTO_memcmp");
{
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");
}

sub gen_random {
my $rdop = shift;
&function_begin_B("OPENSSL_ia32_${rdop}_bytes");
	&push	("edi");
	&push	("ebx");
	&xor	("eax","eax");		# return value
	&mov	("edi",&wparam(0));
	&mov	("ebx",&wparam(1));

	&cmp	("ebx",0);
	&je	(&label("done"));

	&mov	("ecx",8);
&set_label("loop");
	&${rdop}("edx");
	&jc	(&label("break"));
	&loop	(&label("loop"));
	&jmp	(&label("done"));

&set_label("break",16);
	&cmp	("ebx",4);
	&jb	(&label("tail"));
	&mov	(&DWP(0,"edi"),"edx");
	&lea	("edi",&DWP(4,"edi"));
	&add	("eax",4);
	&sub	("ebx",4);
	&jz	(&label("done"));
	&mov	("ecx",8);
	&jmp	(&label("loop"));

&set_label("tail",16);
	&mov	(&BP(0,"edi"),"dl");
	&lea	("edi",&DWP(1,"edi"));
	&inc	("eax");
	&shr	("edx",8);
	&dec	("ebx");
	&jnz	(&label("tail"));

&set_label("done");
	&xor	("edx","edx");		# Clear random value from registers
	&pop	("ebx");
	&pop	("edi");
	&ret	();
&function_end_B("OPENSSL_ia32_${rdop}_bytes");
}
&gen_random("rdrand");
&gen_random("rdseed");

&initseg("OPENSSL_cpuid_setup");

&hidden("OPENSSL_cpuid_setup");
&hidden("OPENSSL_ia32cap_P");

&asm_finish();

close STDOUT or die "error closing STDOUT: $!";
Commit	Line	Data
e0a65194	1	#! /usr/bin/env perl
33388b44	2	# Copyright 2004-2020 The OpenSSL Project Authors. All Rights Reserved.
e0a65194	3	#
0e9725bc	4	# Licensed under the Apache License 2.0 (the "License"). You may not use
e0a65194 RS	5	# this file except in compliance with the License. You can obtain a copy
	6	# in the file LICENSE in the source distribution or at
	7	# https://www.openssl.org/source/license.html
14e21f86	8
4d1f3f7a DSH	9	$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
4d1f3f7a DSH	10	push(@INC, "${dir}perlasm", "perlasm");
14e21f86 AP	11	require "x86asm.pl";
14e21f86 AP	12
1aa89a7a	13	$output = pop and open STDOUT,">$output";
0a4edb93	14
e195c8a2	15	&asm_init($ARGV[0]);
14e21f86	16
c5022659 AP	17	for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
c5022659 AP	18
14e21f86 AP	19	&function_begin("OPENSSL_ia32_cpuid");
	20	&xor ("edx","edx");
	21	&pushf ();
	22	&pop ("eax");
	23	&mov ("ecx","eax");
	24	&xor ("eax",1<<21);
	25	&push ("eax");
	26	&popf ();
	27	&pushf ();
	28	&pop ("eax");
	29	&xor ("ecx","eax");
162f677d	30	&xor ("eax","eax");
1aed5e1a AP	31	&mov ("esi",&wparam(0));
1aed5e1a AP	32	&mov (&DWP(8,"esi"),"eax"); # clear extended feature flags
0f2ece87 AP	33	&bt ("ecx",21);
0f2ece87 AP	34	&jnc (&label("nocpuid"));
162f677d	35	&cpuid ();
761393bb AP	36	&mov ("edi","eax"); # max value for standard query level
761393bb AP	37
162f677d AP	38	&xor ("eax","eax");
	39	&cmp ("ebx",0x756e6547); # "Genu"
	40	&setne (&LB("eax"));
	41	&mov ("ebp","eax");
	42	&cmp ("edx",0x49656e69); # "ineI"
	43	&setne (&LB("eax"));
	44	&or ("ebp","eax");
	45	&cmp ("ecx",0x6c65746e); # "ntel"
	46	&setne (&LB("eax"));
58672313	47	&or ("ebp","eax"); # 0 indicates Intel CPU
58672313 AP	48	&jz (&label("intel"));
	49
	50	&cmp ("ebx",0x68747541); # "Auth"
	51	&setne (&LB("eax"));
	52	&mov ("esi","eax");
	53	&cmp ("edx",0x69746E65); # "enti"
	54	&setne (&LB("eax"));
	55	&or ("esi","eax");
	56	&cmp ("ecx",0x444D4163); # "cAMD"
	57	&setne (&LB("eax"));
	58	&or ("esi","eax"); # 0 indicates AMD CPU
	59	&jnz (&label("intel"));
	60
761393bb	61	# AMD specific
58672313 AP	62	&mov ("eax",0x80000000);
58672313 AP	63	&cpuid ();
b9064221 AP	64	&cmp ("eax",0x80000001);
	65	&jb (&label("intel"));
	66	&mov ("esi","eax");
	67	&mov ("eax",0x80000001);
	68	&cpuid ();
	69	&or ("ebp","ecx");
	70	&and ("ebp",1<<11\|1); # isolate XOP bit
	71	&cmp ("esi",0x80000008);
58672313 AP	72	&jb (&label("intel"));
	73
	74	&mov ("eax",0x80000008);
	75	&cpuid ();
	76	&movz ("esi",&LB("ecx")); # number of cores - 1
	77	&inc ("esi"); # number of cores
	78
761393bb	79	&mov ("eax",1);
5702e965	80	&xor ("ecx","ecx");
761393bb AP	81	&cpuid ();
761393bb AP	82	&bt ("edx",28);
b9064221	83	&jnc (&label("generic"));
761393bb AP	84	&shr ("ebx",16);
	85	&and ("ebx",0xff);
	86	&cmp ("ebx","esi");
b9064221	87	&ja (&label("generic"));
761393bb	88	&and ("edx",0xefffffff); # clear hyper-threading bit
b9064221	89	&jmp (&label("generic"));
609b0852	90
58672313	91	&set_label("intel");
761393bb	92	&cmp ("edi",4);
1aed5e1a	93	&mov ("esi",-1);
761393bb AP	94	&jb (&label("nocacheinfo"));
	95
	96	&mov ("eax",4);
	97	&mov ("ecx",0); # query L1D
	98	&cpuid ();
1aed5e1a AP	99	&mov ("esi","eax");
	100	&shr ("esi",14);
	101	&and ("esi",0xfff); # number of cores -1 per L1D
761393bb AP	102
761393bb AP	103	&set_label("nocacheinfo");
14e21f86	104	&mov ("eax",1);
5702e965	105	&xor ("ecx","ecx");
14e21f86	106	&cpuid ();
18f5603c	107	&and ("edx",0xbfefffff); # force reserved bits #20, #30 to 0
162f677d	108	&cmp ("ebp",0);
4bb90087 AP	109	&jne (&label("notintel"));
4bb90087 AP	110	&or ("edx",1<<30); # set reserved bit#30 on Intel CPUs
46f4e1be	111	&and (&HB("eax"),15); # family ID
3df2eff4	112	&cmp (&HB("eax"),15); # P4?
4bb90087 AP	113	&jne (&label("notintel"));
	114	&or ("edx",1<<20); # set reserved bit#20 to engage RC4_CHAR
	115	&set_label("notintel");
3df2eff4	116	&bt ("edx",28); # test hyper-threading bit
b9064221	117	&jnc (&label("generic"));
761393bb	118	&and ("edx",0xefffffff);
1aed5e1a	119	&cmp ("esi",0);
b9064221	120	&je (&label("generic"));
761393bb AP	121
761393bb AP	122	&or ("edx",0x10000000);
162f677d	123	&shr ("ebx",16);
761393bb	124	&cmp (&LB("ebx"),1);
b9064221	125	&ja (&label("generic"));
b1e8b4e6	126	&and ("edx",0xefffffff); # clear hyper-threading bit if not
b9064221 AP	127
	128	&set_label("generic");
	129	&and ("ebp",1<<11); # isolate AMD XOP flag
18f5603c	130	&and ("ecx",0xfffff7ff); # force 11th bit to 0
1aed5e1a	131	&mov ("esi","edx"); # %ebp:%esi is copy of %ecx:%edx
b9064221 AP	132	&or ("ebp","ecx"); # merge AMD XOP flag
b9064221 AP	133
1aed5e1a AP	134	&cmp ("edi",7);
	135	&mov ("edi",&wparam(0));
	136	&jb (&label("no_extended_info"));
	137	&mov ("eax",7);
	138	&xor ("ecx","ecx");
	139	&cpuid ();
	140	&mov (&DWP(8,"edi"),"ebx"); # save extended feature flag
	141	&set_label("no_extended_info");
	142
	143	&bt ("ebp",27); # check OSXSAVE bit
e879dd43	144	&jnc (&label("clear_avx"));
b9064221 AP	145	&xor ("ecx","ecx");
	146	&data_byte(0x0f,0x01,0xd0); # xgetbv
	147	&and ("eax",6);
	148	&cmp ("eax",6);
	149	&je (&label("done"));
	150	&cmp ("eax",2);
	151	&je (&label("clear_avx"));
	152	&set_label("clear_xmm");
18f5603c AP	153	&and ("ebp",0xfdfffffd); # clear AESNI and PCLMULQDQ bits
18f5603c AP	154	&and ("esi",0xfeffffff); # clear FXSR
b9064221	155	&set_label("clear_avx");
18f5603c	156	&and ("ebp",0xefffe7ff); # clear AVX, FMA and AMD XOP bits
c5cd28bd	157	&and (&DWP(8,"edi"),0xffffffdf); # clear AVX2
3df2eff4	158	&set_label("done");
b9064221 AP	159	&mov ("eax","esi");
b9064221 AP	160	&mov ("edx","ebp");
0f2ece87	161	&set_label("nocpuid");
14e21f86 AP	162	&function_end("OPENSSL_ia32_cpuid");
14e21f86 AP	163
2b247cf8	164	&external_label("OPENSSL_ia32cap_P");
14e21f86	165
2b247cf8	166	&function_begin_B("OPENSSL_rdtsc","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
14e21f86 AP	167	&xor ("eax","eax");
14e21f86 AP	168	&xor ("edx","edx");
2b247cf8	169	&picmeup("ecx","OPENSSL_ia32cap_P");
14e21f86 AP	170	&bt (&DWP(0,"ecx"),4);
	171	&jnc (&label("notsc"));
	172	&rdtsc ();
	173	&set_label("notsc");
	174	&ret ();
	175	&function_end_B("OPENSSL_rdtsc");
	176
f642ebc1	177	# This works in Ring 0 only [read DJGPP+MS-DOS+privileged DPMI host],
c85c5c40 AP	178	# but it's safe to call it on any [supported] 32-bit platform...
	179	# Just check for [non-]zero return value...
	180	&function_begin_B("OPENSSL_instrument_halt","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	181	&picmeup("ecx","OPENSSL_ia32cap_P");
	182	&bt (&DWP(0,"ecx"),4);
	183	&jnc (&label("nohalt")); # no TSC
	184
	185	&data_word(0x9058900e); # push %cs; pop %eax
	186	&and ("eax",3);
	187	&jnz (&label("nohalt")); # not enough privileges
	188
	189	&pushf ();
f9c5e5d9	190	&pop ("eax");
c85c5c40 AP	191	&bt ("eax",9);
	192	&jnc (&label("nohalt")); # interrupts are disabled
	193
	194	&rdtsc ();
	195	&push ("edx");
	196	&push ("eax");
	197	&halt ();
	198	&rdtsc ();
	199
	200	&sub ("eax",&DWP(0,"esp"));
	201	&sbb ("edx",&DWP(4,"esp"));
	202	&add ("esp",8);
	203	&ret ();
	204
	205	&set_label("nohalt");
	206	&xor ("eax","eax");
	207	&xor ("edx","edx");
	208	&ret ();
	209	&function_end_B("OPENSSL_instrument_halt");
	210
f642ebc1 RS	211	# Essentially there is only one use for this function. Under DJGPP:
	212	#
	213	# #include <go32.h>
	214	# ...
	215	# i=OPENSSL_far_spin(_dos_ds,0x46c);
	216	# ...
	217	# to obtain the number of spins till closest timer interrupt.
	218
	219	&function_begin_B("OPENSSL_far_spin");
	220	&pushf ();
	221	&pop ("eax");
	222	&bt ("eax",9);
	223	&jnc (&label("nospin")); # interrupts are disabled
	224
	225	&mov ("eax",&DWP(4,"esp"));
	226	&mov ("ecx",&DWP(8,"esp"));
	227	&data_word (0x90d88e1e); # push %ds, mov %eax,%ds
	228	&xor ("eax","eax");
	229	&mov ("edx",&DWP(0,"ecx"));
	230	&jmp (&label("spin"));
	231
	232	&align (16);
	233	&set_label("spin");
	234	&inc ("eax");
	235	&cmp ("edx",&DWP(0,"ecx"));
	236	&je (&label("spin"));
	237
	238	&data_word (0x1f909090); # pop %ds
	239	&ret ();
	240
	241	&set_label("nospin");
	242	&xor ("eax","eax");
	243	&xor ("edx","edx");
	244	&ret ();
	245	&function_end_B("OPENSSL_far_spin");
cee73df3 AP	246
	247	&function_begin_B("OPENSSL_wipe_cpu","EXTRN\t_OPENSSL_ia32cap_P:DWORD");
	248	&xor ("eax","eax");
	249	&xor ("edx","edx");
	250	&picmeup("ecx","OPENSSL_ia32cap_P");
	251	&mov ("ecx",&DWP(0,"ecx"));
	252	&bt (&DWP(0,"ecx"),1);
	253	&jnc (&label("no_x87"));
c5022659	254	if ($sse2) {
b9064221 AP	255	&and ("ecx",1<<26\|1<<24); # check SSE2 and FXSR bits
	256	&cmp ("ecx",1<<26\|1<<24);
	257	&jne (&label("no_sse2"));
c5022659 AP	258	&pxor ("xmm0","xmm0");
	259	&pxor ("xmm1","xmm1");
	260	&pxor ("xmm2","xmm2");
	261	&pxor ("xmm3","xmm3");
	262	&pxor ("xmm4","xmm4");
	263	&pxor ("xmm5","xmm5");
	264	&pxor ("xmm6","xmm6");
	265	&pxor ("xmm7","xmm7");
	266	&set_label("no_sse2");
	267	}
c49a0aa0 AP	268	# just a bunch of fldz to zap the fp/mm bank followed by finit...
c49a0aa0 AP	269	&data_word(0xeed9eed9,0xeed9eed9,0xeed9eed9,0xeed9eed9,0x90e3db9b);
cee73df3 AP	270	&set_label("no_x87");
	271	&lea ("eax",&DWP(4,"esp"));
	272	&ret ();
	273	&function_end_B("OPENSSL_wipe_cpu");
	274
	275	&function_begin_B("OPENSSL_atomic_add");
	276	&mov ("edx",&DWP(4,"esp")); # fetch the pointer, 1st arg
	277	&mov ("ecx",&DWP(8,"esp")); # fetch the increment, 2nd arg
	278	&push ("ebx");
	279	&nop ();
	280	&mov ("eax",&DWP(0,"edx"));
	281	&set_label("spin");
	282	&lea ("ebx",&DWP(0,"eax","ecx"));
	283	&nop ();
69687aa8	284	&data_word(0x1ab10ff0); # lock; cmpxchg %ebx,(%edx) # %eax is involved and is always reloaded
cee73df3 AP	285	&jne (&label("spin"));
	286	&mov ("eax","ebx"); # OpenSSL expects the new value
	287	&pop ("ebx");
	288	&ret ();
	289	&function_end_B("OPENSSL_atomic_add");
	290
b2dba9bf AP	291	&function_begin_B("OPENSSL_cleanse");
	292	&mov ("edx",&wparam(0));
	293	&mov ("ecx",&wparam(1));
	294	&xor ("eax","eax");
	295	&cmp ("ecx",7);
	296	&jae (&label("lot"));
7676eebf AP	297	&cmp ("ecx",0);
7676eebf AP	298	&je (&label("ret"));
b2dba9bf AP	299	&set_label("little");
	300	&mov (&BP(0,"edx"),"al");
	301	&sub ("ecx",1);
	302	&lea ("edx",&DWP(1,"edx"));
	303	&jnz (&label("little"));
7676eebf	304	&set_label("ret");
b2dba9bf AP	305	&ret ();
	306
	307	&set_label("lot",16);
	308	&test ("edx",3);
	309	&jz (&label("aligned"));
	310	&mov (&BP(0,"edx"),"al");
	311	&lea ("ecx",&DWP(-1,"ecx"));
	312	&lea ("edx",&DWP(1,"edx"));
	313	&jmp (&label("lot"));
	314	&set_label("aligned");
	315	&mov (&DWP(0,"edx"),"eax");
	316	&lea ("ecx",&DWP(-4,"ecx"));
	317	&test ("ecx",-4);
	318	&lea ("edx",&DWP(4,"edx"));
	319	&jnz (&label("aligned"));
	320	&cmp ("ecx",0);
	321	&jne (&label("little"));
	322	&ret ();
	323	&function_end_B("OPENSSL_cleanse");
	324
e33826f0 AP	325	&function_begin_B("CRYPTO_memcmp");
	326	&push ("esi");
	327	&push ("edi");
	328	&mov ("esi",&wparam(0));
	329	&mov ("edi",&wparam(1));
	330	&mov ("ecx",&wparam(2));
	331	&xor ("eax","eax");
	332	&xor ("edx","edx");
	333	&cmp ("ecx",0);
	334	&je (&label("no_data"));
	335	&set_label("loop");
	336	&mov ("dl",&BP(0,"esi"));
	337	&lea ("esi",&DWP(1,"esi"));
	338	&xor ("dl",&BP(0,"edi"));
	339	&lea ("edi",&DWP(1,"edi"));
	340	&or ("al","dl");
	341	&dec ("ecx");
	342	&jnz (&label("loop"));
	343	&neg ("eax");
	344	&shr ("eax",31);
	345	&set_label("no_data");
	346	&pop ("edi");
	347	&pop ("esi");
	348	&ret ();
	349	&function_end_B("CRYPTO_memcmp");
5fabb88a AP	350	{
	351	my $lasttick = "esi";
	352	my $lastdiff = "ebx";
	353	my $out = "edi";
	354	my $cnt = "ecx";
	355	my $max = "ebp";
	356
	357	&function_begin("OPENSSL_instrument_bus");
	358	&mov ("eax",0);
	359	if ($sse2) {
	360	&picmeup("edx","OPENSSL_ia32cap_P");
	361	&bt (&DWP(0,"edx"),4);
	362	&jnc (&label("nogo")); # no TSC
	363	&bt (&DWP(0,"edx"),19);
	364	&jnc (&label("nogo")); # no CLFLUSH
	365
	366	&mov ($out,&wparam(0)); # load arguments
	367	&mov ($cnt,&wparam(1));
	368
	369	# collect 1st tick
	370	&rdtsc ();
	371	&mov ($lasttick,"eax"); # lasttick = tick
	372	&mov ($lastdiff,0); # lastdiff = 0
	373	&clflush(&DWP(0,$out));
b9064221	374	&data_byte(0xf0); # lock
5fabb88a AP	375	&add (&DWP(0,$out),$lastdiff);
	376	&jmp (&label("loop"));
	377
	378	&set_label("loop",16);
	379	&rdtsc ();
	380	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	381	&sub ("eax",$lasttick); # diff
	382	&mov ($lasttick,"edx"); # lasttick = tick
	383	&mov ($lastdiff,"eax"); # lastdiff = diff
	384	&clflush(&DWP(0,$out));
b9064221	385	&data_byte(0xf0); # lock
5fabb88a AP	386	&add (&DWP(0,$out),"eax"); # accumulate diff
	387	&lea ($out,&DWP(4,$out)); # ++$out
	388	&sub ($cnt,1); # --$cnt
	389	&jnz (&label("loop"));
	390
	391	&mov ("eax",&wparam(1));
	392	&set_label("nogo");
	393	}
	394	&function_end("OPENSSL_instrument_bus");
	395
	396	&function_begin("OPENSSL_instrument_bus2");
	397	&mov ("eax",0);
	398	if ($sse2) {
	399	&picmeup("edx","OPENSSL_ia32cap_P");
	400	&bt (&DWP(0,"edx"),4);
	401	&jnc (&label("nogo")); # no TSC
	402	&bt (&DWP(0,"edx"),19);
	403	&jnc (&label("nogo")); # no CLFLUSH
	404
	405	&mov ($out,&wparam(0)); # load arguments
	406	&mov ($cnt,&wparam(1));
	407	&mov ($max,&wparam(2));
	408
	409	&rdtsc (); # collect 1st tick
	410	&mov ($lasttick,"eax"); # lasttick = tick
	411	&mov ($lastdiff,0); # lastdiff = 0
	412
	413	&clflush(&DWP(0,$out));
b9064221	414	&data_byte(0xf0); # lock
5fabb88a AP	415	&add (&DWP(0,$out),$lastdiff);
	416
	417	&rdtsc (); # collect 1st diff
	418	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	419	&sub ("eax",$lasttick); # diff
	420	&mov ($lasttick,"edx"); # lasttick = tick
	421	&mov ($lastdiff,"eax"); # lastdiff = diff
	422	&jmp (&label("loop2"));
	423
	424	&set_label("loop2",16);
	425	&clflush(&DWP(0,$out));
b9064221	426	&data_byte(0xf0); # lock
5fabb88a AP	427	&add (&DWP(0,$out),"eax"); # accumulate diff
	428
	429	&sub ($max,1);
	430	&jz (&label("done2"));
	431
	432	&rdtsc ();
	433	&mov ("edx","eax"); # put aside tick (yes, I neglect edx)
	434	&sub ("eax",$lasttick); # diff
	435	&mov ($lasttick,"edx"); # lasttick = tick
	436	&cmp ("eax",$lastdiff);
	437	&mov ($lastdiff,"eax"); # lastdiff = diff
	438	&mov ("edx",0);
	439	&setne ("dl");
	440	&sub ($cnt,"edx"); # conditional --$cnt
	441	&lea ($out,&DWP(0,$out,"edx",4)); # conditional ++$out
	442	&jnz (&label("loop2"));
	443
	444	&set_label("done2");
	445	&mov ("eax",&wparam(1));
	446	&sub ("eax",$cnt);
	447	&set_label("nogo");
	448	}
	449	&function_end("OPENSSL_instrument_bus2");
	450	}
	451
9c940446 AP	452	sub gen_random {
9c940446 AP	453	my $rdop = shift;
9c940446 AP	454	&function_begin_B("OPENSSL_ia32_${rdop}_bytes");
	455	&push ("edi");
	456	&push ("ebx");
	457	&xor ("eax","eax"); # return value
	458	&mov ("edi",&wparam(0));
	459	&mov ("ebx",&wparam(1));
	460
	461	&cmp ("ebx",0);
	462	&je (&label("done"));
301799b8	463
f4d45640 AP	464	&mov ("ecx",8);
f4d45640 AP	465	&set_label("loop");
9c940446	466	&${rdop}("edx");
f4d45640 AP	467	&jc (&label("break"));
f4d45640 AP	468	&loop (&label("loop"));
9c940446 AP	469	&jmp (&label("done"));
	470
	471	&set_label("break",16);
	472	&cmp ("ebx",4);
	473	&jb (&label("tail"));
	474	&mov (&DWP(0,"edi"),"edx");
	475	&lea ("edi",&DWP(4,"edi"));
	476	&add ("eax",4);
	477	&sub ("ebx",4);
	478	&jz (&label("done"));
	479	&mov ("ecx",8);
	480	&jmp (&label("loop"));
	481
	482	&set_label("tail",16);
	483	&mov (&BP(0,"edi"),"dl");
	484	&lea ("edi",&DWP(1,"edi"));
	485	&inc ("eax");
	486	&shr ("edx",8);
	487	&dec ("ebx");
	488	&jnz (&label("tail"));
	489
	490	&set_label("done");
082193ef	491	&xor ("edx","edx"); # Clear random value from registers
9c940446 AP	492	&pop ("ebx");
9c940446 AP	493	&pop ("edi");
f4d45640	494	&ret ();
9c940446 AP	495	&function_end_B("OPENSSL_ia32_${rdop}_bytes");
	496	}
	497	&gen_random("rdrand");
	498	&gen_random("rdseed");
f4d45640	499
2b247cf8	500	&initseg("OPENSSL_cpuid_setup");
14e21f86	501
6206682a AP	502	&hidden("OPENSSL_cpuid_setup");
	503	&hidden("OPENSSL_ia32cap_P");
	504
14e21f86	505	&asm_finish();
0a4edb93	506
a21314db	507	close STDOUT or die "error closing STDOUT: $!";