Extend OPENSSL_ia32cap_P with extra word to accomodate AVX2 capability.

crypto/cryptlib.c

@@ -125,7 +125,7 @@ static double SSLeay_MSVC5_hack=0.0; /* and for VC1.5 */
 	defined(__INTEL__) || \
 	defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || defined(_M_X64)
 
-extern unsigned int  OPENSSL_ia32cap_P[2];
+extern unsigned int  OPENSSL_ia32cap_P[4];
 unsigned int *OPENSSL_ia32cap_loc(void) { return OPENSSL_ia32cap_P; }
 
 #if defined(OPENSSL_CPUID_OBJ) && !defined(OPENSSL_NO_ASM) && !defined(I386_ONLY)
@@ -137,7 +137,7 @@ typedef unsigned long long IA32CAP;
 #endif
 void OPENSSL_cpuid_setup(void)
 { static int trigger=0;
-  IA32CAP OPENSSL_ia32_cpuid(void);
+  IA32CAP OPENSSL_ia32_cpuid(unsigned int *);
   IA32CAP vec;
   char *env;
 
@@ -151,10 +151,18 @@ void OPENSSL_cpuid_setup(void)
 #else
 	if (!sscanf(env+off,"%lli",(long long *)&vec)) vec = strtoul(env+off,NULL,0);
 #endif
-	if (off) vec = OPENSSL_ia32_cpuid()&~vec;
+	if (off) vec = OPENSSL_ia32_cpuid(OPENSSL_ia32cap_P)&~vec;
+
+	OPENSSL_ia32cap_P[2] = 0;
+	if ((env=strchr(env,':'))) {
+	    off = (env[1]=='~')?2:1;
+	    vec = strtoul(env+off,NULL,0);
+	    if (off>1)	OPENSSL_ia32cap_P[2] &= ~vec;
+	    else	OPENSSL_ia32cap_P[2] = vec;
+	}
     }
     else
-	vec = OPENSSL_ia32_cpuid();
+	vec = OPENSSL_ia32_cpuid(OPENSSL_ia32cap_P);
 
     /*
      * |(1<<10) sets a reserved bit to signal that variable
@@ -165,7 +173,7 @@ void OPENSSL_cpuid_setup(void)
     OPENSSL_ia32cap_P[1] = (unsigned int)(vec>>32);
 }
 #else
-unsigned int OPENSSL_ia32cap_P[2];
+unsigned int OPENSSL_ia32cap_P[4];
 #endif
 
 #else
@@ -173,7 +181,7 @@ unsigned int *OPENSSL_ia32cap_loc(void) { return NULL; }
 #endif
 int OPENSSL_NONPIC_relocated = 0;
 #if !defined(OPENSSL_CPUID_SETUP) && !defined(OPENSSL_CPUID_OBJ)
-void OPENSSL_cpuid_setup(void) {}
+void OPENSSL_cpuid_setup(unsigned int *) {}
 #endif
 
 #if (defined(_WIN32) || defined(__CYGWIN__)) && defined(_WINDLL)

crypto/perlasm/x86asm.pl

@@ -131,6 +131,32 @@ sub ::rdrand
     {	&::generic("rdrand",@_);	}
 }
 
+sub rxb {
+ local *opcode=shift;
+ my ($dst,$src1,$src2,$rxb)=@_;
+
+   $rxb|=0x7<<5;
+   $rxb&=~(0x04<<5) if($dst>=8);
+   $rxb&=~(0x01<<5) if($src1>=8);
+   $rxb&=~(0x02<<5) if($src2>=8);
+   push @opcode,$rxb;
+}
+
+sub ::vprotd
+{ my $args=join(',',@_);
+    if ($args =~ /xmm([0-7]),xmm([0-7]),([x0-9a-f]+)/)
+    { my @opcode=(0x8f);
+	rxb(\@opcode,$1,$2,-1,0x08);
+	push @opcode,0x78,0xc2;
+	push @opcode,0xc0|($2&7)|(($1&7)<<3);		# ModR/M
+	my $c=$3;
+	push @opcode,$c=~/^0/?oct($c):$c;
+	&::data_byte(@opcode);
+    }
+    else
+    {	&::generic("vprotd",@_);	}
+}
+
 # label management
 $lbdecor="L";		# local label decoration, set by package
 $label="000";

crypto/perlasm/x86gas.pl

@@ -70,6 +70,8 @@ sub ::DWP
 { my($addr,$reg1,$reg2,$idx)=@_;
   my $ret="";
 
+    if (!defined($idx) && 1*$reg2) { $idx=$reg2; $reg2=$reg1; undef $reg1; }
+
     $addr =~ s/^\s+//;
     # prepend global references with optional underscore
     $addr =~ s/^([^\+\-0-9][^\+\-]*)/&::islabel($1) or "$nmdecor$1"/ige;
@@ -157,7 +159,7 @@ sub ::file_end
 	}
     }
     if (grep {/\b${nmdecor}OPENSSL_ia32cap_P\b/i} @out) {
-	my $tmp=".comm\t${nmdecor}OPENSSL_ia32cap_P,8";
+	my $tmp=".comm\t${nmdecor}OPENSSL_ia32cap_P,16";
 	if ($::macosx)	{ push (@out,"$tmp,2\n"); }
 	elsif ($::elf)	{ push (@out,"$tmp,4\n"); }
 	else		{ push (@out,"$tmp\n"); }

crypto/perlasm/x86masm.pl

@@ -39,6 +39,8 @@ sub get_mem
 { my($size,$addr,$reg1,$reg2,$idx)=@_;
   my($post,$ret);
 
+    if (!defined($idx) && 1*$reg2) { $idx=$reg2; $reg2=$reg1; undef $reg1; }
+
     $ret .= "$size PTR " if ($size ne "");
 
     $addr =~ s/^\s+//;
@@ -133,7 +135,7 @@ ___
     if (grep {/\b${nmdecor}OPENSSL_ia32cap_P\b/i} @out)
     {	my $comm=<<___;
 .bss	SEGMENT 'BSS'
-COMM	${nmdecor}OPENSSL_ia32cap_P:QWORD
+COMM	${nmdecor}OPENSSL_ia32cap_P:DWORD:4
 .bss	ENDS
 ___
 	# comment out OPENSSL_ia32cap_P declarations

crypto/perlasm/x86nasm.pl

@@ -36,6 +36,8 @@ sub get_mem
 { my($size,$addr,$reg1,$reg2,$idx)=@_;
   my($post,$ret);
 
+    if (!defined($idx) && 1*$reg2) { $idx=$reg2; $reg2=$reg1; undef $reg1; }
+
     if ($size ne "")
     {	$ret .= "$size";
 	$ret .= " PTR" if ($::mwerks);
@@ -117,7 +119,7 @@ sub ::file_end
 {   if (grep {/\b${nmdecor}OPENSSL_ia32cap_P\b/i} @out)
     {	my $comm=<<___;
 ${drdecor}segment	.bss
-${drdecor}common	${nmdecor}OPENSSL_ia32cap_P 8
+${drdecor}common	${nmdecor}OPENSSL_ia32cap_P 16
 ___
 	# comment out OPENSSL_ia32cap_P declarations
 	grep {s/(^extern\s+${nmdecor}OPENSSL_ia32cap_P)/\;$1/} @out;

crypto/x86_64cpuid.pl

@@ -23,7 +23,7 @@ print<<___;
 	call	OPENSSL_cpuid_setup
 
 .hidden	OPENSSL_ia32cap_P
-.comm	OPENSSL_ia32cap_P,8,4
+.comm	OPENSSL_ia32cap_P,16,4
 
 .text
 
@@ -52,12 +52,13 @@ OPENSSL_rdtsc:
 .size	OPENSSL_rdtsc,.-OPENSSL_rdtsc
 
 .globl	OPENSSL_ia32_cpuid
-.type	OPENSSL_ia32_cpuid,\@abi-omnipotent
+.type	OPENSSL_ia32_cpuid,\@function,1
 .align	16
 OPENSSL_ia32_cpuid:
 	mov	%rbx,%r8		# save %rbx
 
 	xor	%eax,%eax
+	mov	%eax,8(%rdi)		# clear 3rd word
 	cpuid
 	mov	%eax,%r11d		# max value for standard query level
 
@@ -125,6 +126,14 @@ OPENSSL_ia32_cpuid:
 	shr	\$14,%r10d
 	and	\$0xfff,%r10d		# number of cores -1 per L1D
 
+	cmp	\$7,%r11d
+	jb	.Lnocacheinfo
+
+	mov	\$7,%eax
+	xor	%ecx,%ecx
+	cpuid
+	mov	%ebx,8(%rdi)
+
 .Lnocacheinfo:
 	mov	\$1,%eax
 	cpuid
@@ -164,6 +173,7 @@ OPENSSL_ia32_cpuid:
 .Lclear_avx:
 	mov	\$0xefffe7ff,%eax	# ~(1<<28|1<<12|1<<11)
 	and	%eax,%r9d		# clear AVX, FMA and AMD XOP bits
+	andl	\$0xffffffdf,8(%rdi)	# cleax AVX2, ~(1<<5)
 .Ldone:
 	shl	\$32,%r9
 	mov	%r10d,%eax

crypto/x86cpuid.pl

@@ -22,6 +22,8 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
 	&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
 
@@ -89,6 +91,15 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
 	&shr	("edi",14);
 	&and	("edi",0xfff);		# number of cores -1 per L1D
 
+	&cmp	("edi",7);
+	&jb	(&label("nocacheinfo"));
+
+	&mov	("esi",&wparam(0));
+	&mov	("eax",7);
+	&xor	("ecx","ecx");
+	&cpuid	();
+	&mov	(&DWP(8,"esi"),"ebx");
+
 &set_label("nocacheinfo");
 	&mov	("eax",1);
 	&cpuid	();
@@ -133,6 +144,8 @@ for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
 	&and	("esi",0xfeffffff);	# clear FXSR
 &set_label("clear_avx");
 	&and	("ebp",0xefffe7ff);	# clear AVX, FMA and AMD XOP bits
+	&mov	("edi",&wparam(0));
+	&and	(&DWP(8,"edi"),0xffffffdf);	# clear AVX2
 &set_label("done");
 	&mov	("eax","esi");
 	&mov	("edx","ebp");

doc/crypto/OPENSSL_ia32cap.pod

@@ -72,3 +72,17 @@ the data cache is actually shared between logical cores. This in turn
 affects the decision on whether or not expensive countermeasures
 against cache-timing attacks are applied, most notably in AES assembler
 module.
+
+The vector is further extended with EBX value returned by CPUID with
+EAX=7 and ECX=0 as input. Following bits are significant:
+
+=item bit #64+3 denoting availability of BMI1 instructions, e.g. ANDN;
+
+=item bit #64+5 denoting availability of AVX2 instructions;
+
+=item bit #64+8 denoting availability of BMI2 instructions, e.g. MUXL
+      and RORX;
+
+=item bit #64+18 denoting availability of RDSEED instruction;
+
+=itme bit #64+19 denoting availability of ADCX and ADOX instructions;