crypto/x86[_64]cpuid.pl: add OPENSSL_ia32_rd[rand|seed]_bytes.

Reviewed-by: Richard Levitte <levitte@openssl.org>

3 files changed, 98 insertions, 32 deletions

diff --git a/crypto/perlasm/x86_64-xlate.pl b/crypto/perlasm/x86_64-xlate.pl
index 60b98d7407..617adf9a38 100755
--- a/crypto/perlasm/x86_64-xlate.pl
+++ b/crypto/perlasm/x86_64-xlate.pl
@@ -805,7 +805,7 @@ my $rdrand = sub {
       my @opcode=();
       my $dst=$1;
 	if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
-	rex(\@opcode,0,$1,8);
+	rex(\@opcode,0,$dst,8);
 	push @opcode,0x0f,0xc7,0xf0|($dst&7);
 	@opcode;
     } else {
@@ -818,7 +818,7 @@ my $rdseed = sub {
       my @opcode=();
       my $dst=$1;
 	if ($dst !~ /[0-9]+/) { $dst = $regrm{"%e$dst"}; }
-	rex(\@opcode,0,$1,8);
+	rex(\@opcode,0,$dst,8);
 	push @opcode,0x0f,0xc7,0xf8|($dst&7);
 	@opcode;
     } else {
diff --git a/crypto/x86_64cpuid.pl b/crypto/x86_64cpuid.pl
index 285ad1ba7d..6cb152148b 100644
--- a/crypto/x86_64cpuid.pl
+++ b/crypto/x86_64cpuid.pl
@@ -393,36 +393,67 @@ OPENSSL_instrument_bus2:
 ___
 }
 
+sub gen_random {
+my $rdop = shift;
 print<<___;
-.globl	OPENSSL_ia32_rdrand
-.type	OPENSSL_ia32_rdrand,\@abi-omnipotent
+.globl	OPENSSL_ia32_${rdop}
+.type	OPENSSL_ia32_${rdop},\@abi-omnipotent
 .align	16
-OPENSSL_ia32_rdrand:
+OPENSSL_ia32_${rdop}:
 	mov	\$8,%ecx
-.Loop_rdrand:
-	rdrand	%rax
-	jc	.Lbreak_rdrand
-	loop	.Loop_rdrand
-.Lbreak_rdrand:
+.Loop_${rdop}:
+	${rdop}	%rax
+	jc	.Lbreak_${rdop}
+	loop	.Loop_${rdop}
+.Lbreak_${rdop}:
 	cmp	\$0,%rax
 	cmove	%rcx,%rax
 	ret
-.size	OPENSSL_ia32_rdrand,.-OPENSSL_ia32_rdrand
+.size	OPENSSL_ia32_${rdop},.-OPENSSL_ia32_${rdop}
 
-.globl	OPENSSL_ia32_rdseed
-.type	OPENSSL_ia32_rdseed,\@abi-omnipotent
+.globl	OPENSSL_ia32_${rdop}_bytes
+.type	OPENSSL_ia32_${rdop}_bytes,\@abi-omnipotent
 .align	16
-OPENSSL_ia32_rdseed:
-	mov	\$8,%ecx
-.Loop_rdseed:
-	rdseed	%rax
-	jc	.Lbreak_rdseed
-	loop	.Loop_rdseed
-.Lbreak_rdseed:
-	cmp	\$0,%rax
-	cmove	%rcx,%rax
+OPENSSL_ia32_${rdop}_bytes:
+	xor	%rax, %rax	# return value
+	cmp	\$0,$arg2
+	je	.Ldone_${rdop}_bytes
+
+	mov	\$8,%r11
+.Loop_${rdop}_bytes:
+	${rdop}	%r10
+	jc	.Lbreak_${rdop}_bytes
+	dec	%r11
+	jnz	.Loop_${rdop}_bytes
+	jmp	.Ldone_${rdop}_bytes
+
+.align	16
+.Lbreak_${rdop}_bytes:
+	cmp	\$8,$arg2
+	jb	.Ltail_${rdop}_bytes
+	mov	%r10,($arg1)
+	lea	8($arg1),$arg1
+	add	\$8,%rax
+	sub	\$8,$arg2
+	jz	.Ldone_${rdop}_bytes
+	mov	\$8,%r11
+	jmp	.Loop_${rdop}_bytes
+
+.align	16
+.Ltail_${rdop}_bytes:
+	mov	%r10b,($arg1)
+	lea	1($arg1),$arg1
+	inc	%rax
+	shr	\$8,%r8
+	dec	$arg2
+	jnz	.Ltail_${rdop}_bytes
+
+.Ldone_${rdop}_bytes:
 	ret
-.size	OPENSSL_ia32_rdseed,.-OPENSSL_ia32_rdseed
+.size	OPENSSL_ia32_${rdop}_bytes,.-OPENSSL_ia32_${rdop}_bytes
 ___
+}
+gen_random("rdrand");
+gen_random("rdseed");
 
 close STDOUT;	# flush
diff --git a/crypto/x86cpuid.pl b/crypto/x86cpuid.pl
index 99ffa1d2fb..c45b18356b 100644
--- a/crypto/x86cpuid.pl
+++ b/crypto/x86cpuid.pl
@@ -492,29 +492,64 @@ my $max = "ebp";
 &function_end("OPENSSL_instrument_bus2");
 }
 
-&function_begin_B("OPENSSL_ia32_rdrand");
+sub gen_random {
+my $rdop = shift;
+&function_begin_B("OPENSSL_ia32_${rdop}");
 	&mov	("ecx",8);
 &set_label("loop");
-	&rdrand	("eax");
+	&${rdop}("eax");
 	&jc	(&label("break"));
 	&loop	(&label("loop"));
 &set_label("break");
 	&cmp	("eax",0);
 	&cmove	("eax","ecx");
 	&ret	();
-&function_end_B("OPENSSL_ia32_rdrand");
+&function_end_B("OPENSSL_ia32_${rdop}");
+
+&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"));
 
-&function_begin_B("OPENSSL_ia32_rdseed");
 	&mov	("ecx",8);
 &set_label("loop");
-	&rdseed	("eax");
+	&${rdop}("edx");
 	&jc	(&label("break"));
 	&loop	(&label("loop"));
-&set_label("break");
-	&cmp	("eax",0);
-	&cmove	("eax","ecx");
+	&jmp	(&label("done"));
+
+&set_label("break",16);
+	&cmp	("ebx",4);
+	&jb	(&label("tail"));
+	&mov	(&DWP(0,"edi"),"edx");
+	&lea	("edi",&DWP(4,"edi"));
+	&add	("eax",4);
+	&sub	("ebx",4);
+	&jz	(&label("done"));
+	&mov	("ecx",8);
+	&jmp	(&label("loop"));
+
+&set_label("tail",16);
+	&mov	(&BP(0,"edi"),"dl");
+	&lea	("edi",&DWP(1,"edi"));
+	&inc	("eax");
+	&shr	("edx",8);
+	&dec	("ebx");
+	&jnz	(&label("tail"));
+
+&set_label("done");
+	&pop	("ebx");
+	&pop	("edi");
 	&ret	();
-&function_end_B("OPENSSL_ia32_rdseed");
+&function_end_B("OPENSSL_ia32_${rdop}_bytes");
+}
+&gen_random("rdrand");
+&gen_random("rdseed");
 
 &initseg("OPENSSL_cpuid_setup");