Remove o_init.o special case from Makefile: this doesn't work.
[openssl.git] / crypto / x86_64cpuid.pl
index 84b1cbe85f59984d99e5ecda20b0449c057d1729..0d233a10ea9e4179a0ced04a5c671ac0044bf17f 100644 (file)
@@ -7,7 +7,11 @@ if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
 
 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
-open STDOUT,"| $^X ${dir}perlasm/x86_64-xlate.pl $flavour $output";
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open STDOUT,"| $^X $xlate $flavour $output";
 
 ($arg1,$arg2,$arg3,$arg4)=$win64?("%rcx","%rdx","%r8", "%r9") :        # Win64 order
                                 ("%rdi","%rsi","%rdx","%rcx"); # Unix order
@@ -124,13 +128,14 @@ OPENSSL_ia32_cpuid:
 .Lnocacheinfo:
        mov     \$1,%eax
        cpuid
+       and     \$0xbfefffff,%edx       # force reserved bits to 0
        cmp     \$0,%r9d
        jne     .Lnotintel
-       or      \$0x00100000,%edx       # use reserved 20th bit to engage RC4_CHAR
+       or      \$0x40000000,%edx       # set reserved bit#30 on Intel CPUs
        and     \$15,%ah
        cmp     \$15,%ah                # examine Family ID
-       j     .Lnotintel
-       or      \$0x40000000,%edx       # use reserved bit to skip unrolled loop
+       jne     .Lnotintel
+       or      \$0x00100000,%edx       # set reserved bit#20 to engage RC4_CHAR
 .Lnotintel:
        bt      \$28,%edx               # test hyper-threading bit
        jnc     .Lgeneric
@@ -348,4 +353,21 @@ OPENSSL_instrument_bus2:
 ___
 }
 
+print<<___;
+.globl OPENSSL_ia32_rdrand
+.type  OPENSSL_ia32_rdrand,\@abi-omnipotent
+.align 16
+OPENSSL_ia32_rdrand:
+       mov     \$8,%ecx
+.Loop_rdrand:
+       rdrand  %rax
+       jc      .Lbreak_rdrand
+       loop    .Loop_rdrand
+.Lbreak_rdrand:
+       cmp     \$0,%rax
+       cmove   %rcx,%rax
+       ret
+.size  OPENSSL_ia32_rdrand,.-OPENSSL_ia32_rdrand
+___
+
 close STDOUT;  # flush