fix some cause of bad TEXTREL on i386 and amd64

- global function calls in .init sections (diff makes them via PLT)
- calls to global functions in aes-586.S (made static or local)
- global variable accesses in rc4-x86_64.S (now made via GOT)
from djm@large; ok miod@

10 files changed, 128 insertions, 28 deletions

diff --git a/src/lib/libcrypto/aes/asm/aes-586.pl b/src/lib/libcrypto/aes/asm/aes-586.pl
index 89fa261794..3da307bef9 100644
--- a/src/lib/libcrypto/aes/asm/aes-586.pl
+++ b/src/lib/libcrypto/aes/asm/aes-586.pl
@@ -250,7 +250,7 @@ sub enclast()
 sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 
 &public_label("AES_Te");
-&function_begin_B("_x86_AES_encrypt");
+&function_begin_C("_x86_AES_encrypt");
         if ($vertical_spin) {
                 # I need high parts of volatile registers to be accessible...
                 &exch   ($s1="edi",$key="ebx");
@@ -539,7 +539,7 @@ sub declast()
 }
 
 &public_label("AES_Td");
-&function_begin_B("_x86_AES_decrypt");
+&function_begin_C("_x86_AES_decrypt");
         # note that caller is expected to allocate stack frame for me!
         &mov    (&DWP(12,"esp"),$key);          # save key
 
@@ -1240,7 +1240,7 @@ sub enckey()
 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
 #                        AES_KEY *key)
 &public_label("AES_Te");
-&function_begin("AES_set_encrypt_key");
+&function_begin("AES_set_encrypt_key", "", "_x86_AES_set_encrypt_key");
         &mov    ("esi",&wparam(0));             # user supplied key
         &mov    ("edi",&wparam(2));             # private key schedule
 
@@ -1467,7 +1467,7 @@ sub deckey()
         &mov    (&DWP(0,"esp"),"eax");
         &mov    (&DWP(4,"esp"),"ecx");
         &mov    (&DWP(8,"esp"),"edx");
-        &call   ("AES_set_encrypt_key");
+        &call   ("_x86_AES_set_encrypt_key");
         &add    ("esp",12);
         &cmp    ("eax",0);
         &je     (&label("proceed"));
diff --git a/src/lib/libcrypto/perlasm/x86_64-xlate.pl b/src/lib/libcrypto/perlasm/x86_64-xlate.pl
index a4af769b4a..74153b017d 100755
--- a/src/lib/libcrypto/perlasm/x86_64-xlate.pl
+++ b/src/lib/libcrypto/perlasm/x86_64-xlate.pl
@@ -163,7 +163,8 @@ my $current_function;
         local   *line = shift;
         undef   $ret;
 
-        if ($line =~ /^([^\(,]*)\(([%\w,]+)\)/) {
+        if ($line =~ /^([^\(,]*)\(([%\w,]+)\)/ &&
+            !($line =~ /^PIC_(GOT|PLT)/)) {
             $self->{label} = $1;
             ($self->{base},$self->{index},$self->{scale})=split(/,/,$2);
             $self->{scale} = 1 if (!defined($self->{scale}));
@@ -429,6 +430,8 @@ my $current_function;
     }
 }
 
+print "#include <machine/asm.h>\n";
+
 while($line=<>) {
 
     chomp($line);
diff --git a/src/lib/libcrypto/perlasm/x86unix.pl b/src/lib/libcrypto/perlasm/x86unix.pl
index 02d72a32bc..ae8f0964dc 100644
--- a/src/lib/libcrypto/perlasm/x86unix.pl
+++ b/src/lib/libcrypto/perlasm/x86unix.pl
@@ -345,7 +345,7 @@ sub main'file
         local($file)=@_;
 
         if ($main'openbsd)
-                { push(@out,"#include <machine/asm.h>\n"); return; }
+                { push(@out,"#include <machine/asm.h>\n"); }
 
         local($tmp)=<<"EOF";
         .file   "$file.s"
@@ -355,13 +355,17 @@ EOF
 
 sub main'function_begin
         {
-        local($func)=@_;
+        local($func,$junk,$llabel)=@_;
 
         &main'external_label($func);
         $func=$under.$func;
 
         if ($main'openbsd)
-                { push (@out, "\nENTRY($func)\n"); goto skip; }
+                {
+                push (@out, "\nENTRY($func)\n");
+                push (@out, "$llabel:\n") if $llabel;
+                goto skip;
+                }
 
         local($tmp)=<<"EOF";
 .text
@@ -417,6 +421,44 @@ skip:
         $stack=4;
         }
 
+# Like function_begin_B but with static linkage
+sub main'function_begin_C
+        {
+        local($func,$extra)=@_;
+
+        &main'external_label($func);
+        $func=$under.$func;
+
+        if ($main'openbsd)
+                {
+                        local($tmp)=<<"EOF";
+.text
+_ALIGN_TEXT
+.type $func,\@function
+$func:
+EOF
+                        push(@out, $tmp);
+                        goto skip;
+                }
+
+        local($tmp)=<<"EOF";
+.text
+.globl  $func
+EOF
+        push(@out,$tmp);
+        if ($main'cpp)
+                { push(@out,"TYPE($func,\@function)\n"); }
+        elsif ($main'coff)
+                { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
+        elsif ($main'aout and !$main'pic)
+                { }
+        else    { push(@out,".type      $func,\@function\n"); }
+        push(@out,".align\t$align\n");
+        push(@out,"$func:\n");
+skip:
+        $stack=4;
+        }
+
 sub main'function_end
         {
         local($func)=@_;
@@ -474,6 +516,8 @@ sub main'function_end_B
         %label=();
         }
 
+sub main'function_end_C { function_end_B(@_); }
+
 sub main'wparam
         {
         local($num)=@_;
@@ -510,7 +554,7 @@ sub main'swtmp
 
 sub main'comment
         {
-        if (!defined($com_start) or $main'elf)
+        if (!defined($com_start) or (!$main'openbsd && $main'elf))
                 {       # Regarding $main'elf above...
                         # GNU and SVR4 as'es use different comment delimiters,
                 push(@out,"\n");        # so we just skip ELF comments...
@@ -731,7 +775,9 @@ sub main'initseg
                 {
                 $tmp=<<___;
 .section        .init
-        call    $under$f
+                PIC_PROLOGUE
+                call    PIC_PLT($under$f)
+                PIC_EPILOGUE
         jmp     .Linitalign
 .align  $align
 .Linitalign:
diff --git a/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl b/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
index 2d47320485..92c52f3433 100755
--- a/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
+++ b/src/lib/libcrypto/rc4/asm/rc4-x86_64.pl
@@ -269,8 +269,7 @@ RC4_set_key:
         xor     $ido,$ido
         xor     %r10,%r10
         xor     %r11,%r11
-
-        mov     OPENSSL_ia32cap_P(%rip),$idx#d
+        mov     PIC_GOT(OPENSSL_ia32cap_P),$idx#d
         bt      \$20,$idx#d
         jnc     .Lw1stloop
         bt      \$30,$idx#d
@@ -338,7 +337,7 @@ RC4_set_key:
 RC4_options:
         .picmeup %rax
         lea     .Lopts-.(%rax),%rax
-        mov     OPENSSL_ia32cap_P(%rip),%edx
+        mov     PIC_GOT(OPENSSL_ia32cap_P),%edx
         bt      \$20,%edx
         jnc     .Ldone
         add     \$12,%rax
diff --git a/src/lib/libcrypto/x86_64cpuid.pl b/src/lib/libcrypto/x86_64cpuid.pl
index 2616a03da6..8946b464a8 100644
--- a/src/lib/libcrypto/x86_64cpuid.pl
+++ b/src/lib/libcrypto/x86_64cpuid.pl
@@ -47,6 +47,8 @@ CRT\$XIU	ENDS
 
 ___
 print<<___ if(!defined($masm));
+#include <machine/asm.h>
+
 .text
 
 .globl  OPENSSL_atomic_add
@@ -95,7 +97,7 @@ OPENSSL_wipe_cpu:
 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 
 .section        .init
-        call    OPENSSL_cpuid_setup
+        call    PIC_PLT(OPENSSL_cpuid_setup)
 
 ___
 
diff --git a/src/lib/libssl/src/crypto/aes/asm/aes-586.pl b/src/lib/libssl/src/crypto/aes/asm/aes-586.pl
index 89fa261794..3da307bef9 100644
--- a/src/lib/libssl/src/crypto/aes/asm/aes-586.pl
+++ b/src/lib/libssl/src/crypto/aes/asm/aes-586.pl
@@ -250,7 +250,7 @@ sub enclast()
 sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 
 &public_label("AES_Te");
-&function_begin_B("_x86_AES_encrypt");
+&function_begin_C("_x86_AES_encrypt");
         if ($vertical_spin) {
                 # I need high parts of volatile registers to be accessible...
                 &exch   ($s1="edi",$key="ebx");
@@ -539,7 +539,7 @@ sub declast()
 }
 
 &public_label("AES_Td");
-&function_begin_B("_x86_AES_decrypt");
+&function_begin_C("_x86_AES_decrypt");
         # note that caller is expected to allocate stack frame for me!
         &mov    (&DWP(12,"esp"),$key);          # save key
 
@@ -1240,7 +1240,7 @@ sub enckey()
 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
 #                        AES_KEY *key)
 &public_label("AES_Te");
-&function_begin("AES_set_encrypt_key");
+&function_begin("AES_set_encrypt_key", "", "_x86_AES_set_encrypt_key");
         &mov    ("esi",&wparam(0));             # user supplied key
         &mov    ("edi",&wparam(2));             # private key schedule
 
@@ -1467,7 +1467,7 @@ sub deckey()
         &mov    (&DWP(0,"esp"),"eax");
         &mov    (&DWP(4,"esp"),"ecx");
         &mov    (&DWP(8,"esp"),"edx");
-        &call   ("AES_set_encrypt_key");
+        &call   ("_x86_AES_set_encrypt_key");
         &add    ("esp",12);
         &cmp    ("eax",0);
         &je     (&label("proceed"));
diff --git a/src/lib/libssl/src/crypto/perlasm/x86_64-xlate.pl b/src/lib/libssl/src/crypto/perlasm/x86_64-xlate.pl
index a4af769b4a..74153b017d 100755
--- a/src/lib/libssl/src/crypto/perlasm/x86_64-xlate.pl
+++ b/src/lib/libssl/src/crypto/perlasm/x86_64-xlate.pl
@@ -163,7 +163,8 @@ my $current_function;
         local   *line = shift;
         undef   $ret;
 
-        if ($line =~ /^([^\(,]*)\(([%\w,]+)\)/) {
+        if ($line =~ /^([^\(,]*)\(([%\w,]+)\)/ &&
+            !($line =~ /^PIC_(GOT|PLT)/)) {
             $self->{label} = $1;
             ($self->{base},$self->{index},$self->{scale})=split(/,/,$2);
             $self->{scale} = 1 if (!defined($self->{scale}));
@@ -429,6 +430,8 @@ my $current_function;
     }
 }
 
+print "#include <machine/asm.h>\n";
+
 while($line=<>) {
 
     chomp($line);
diff --git a/src/lib/libssl/src/crypto/perlasm/x86unix.pl b/src/lib/libssl/src/crypto/perlasm/x86unix.pl
index 02d72a32bc..ae8f0964dc 100644
--- a/src/lib/libssl/src/crypto/perlasm/x86unix.pl
+++ b/src/lib/libssl/src/crypto/perlasm/x86unix.pl
@@ -345,7 +345,7 @@ sub main'file
         local($file)=@_;
 
         if ($main'openbsd)
-                { push(@out,"#include <machine/asm.h>\n"); return; }
+                { push(@out,"#include <machine/asm.h>\n"); }
 
         local($tmp)=<<"EOF";
         .file   "$file.s"
@@ -355,13 +355,17 @@ EOF
 
 sub main'function_begin
         {
-        local($func)=@_;
+        local($func,$junk,$llabel)=@_;
 
         &main'external_label($func);
         $func=$under.$func;
 
         if ($main'openbsd)
-                { push (@out, "\nENTRY($func)\n"); goto skip; }
+                {
+                push (@out, "\nENTRY($func)\n");
+                push (@out, "$llabel:\n") if $llabel;
+                goto skip;
+                }
 
         local($tmp)=<<"EOF";
 .text
@@ -417,6 +421,44 @@ skip:
         $stack=4;
         }
 
+# Like function_begin_B but with static linkage
+sub main'function_begin_C
+        {
+        local($func,$extra)=@_;
+
+        &main'external_label($func);
+        $func=$under.$func;
+
+        if ($main'openbsd)
+                {
+                        local($tmp)=<<"EOF";
+.text
+_ALIGN_TEXT
+.type $func,\@function
+$func:
+EOF
+                        push(@out, $tmp);
+                        goto skip;
+                }
+
+        local($tmp)=<<"EOF";
+.text
+.globl  $func
+EOF
+        push(@out,$tmp);
+        if ($main'cpp)
+                { push(@out,"TYPE($func,\@function)\n"); }
+        elsif ($main'coff)
+                { $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
+        elsif ($main'aout and !$main'pic)
+                { }
+        else    { push(@out,".type      $func,\@function\n"); }
+        push(@out,".align\t$align\n");
+        push(@out,"$func:\n");
+skip:
+        $stack=4;
+        }
+
 sub main'function_end
         {
         local($func)=@_;
@@ -474,6 +516,8 @@ sub main'function_end_B
         %label=();
         }
 
+sub main'function_end_C { function_end_B(@_); }
+
 sub main'wparam
         {
         local($num)=@_;
@@ -510,7 +554,7 @@ sub main'swtmp
 
 sub main'comment
         {
-        if (!defined($com_start) or $main'elf)
+        if (!defined($com_start) or (!$main'openbsd && $main'elf))
                 {       # Regarding $main'elf above...
                         # GNU and SVR4 as'es use different comment delimiters,
                 push(@out,"\n");        # so we just skip ELF comments...
@@ -731,7 +775,9 @@ sub main'initseg
                 {
                 $tmp=<<___;
 .section        .init
-        call    $under$f
+                PIC_PROLOGUE
+                call    PIC_PLT($under$f)
+                PIC_EPILOGUE
         jmp     .Linitalign
 .align  $align
 .Linitalign:
diff --git a/src/lib/libssl/src/crypto/rc4/asm/rc4-x86_64.pl b/src/lib/libssl/src/crypto/rc4/asm/rc4-x86_64.pl
index 2d47320485..92c52f3433 100755
--- a/src/lib/libssl/src/crypto/rc4/asm/rc4-x86_64.pl
+++ b/src/lib/libssl/src/crypto/rc4/asm/rc4-x86_64.pl
@@ -269,8 +269,7 @@ RC4_set_key:
         xor     $ido,$ido
         xor     %r10,%r10
         xor     %r11,%r11
-
-        mov     OPENSSL_ia32cap_P(%rip),$idx#d
+        mov     PIC_GOT(OPENSSL_ia32cap_P),$idx#d
         bt      \$20,$idx#d
         jnc     .Lw1stloop
         bt      \$30,$idx#d
@@ -338,7 +337,7 @@ RC4_set_key:
 RC4_options:
         .picmeup %rax
         lea     .Lopts-.(%rax),%rax
-        mov     OPENSSL_ia32cap_P(%rip),%edx
+        mov     PIC_GOT(OPENSSL_ia32cap_P),%edx
         bt      \$20,%edx
         jnc     .Ldone
         add     \$12,%rax
diff --git a/src/lib/libssl/src/crypto/x86_64cpuid.pl b/src/lib/libssl/src/crypto/x86_64cpuid.pl
index 2616a03da6..8946b464a8 100644
--- a/src/lib/libssl/src/crypto/x86_64cpuid.pl
+++ b/src/lib/libssl/src/crypto/x86_64cpuid.pl
@@ -47,6 +47,8 @@ CRT\$XIU	ENDS
 
 ___
 print<<___ if(!defined($masm));
+#include <machine/asm.h>
+
 .text
 
 .globl  OPENSSL_atomic_add
@@ -95,7 +97,7 @@ OPENSSL_wipe_cpu:
 .size   OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 
 .section        .init
-        call    OPENSSL_cpuid_setup
+        call    PIC_PLT(OPENSSL_cpuid_setup)
 
 ___