Add endbr64 where needed by inspection. Passes regresson tests.

ok jsing, and kind of tb an earlier version

5 files changed, 60 insertions, 0 deletions

diff --git a/src/lib/libcrypto/aes/asm/aes-x86_64.pl b/src/lib/libcrypto/aes/asm/aes-x86_64.pl
index b7399b552a..78ba20ca59 100755
--- a/src/lib/libcrypto/aes/asm/aes-x86_64.pl
+++ b/src/lib/libcrypto/aes/asm/aes-x86_64.pl
@@ -318,6 +318,7 @@ $code.=<<___;
 .type   _x86_64_AES_encrypt,\@abi-omnipotent
 .align  16
 _x86_64_AES_encrypt:
+        endbr64
         xor     0($key),$s0                     # xor with key
         xor     4($key),$s1
         xor     8($key),$s2
@@ -548,6 +549,7 @@ $code.=<<___;
 .type   _x86_64_AES_encrypt_compact,\@abi-omnipotent
 .align  16
 _x86_64_AES_encrypt_compact:
+        endbr64
         lea     128($sbox),$inp                 # size optimization
         mov     0-128($inp),$acc1               # prefetch Te4
         mov     32-128($inp),$acc2
@@ -593,6 +595,7 @@ $code.=<<___;
 .hidden asm_AES_encrypt
 asm_AES_encrypt:
 AES_encrypt:
+        endbr64
         push    %rbx
         push    %rbp
         push    %r12
@@ -884,6 +887,7 @@ $code.=<<___;
 .type   _x86_64_AES_decrypt,\@abi-omnipotent
 .align  16
 _x86_64_AES_decrypt:
+        endbr64
         xor     0($key),$s0                     # xor with key
         xor     4($key),$s1
         xor     8($key),$s2
@@ -1138,6 +1142,7 @@ $code.=<<___;
 .type   _x86_64_AES_decrypt_compact,\@abi-omnipotent
 .align  16
 _x86_64_AES_decrypt_compact:
+        endbr64
         lea     128($sbox),$inp                 # size optimization
         mov     0-128($inp),$acc1               # prefetch Td4
         mov     32-128($inp),$acc2
@@ -1192,6 +1197,7 @@ $code.=<<___;
 .hidden asm_AES_decrypt
 asm_AES_decrypt:
 AES_decrypt:
+        endbr64
         push    %rbx
         push    %rbp
         push    %r12
@@ -1291,6 +1297,7 @@ $code.=<<___;
 .type   AES_set_encrypt_key,\@function,3
 .align  16
 AES_set_encrypt_key:
+        endbr64
         push    %rbx
         push    %rbp
         push    %r12                    # redundant, but allows to share 
@@ -1316,6 +1323,7 @@ AES_set_encrypt_key:
 .type   _x86_64_AES_set_encrypt_key,\@abi-omnipotent
 .align  16
 _x86_64_AES_set_encrypt_key:
+        endbr64
         mov     %esi,%ecx                       # %ecx=bits
         mov     %rdi,%rsi                       # %rsi=userKey
         mov     %rdx,%rdi                       # %rdi=key
@@ -1561,6 +1569,7 @@ $code.=<<___;
 .type   AES_set_decrypt_key,\@function,3
 .align  16
 AES_set_decrypt_key:
+        endbr64
         push    %rbx
         push    %rbp
         push    %r12
@@ -1660,6 +1669,7 @@ $code.=<<___;
 .hidden asm_AES_cbc_encrypt
 asm_AES_cbc_encrypt:
 AES_cbc_encrypt:
+        endbr64
         cmp     \$0,%rdx        # check length
         je      .Lcbc_epilogue
         pushfq
@@ -2551,6 +2561,7 @@ $code.=<<___;
 .type   block_se_handler,\@abi-omnipotent
 .align  16
 block_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2609,6 +2620,7 @@ block_se_handler:
 .type   key_se_handler,\@abi-omnipotent
 .align  16
 key_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2666,6 +2678,7 @@ key_se_handler:
 .type   cbc_se_handler,\@abi-omnipotent
 .align  16
 cbc_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
diff --git a/src/lib/libcrypto/aes/asm/aesni-sha1-x86_64.pl b/src/lib/libcrypto/aes/asm/aesni-sha1-x86_64.pl
index bafa906a05..879d16793f 100644
--- a/src/lib/libcrypto/aes/asm/aesni-sha1-x86_64.pl
+++ b/src/lib/libcrypto/aes/asm/aesni-sha1-x86_64.pl
@@ -89,6 +89,7 @@ $code.=<<___;
 .type   aesni_cbc_sha1_enc,\@abi-omnipotent
 .align  16
 aesni_cbc_sha1_enc:
+        endbr64
         # caller should check for SSSE3 and AES-NI bits
         mov     OPENSSL_ia32cap_P+0(%rip),%r10d
         mov     OPENSSL_ia32cap_P+4(%rip),%r11d
@@ -132,6 +133,7 @@ $code.=<<___;
 .type   aesni_cbc_sha1_enc_ssse3,\@function,6
 .align  16
 aesni_cbc_sha1_enc_ssse3:
+        endbr64
         mov     `($win64?56:8)`(%rsp),$inp      # load 7th argument
         #shr    \$6,$len                        # debugging artefact
         #jz     .Lepilogue_ssse3                # debugging artefact
@@ -650,6 +652,7 @@ $code.=<<___;
 .type   aesni_cbc_sha1_enc_avx,\@function,6
 .align  16
 aesni_cbc_sha1_enc_avx:
+        endbr64
         mov     `($win64?56:8)`(%rsp),$inp      # load 7th argument
         #shr    \$6,$len                        # debugging artefact
         #jz     .Lepilogue_avx                  # debugging artefact
@@ -1100,6 +1103,7 @@ $code.=<<___;
 .type   ssse3_handler,\@abi-omnipotent
 .align  16
 ssse3_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
diff --git a/src/lib/libcrypto/aes/asm/aesni-x86_64.pl b/src/lib/libcrypto/aes/asm/aesni-x86_64.pl
index d394e6f696..c261a37958 100644
--- a/src/lib/libcrypto/aes/asm/aesni-x86_64.pl
+++ b/src/lib/libcrypto/aes/asm/aesni-x86_64.pl
@@ -242,6 +242,7 @@ $code.=<<___;
 .type   ${PREFIX}_encrypt,\@abi-omnipotent
 .align  16
 ${PREFIX}_encrypt:
+        endbr64
         movups  ($inp),$inout0          # load input
         mov     240($key),$rounds       # key->rounds
 ___
@@ -255,6 +256,7 @@ $code.=<<___;
 .type   ${PREFIX}_decrypt,\@abi-omnipotent
 .align  16
 ${PREFIX}_decrypt:
+        endbr64
         movups  ($inp),$inout0          # load input
         mov     240($key),$rounds       # key->rounds
 ___
@@ -284,6 +286,7 @@ $code.=<<___;
 .type   _aesni_${dir}rypt3,\@abi-omnipotent
 .align  16
 _aesni_${dir}rypt3:
+        endbr64
         $movkey ($key),$rndkey0
         shr     \$1,$rounds
         $movkey 16($key),$rndkey1
@@ -328,6 +331,7 @@ $code.=<<___;
 .type   _aesni_${dir}rypt4,\@abi-omnipotent
 .align  16
 _aesni_${dir}rypt4:
+        endbr64
         $movkey ($key),$rndkey0
         shr     \$1,$rounds
         $movkey 16($key),$rndkey1
@@ -373,6 +377,7 @@ $code.=<<___;
 .type   _aesni_${dir}rypt6,\@abi-omnipotent
 .align  16
 _aesni_${dir}rypt6:
+        endbr64
         $movkey         ($key),$rndkey0
         shr             \$1,$rounds
         $movkey         16($key),$rndkey1
@@ -437,6 +442,7 @@ $code.=<<___;
 .type   _aesni_${dir}rypt8,\@abi-omnipotent
 .align  16
 _aesni_${dir}rypt8:
+        endbr64
         $movkey         ($key),$rndkey0
         shr             \$1,$rounds
         $movkey         16($key),$rndkey1
@@ -525,6 +531,7 @@ $code.=<<___;
 .type   aesni_ecb_encrypt,\@function,5
 .align  16
 aesni_ecb_encrypt:
+        endbr64
         and     \$-16,$len
         jz      .Lecb_ret
 
@@ -830,6 +837,7 @@ $code.=<<___;
 .type   aesni_ccm64_encrypt_blocks,\@function,6
 .align  16
 aesni_ccm64_encrypt_blocks:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         lea     -0x58(%rsp),%rsp
@@ -2478,6 +2486,7 @@ $code.=<<___;
 .type   ${PREFIX}_set_decrypt_key,\@abi-omnipotent
 .align  16
 ${PREFIX}_set_decrypt_key:
+        endbr64
         sub     \$8,%rsp
         call    __aesni_set_encrypt_key
         shl     \$4,$bits               # rounds-1 after _aesni_set_encrypt_key
@@ -2528,6 +2537,7 @@ $code.=<<___;
 .type   ${PREFIX}_set_encrypt_key,\@abi-omnipotent
 .align  16
 ${PREFIX}_set_encrypt_key:
+        endbr64
 __aesni_set_encrypt_key:
         sub     \$8,%rsp
         mov     \$-1,%rax
@@ -2749,6 +2759,7 @@ $code.=<<___ if ($PREFIX eq "aesni");
 .type   ecb_se_handler,\@abi-omnipotent
 .align  16
 ecb_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2768,6 +2779,7 @@ ecb_se_handler:
 .type   ccm64_se_handler,\@abi-omnipotent
 .align  16
 ccm64_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2809,6 +2821,7 @@ ccm64_se_handler:
 .type   ctr32_se_handler,\@abi-omnipotent
 .align  16
 ctr32_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2844,6 +2857,7 @@ ctr32_se_handler:
 .type   xts_se_handler,\@abi-omnipotent
 .align  16
 xts_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
@@ -2885,6 +2899,7 @@ $code.=<<___;
 .type   cbc_se_handler,\@abi-omnipotent
 .align  16
 cbc_se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
diff --git a/src/lib/libcrypto/aes/asm/bsaes-x86_64.pl b/src/lib/libcrypto/aes/asm/bsaes-x86_64.pl
index 20e9e1f71f..7098ba27f4 100644
--- a/src/lib/libcrypto/aes/asm/bsaes-x86_64.pl
+++ b/src/lib/libcrypto/aes/asm/bsaes-x86_64.pl
@@ -813,6 +813,7 @@ $code.=<<___;
 .type   _bsaes_encrypt8,\@abi-omnipotent
 .align  64
 _bsaes_encrypt8:
+        endbr64
         lea     .LBS0(%rip), $const     # constants table
 
         movdqa  ($key), @XMM[9]         # round 0 key
@@ -877,6 +878,7 @@ $code.=<<___;
 .type   _bsaes_decrypt8,\@abi-omnipotent
 .align  64
 _bsaes_decrypt8:
+        endbr64
         lea     .LBS0(%rip), $const     # constants table
 
         movdqa  ($key), @XMM[9]         # round 0 key
@@ -968,6 +970,7 @@ $code.=<<___;
 .type   _bsaes_key_convert,\@abi-omnipotent
 .align  16
 _bsaes_key_convert:
+        endbr64
         lea     .Lmasks(%rip), $const
         movdqu  ($inp), %xmm7           # load round 0 key
         lea     0x10($inp), $inp
@@ -1057,6 +1060,7 @@ $code.=<<___;
 .type   bsaes_enc_key_convert,\@function,2
 .align  16
 bsaes_enc_key_convert:
+        endbr64
         mov     240($inp),%r10d         # pass rounds
         mov     $inp,%rcx               # pass key
         mov     $out,%rax               # pass key schedule
@@ -1071,6 +1075,7 @@ bsaes_enc_key_convert:
 .align  16
 bsaes_encrypt_128:
 .Lenc128_loop:
+        endbr64
         movdqu  0x00($inp), @XMM[0]     # load input
         movdqu  0x10($inp), @XMM[1]
         movdqu  0x20($inp), @XMM[2]
@@ -1103,6 +1108,7 @@ bsaes_encrypt_128:
 .type   bsaes_dec_key_convert,\@function,2
 .align  16
 bsaes_dec_key_convert:
+        endbr64
         mov     240($inp),%r10d         # pass rounds
         mov     $inp,%rcx               # pass key
         mov     $out,%rax               # pass key schedule
@@ -1117,6 +1123,7 @@ bsaes_dec_key_convert:
 .type   bsaes_decrypt_128,\@function,4
 .align  16
 bsaes_decrypt_128:
+        endbr64
 .Ldec128_loop:
         movdqu  0x00($inp), @XMM[0]     # load input
         movdqu  0x10($inp), @XMM[1]
@@ -1162,6 +1169,7 @@ $code.=<<___;
 .type   bsaes_ecb_encrypt_blocks,\@abi-omnipotent
 .align  16
 bsaes_ecb_encrypt_blocks:
+        endbr64
         mov     %rsp, %rax
 .Lecb_enc_prologue:
         push    %rbp
@@ -1363,6 +1371,7 @@ $code.=<<___;
 .type   bsaes_ecb_decrypt_blocks,\@abi-omnipotent
 .align  16
 bsaes_ecb_decrypt_blocks:
+        endbr64
         mov     %rsp, %rax
 .Lecb_dec_prologue:
         push    %rbp
@@ -1568,6 +1577,7 @@ $code.=<<___;
 .type   bsaes_cbc_encrypt,\@abi-omnipotent
 .align  16
 bsaes_cbc_encrypt:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         mov     48(%rsp),$arg6          # pull direction flag
@@ -1855,6 +1865,7 @@ $code.=<<___;
 .type   bsaes_ctr32_encrypt_blocks,\@abi-omnipotent
 .align  16
 bsaes_ctr32_encrypt_blocks:
+        endbr64
         mov     %rsp, %rax
 .Lctr_enc_prologue:
         push    %rbp
@@ -2096,6 +2107,7 @@ $code.=<<___;
 .type   bsaes_xts_encrypt,\@abi-omnipotent
 .align  16
 bsaes_xts_encrypt:
+        endbr64
         mov     %rsp, %rax
 .Lxts_enc_prologue:
         push    %rbp
@@ -2477,6 +2489,7 @@ $code.=<<___;
 .type   bsaes_xts_decrypt,\@abi-omnipotent
 .align  16
 bsaes_xts_decrypt:
+        endbr64
         mov     %rsp, %rax
 .Lxts_dec_prologue:
         push    %rbp
@@ -2953,6 +2966,7 @@ $code.=<<___;
 .type   se_handler,\@abi-omnipotent
 .align  16
 se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx
diff --git a/src/lib/libcrypto/aes/asm/vpaes-x86_64.pl b/src/lib/libcrypto/aes/asm/vpaes-x86_64.pl
index 3ffb1a3038..8ff8d8602b 100644
--- a/src/lib/libcrypto/aes/asm/vpaes-x86_64.pl
+++ b/src/lib/libcrypto/aes/asm/vpaes-x86_64.pl
@@ -82,6 +82,7 @@ $code.=<<___;
 .type   _vpaes_encrypt_core,\@abi-omnipotent
 .align 16
 _vpaes_encrypt_core:
+        endbr64
         mov     %rdx,   %r9
         mov     \$16,   %r11
         mov     240(%rdx),%eax
@@ -172,6 +173,7 @@ _vpaes_encrypt_core:
 .type   _vpaes_decrypt_core,\@abi-omnipotent
 .align  16
 _vpaes_decrypt_core:
+        endbr64
         mov     %rdx,   %r9             # load key
         mov     240(%rdx),%eax
         movdqa  %xmm9,  %xmm1
@@ -279,6 +281,7 @@ _vpaes_decrypt_core:
 .type   _vpaes_schedule_core,\@abi-omnipotent
 .align  16
 _vpaes_schedule_core:
+        endbr64
         # rdi = key
         # rsi = size in bits
         # rdx = buffer
@@ -464,6 +467,7 @@ _vpaes_schedule_core:
 .type   _vpaes_schedule_192_smear,\@abi-omnipotent
 .align  16
 _vpaes_schedule_192_smear:
+        endbr64
         pshufd  \$0x80, %xmm6,  %xmm0   # d c 0 0 -> c 0 0 0
         pxor    %xmm0,  %xmm6           # -> c+d c 0 0
         pshufd  \$0xFE, %xmm7,  %xmm0   # b a _ _ -> b b b a
@@ -495,6 +499,7 @@ _vpaes_schedule_192_smear:
 .type   _vpaes_schedule_round,\@abi-omnipotent
 .align  16
 _vpaes_schedule_round:
+        endbr64
         # extract rcon from xmm8
         pxor    %xmm1,  %xmm1
         palignr \$15,   %xmm8,  %xmm1
@@ -562,6 +567,7 @@ _vpaes_schedule_low_round:
 .type   _vpaes_schedule_transform,\@abi-omnipotent
 .align  16
 _vpaes_schedule_transform:
+        endbr64
         movdqa  %xmm9,  %xmm1
         pandn   %xmm0,  %xmm1
         psrld   \$4,    %xmm1
@@ -600,6 +606,7 @@ _vpaes_schedule_transform:
 .type   _vpaes_schedule_mangle,\@abi-omnipotent
 .align  16
 _vpaes_schedule_mangle:
+        endbr64
         movdqa  %xmm0,  %xmm4   # save xmm0 for later
         movdqa  .Lk_mc_forward(%rip),%xmm5
         test    %rcx,   %rcx
@@ -673,6 +680,7 @@ _vpaes_schedule_mangle:
 .type   ${PREFIX}_set_encrypt_key,\@function,3
 .align  16
 ${PREFIX}_set_encrypt_key:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         lea     -0xb8(%rsp),%rsp
@@ -721,6 +729,7 @@ $code.=<<___;
 .type   ${PREFIX}_set_decrypt_key,\@function,3
 .align  16
 ${PREFIX}_set_decrypt_key:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         lea     -0xb8(%rsp),%rsp
@@ -774,6 +783,7 @@ $code.=<<___;
 .type   ${PREFIX}_encrypt,\@function,3
 .align  16
 ${PREFIX}_encrypt:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         lea     -0xb8(%rsp),%rsp
@@ -817,6 +827,7 @@ $code.=<<___;
 .type   ${PREFIX}_decrypt,\@function,3
 .align  16
 ${PREFIX}_decrypt:
+        endbr64
 ___
 $code.=<<___ if ($win64);
         lea     -0xb8(%rsp),%rsp
@@ -866,6 +877,7 @@ $code.=<<___;
 .type   ${PREFIX}_cbc_encrypt,\@function,6
 .align  16
 ${PREFIX}_cbc_encrypt:
+        endbr64
         xchg    $key,$len
 ___
 ($len,$key)=($key,$len);
@@ -949,6 +961,7 @@ $code.=<<___;
 .type   _vpaes_preheat,\@abi-omnipotent
 .align  16
 _vpaes_preheat:
+        endbr64
         lea     .Lk_s0F(%rip), %r10
         movdqa  -0x20(%r10), %xmm10     # .Lk_inv
         movdqa  -0x10(%r10), %xmm11     # .Lk_inv+16
@@ -1079,6 +1092,7 @@ $code.=<<___;
 .type   se_handler,\@abi-omnipotent
 .align  16
 se_handler:
+        endbr64
         push    %rsi
         push    %rdi
         push    %rbx