openssl/openssl-1.1.1-intel-cet.patch

diff -up openssl-1.1.1d/crypto/aes/asm/aesni-x86_64.pl.intel-cet openssl-1.1.1d/crypto/aes/asm/aesni-x86_64.pl
--- openssl-1.1.1d/crypto/aes/asm/aesni-x86_64.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/aes/asm/aesni-x86_64.pl	2020-02-17 12:00:19.011235601 +0100
@@ -275,6 +275,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_encrypt:
 .cfi_startproc
+	endbranch
 	movups	($inp),$inout0		# load input
 	mov	240($key),$rounds	# key->rounds
 ___
@@ -293,6 +294,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_decrypt:
 .cfi_startproc
+	endbranch
 	movups	($inp),$inout0		# load input
 	mov	240($key),$rounds	# key->rounds
 ___
@@ -613,6 +615,7 @@ $code.=<<___;
 .align	16
 aesni_ecb_encrypt:
 .cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0x58(%rsp),%rsp
@@ -984,6 +987,8 @@ $code.=<<___;
 .type	aesni_ccm64_encrypt_blocks,\@function,6
 .align	16
 aesni_ccm64_encrypt_blocks:
+.cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0x58(%rsp),%rsp
@@ -1066,6 +1071,7 @@ $code.=<<___ if ($win64);
 ___
 $code.=<<___;
 	ret
+.cfi_endproc
 .size	aesni_ccm64_encrypt_blocks,.-aesni_ccm64_encrypt_blocks
 ___
 ######################################################################
@@ -1074,6 +1080,8 @@ $code.=<<___;
 .type	aesni_ccm64_decrypt_blocks,\@function,6
 .align	16
 aesni_ccm64_decrypt_blocks:
+.cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0x58(%rsp),%rsp
@@ -1173,6 +1181,7 @@ $code.=<<___ if ($win64);
 ___
 $code.=<<___;
 	ret
+.cfi_endproc
 .size	aesni_ccm64_decrypt_blocks,.-aesni_ccm64_decrypt_blocks
 ___
 }
@@ -1199,6 +1208,7 @@ $code.=<<___;
 .align	16
 aesni_ctr32_encrypt_blocks:
 .cfi_startproc
+	endbranch
 	cmp	\$1,$len
 	jne	.Lctr32_bulk
 
@@ -1771,6 +1781,7 @@ $code.=<<___;
 .align	16
 aesni_xts_encrypt:
 .cfi_startproc
+	endbranch
 	lea	(%rsp),%r11			# frame pointer
 .cfi_def_cfa_register	%r11
 	push	%rbp
@@ -2254,6 +2265,7 @@ $code.=<<___;
 .align	16
 aesni_xts_decrypt:
 .cfi_startproc
+	endbranch
 	lea	(%rsp),%r11			# frame pointer
 .cfi_def_cfa_register	%r11
 	push	%rbp
@@ -2779,6 +2791,7 @@ $code.=<<___;
 .align	32
 aesni_ocb_encrypt:
 .cfi_startproc
+	endbranch
 	lea	(%rsp),%rax
 	push	%rbx
 .cfi_push	%rbx
@@ -3031,6 +3044,7 @@ $code.=<<___;
 .type	__ocb_encrypt6,\@abi-omnipotent
 .align	32
 __ocb_encrypt6:
+.cfi_startproc
 	 pxor		$rndkey0l,@offset[5]	# offset_i ^ round[0]
 	 movdqu		($L_p,$i1),@offset[1]
 	 movdqa		@offset[0],@offset[2]
@@ -3128,11 +3142,13 @@ __ocb_encrypt6:
 	aesenclast	@offset[4],$inout4
 	aesenclast	@offset[5],$inout5
 	ret
+.cfi_endproc
 .size	__ocb_encrypt6,.-__ocb_encrypt6
 
 .type	__ocb_encrypt4,\@abi-omnipotent
 .align	32
 __ocb_encrypt4:
+.cfi_startproc
 	 pxor		$rndkey0l,@offset[5]	# offset_i ^ round[0]
 	 movdqu		($L_p,$i1),@offset[1]
 	 movdqa		@offset[0],@offset[2]
@@ -3197,11 +3213,13 @@ __ocb_encrypt4:
 	aesenclast	@offset[2],$inout2
 	aesenclast	@offset[3],$inout3
 	ret
+.cfi_endproc
 .size	__ocb_encrypt4,.-__ocb_encrypt4
 
 .type	__ocb_encrypt1,\@abi-omnipotent
 .align	32
 __ocb_encrypt1:
+.cfi_startproc
 	 pxor		@offset[5],$inout5	# offset_i
 	 pxor		$rndkey0l,$inout5	# offset_i ^ round[0]
 	pxor		$inout0,$checksum	# accumulate checksum
@@ -3232,6 +3250,7 @@ __ocb_encrypt1:
 
 	aesenclast	$inout5,$inout0
 	ret
+.cfi_endproc
 .size	__ocb_encrypt1,.-__ocb_encrypt1
 
 .globl	aesni_ocb_decrypt
@@ -3239,6 +3258,7 @@ __ocb_encrypt1:
 .align	32
 aesni_ocb_decrypt:
 .cfi_startproc
+	endbranch
 	lea	(%rsp),%rax
 	push	%rbx
 .cfi_push	%rbx
@@ -3513,6 +3533,7 @@ $code.=<<___;
 .type	__ocb_decrypt6,\@abi-omnipotent
 .align	32
 __ocb_decrypt6:
+.cfi_startproc
 	 pxor		$rndkey0l,@offset[5]	# offset_i ^ round[0]
 	 movdqu		($L_p,$i1),@offset[1]
 	 movdqa		@offset[0],@offset[2]
@@ -3604,11 +3625,13 @@ __ocb_decrypt6:
 	aesdeclast	@offset[4],$inout4
 	aesdeclast	@offset[5],$inout5
 	ret
+.cfi_endproc
 .size	__ocb_decrypt6,.-__ocb_decrypt6
 
 .type	__ocb_decrypt4,\@abi-omnipotent
 .align	32
 __ocb_decrypt4:
+.cfi_startproc
 	 pxor		$rndkey0l,@offset[5]	# offset_i ^ round[0]
 	 movdqu		($L_p,$i1),@offset[1]
 	 movdqa		@offset[0],@offset[2]
@@ -3669,11 +3692,13 @@ __ocb_decrypt4:
 	aesdeclast	@offset[2],$inout2
 	aesdeclast	@offset[3],$inout3
 	ret
+.cfi_endproc
 .size	__ocb_decrypt4,.-__ocb_decrypt4
 
 .type	__ocb_decrypt1,\@abi-omnipotent
 .align	32
 __ocb_decrypt1:
+.cfi_startproc
 	 pxor		@offset[5],$inout5	# offset_i
 	 pxor		$rndkey0l,$inout5	# offset_i ^ round[0]
 	pxor		$inout5,$inout0		# input ^ round[0] ^ offset_i
@@ -3703,6 +3728,7 @@ __ocb_decrypt1:
 
 	aesdeclast	$inout5,$inout0
 	ret
+.cfi_endproc
 .size	__ocb_decrypt1,.-__ocb_decrypt1
 ___
 } }}
@@ -3721,6 +3747,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_cbc_encrypt:
 .cfi_startproc
+	endbranch
 	test	$len,$len		# check length
 	jz	.Lcbc_ret
 
@@ -4637,7 +4664,6 @@ __aesni_set_encrypt_key:
 	add	\$8,%rsp
 .cfi_adjust_cfa_offset	-8
 	ret
-.cfi_endproc
 .LSEH_end_set_encrypt_key:
 
 .align	16
@@ -4708,6 +4734,7 @@ __aesni_set_encrypt_key:
 	shufps	\$0b10101010,%xmm1,%xmm1	# critical path
 	xorps	%xmm1,%xmm2
 	ret
+.cfi_endproc
 .size	${PREFIX}_set_encrypt_key,.-${PREFIX}_set_encrypt_key
 .size	__aesni_set_encrypt_key,.-__aesni_set_encrypt_key
 ___
diff -up openssl-1.1.1d/crypto/aes/asm/vpaes-x86_64.pl.intel-cet openssl-1.1.1d/crypto/aes/asm/vpaes-x86_64.pl
--- openssl-1.1.1d/crypto/aes/asm/vpaes-x86_64.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/aes/asm/vpaes-x86_64.pl	2020-02-17 11:55:07.374557249 +0100
@@ -696,6 +696,7 @@ _vpaes_schedule_mangle:
 .align	16
 ${PREFIX}_set_encrypt_key:
 .cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0xb8(%rsp),%rsp
@@ -746,6 +747,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_set_decrypt_key:
 .cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0xb8(%rsp),%rsp
@@ -801,6 +803,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_encrypt:
 .cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0xb8(%rsp),%rsp
@@ -846,6 +849,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_decrypt:
 .cfi_startproc
+	endbranch
 ___
 $code.=<<___ if ($win64);
 	lea	-0xb8(%rsp),%rsp
@@ -897,6 +901,7 @@ $code.=<<___;
 .align	16
 ${PREFIX}_cbc_encrypt:
 .cfi_startproc
+	endbranch
 	xchg	$key,$len
 ___
 ($len,$key)=($key,$len);
diff -up openssl-1.1.1d/crypto/async/arch/async_posix.c.intel-cet openssl-1.1.1d/crypto/async/arch/async_posix.c
--- openssl-1.1.1d/crypto/async/arch/async_posix.c.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/async/arch/async_posix.c	2020-02-17 11:55:07.374557249 +0100
@@ -34,7 +34,9 @@ void async_local_cleanup(void)
 
 int async_fibre_makecontext(async_fibre *fibre)
 {
+#ifndef USE_SWAPCONTEXT
     fibre->env_init = 0;
+#endif
     if (getcontext(&fibre->fibre) == 0) {
         fibre->fibre.uc_stack.ss_sp = OPENSSL_malloc(STACKSIZE);
         if (fibre->fibre.uc_stack.ss_sp != NULL) {
diff -up openssl-1.1.1d/crypto/async/arch/async_posix.h.intel-cet openssl-1.1.1d/crypto/async/arch/async_posix.h
--- openssl-1.1.1d/crypto/async/arch/async_posix.h.intel-cet	2020-02-17 11:55:06.600570492 +0100
+++ openssl-1.1.1d/crypto/async/arch/async_posix.h	2020-02-17 11:55:07.374557249 +0100
@@ -25,17 +25,33 @@
 #  define ASYNC_POSIX
 #  define ASYNC_ARCH
 
+#  ifdef __CET__
+/*
+ * When Intel CET is enabled, makecontext will create a different
+ * shadow stack for each context.  async_fibre_swapcontext cannot
+ * use _longjmp.  It must call swapcontext to swap shadow stack as
+ * well as normal stack.
+ */
+#   define USE_SWAPCONTEXT
+#  endif
 #  include <ucontext.h>
-#  include <setjmp.h>
+#  ifndef USE_SWAPCONTEXT
+#   include <setjmp.h>
+#  endif
 
 typedef struct async_fibre_st {
     ucontext_t fibre;
+#  ifndef USE_SWAPCONTEXT
     jmp_buf env;
     int env_init;
+#  endif
 } async_fibre;
 
 static ossl_inline int async_fibre_swapcontext(async_fibre *o, async_fibre *n, int r)
 {
+#  ifdef USE_SWAPCONTEXT
+    swapcontext(&o->fibre, &n->fibre);
+#  else
     o->env_init = 1;
 
     if (!r || !_setjmp(o->env)) {
@@ -44,6 +60,7 @@ static ossl_inline int async_fibre_swapc
         else
             setcontext(&n->fibre);
     }
+#  endif
 
     return 1;
 }
diff -up openssl-1.1.1d/crypto/camellia/asm/cmll-x86_64.pl.intel-cet openssl-1.1.1d/crypto/camellia/asm/cmll-x86_64.pl
--- openssl-1.1.1d/crypto/camellia/asm/cmll-x86_64.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/camellia/asm/cmll-x86_64.pl	2020-02-17 11:55:07.375557232 +0100
@@ -677,6 +677,7 @@ $code.=<<___;
 .align	16
 Camellia_cbc_encrypt:
 .cfi_startproc
+	endbranch
 	cmp	\$0,%rdx
 	je	.Lcbc_abort
 	push	%rbx
diff -up openssl-1.1.1d/crypto/modes/asm/ghash-x86_64.pl.intel-cet openssl-1.1.1d/crypto/modes/asm/ghash-x86_64.pl
--- openssl-1.1.1d/crypto/modes/asm/ghash-x86_64.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/modes/asm/ghash-x86_64.pl	2020-02-17 11:55:07.375557232 +0100
@@ -239,6 +239,7 @@ $code=<<___;
 .align	16
 gcm_gmult_4bit:
 .cfi_startproc
+	endbranch
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp		# %rbp and others are pushed exclusively in
@@ -286,6 +287,7 @@ $code.=<<___;
 .align	16
 gcm_ghash_4bit:
 .cfi_startproc
+	endbranch
 	push	%rbx
 .cfi_push	%rbx
 	push	%rbp
@@ -612,6 +614,7 @@ $code.=<<___;
 .align	16
 gcm_gmult_clmul:
 .cfi_startproc
+	endbranch
 .L_gmult_clmul:
 	movdqu		($Xip),$Xi
 	movdqa		.Lbswap_mask(%rip),$T3
@@ -663,6 +666,7 @@ $code.=<<___;
 .align	32
 gcm_ghash_clmul:
 .cfi_startproc
+	endbranch
 .L_ghash_clmul:
 ___
 $code.=<<___ if ($win64);
@@ -1166,6 +1170,7 @@ $code.=<<___;
 .align	32
 gcm_gmult_avx:
 .cfi_startproc
+	endbranch
 	jmp	.L_gmult_clmul
 .cfi_endproc
 .size	gcm_gmult_avx,.-gcm_gmult_avx
@@ -1177,6 +1182,7 @@ $code.=<<___;
 .align	32
 gcm_ghash_avx:
 .cfi_startproc
+	endbranch
 ___
 if ($avx) {
 my ($Xip,$Htbl,$inp,$len)=@_4args;
diff -up openssl-1.1.1d/crypto/perlasm/cbc.pl.intel-cet openssl-1.1.1d/crypto/perlasm/cbc.pl
--- openssl-1.1.1d/crypto/perlasm/cbc.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/perlasm/cbc.pl	2020-02-17 11:55:07.375557232 +0100
@@ -165,21 +165,28 @@ sub cbc
 	&jmp_ptr($count);
 
 &set_label("ej7");
+	&endbranch()
 	&movb(&HB("edx"),	&BP(6,$in,"",0));
 	&shl("edx",8);
 &set_label("ej6");
+	&endbranch()
 	&movb(&HB("edx"),	&BP(5,$in,"",0));
 &set_label("ej5");
+	&endbranch()
 	&movb(&LB("edx"),	&BP(4,$in,"",0));
 &set_label("ej4");
+	&endbranch()
 	&mov("ecx",		&DWP(0,$in,"",0));
 	&jmp(&label("ejend"));
 &set_label("ej3");
+	&endbranch()
 	&movb(&HB("ecx"),	&BP(2,$in,"",0));
 	&shl("ecx",8);
 &set_label("ej2");
+	&endbranch()
 	&movb(&HB("ecx"),	&BP(1,$in,"",0));
 &set_label("ej1");
+	&endbranch()
 	&movb(&LB("ecx"),	&BP(0,$in,"",0));
 &set_label("ejend");
 
diff -up openssl-1.1.1d/crypto/perlasm/x86_64-xlate.pl.intel-cet openssl-1.1.1d/crypto/perlasm/x86_64-xlate.pl
--- openssl-1.1.1d/crypto/perlasm/x86_64-xlate.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/perlasm/x86_64-xlate.pl	2020-02-17 11:55:07.375557232 +0100
@@ -101,6 +101,33 @@ elsif (!$gas)
     $decor="\$L\$";
 }
 
+my $cet_property;
+if ($flavour =~ /elf/) {
+	# Always generate .note.gnu.property section for ELF outputs to
+	# mark Intel CET support since all input files must be marked
+	# with Intel CET support in order for linker to mark output with
+	# Intel CET support.
+	my $p2align=3; $p2align=2 if ($flavour eq "elf32");
+	$cet_property = <<_____;
+	.section ".note.gnu.property", "a"
+	.p2align $p2align
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz "GNU"
+1:
+	.p2align $p2align
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align $p2align
+4:
+_____
+}
+
 my $current_segment;
 my $current_function;
 my %globals;
@@ -1213,6 +1240,7 @@ while(defined(my $line=<>)) {
     print $line,"\n";
 }
 
+print "$cet_property"			if ($cet_property);
 print "\n$current_segment\tENDS\n"	if ($current_segment && $masm);
 print "END\n"				if ($masm);
 
diff -up openssl-1.1.1d/crypto/perlasm/x86gas.pl.intel-cet openssl-1.1.1d/crypto/perlasm/x86gas.pl
--- openssl-1.1.1d/crypto/perlasm/x86gas.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/perlasm/x86gas.pl	2020-02-17 11:55:07.376557215 +0100
@@ -124,6 +124,7 @@ sub ::function_begin_B
     push(@out,".align\t$align\n");
     push(@out,"$func:\n");
     push(@out,"$begin:\n")		if ($global);
+    &::endbranch();
     $::stack=4;
 }
 
@@ -172,6 +173,26 @@ sub ::file_end
 	else		{ push (@out,"$tmp\n"); }
     }
     push(@out,$initseg) if ($initseg);
+    if ($::elf) {
+	push(@out,"
+	.section \".note.gnu.property\", \"a\"
+	.p2align 2
+	.long 1f - 0f
+	.long 4f - 1f
+	.long 5
+0:
+	.asciz \"GNU\"
+1:
+	.p2align 2
+	.long 0xc0000002
+	.long 3f - 2f
+2:
+	.long 3
+3:
+	.p2align 2
+4:
+");
+    }
 }
 
 sub ::data_byte	{   push(@out,".byte\t".join(',',@_)."\n");   }
diff -up openssl-1.1.1d/crypto/poly1305/asm/poly1305-x86_64.pl.intel-cet openssl-1.1.1d/crypto/poly1305/asm/poly1305-x86_64.pl
--- openssl-1.1.1d/crypto/poly1305/asm/poly1305-x86_64.pl.intel-cet	2020-02-17 11:55:07.376557215 +0100
+++ openssl-1.1.1d/crypto/poly1305/asm/poly1305-x86_64.pl	2020-02-17 12:02:12.295308065 +0100
@@ -90,7 +90,7 @@ if (!$avx && $win64 && ($flavour =~ /mas
 	$avx = ($1>=10) + ($1>=12);
 }
 
-if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
+if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([0-9]+\.[0-9]+)/) {
 	$avx = ($2>=3.0) + ($2>3.0);
 }
 
@@ -168,6 +168,7 @@ $code.=<<___;
 .type	poly1305_init,\@function,3
 .align	32
 poly1305_init:
+.cfi_startproc
 	xor	%rax,%rax
 	mov	%rax,0($ctx)		# initialize hash value
 	mov	%rax,8($ctx)
@@ -219,6 +220,7 @@ $code.=<<___;
 	mov	\$1,%eax
 .Lno_key:
 	ret
+.cfi_endproc
 .size	poly1305_init,.-poly1305_init
 
 .type	poly1305_blocks,\@function,4
@@ -298,6 +300,7 @@ $code.=<<___;
 .type	poly1305_emit,\@function,3
 .align	32
 poly1305_emit:
+.cfi_startproc
 .Lemit:
 	mov	0($ctx),%r8	# load hash value
 	mov	8($ctx),%r9
@@ -318,6 +321,7 @@ poly1305_emit:
 	mov	%rcx,8($mac)
 
 	ret
+.cfi_endproc
 .size	poly1305_emit,.-poly1305_emit
 ___
 if ($avx) {
@@ -342,15 +346,18 @@ $code.=<<___;
 .type	__poly1305_block,\@abi-omnipotent
 .align	32
 __poly1305_block:
+.cfi_startproc
 ___
 	&poly1305_iteration();
 $code.=<<___;
 	ret
+.cfi_endproc
 .size	__poly1305_block,.-__poly1305_block
 
 .type	__poly1305_init_avx,\@abi-omnipotent
 .align	32
 __poly1305_init_avx:
+.cfi_startproc
 	mov	$r0,$h0
 	mov	$r1,$h1
 	xor	$h2,$h2
@@ -508,6 +515,7 @@ __poly1305_init_avx:
 
 	lea	-48-64($ctx),$ctx	# size [de-]optimization
 	ret
+.cfi_endproc
 .size	__poly1305_init_avx,.-__poly1305_init_avx
 
 .type	poly1305_blocks_avx,\@function,4
@@ -1373,6 +1381,7 @@ $code.=<<___;
 .type	poly1305_emit_avx,\@function,3
 .align	32
 poly1305_emit_avx:
+.cfi_startproc
 	cmpl	\$0,20($ctx)	# is_base2_26?
 	je	.Lemit
 
@@ -1423,6 +1432,7 @@ poly1305_emit_avx:
 	mov	%rcx,8($mac)
 
 	ret
+.cfi_endproc
 .size	poly1305_emit_avx,.-poly1305_emit_avx
 ___
 
@@ -2741,6 +2751,7 @@ $code.=<<___;
 .type	poly1305_init_base2_44,\@function,3
 .align	32
 poly1305_init_base2_44:
+.cfi_startproc
 	xor	%rax,%rax
 	mov	%rax,0($ctx)		# initialize hash value
 	mov	%rax,8($ctx)
@@ -2782,6 +2793,7 @@ ___
 $code.=<<___;
 	mov	\$1,%eax
 	ret
+.cfi_endproc
 .size	poly1305_init_base2_44,.-poly1305_init_base2_44
 ___
 {
@@ -2793,6 +2805,8 @@ $code.=<<___;
 .type	poly1305_blocks_vpmadd52,\@function,4
 .align	32
 poly1305_blocks_vpmadd52:
+.cfi_startproc
+	endbranch
 	shr	\$4,$len
 	jz	.Lno_data_vpmadd52		# too short
 
@@ -2899,6 +2913,7 @@ poly1305_blocks_vpmadd52:
 
 .Lno_data_vpmadd52:
 	ret
+.cfi_endproc
 .size	poly1305_blocks_vpmadd52,.-poly1305_blocks_vpmadd52
 ___
 }
@@ -2916,6 +2931,7 @@ $code.=<<___;
 .type	poly1305_blocks_vpmadd52_4x,\@function,4
 .align	32
 poly1305_blocks_vpmadd52_4x:
+.cfi_startproc
 	shr	\$4,$len
 	jz	.Lno_data_vpmadd52_4x		# too short
 
@@ -3340,6 +3356,7 @@ poly1305_blocks_vpmadd52_4x:
 
 .Lno_data_vpmadd52_4x:
 	ret
+.cfi_endproc
 .size	poly1305_blocks_vpmadd52_4x,.-poly1305_blocks_vpmadd52_4x
 ___
 }
@@ -3358,6 +3375,7 @@ $code.=<<___;
 .type	poly1305_blocks_vpmadd52_8x,\@function,4
 .align	32
 poly1305_blocks_vpmadd52_8x:
+.cfi_startproc
 	shr	\$4,$len
 	jz	.Lno_data_vpmadd52_8x		# too short
 
@@ -3713,6 +3731,7 @@ $code.=<<___;
 
 .Lno_data_vpmadd52_8x:
 	ret
+.cfi_endproc
 .size	poly1305_blocks_vpmadd52_8x,.-poly1305_blocks_vpmadd52_8x
 ___
 }
@@ -3720,6 +3739,8 @@ $code.=<<___;
 .type	poly1305_emit_base2_44,\@function,3
 .align	32
 poly1305_emit_base2_44:
+.cfi_startproc
+	endbranch
 	mov	0($ctx),%r8	# load hash value
 	mov	8($ctx),%r9
 	mov	16($ctx),%r10
@@ -3750,6 +3771,7 @@ poly1305_emit_base2_44:
 	mov	%rcx,8($mac)
 
 	ret
+.cfi_endproc
 .size	poly1305_emit_base2_44,.-poly1305_emit_base2_44
 ___
 }	}	}
@@ -3800,6 +3822,7 @@ $code.=<<___;
 .type	xor128_encrypt_n_pad,\@abi-omnipotent
 .align	16
 xor128_encrypt_n_pad:
+.cfi_startproc
 	sub	$otp,$inp
 	sub	$otp,$out
 	mov	$len,%r10		# put len aside
@@ -3841,12 +3864,14 @@ xor128_encrypt_n_pad:
 .Ldone_enc:
 	mov	$otp,%rax
 	ret
+.cfi_endproc
 .size	xor128_encrypt_n_pad,.-xor128_encrypt_n_pad
 
 .globl	xor128_decrypt_n_pad
 .type	xor128_decrypt_n_pad,\@abi-omnipotent
 .align	16
 xor128_decrypt_n_pad:
+.cfi_startproc
 	sub	$otp,$inp
 	sub	$otp,$out
 	mov	$len,%r10		# put len aside
@@ -3892,6 +3917,7 @@ xor128_decrypt_n_pad:
 .Ldone_dec:
 	mov	$otp,%rax
 	ret
+.cfi_endproc
 .size	xor128_decrypt_n_pad,.-xor128_decrypt_n_pad
 ___
 }
diff -up openssl-1.1.1d/crypto/rc4/asm/rc4-x86_64.pl.intel-cet openssl-1.1.1d/crypto/rc4/asm/rc4-x86_64.pl
--- openssl-1.1.1d/crypto/rc4/asm/rc4-x86_64.pl.intel-cet	2020-02-17 11:55:07.377557198 +0100
+++ openssl-1.1.1d/crypto/rc4/asm/rc4-x86_64.pl	2020-02-17 12:03:09.117341235 +0100
@@ -138,11 +138,13 @@ $code=<<___;
 .globl	RC4
 .type	RC4,\@function,4
 .align	16
-RC4:	or	$len,$len
+RC4:
+.cfi_startproc
+	endbranch
+	or	$len,$len
 	jne	.Lentry
 	ret
 .Lentry:
-.cfi_startproc
 	push	%rbx
 .cfi_push	%rbx
 	push	%r12
@@ -453,6 +455,8 @@ $code.=<<___;
 .type	RC4_set_key,\@function,3
 .align	16
 RC4_set_key:
+.cfi_startproc
+	endbranch
 	lea	8($dat),$dat
 	lea	($inp,$len),$inp
 	neg	$len
@@ -519,12 +523,15 @@ RC4_set_key:
 	mov	%eax,-8($dat)
 	mov	%eax,-4($dat)
 	ret
+.cfi_endproc
 .size	RC4_set_key,.-RC4_set_key
 
 .globl	RC4_options
 .type	RC4_options,\@abi-omnipotent
 .align	16
 RC4_options:
+.cfi_startproc
+	endbranch
 	lea	.Lopts(%rip),%rax
 	mov	OPENSSL_ia32cap_P(%rip),%edx
 	bt	\$20,%edx
@@ -537,6 +544,7 @@ RC4_options:
 	add	\$12,%rax
 .Ldone:
 	ret
+.cfi_endproc
 .align	64
 .Lopts:
 .asciz	"rc4(8x,int)"
diff -up openssl-1.1.1d/crypto/x86_64cpuid.pl.intel-cet openssl-1.1.1d/crypto/x86_64cpuid.pl
--- openssl-1.1.1d/crypto/x86_64cpuid.pl.intel-cet	2019-09-10 15:13:07.000000000 +0200
+++ openssl-1.1.1d/crypto/x86_64cpuid.pl	2020-02-17 12:04:04.921391729 +0100
@@ -39,6 +39,8 @@ print<<___;
 .type	OPENSSL_atomic_add,\@abi-omnipotent
 .align	16
 OPENSSL_atomic_add:
+.cfi_startproc
+	endbranch
 	movl	($arg1),%eax
 .Lspin:	leaq	($arg2,%rax),%r8
 	.byte	0xf0		# lock
@@ -47,16 +49,20 @@ OPENSSL_atomic_add:
 	movl	%r8d,%eax
 	.byte	0x48,0x98	# cltq/cdqe
 	ret
+.cfi_endproc
 .size	OPENSSL_atomic_add,.-OPENSSL_atomic_add
 
 .globl	OPENSSL_rdtsc
 .type	OPENSSL_rdtsc,\@abi-omnipotent
 .align	16
 OPENSSL_rdtsc:
+.cfi_startproc
+	endbranch
 	rdtsc
 	shl	\$32,%rdx
 	or	%rdx,%rax
 	ret
+.cfi_endproc
 .size	OPENSSL_rdtsc,.-OPENSSL_rdtsc
 
 .globl	OPENSSL_ia32_cpuid
@@ -64,6 +70,7 @@ OPENSSL_rdtsc:
 .align	16
 OPENSSL_ia32_cpuid:
 .cfi_startproc
+	endbranch
 	mov	%rbx,%r8		# save %rbx
 .cfi_register	%rbx,%r8
 
@@ -232,6 +239,8 @@ OPENSSL_ia32_cpuid:
 .type   OPENSSL_cleanse,\@abi-omnipotent
 .align  16
 OPENSSL_cleanse:
+.cfi_startproc
+	endbranch
 	xor	%rax,%rax
 	cmp	\$15,$arg2
 	jae	.Lot
@@ -261,12 +270,15 @@ OPENSSL_cleanse:
 	cmp	\$0,$arg2
 	jne	.Little
 	ret
+.cfi_endproc
 .size	OPENSSL_cleanse,.-OPENSSL_cleanse
 
 .globl  CRYPTO_memcmp
 .type   CRYPTO_memcmp,\@abi-omnipotent
 .align  16
 CRYPTO_memcmp:
+.cfi_startproc
+	endbranch
 	xor	%rax,%rax
 	xor	%r10,%r10
 	cmp	\$0,$arg3
@@ -295,6 +307,7 @@ CRYPTO_memcmp:
 	shr	\$63,%rax
 .Lno_data:
 	ret
+.cfi_endproc
 .size	CRYPTO_memcmp,.-CRYPTO_memcmp
 ___
 
@@ -303,6 +316,8 @@ print<<___ if (!$win64);
 .type	OPENSSL_wipe_cpu,\@abi-omnipotent
 .align	16
 OPENSSL_wipe_cpu:
+.cfi_startproc
+	endbranch
 	pxor	%xmm0,%xmm0
 	pxor	%xmm1,%xmm1
 	pxor	%xmm2,%xmm2
@@ -329,6 +344,7 @@ OPENSSL_wipe_cpu:
 	xorq	%r11,%r11
 	leaq	8(%rsp),%rax
 	ret
+.cfi_endproc
 .size	OPENSSL_wipe_cpu,.-OPENSSL_wipe_cpu
 ___
 print<<___ if ($win64);
@@ -365,6 +381,8 @@ print<<___;
 .type	OPENSSL_instrument_bus,\@abi-omnipotent
 .align	16
 OPENSSL_instrument_bus:
+.cfi_startproc
+	endbranch
 	mov	$arg1,$out	# tribute to Win64
 	mov	$arg2,$cnt
 	mov	$arg2,$max
@@ -391,12 +409,15 @@ OPENSSL_instrument_bus:
 
 	mov	$max,%rax
 	ret
+.cfi_endproc
 .size	OPENSSL_instrument_bus,.-OPENSSL_instrument_bus
 
 .globl	OPENSSL_instrument_bus2
 .type	OPENSSL_instrument_bus2,\@abi-omnipotent
 .align	16
 OPENSSL_instrument_bus2:
+.cfi_startproc
+	endbranch
 	mov	$arg1,$out	# tribute to Win64
 	mov	$arg2,$cnt
 	mov	$arg3,$max
@@ -439,6 +460,7 @@ OPENSSL_instrument_bus2:
 	mov	$redzone(%rsp),%rax
 	sub	$cnt,%rax
 	ret
+.cfi_endproc
 .size	OPENSSL_instrument_bus2,.-OPENSSL_instrument_bus2
 ___
 }
@@ -450,6 +472,8 @@ print<<___;
 .type	OPENSSL_ia32_${rdop}_bytes,\@abi-omnipotent
 .align	16
 OPENSSL_ia32_${rdop}_bytes:
+.cfi_startproc
+	endbranch
 	xor	%rax, %rax	# return value
 	cmp	\$0,$arg2
 	je	.Ldone_${rdop}_bytes
@@ -486,6 +510,7 @@ OPENSSL_ia32_${rdop}_bytes:
 .Ldone_${rdop}_bytes:
 	xor	%r10,%r10	# Clear sensitive data from register
 	ret
+.cfi_endproc
 .size	OPENSSL_ia32_${rdop}_bytes,.-OPENSSL_ia32_${rdop}_bytes
 ___
 }