2 # Copyright 2018 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the Apache License 2.0 (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
9 # ====================================================================
10 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
11 # project. The module is, however, dual licensed under OpenSSL and
12 # CRYPTOGAMS licenses depending on where you obtain it. For further
13 # details see http://www.openssl.org/~appro/cryptogams/.
14 # ====================================================================
16 # X25519 lower-level primitives for x86_64.
20 # This module implements radix 2^51 multiplication and squaring, and
21 # radix 2^64 multiplication, squaring, addition, subtraction and final
22 # reduction. Latter radix is used on ADCX/ADOX-capable processors such
23 # as Broadwell. On related note one should mention that there are
24 # vector implementations that provide significantly better performance
25 # on some processors(*), but they are large and overly complex. Which
26 # in combination with them being effectively processor-specific makes
27 # the undertaking hard to justify. The goal for this implementation
28 # is rather versatility and simplicity [and ultimately formal
31 # (*) For example sandy2x should provide ~30% improvement on Sandy
32 # Bridge, but only nominal ~5% on Haswell [and big loss on
33 # Broadwell and successors].
35 ######################################################################
36 # Improvement coefficients:
38 # amd64-51(*) gcc-5.x(**)
41 # Sandy Bridge -3% +11%
43 # Broadwell(***) +30% +35%
44 # Skylake(***) +33% +47%
45 # Silvermont +20% +26%
48 # Ryzen(***) +43% +40%
51 # (*) amd64-51 is popular assembly implementation with 2^51 radix,
52 # only multiplication and squaring subroutines were linked
53 # for comparison, but not complete ladder step; gain on most
54 # processors is because this module refrains from shld, and
55 # minor regression on others is because this does result in
56 # higher instruction count;
57 # (**) compiler is free to inline functions, in assembly one would
58 # need to implement ladder step to do that, and it will improve
59 # performance by several percent;
60 # (***) ADCX/ADOX result for 2^64 radix, there is no corresponding
61 # C implementation, so that comparison is always against
64 # $output is the last argument if it looks like a file (it has an extension)
65 # $flavour is the first argument if it doesn't look like a file
66 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
67 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
69 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
71 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
72 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
73 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
74 die "can't locate x86_64-xlate.pl";
76 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
77 or die "can't call $xlate: $!";
80 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
81 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
85 if (!$addx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
86 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
90 if (!$addx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
91 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
95 if (!$addx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9])\.([0-9]+)/) {
96 my $ver = $2 + $3/100.0; # 3.1->3.01, 3.10->3.10
103 .globl x25519_fe51_mul
104 .type x25519_fe51_mul,\@function,3
121 .cfi_adjust_cfa_offset 40
124 mov 8*0(%rsi),%rax # f[0]
125 mov 8*0(%rdx),%r11 # load g[0-4]
131 mov %rdi,8*4(%rsp) # offload 1st argument
133 mulq %r11 # f[0]*g[0]
134 mov %r11,8*0(%rsp) # offload g[0]
135 mov %rax,%rbx # %rbx:%rcx = h0
138 mulq %r12 # f[0]*g[1]
139 mov %r12,8*1(%rsp) # offload g[1]
140 mov %rax,%r8 # %r8:%r9 = h1
142 lea (%r14,%r14,8),%r15
144 mulq %r13 # f[0]*g[2]
145 mov %r13,8*2(%rsp) # offload g[2]
146 mov %rax,%r10 # %r10:%r11 = h2
148 lea (%r14,%r15,2),%rdi # g[4]*19
150 mulq %rbp # f[0]*g[3]
151 mov %rax,%r12 # %r12:%r13 = h3
152 mov 8*0(%rsi),%rax # f[0]
154 mulq %r14 # f[0]*g[4]
155 mov %rax,%r14 # %r14:%r15 = h4
156 mov 8*1(%rsi),%rax # f[1]
159 mulq %rdi # f[1]*g[4]*19
161 mov 8*2(%rsi),%rax # f[2]
163 mulq %rdi # f[2]*g[4]*19
165 mov 8*3(%rsi),%rax # f[3]
167 mulq %rdi # f[3]*g[4]*19
169 mov 8*4(%rsi),%rax # f[4]
171 mulq %rdi # f[4]*g[4]*19
172 imulq \$19,%rbp,%rdi # g[3]*19
174 mov 8*1(%rsi),%rax # f[1]
176 mulq %rbp # f[1]*g[3]
177 mov 8*2(%rsp),%rbp # g[2]
179 mov 8*2(%rsi),%rax # f[2]
182 mulq %rdi # f[2]*g[3]*19
184 mov 8*3(%rsi),%rax # f[3]
186 mulq %rdi # f[3]*g[3]*19
188 mov 8*4(%rsi),%rax # f[4]
190 mulq %rdi # f[4]*g[3]*19
191 imulq \$19,%rbp,%rdi # g[2]*19
193 mov 8*1(%rsi),%rax # f[1]
195 mulq %rbp # f[1]*g[2]
197 mov 8*2(%rsi),%rax # f[2]
199 mulq %rbp # f[2]*g[2]
200 mov 8*1(%rsp),%rbp # g[1]
202 mov 8*3(%rsi),%rax # f[3]
205 mulq %rdi # f[3]*g[2]*19
207 mov 8*4(%rsi),%rax # f[3]
209 mulq %rdi # f[4]*g[2]*19
211 mov 8*1(%rsi),%rax # f[1]
213 mulq %rbp # f[1]*g[1]
216 mov 8*2(%rsi),%rax # f[2]
218 mulq %rbp # f[2]*g[1]
220 mov 8*3(%rsi),%rax # f[3]
222 mulq %rbp # f[3]*g[1]
223 mov 8*0(%rsp),%rbp # g[0]
225 mov 8*4(%rsi),%rax # f[4]
228 mulq %rdi # f[4]*g[1]*19
230 mov 8*1(%rsi),%rax # f[1]
234 mov 8*2(%rsi),%rax # f[2]
238 mov 8*3(%rsi),%rax # f[3]
242 mov 8*4(%rsi),%rax # f[4]
244 mulq %rbp # f[4]*g[0]
248 mov 8*4(%rsp),%rdi # restore 1st argument
252 .size x25519_fe51_mul,.-x25519_fe51_mul
254 .globl x25519_fe51_sqr
255 .type x25519_fe51_sqr,\@function,2
272 .cfi_adjust_cfa_offset 40
275 mov 8*0(%rsi),%rax # g[0]
276 mov 8*2(%rsi),%r15 # g[2]
277 mov 8*4(%rsi),%rbp # g[4]
279 mov %rdi,8*4(%rsp) # offload 1st argument
281 mulq %rax # g[0]*g[0]
283 mov 8*1(%rsi),%rax # g[1]
285 mulq %r14 # 2*g[0]*g[1]
288 mov %r15,8*0(%rsp) # offload g[2]
290 mulq %r14 # 2*g[0]*g[2]
294 imulq \$19,%rbp,%rdi # g[4]*19
295 mulq %r14 # 2*g[0]*g[3]
299 mulq %r14 # 2*g[0]*g[4]
304 mulq %rdi # g[4]*g[4]*19
306 mov 8*1(%rsi),%rax # g[1]
309 mov 8*3(%rsi),%rsi # g[3]
311 mulq %rax # g[1]*g[1]
313 mov 8*0(%rsp),%rax # g[2]
315 mulq %rbp # 2*g[1]*g[2]
319 mulq %rsi # 2*g[1]*g[3]
323 imulq \$19,%rsi,%rbp # g[3]*19
324 mulq %rdi # 2*g[1]*g[4]*19
329 mulq %rdi # 2*g[3]*g[4]*19
333 mulq %rbp # g[3]*g[3]*19
335 mov 8*0(%rsp),%rax # g[2]
339 mulq %rax # g[2]*g[2]
343 mulq %rsi # 2*g[2]*g[3]*19
347 mulq %rdi # 2*g[2]*g[4]*19
351 mov 8*4(%rsp),%rdi # restore 1st argument
356 mov \$0x7ffffffffffff,%rbp
361 and %rbp,%rdx # %rdx = g2 = h2 & mask
362 or %r10,%r11 # h2>>51
364 adc \$0,%r13 # h3 += h2>>51
369 and %rbp,%rax # %rax = g0 = h0 & mask
370 or %rbx,%rcx # h0>>51
371 add %rcx,%r8 # h1 += h0>>51
377 and %rbp,%rbx # %rbx = g3 = h3 & mask
378 or %r12,%r13 # h3>>51
379 add %r13,%r14 # h4 += h3>>51
385 and %rbp,%rcx # %rcx = g1 = h1 & mask
387 add %r9,%rdx # g2 += h1>>51
392 and %rbp,%r10 # %r10 = g4 = h0 & mask
393 or %r14,%r15 # h0>>51
395 lea (%r15,%r15,8),%r14
396 lea (%r15,%r14,2),%r15
397 add %r15,%rax # g0 += (h0>>51)*19
400 and %rbp,%rdx # g2 &= mask
402 add %r8,%rbx # g3 += g2>>51
405 and %rbp,%rax # g0 &= mask
407 add %r9,%rcx # g1 += g0>>51
409 mov %rax,8*0(%rdi) # save the result
428 .cfi_adjust_cfa_offset 88
432 .size x25519_fe51_sqr,.-x25519_fe51_sqr
434 .globl x25519_fe51_mul121666
435 .type x25519_fe51_mul121666,\@function,2
437 x25519_fe51_mul121666:
452 .cfi_adjust_cfa_offset 40
453 .Lfe51_mul121666_body:
457 mov %rax,%rbx # %rbx:%rcx = h0
461 mov %rax,%r8 # %r8:%r9 = h1
465 mov %rax,%r10 # %r10:%r11 = h2
469 mov %rax,%r12 # %r12:%r13 = h3
470 mov \$121666,%eax # f[0]
473 mov %rax,%r14 # %r14:%r15 = h4
477 .Lfe51_mul121666_epilogue:
479 .size x25519_fe51_mul121666,.-x25519_fe51_mul121666
481 ########################################################################
482 # Base 2^64 subroutines modulo 2*(2^255-19)
485 my ($acc0,$acc1,$acc2,$acc3,$acc4,$acc5,$acc6,$acc7) = map("%r$_",(8..15));
488 .extern OPENSSL_ia32cap_P
489 .globl x25519_fe64_eligible
490 .type x25519_fe64_eligible,\@abi-omnipotent
492 x25519_fe64_eligible:
493 mov OPENSSL_ia32cap_P+8(%rip),%ecx
499 .size x25519_fe64_eligible,.-x25519_fe64_eligible
501 .globl x25519_fe64_mul
502 .type x25519_fe64_mul,\@function,3
518 push %rdi # offload dst
521 .cfi_adjust_cfa_offset 16
525 mov 8*0(%rdx),%rbp # b[0]
526 mov 8*0(%rsi),%rdx # a[0]
527 mov 8*1(%rax),%rcx # b[1]
528 mov 8*2(%rax),$acc6 # b[2]
529 mov 8*3(%rax),$acc7 # b[3]
531 mulx %rbp,$acc0,%rax # a[0]*b[0]
532 xor %edi,%edi # cf=0,of=0
533 mulx %rcx,$acc1,%rbx # a[0]*b[1]
535 mulx $acc6,$acc2,%rax # a[0]*b[2]
537 mulx $acc7,$acc3,$acc4 # a[0]*b[3]
538 mov 8*1(%rsi),%rdx # a[1]
540 mov $acc6,(%rsp) # offload b[2]
541 adcx %rdi,$acc4 # cf=0
543 mulx %rbp,%rax,%rbx # a[1]*b[0]
546 mulx %rcx,%rax,%rbx # a[1]*b[1]
549 mulx $acc6,%rax,%rbx # a[1]*b[2]
552 mulx $acc7,%rax,$acc5 # a[1]*b[3]
553 mov 8*2(%rsi),%rdx # a[2]
555 adcx %rdi,$acc5 # cf=0
556 adox %rdi,$acc5 # of=0
558 mulx %rbp,%rax,%rbx # a[2]*b[0]
561 mulx %rcx,%rax,%rbx # a[2]*b[1]
564 mulx $acc6,%rax,%rbx # a[2]*b[2]
567 mulx $acc7,%rax,$acc6 # a[2]*b[3]
568 mov 8*3(%rsi),%rdx # a[3]
570 adox %rdi,$acc6 # of=0
571 adcx %rdi,$acc6 # cf=0
573 mulx %rbp,%rax,%rbx # a[3]*b[0]
576 mulx %rcx,%rax,%rbx # a[3]*b[1]
579 mulx (%rsp),%rax,%rbx # a[3]*b[2]
582 mulx $acc7,%rax,$acc7 # a[3]*b[3]
585 adcx %rdi,$acc7 # cf=0
586 adox %rdi,$acc7 # of=0
591 .size x25519_fe64_mul,.-x25519_fe64_mul
593 .globl x25519_fe64_sqr
594 .type x25519_fe64_sqr,\@function,2
610 push %rdi # offload dst
613 .cfi_adjust_cfa_offset 16
616 mov 8*0(%rsi),%rdx # a[0]
617 mov 8*1(%rsi),%rcx # a[1]
618 mov 8*2(%rsi),%rbp # a[2]
619 mov 8*3(%rsi),%rsi # a[3]
621 ################################################################
622 mulx %rdx,$acc0,$acc7 # a[0]*a[0]
623 mulx %rcx,$acc1,%rax # a[0]*a[1]
624 xor %edi,%edi # cf=0,of=0
625 mulx %rbp,$acc2,%rbx # a[0]*a[2]
627 mulx %rsi,$acc3,$acc4 # a[0]*a[3]
630 adcx %rdi,$acc4 # cf=0
632 ################################################################
633 mulx %rbp,%rax,%rbx # a[1]*a[2]
636 mulx %rsi,%rax,$acc5 # a[1]*a[3]
641 ################################################################
642 mulx %rsi,%rax,$acc6 # a[2]*a[3]
645 adcx %rdi,$acc6 # cf=0
646 adox %rdi,$acc6 # of=0
648 adcx $acc1,$acc1 # acc1:6<<1
651 mulx %rdx,%rax,%rbx # a[1]*a[1]
657 mulx %rdx,%rax,%rbx # a[2]*a[2]
663 mulx %rdx,%rax,$acc7 # a[3]*a[3]
666 adcx %rdi,$acc7 # cf=0
667 adox %rdi,$acc7 # of=0
681 mulx $acc7,%rax,$acc4
686 mov 8*2(%rsp),%rdi # restore dst
694 sbb %rax,%rax # cf -> mask
716 .cfi_adjust_cfa_offset 88
720 .size x25519_fe64_sqr,.-x25519_fe64_sqr
722 .globl x25519_fe64_mul121666
723 .type x25519_fe64_mul121666,\@function,2
725 x25519_fe64_mul121666:
726 .Lfe64_mul121666_body:
728 mulx 8*0(%rsi),$acc0,%rcx
729 mulx 8*1(%rsi),$acc1,%rax
731 mulx 8*2(%rsi),$acc2,%rcx
733 mulx 8*3(%rsi),$acc3,%rax
744 sbb %rax,%rax # cf -> mask
753 .Lfe64_mul121666_epilogue:
755 .size x25519_fe64_mul121666,.-x25519_fe64_mul121666
757 .globl x25519_fe64_add
758 .type x25519_fe64_add,\@function,3
772 sbb %rax,%rax # cf -> mask
781 sbb %rax,%rax # cf -> mask
790 .size x25519_fe64_add,.-x25519_fe64_add
792 .globl x25519_fe64_sub
793 .type x25519_fe64_sub,\@function,3
807 sbb %rax,%rax # cf -> mask
816 sbb %rax,%rax # cf -> mask
825 .size x25519_fe64_sub,.-x25519_fe64_sub
827 .globl x25519_fe64_tobytes
828 .type x25519_fe64_tobytes,\@function,2
837 ################################# reduction modulo 2^255-19
838 lea ($acc3,$acc3),%rax
839 sar \$63,$acc3 # most significant bit -> mask
840 shr \$1,%rax # most significant bit cleared
842 add \$19,$acc3 # compare to modulus in the same go
849 lea (%rax,%rax),$acc3
850 sar \$63,%rax # most significant bit -> mask
851 shr \$1,$acc3 # most significant bit cleared
867 .size x25519_fe64_tobytes,.-x25519_fe64_tobytes
871 .globl x25519_fe64_eligible
872 .type x25519_fe64_eligible,\@abi-omnipotent
874 x25519_fe64_eligible:
877 .size x25519_fe64_eligible,.-x25519_fe64_eligible
879 .globl x25519_fe64_mul
880 .type x25519_fe64_mul,\@abi-omnipotent
881 .globl x25519_fe64_sqr
882 .globl x25519_fe64_mul121666
883 .globl x25519_fe64_add
884 .globl x25519_fe64_sub
885 .globl x25519_fe64_tobytes
888 x25519_fe64_mul121666:
892 .byte 0x0f,0x0b # ud2
894 .size x25519_fe64_mul,.-x25519_fe64_mul
898 .asciz "X25519 primitives for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
901 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
902 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
910 .extern __imp_RtlVirtualUnwind
912 .type short_handler,\@abi-omnipotent
926 mov 120($context),%rax # pull context->Rax
927 mov 248($context),%rbx # pull context->Rip
929 mov 8($disp),%rsi # disp->ImageBase
930 mov 56($disp),%r11 # disp->HandlerData
932 mov 0(%r11),%r10d # HandlerData[0]
933 lea (%rsi,%r10),%r10 # end of prologue label
934 cmp %r10,%rbx # context->Rip<end of prologue label
937 mov 152($context),%rax # pull context->Rsp
938 jmp .Lcommon_seh_tail
939 .size short_handler,.-short_handler
941 .type full_handler,\@abi-omnipotent
955 mov 120($context),%rax # pull context->Rax
956 mov 248($context),%rbx # pull context->Rip
958 mov 8($disp),%rsi # disp->ImageBase
959 mov 56($disp),%r11 # disp->HandlerData
961 mov 0(%r11),%r10d # HandlerData[0]
962 lea (%rsi,%r10),%r10 # end of prologue label
963 cmp %r10,%rbx # context->Rip<end of prologue label
966 mov 152($context),%rax # pull context->Rsp
968 mov 4(%r11),%r10d # HandlerData[1]
969 lea (%rsi,%r10),%r10 # epilogue label
970 cmp %r10,%rbx # context->Rip>=epilogue label
971 jae .Lcommon_seh_tail
973 mov 8(%r11),%r10d # HandlerData[2]
982 mov %rbx,144($context) # restore context->Rbx
983 mov %rbp,160($context) # restore context->Rbp
984 mov %r12,216($context) # restore context->R12
985 mov %r13,224($context) # restore context->R13
986 mov %r14,232($context) # restore context->R14
987 mov %r15,240($context) # restore context->R15
992 mov %rax,152($context) # restore context->Rsp
993 mov %rsi,168($context) # restore context->Rsi
994 mov %rdi,176($context) # restore context->Rdi
996 mov 40($disp),%rdi # disp->ContextRecord
997 mov $context,%rsi # context
998 mov \$154,%ecx # sizeof(CONTEXT)
999 .long 0xa548f3fc # cld; rep movsq
1002 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1003 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1004 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1005 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1006 mov 40(%rsi),%r10 # disp->ContextRecord
1007 lea 56(%rsi),%r11 # &disp->HandlerData
1008 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1009 mov %r10,32(%rsp) # arg5
1010 mov %r11,40(%rsp) # arg6
1011 mov %r12,48(%rsp) # arg7
1012 mov %rcx,56(%rsp) # arg8, (NULL)
1013 call *__imp_RtlVirtualUnwind(%rip)
1015 mov \$1,%eax # ExceptionContinueSearch
1027 .size full_handler,.-full_handler
1031 .rva .LSEH_begin_x25519_fe51_mul
1032 .rva .LSEH_end_x25519_fe51_mul
1033 .rva .LSEH_info_x25519_fe51_mul
1035 .rva .LSEH_begin_x25519_fe51_sqr
1036 .rva .LSEH_end_x25519_fe51_sqr
1037 .rva .LSEH_info_x25519_fe51_sqr
1039 .rva .LSEH_begin_x25519_fe51_mul121666
1040 .rva .LSEH_end_x25519_fe51_mul121666
1041 .rva .LSEH_info_x25519_fe51_mul121666
1043 $code.=<<___ if ($addx);
1044 .rva .LSEH_begin_x25519_fe64_mul
1045 .rva .LSEH_end_x25519_fe64_mul
1046 .rva .LSEH_info_x25519_fe64_mul
1048 .rva .LSEH_begin_x25519_fe64_sqr
1049 .rva .LSEH_end_x25519_fe64_sqr
1050 .rva .LSEH_info_x25519_fe64_sqr
1052 .rva .LSEH_begin_x25519_fe64_mul121666
1053 .rva .LSEH_end_x25519_fe64_mul121666
1054 .rva .LSEH_info_x25519_fe64_mul121666
1056 .rva .LSEH_begin_x25519_fe64_add
1057 .rva .LSEH_end_x25519_fe64_add
1058 .rva .LSEH_info_x25519_fe64_add
1060 .rva .LSEH_begin_x25519_fe64_sub
1061 .rva .LSEH_end_x25519_fe64_sub
1062 .rva .LSEH_info_x25519_fe64_sub
1064 .rva .LSEH_begin_x25519_fe64_tobytes
1065 .rva .LSEH_end_x25519_fe64_tobytes
1066 .rva .LSEH_info_x25519_fe64_tobytes
1071 .LSEH_info_x25519_fe51_mul:
1074 .rva .Lfe51_mul_body,.Lfe51_mul_epilogue # HandlerData[]
1076 .LSEH_info_x25519_fe51_sqr:
1079 .rva .Lfe51_sqr_body,.Lfe51_sqr_epilogue # HandlerData[]
1081 .LSEH_info_x25519_fe51_mul121666:
1084 .rva .Lfe51_mul121666_body,.Lfe51_mul121666_epilogue # HandlerData[]
1087 $code.=<<___ if ($addx);
1088 .LSEH_info_x25519_fe64_mul:
1091 .rva .Lfe64_mul_body,.Lfe64_mul_epilogue # HandlerData[]
1093 .LSEH_info_x25519_fe64_sqr:
1096 .rva .Lfe64_sqr_body,.Lfe64_sqr_epilogue # HandlerData[]
1098 .LSEH_info_x25519_fe64_mul121666:
1101 .rva .Lfe64_mul121666_body,.Lfe64_mul121666_epilogue # HandlerData[]
1102 .LSEH_info_x25519_fe64_add:
1105 .rva .Lfe64_add_body,.Lfe64_add_epilogue # HandlerData[]
1106 .LSEH_info_x25519_fe64_sub:
1109 .rva .Lfe64_sub_body,.Lfe64_sub_epilogue # HandlerData[]
1110 .LSEH_info_x25519_fe64_tobytes:
1113 .rva .Lfe64_to_body,.Lfe64_to_epilogue # HandlerData[]
1117 $code =~ s/\`([^\`]*)\`/eval $1/gem;