1# Copyright 2020-2022 The OpenSSL Project Authors. All Rights Reserved.
2# Copyright (c) 2020, Intel Corporation. All Rights Reserved.
3#
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
8#
9#
10# Originally written by Ilya Albrekht, Sergey Kirillov and Andrey Matyukov
11# Intel Corporation
12#
13# December 2020
14#
15# Initial release.
16#
17# Implementation utilizes 256-bit (ymm) registers to avoid frequency scaling issues.
18#
19# IceLake-Client @ 1.3GHz
20# |---------+----------------------+--------------+-------------|
21# |         | OpenSSL 3.0.0-alpha9 | this         | Unit        |
22# |---------+----------------------+--------------+-------------|
23# | rsa2048 | 2 127 659            | 1 015 625    | cycles/sign |
24# |         | 611                  | 1280 / +109% | sign/s      |
25# |---------+----------------------+--------------+-------------|
26#
27
28# $output is the last argument if it looks like a file (it has an extension)
29# $flavour is the first argument if it doesn't look like a file
30$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
31$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
32
33$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
34$avx512ifma=0;
35
36$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
37( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
38( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
39die "can't locate x86_64-xlate.pl";
40
41if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
42        =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
43    $avx512ifma = ($1>=2.26);
44}
45
46if (!$avx512 && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
47       `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)(?:\.([0-9]+))?/) {
48    $avx512ifma = ($1==2.11 && $2>=8) + ($1>=2.12);
49}
50
51if (!$avx512 && `$ENV{CC} -v 2>&1`
52    =~ /(Apple)?\s*((?:clang|LLVM) version|.*based on LLVM) ([0-9]+)\.([0-9]+)\.([0-9]+)?/) {
53    my $ver = $3 + $4/100.0 + $5/10000.0; # 3.1.0->3.01, 3.10.1->3.1001
54    if ($1) {
55        # Apple conditions, they use a different version series, see
56        # https://en.wikipedia.org/wiki/Xcode#Xcode_7.0_-_10.x_(since_Free_On-Device_Development)_2
57        # clang 7.0.0 is Apple clang 10.0.1
58        $avx512ifma = ($ver>=10.0001)
59    } else {
60        $avx512ifma = ($3>=7.0);
61    }
62}
63
64open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
65    or die "can't call $xlate: $!";
66*STDOUT=*OUT;
67
68if ($avx512ifma>0) {{{
69@_6_args_universal_ABI = ("%rdi","%rsi","%rdx","%rcx","%r8","%r9");
70
71$code.=<<___;
72.extern OPENSSL_ia32cap_P
73.globl  ossl_rsaz_avx512ifma_eligible
74.type   ossl_rsaz_avx512ifma_eligible,\@abi-omnipotent
75.align  32
76ossl_rsaz_avx512ifma_eligible:
77    mov OPENSSL_ia32cap_P+8(%rip), %ecx
78    xor %eax,%eax
79    and \$`1<<31|1<<21|1<<17|1<<16`, %ecx     # avx512vl + avx512ifma + avx512dq + avx512f
80    cmp \$`1<<31|1<<21|1<<17|1<<16`, %ecx
81    cmove %ecx,%eax
82    ret
83.size   ossl_rsaz_avx512ifma_eligible, .-ossl_rsaz_avx512ifma_eligible
84___
85
86###############################################################################
87# Almost Montgomery Multiplication (AMM) for 20-digit number in radix 2^52.
88#
89# AMM is defined as presented in the paper
90# "Efficient Software Implementations of Modular Exponentiation" by Shay Gueron.
91#
92# The input and output are presented in 2^52 radix domain, i.e.
93#   |res|, |a|, |b|, |m| are arrays of 20 64-bit qwords with 12 high bits zeroed.
94#   |k0| is a Montgomery coefficient, which is here k0 = -1/m mod 2^64
95#        (note, the implementation counts only 52 bits from it).
96#
97# NB: the AMM implementation does not perform "conditional" subtraction step as
98# specified in the original algorithm as according to the paper "Enhanced Montgomery
99# Multiplication" by Shay Gueron (see Lemma 1), the result will be always < 2*2^1024
100# and can be used as a direct input to the next AMM iteration.
101# This post-condition is true, provided the correct parameter |s| is choosen, i.e.
102# s >= n + 2 * k, which matches our case: 1040 > 1024 + 2 * 1.
103#
104# void ossl_rsaz_amm52x20_x1_256(BN_ULONG *res,
105#                           const BN_ULONG *a,
106#                           const BN_ULONG *b,
107#                           const BN_ULONG *m,
108#                           BN_ULONG k0);
109###############################################################################
110{
111# input parameters ("%rdi","%rsi","%rdx","%rcx","%r8")
112my ($res,$a,$b,$m,$k0) = @_6_args_universal_ABI;
113
114my $mask52     = "%rax";
115my $acc0_0     = "%r9";
116my $acc0_0_low = "%r9d";
117my $acc0_1     = "%r15";
118my $acc0_1_low = "%r15d";
119my $b_ptr      = "%r11";
120
121my $iter = "%ebx";
122
123my $zero = "%ymm0";
124my ($R0_0,$R0_0h,$R1_0,$R1_0h,$R2_0) = ("%ymm1", map("%ymm$_",(16..19)));
125my ($R0_1,$R0_1h,$R1_1,$R1_1h,$R2_1) = ("%ymm2", map("%ymm$_",(20..23)));
126my $Bi = "%ymm3";
127my $Yi = "%ymm4";
128
129# Registers mapping for normalization.
130# We can reuse Bi, Yi registers here.
131my $TMP = $Bi;
132my $mask52x4 = $Yi;
133my ($T0,$T0h,$T1,$T1h,$T2) = map("%ymm$_", (24..28));
134
135sub amm52x20_x1() {
136# _data_offset - offset in the |a| or |m| arrays pointing to the beginning
137#                of data for corresponding AMM operation;
138# _b_offset    - offset in the |b| array pointing to the next qword digit;
139my ($_data_offset,$_b_offset,$_acc,$_R0,$_R0h,$_R1,$_R1h,$_R2,$_k0) = @_;
140my $_R0_xmm = $_R0;
141$_R0_xmm =~ s/%y/%x/;
142$code.=<<___;
143    movq    $_b_offset($b_ptr), %r13             # b[i]
144
145    vpbroadcastq    %r13, $Bi                    # broadcast b[i]
146    movq    $_data_offset($a), %rdx
147    mulx    %r13, %r13, %r12                     # a[0]*b[i] = (t0,t2)
148    addq    %r13, $_acc                          # acc += t0
149    movq    %r12, %r10
150    adcq    \$0, %r10                            # t2 += CF
151
152    movq    $_k0, %r13
153    imulq   $_acc, %r13                          # acc * k0
154    andq    $mask52, %r13                        # yi = (acc * k0) & mask52
155
156    vpbroadcastq    %r13, $Yi                    # broadcast y[i]
157    movq    $_data_offset($m), %rdx
158    mulx    %r13, %r13, %r12                     # yi * m[0] = (t0,t1)
159    addq    %r13, $_acc                          # acc += t0
160    adcq    %r12, %r10                           # t2 += (t1 + CF)
161
162    shrq    \$52, $_acc
163    salq    \$12, %r10
164    or      %r10, $_acc                          # acc = ((acc >> 52) | (t2 << 12))
165
166    vpmadd52luq `$_data_offset+64*0`($a), $Bi, $_R0
167    vpmadd52luq `$_data_offset+64*0+32`($a), $Bi, $_R0h
168    vpmadd52luq `$_data_offset+64*1`($a), $Bi, $_R1
169    vpmadd52luq `$_data_offset+64*1+32`($a), $Bi, $_R1h
170    vpmadd52luq `$_data_offset+64*2`($a), $Bi, $_R2
171
172    vpmadd52luq `$_data_offset+64*0`($m), $Yi, $_R0
173    vpmadd52luq `$_data_offset+64*0+32`($m), $Yi, $_R0h
174    vpmadd52luq `$_data_offset+64*1`($m), $Yi, $_R1
175    vpmadd52luq `$_data_offset+64*1+32`($m), $Yi, $_R1h
176    vpmadd52luq `$_data_offset+64*2`($m), $Yi, $_R2
177
178    # Shift accumulators right by 1 qword, zero extending the highest one
179    valignq     \$1, $_R0, $_R0h, $_R0
180    valignq     \$1, $_R0h, $_R1, $_R0h
181    valignq     \$1, $_R1, $_R1h, $_R1
182    valignq     \$1, $_R1h, $_R2, $_R1h
183    valignq     \$1, $_R2, $zero, $_R2
184
185    vmovq   $_R0_xmm, %r13
186    addq    %r13, $_acc    # acc += R0[0]
187
188    vpmadd52huq `$_data_offset+64*0`($a), $Bi, $_R0
189    vpmadd52huq `$_data_offset+64*0+32`($a), $Bi, $_R0h
190    vpmadd52huq `$_data_offset+64*1`($a), $Bi, $_R1
191    vpmadd52huq `$_data_offset+64*1+32`($a), $Bi, $_R1h
192    vpmadd52huq `$_data_offset+64*2`($a), $Bi, $_R2
193
194    vpmadd52huq `$_data_offset+64*0`($m), $Yi, $_R0
195    vpmadd52huq `$_data_offset+64*0+32`($m), $Yi, $_R0h
196    vpmadd52huq `$_data_offset+64*1`($m), $Yi, $_R1
197    vpmadd52huq `$_data_offset+64*1+32`($m), $Yi, $_R1h
198    vpmadd52huq `$_data_offset+64*2`($m), $Yi, $_R2
199___
200}
201
202# Normalization routine: handles carry bits in R0..R2 QWs and
203# gets R0..R2 back to normalized 2^52 representation.
204#
205# Uses %r8-14,%e[bcd]x
206sub amm52x20_x1_norm {
207my ($_acc,$_R0,$_R0h,$_R1,$_R1h,$_R2) = @_;
208$code.=<<___;
209    # Put accumulator to low qword in R0
210    vpbroadcastq    $_acc, $TMP
211    vpblendd \$3, $TMP, $_R0, $_R0
212
213    # Extract "carries" (12 high bits) from each QW of R0..R2
214    # Save them to LSB of QWs in T0..T2
215    vpsrlq    \$52, $_R0,   $T0
216    vpsrlq    \$52, $_R0h,  $T0h
217    vpsrlq    \$52, $_R1,   $T1
218    vpsrlq    \$52, $_R1h,  $T1h
219    vpsrlq    \$52, $_R2,   $T2
220
221    # "Shift left" T0..T2 by 1 QW
222    valignq \$3, $T1h,  $T2,  $T2
223    valignq \$3, $T1,   $T1h, $T1h
224    valignq \$3, $T0h,  $T1,  $T1
225    valignq \$3, $T0,   $T0h, $T0h
226    valignq \$3, $zero, $T0,  $T0
227
228    # Drop "carries" from R0..R2 QWs
229    vpandq    $mask52x4, $_R0,  $_R0
230    vpandq    $mask52x4, $_R0h, $_R0h
231    vpandq    $mask52x4, $_R1,  $_R1
232    vpandq    $mask52x4, $_R1h, $_R1h
233    vpandq    $mask52x4, $_R2,  $_R2
234
235    # Sum R0..R2 with corresponding adjusted carries
236    vpaddq  $T0,  $_R0,  $_R0
237    vpaddq  $T0h, $_R0h, $_R0h
238    vpaddq  $T1,  $_R1,  $_R1
239    vpaddq  $T1h, $_R1h, $_R1h
240    vpaddq  $T2,  $_R2,  $_R2
241
242    # Now handle carry bits from this addition
243    # Get mask of QWs which 52-bit parts overflow...
244    vpcmpuq   \$1, $_R0,  $mask52x4, %k1 # OP=lt
245    vpcmpuq   \$1, $_R0h, $mask52x4, %k2
246    vpcmpuq   \$1, $_R1,  $mask52x4, %k3
247    vpcmpuq   \$1, $_R1h, $mask52x4, %k4
248    vpcmpuq   \$1, $_R2,  $mask52x4, %k5
249    kmovb   %k1, %r14d                   # k1
250    kmovb   %k2, %r13d                   # k1h
251    kmovb   %k3, %r12d                   # k2
252    kmovb   %k4, %r11d                   # k2h
253    kmovb   %k5, %r10d                   # k3
254
255    # ...or saturated
256    vpcmpuq   \$0, $_R0,  $mask52x4, %k1 # OP=eq
257    vpcmpuq   \$0, $_R0h, $mask52x4, %k2
258    vpcmpuq   \$0, $_R1,  $mask52x4, %k3
259    vpcmpuq   \$0, $_R1h, $mask52x4, %k4
260    vpcmpuq   \$0, $_R2,  $mask52x4, %k5
261    kmovb   %k1, %r9d                    # k4
262    kmovb   %k2, %r8d                    # k4h
263    kmovb   %k3, %ebx                    # k5
264    kmovb   %k4, %ecx                    # k5h
265    kmovb   %k5, %edx                    # k6
266
267    # Get mask of QWs where carries shall be propagated to.
268    # Merge 4-bit masks to 8-bit values to use add with carry.
269    shl   \$4, %r13b
270    or    %r13b, %r14b
271    shl   \$4, %r11b
272    or    %r11b, %r12b
273
274    add   %r14b, %r14b
275    adc   %r12b, %r12b
276    adc   %r10b, %r10b
277
278    shl   \$4, %r8b
279    or    %r8b,%r9b
280    shl   \$4, %cl
281    or    %cl, %bl
282
283    add   %r9b, %r14b
284    adc   %bl, %r12b
285    adc   %dl, %r10b
286
287    xor   %r9b, %r14b
288    xor   %bl, %r12b
289    xor   %dl, %r10b
290
291    kmovb   %r14d, %k1
292    shr     \$4, %r14b
293    kmovb   %r14d, %k2
294    kmovb   %r12d, %k3
295    shr     \$4, %r12b
296    kmovb   %r12d, %k4
297    kmovb   %r10d, %k5
298
299    # Add carries according to the obtained mask
300    vpsubq  $mask52x4, $_R0,  ${_R0}{%k1}
301    vpsubq  $mask52x4, $_R0h, ${_R0h}{%k2}
302    vpsubq  $mask52x4, $_R1,  ${_R1}{%k3}
303    vpsubq  $mask52x4, $_R1h, ${_R1h}{%k4}
304    vpsubq  $mask52x4, $_R2,  ${_R2}{%k5}
305
306    vpandq   $mask52x4, $_R0,  $_R0
307    vpandq   $mask52x4, $_R0h, $_R0h
308    vpandq   $mask52x4, $_R1,  $_R1
309    vpandq   $mask52x4, $_R1h, $_R1h
310    vpandq   $mask52x4, $_R2,  $_R2
311___
312}
313
314$code.=<<___;
315.text
316
317.globl  ossl_rsaz_amm52x20_x1_256
318.type   ossl_rsaz_amm52x20_x1_256,\@function,5
319.align 32
320ossl_rsaz_amm52x20_x1_256:
321.cfi_startproc
322    endbranch
323    push    %rbx
324.cfi_push   %rbx
325    push    %rbp
326.cfi_push   %rbp
327    push    %r12
328.cfi_push   %r12
329    push    %r13
330.cfi_push   %r13
331    push    %r14
332.cfi_push   %r14
333    push    %r15
334.cfi_push   %r15
335.Lrsaz_amm52x20_x1_256_body:
336
337    # Zeroing accumulators
338    vpxord   $zero, $zero, $zero
339    vmovdqa64   $zero, $R0_0
340    vmovdqa64   $zero, $R0_0h
341    vmovdqa64   $zero, $R1_0
342    vmovdqa64   $zero, $R1_0h
343    vmovdqa64   $zero, $R2_0
344
345    xorl    $acc0_0_low, $acc0_0_low
346
347    movq    $b, $b_ptr                       # backup address of b
348    movq    \$0xfffffffffffff, $mask52       # 52-bit mask
349
350    # Loop over 20 digits unrolled by 4
351    mov     \$5, $iter
352
353.align 32
354.Lloop5:
355___
356    foreach my $idx (0..3) {
357        &amm52x20_x1(0,8*$idx,$acc0_0,$R0_0,$R0_0h,$R1_0,$R1_0h,$R2_0,$k0);
358    }
359$code.=<<___;
360    lea    `4*8`($b_ptr), $b_ptr
361    dec    $iter
362    jne    .Lloop5
363
364    vmovdqa64   .Lmask52x4(%rip), $mask52x4
365___
366    &amm52x20_x1_norm($acc0_0,$R0_0,$R0_0h,$R1_0,$R1_0h,$R2_0);
367$code.=<<___;
368
369    vmovdqu64   $R0_0, ($res)
370    vmovdqu64   $R0_0h, 32($res)
371    vmovdqu64   $R1_0, 64($res)
372    vmovdqu64   $R1_0h, 96($res)
373    vmovdqu64   $R2_0, 128($res)
374
375    vzeroupper
376    mov  0(%rsp),%r15
377.cfi_restore    %r15
378    mov  8(%rsp),%r14
379.cfi_restore    %r14
380    mov  16(%rsp),%r13
381.cfi_restore    %r13
382    mov  24(%rsp),%r12
383.cfi_restore    %r12
384    mov  32(%rsp),%rbp
385.cfi_restore    %rbp
386    mov  40(%rsp),%rbx
387.cfi_restore    %rbx
388    lea  48(%rsp),%rsp
389.cfi_adjust_cfa_offset  -48
390.Lrsaz_amm52x20_x1_256_epilogue:
391    ret
392.cfi_endproc
393.size   ossl_rsaz_amm52x20_x1_256, .-ossl_rsaz_amm52x20_x1_256
394___
395
396$code.=<<___;
397.data
398.align 32
399.Lmask52x4:
400    .quad   0xfffffffffffff
401    .quad   0xfffffffffffff
402    .quad   0xfffffffffffff
403    .quad   0xfffffffffffff
404___
405
406###############################################################################
407# Dual Almost Montgomery Multiplication for 20-digit number in radix 2^52
408#
409# See description of ossl_rsaz_amm52x20_x1_256() above for details about Almost
410# Montgomery Multiplication algorithm and function input parameters description.
411#
412# This function does two AMMs for two independent inputs, hence dual.
413#
414# void ossl_rsaz_amm52x20_x2_256(BN_ULONG out[2][20],
415#                           const BN_ULONG a[2][20],
416#                           const BN_ULONG b[2][20],
417#                           const BN_ULONG m[2][20],
418#                           const BN_ULONG k0[2]);
419###############################################################################
420
421$code.=<<___;
422.text
423
424.globl  ossl_rsaz_amm52x20_x2_256
425.type   ossl_rsaz_amm52x20_x2_256,\@function,5
426.align 32
427ossl_rsaz_amm52x20_x2_256:
428.cfi_startproc
429    endbranch
430    push    %rbx
431.cfi_push   %rbx
432    push    %rbp
433.cfi_push   %rbp
434    push    %r12
435.cfi_push   %r12
436    push    %r13
437.cfi_push   %r13
438    push    %r14
439.cfi_push   %r14
440    push    %r15
441.cfi_push   %r15
442.Lrsaz_amm52x20_x2_256_body:
443
444    # Zeroing accumulators
445    vpxord   $zero, $zero, $zero
446    vmovdqa64   $zero, $R0_0
447    vmovdqa64   $zero, $R0_0h
448    vmovdqa64   $zero, $R1_0
449    vmovdqa64   $zero, $R1_0h
450    vmovdqa64   $zero, $R2_0
451    vmovdqa64   $zero, $R0_1
452    vmovdqa64   $zero, $R0_1h
453    vmovdqa64   $zero, $R1_1
454    vmovdqa64   $zero, $R1_1h
455    vmovdqa64   $zero, $R2_1
456
457    xorl    $acc0_0_low, $acc0_0_low
458    xorl    $acc0_1_low, $acc0_1_low
459
460    movq    $b, $b_ptr                       # backup address of b
461    movq    \$0xfffffffffffff, $mask52       # 52-bit mask
462
463    mov    \$20, $iter
464
465.align 32
466.Lloop20:
467___
468    &amm52x20_x1(   0,   0,$acc0_0,$R0_0,$R0_0h,$R1_0,$R1_0h,$R2_0,"($k0)");
469    # 20*8 = offset of the next dimension in two-dimension array
470    &amm52x20_x1(20*8,20*8,$acc0_1,$R0_1,$R0_1h,$R1_1,$R1_1h,$R2_1,"8($k0)");
471$code.=<<___;
472    lea    8($b_ptr), $b_ptr
473    dec    $iter
474    jne    .Lloop20
475
476    vmovdqa64   .Lmask52x4(%rip), $mask52x4
477___
478    &amm52x20_x1_norm($acc0_0,$R0_0,$R0_0h,$R1_0,$R1_0h,$R2_0);
479    &amm52x20_x1_norm($acc0_1,$R0_1,$R0_1h,$R1_1,$R1_1h,$R2_1);
480$code.=<<___;
481
482    vmovdqu64   $R0_0, ($res)
483    vmovdqu64   $R0_0h, 32($res)
484    vmovdqu64   $R1_0, 64($res)
485    vmovdqu64   $R1_0h, 96($res)
486    vmovdqu64   $R2_0, 128($res)
487
488    vmovdqu64   $R0_1, 160($res)
489    vmovdqu64   $R0_1h, 192($res)
490    vmovdqu64   $R1_1, 224($res)
491    vmovdqu64   $R1_1h, 256($res)
492    vmovdqu64   $R2_1, 288($res)
493
494    vzeroupper
495    mov  0(%rsp),%r15
496.cfi_restore    %r15
497    mov  8(%rsp),%r14
498.cfi_restore    %r14
499    mov  16(%rsp),%r13
500.cfi_restore    %r13
501    mov  24(%rsp),%r12
502.cfi_restore    %r12
503    mov  32(%rsp),%rbp
504.cfi_restore    %rbp
505    mov  40(%rsp),%rbx
506.cfi_restore    %rbx
507    lea  48(%rsp),%rsp
508.cfi_adjust_cfa_offset  -48
509.Lrsaz_amm52x20_x2_256_epilogue:
510    ret
511.cfi_endproc
512.size   ossl_rsaz_amm52x20_x2_256, .-ossl_rsaz_amm52x20_x2_256
513___
514}
515
516###############################################################################
517# Constant time extraction from the precomputed table of powers base^i, where
518#    i = 0..2^EXP_WIN_SIZE-1
519#
520# The input |red_table| contains precomputations for two independent base values,
521# so the |tbl_idx| indicates for which base shall we extract the value.
522# |red_table_idx| is a power index.
523#
524# Extracted value (output) is 20 digit number in 2^52 radix.
525#
526# void ossl_extract_multiplier_2x20_win5(BN_ULONG *red_Y,
527#                                        const BN_ULONG red_table[1 << EXP_WIN_SIZE][2][20],
528#                                        int red_table_idx,
529#                                        int tbl_idx);           # 0 or 1
530#
531# EXP_WIN_SIZE = 5
532###############################################################################
533{
534# input parameters
535my ($out,$red_tbl,$red_tbl_idx,$tbl_idx) = @_6_args_universal_ABI;
536
537my ($t0,$t1,$t2,$t3,$t4) = map("%ymm$_", (0..4));
538my $t4xmm = $t4;
539$t4xmm =~ s/%y/%x/;
540my ($tmp0,$tmp1,$tmp2,$tmp3,$tmp4) = map("%ymm$_", (16..20));
541my ($cur_idx,$idx,$ones) = map("%ymm$_", (21..23));
542
543$code.=<<___;
544.text
545
546.align 32
547.globl  ossl_extract_multiplier_2x20_win5
548.type   ossl_extract_multiplier_2x20_win5,\@function,4
549ossl_extract_multiplier_2x20_win5:
550.cfi_startproc
551    endbranch
552    leaq    ($tbl_idx,$tbl_idx,4), %rax
553    salq    \$5, %rax
554    addq    %rax, $red_tbl
555
556    vmovdqa64   .Lones(%rip), $ones         # broadcast ones
557    vpbroadcastq    $red_tbl_idx, $idx
558    leaq   `(1<<5)*2*20*8`($red_tbl), %rax  # holds end of the tbl
559
560    vpxor   $t4xmm, $t4xmm, $t4xmm
561    vmovdqa64   $t4, $t3                    # zeroing t0..4, cur_idx
562    vmovdqa64   $t4, $t2
563    vmovdqa64   $t4, $t1
564    vmovdqa64   $t4, $t0
565    vmovdqa64   $t4, $cur_idx
566
567.align 32
568.Lloop:
569    vpcmpq  \$0, $cur_idx, $idx, %k1        # mask of (idx == cur_idx)
570    addq    \$320, $red_tbl                 # 320 = 2 * 20 digits * 8 bytes
571    vpaddq  $ones, $cur_idx, $cur_idx       # increment cur_idx
572    vmovdqu64  -320($red_tbl), $tmp0        # load data from red_tbl
573    vmovdqu64  -288($red_tbl), $tmp1
574    vmovdqu64  -256($red_tbl), $tmp2
575    vmovdqu64  -224($red_tbl), $tmp3
576    vmovdqu64  -192($red_tbl), $tmp4
577    vpblendmq  $tmp0, $t0, ${t0}{%k1}       # extract data when mask is not zero
578    vpblendmq  $tmp1, $t1, ${t1}{%k1}
579    vpblendmq  $tmp2, $t2, ${t2}{%k1}
580    vpblendmq  $tmp3, $t3, ${t3}{%k1}
581    vpblendmq  $tmp4, $t4, ${t4}{%k1}
582    cmpq    $red_tbl, %rax
583    jne .Lloop
584
585    vmovdqu64   $t0, ($out)                 # store t0..4
586    vmovdqu64   $t1, 32($out)
587    vmovdqu64   $t2, 64($out)
588    vmovdqu64   $t3, 96($out)
589    vmovdqu64   $t4, 128($out)
590
591    ret
592.cfi_endproc
593.size   ossl_extract_multiplier_2x20_win5, .-ossl_extract_multiplier_2x20_win5
594___
595$code.=<<___;
596.data
597.align 32
598.Lones:
599    .quad   1,1,1,1
600___
601}
602
603if ($win64) {
604$rec="%rcx";
605$frame="%rdx";
606$context="%r8";
607$disp="%r9";
608
609$code.=<<___
610.extern     __imp_RtlVirtualUnwind
611.type   rsaz_def_handler,\@abi-omnipotent
612.align  16
613rsaz_def_handler:
614    push    %rsi
615    push    %rdi
616    push    %rbx
617    push    %rbp
618    push    %r12
619    push    %r13
620    push    %r14
621    push    %r15
622    pushfq
623    sub     \$64,%rsp
624
625    mov     120($context),%rax # pull context->Rax
626    mov     248($context),%rbx # pull context->Rip
627
628    mov     8($disp),%rsi      # disp->ImageBase
629    mov     56($disp),%r11     # disp->HandlerData
630
631    mov     0(%r11),%r10d      # HandlerData[0]
632    lea     (%rsi,%r10),%r10   # prologue label
633    cmp     %r10,%rbx          # context->Rip<.Lprologue
634    jb  .Lcommon_seh_tail
635
636    mov     152($context),%rax # pull context->Rsp
637
638    mov     4(%r11),%r10d      # HandlerData[1]
639    lea     (%rsi,%r10),%r10   # epilogue label
640    cmp     %r10,%rbx          # context->Rip>=.Lepilogue
641    jae     .Lcommon_seh_tail
642
643    lea     48(%rax),%rax
644
645    mov     -8(%rax),%rbx
646    mov     -16(%rax),%rbp
647    mov     -24(%rax),%r12
648    mov     -32(%rax),%r13
649    mov     -40(%rax),%r14
650    mov     -48(%rax),%r15
651    mov     %rbx,144($context) # restore context->Rbx
652    mov     %rbp,160($context) # restore context->Rbp
653    mov     %r12,216($context) # restore context->R12
654    mov     %r13,224($context) # restore context->R13
655    mov     %r14,232($context) # restore context->R14
656    mov     %r15,240($context) # restore context->R14
657
658.Lcommon_seh_tail:
659    mov     8(%rax),%rdi
660    mov     16(%rax),%rsi
661    mov     %rax,152($context) # restore context->Rsp
662    mov     %rsi,168($context) # restore context->Rsi
663    mov     %rdi,176($context) # restore context->Rdi
664
665    mov     40($disp),%rdi     # disp->ContextRecord
666    mov     $context,%rsi      # context
667    mov     \$154,%ecx         # sizeof(CONTEXT)
668    .long   0xa548f3fc         # cld; rep movsq
669
670    mov     $disp,%rsi
671    xor     %rcx,%rcx          # arg1, UNW_FLAG_NHANDLER
672    mov     8(%rsi),%rdx       # arg2, disp->ImageBase
673    mov     0(%rsi),%r8        # arg3, disp->ControlPc
674    mov     16(%rsi),%r9       # arg4, disp->FunctionEntry
675    mov     40(%rsi),%r10      # disp->ContextRecord
676    lea     56(%rsi),%r11      # &disp->HandlerData
677    lea     24(%rsi),%r12      # &disp->EstablisherFrame
678    mov     %r10,32(%rsp)      # arg5
679    mov     %r11,40(%rsp)      # arg6
680    mov     %r12,48(%rsp)      # arg7
681    mov     %rcx,56(%rsp)      # arg8, (NULL)
682    call    *__imp_RtlVirtualUnwind(%rip)
683
684    mov     \$1,%eax           # ExceptionContinueSearch
685    add     \$64,%rsp
686    popfq
687    pop     %r15
688    pop     %r14
689    pop     %r13
690    pop     %r12
691    pop     %rbp
692    pop     %rbx
693    pop     %rdi
694    pop     %rsi
695    ret
696.size   rsaz_def_handler,.-rsaz_def_handler
697
698.section    .pdata
699.align  4
700    .rva    .LSEH_begin_ossl_rsaz_amm52x20_x1_256
701    .rva    .LSEH_end_ossl_rsaz_amm52x20_x1_256
702    .rva    .LSEH_info_ossl_rsaz_amm52x20_x1_256
703
704    .rva    .LSEH_begin_ossl_rsaz_amm52x20_x2_256
705    .rva    .LSEH_end_ossl_rsaz_amm52x20_x2_256
706    .rva    .LSEH_info_ossl_rsaz_amm52x20_x2_256
707
708    .rva    .LSEH_begin_ossl_extract_multiplier_2x20_win5
709    .rva    .LSEH_end_ossl_extract_multiplier_2x20_win5
710    .rva    .LSEH_info_ossl_extract_multiplier_2x20_win5
711
712.section    .xdata
713.align  8
714.LSEH_info_ossl_rsaz_amm52x20_x1_256:
715    .byte   9,0,0,0
716    .rva    rsaz_def_handler
717    .rva    .Lrsaz_amm52x20_x1_256_body,.Lrsaz_amm52x20_x1_256_epilogue
718.LSEH_info_ossl_rsaz_amm52x20_x2_256:
719    .byte   9,0,0,0
720    .rva    rsaz_def_handler
721    .rva    .Lrsaz_amm52x20_x2_256_body,.Lrsaz_amm52x20_x2_256_epilogue
722.LSEH_info_ossl_extract_multiplier_2x20_win5:
723    .byte   9,0,0,0
724    .rva    rsaz_def_handler
725    .rva    .LSEH_begin_ossl_extract_multiplier_2x20_win5,.LSEH_begin_ossl_extract_multiplier_2x20_win5
726___
727}
728}}} else {{{                # fallback for old assembler
729$code.=<<___;
730.text
731
732.globl  ossl_rsaz_avx512ifma_eligible
733.type   ossl_rsaz_avx512ifma_eligible,\@abi-omnipotent
734ossl_rsaz_avx512ifma_eligible:
735    xor     %eax,%eax
736    ret
737.size   ossl_rsaz_avx512ifma_eligible, .-ossl_rsaz_avx512ifma_eligible
738
739.globl  ossl_rsaz_amm52x20_x1_256
740.globl  ossl_rsaz_amm52x20_x2_256
741.globl  ossl_extract_multiplier_2x20_win5
742.type   ossl_rsaz_amm52x20_x1_256,\@abi-omnipotent
743ossl_rsaz_amm52x20_x1_256:
744ossl_rsaz_amm52x20_x2_256:
745ossl_extract_multiplier_2x20_win5:
746    .byte   0x0f,0x0b    # ud2
747    ret
748.size   ossl_rsaz_amm52x20_x1_256, .-ossl_rsaz_amm52x20_x1_256
749___
750}}}
751
752$code =~ s/\`([^\`]*)\`/eval $1/gem;
753print $code;
754close STDOUT or die "error closing STDOUT: $!";
755