Add ec/asm/ecp_nistz256-x86.pl module.
authorAndy Polyakov <appro@openssl.org>
Mon, 9 Feb 2015 22:19:16 +0000 (23:19 +0100)
committerAndy Polyakov <appro@openssl.org>
Mon, 9 Feb 2015 22:19:16 +0000 (23:19 +0100)
Reviewed-by: Emilia Käsper <emilia@openssl.org>
crypto/ec/Makefile
crypto/ec/asm/ecp_nistz256-x86.pl [new file with mode: 0755]

index 852183af59328255ee8a3001c62f2d5d2cf0d556..319e0039db71546d0f852db6421d3220c9ac8e31 100644 (file)
@@ -48,6 +48,9 @@ lib:  $(LIBOBJ)
        $(RANLIB) $(LIB) || echo Never mind.
        @touch lib
 
+ecp_nistz256-x86.s:    asm/ecp_nistz256-x86.pl
+       $(PERL) asm/ecp_nistz256-x86.pl $(PERLASM_SCHEME) $(CFLAGS) $(PROCESSOR) > $@
+
 ecp_nistz256-x86_64.s: asm/ecp_nistz256-x86_64.pl
        $(PERL) asm/ecp_nistz256-x86_64.pl $(PERLASM_SCHEME) > $@
 
diff --git a/crypto/ec/asm/ecp_nistz256-x86.pl b/crypto/ec/asm/ecp_nistz256-x86.pl
new file mode 100755 (executable)
index 0000000..0670c69
--- /dev/null
@@ -0,0 +1,1823 @@
+#!/usr/bin/env perl
+
+# ====================================================================
+# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
+# project. The module is, however, dual licensed under OpenSSL and
+# CRYPTOGAMS licenses depending on where you obtain it. For further
+# details see http://www.openssl.org/~appro/cryptogams/.
+# ====================================================================
+#
+# ECP_NISTZ256 module for x86/SSE2.
+#
+# October 2014.
+#
+# Original ECP_NISTZ256 submission targeting x86_64 is detailed in
+# http://eprint.iacr.org/2013/816. In the process of adaptation
+# original .c module was made 32-bit savvy in order to make this
+# implementation possible.
+#
+#              with/without -DECP_NISTZ256_ASM
+# Pentium      +66-163%
+# PIII         +72-172%
+# P4           +65-132%
+# Core2                +90-215%
+# Sandy Bridge +105-265% (contemporary i[57]-* are all close to this)
+# Atom         +65-155%
+# Opteron      +54-110%
+# Bulldozer    +99-240%
+# VIA Nano     +93-290%
+#
+# Ranges denote minimum and maximum improvement coefficients depending
+# on benchmark. Lower coefficients are for ECDSA sign, server-side
+# operation. Keep in mind that +200% means 3x improvement.
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+push(@INC,"${dir}","${dir}../../perlasm");
+require "x86asm.pl";
+
+&asm_init($ARGV[0],"ecp_nistz256-x86.pl",$ARGV[$#ARGV] eq "386");
+
+$sse2=0;
+for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
+
+&external_label("OPENSSL_ia32cap_P") if ($sse2);
+
+
+########################################################################
+# Convert ecp_nistz256_table.c to layout expected by ecp_nistz_gather_w7
+#
+open TABLE,"<ecp_nistz256_table.c"             or
+open TABLE,"<${dir}../ecp_nistz256_table.c"    or
+die "failed to open ecp_nistz256_table.c:",$!;
+
+use integer;
+
+foreach(<TABLE>) {
+       s/TOBN\(\s*(0x[0-9a-f]+),\s*(0x[0-9a-f]+)\s*\)/push @arr,hex($2),hex($1)/geo;
+}
+close TABLE;
+
+# See ecp_nistz256_table.c for explanation for why it's 64*16*37.
+# 64*16*37-1 is because $#arr returns last valid index or @arr, not
+# amount of elements.
+die "insane number of elements" if ($#arr != 64*16*37-1);
+
+&public_label("ecp_nistz256_precomputed");
+&align(4096);
+&set_label("ecp_nistz256_precomputed");
+
+########################################################################
+# this conversion smashes P256_POINT_AFFINE by individual bytes with
+# 64 byte interval, similar to
+#      1111222233334444
+#      1234123412341234
+for(1..37) {
+       @tbl = splice(@arr,0,64*16);
+       for($i=0;$i<64;$i++) {
+               undef @line;
+               for($j=0;$j<64;$j++) {
+                       push @line,(@tbl[$j*16+$i/4]>>(($i%4)*8))&0xff;
+               }
+               &data_byte(join(',',map { sprintf "0x%02x",$_} @line));
+       }
+}
+
+########################################################################
+# Keep in mind that constants are stored least to most significant word
+&static_label("RR");
+&set_label("RR",64);
+&data_word(3,0,-1,-5,-2,-1,-3,4);      # 2^512 mod P-256
+
+&static_label("ONE_mont");
+&set_label("ONE_mont");
+&data_word(1,0,0,-1,-1,-1,-2,0);
+
+&static_label("ONE");
+&set_label("ONE");
+&data_word(1,0,0,0,0,0,0,0);
+&asciz("ECP_NISZ256 for x86/SSE2, CRYPTOGAMS by <appro\@openssl.org>");
+&align(64);
+
+########################################################################
+# void ecp_nistz256_mul_by_2(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_mul_by_2");
+       &mov    ("esi",&wparam(1));
+       &mov    ("edi",&wparam(0));
+       &mov    ("ebp","esi");
+########################################################################
+# common pattern for internal functions is that %edi is result pointer,
+# %esi and %ebp are input ones, %ebp being optional. %edi is preserved.
+       &call   ("_ecp_nistz256_add");
+&function_end("ecp_nistz256_mul_by_2");
+
+########################################################################
+# void ecp_nistz256_mul_by_3(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_mul_by_3");
+       &mov    ("esi",&wparam(1));
+                                       # multiplication by 3 is performed
+                                       # as 2*n+n, but we can't use output
+                                       # to store 2*n, because if output
+                                       # pointer equals to input, then
+                                       # we'll get 2*n+2*n.
+       &stack_push(8);                 # therefore we need to allocate
+                                       # 256-bit intermediate buffer.
+       &mov    ("edi","esp");
+       &mov    ("ebp","esi");
+       &call   ("_ecp_nistz256_add");
+       &lea    ("esi",&DWP(0,"edi"));
+       &mov    ("ebp",&wparam(1));
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_add");
+       &stack_pop(8);
+&function_end("ecp_nistz256_mul_by_3");
+
+########################################################################
+# void ecp_nistz256_div_by_2(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_div_by_2");
+       &mov    ("esi",&wparam(1));
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_div_by_2");
+&function_end("ecp_nistz256_div_by_2");
+
+&function_begin_B("_ecp_nistz256_div_by_2");
+       # tmp = a is odd ? a+mod : a
+       #
+       # note that because mod has special form, i.e. consists of
+       # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
+       # assigning least significant bit of input to one register,
+       # %ebp, and its negative to another, %edx.
+
+       &mov    ("ebp",&DWP(0,"esi"));
+       &xor    ("edx","edx");
+       &mov    ("ebx",&DWP(4,"esi"));
+       &mov    ("eax","ebp");
+       &and    ("ebp",1);
+       &mov    ("ecx",&DWP(8,"esi"));
+       &sub    ("edx","ebp");
+
+       &add    ("eax","edx");
+       &adc    ("ebx","edx");
+       &mov    (&DWP(0,"edi"),"eax");
+       &adc    ("ecx","edx");
+       &mov    (&DWP(4,"edi"),"ebx");
+       &mov    (&DWP(8,"edi"),"ecx");
+
+       &mov    ("eax",&DWP(12,"esi"));
+       &mov    ("ebx",&DWP(16,"esi"));
+       &adc    ("eax",0);
+       &mov    ("ecx",&DWP(20,"esi"));
+       &adc    ("ebx",0);
+       &mov    (&DWP(12,"edi"),"eax");
+       &adc    ("ecx",0);
+       &mov    (&DWP(16,"edi"),"ebx");
+       &mov    (&DWP(20,"edi"),"ecx");
+
+       &mov    ("eax",&DWP(24,"esi"));
+       &mov    ("ebx",&DWP(28,"esi"));
+       &adc    ("eax","ebp");
+       &adc    ("ebx","edx");
+       &mov    (&DWP(24,"edi"),"eax");
+       &sbb    ("esi","esi");                  # broadcast carry bit
+       &mov    (&DWP(28,"edi"),"ebx");
+
+       # ret = tmp >> 1
+
+       &mov    ("eax",&DWP(0,"edi"));
+       &mov    ("ebx",&DWP(4,"edi"));
+       &mov    ("ecx",&DWP(8,"edi"));
+       &mov    ("edx",&DWP(12,"edi"));
+
+       &shr    ("eax",1);
+       &mov    ("ebp","ebx");
+       &shl    ("ebx",31);
+       &or     ("eax","ebx");
+
+       &shr    ("ebp",1);
+       &mov    ("ebx","ecx");
+       &shl    ("ecx",31);
+       &mov    (&DWP(0,"edi"),"eax");
+       &or     ("ebp","ecx");
+       &mov    ("eax",&DWP(16,"edi"));
+
+       &shr    ("ebx",1);
+       &mov    ("ecx","edx");
+       &shl    ("edx",31);
+       &mov    (&DWP(4,"edi"),"ebp");
+       &or     ("ebx","edx");
+       &mov    ("ebp",&DWP(20,"edi"));
+
+       &shr    ("ecx",1);
+       &mov    ("edx","eax");
+       &shl    ("eax",31);
+       &mov    (&DWP(8,"edi"),"ebx");
+       &or     ("ecx","eax");
+       &mov    ("ebx",&DWP(24,"edi"));
+
+       &shr    ("edx",1);
+       &mov    ("eax","ebp");
+       &shl    ("ebp",31);
+       &mov    (&DWP(12,"edi"),"ecx");
+       &or     ("edx","ebp");
+       &mov    ("ecx",&DWP(28,"edi"));
+
+       &shr    ("eax",1);
+       &mov    ("ebp","ebx");
+       &shl    ("ebx",31);
+       &mov    (&DWP(16,"edi"),"edx");
+       &or     ("eax","ebx");
+
+       &shr    ("ebp",1);
+       &mov    ("ebx","ecx");
+       &shl    ("ecx",31);
+       &mov    (&DWP(20,"edi"),"eax");
+       &or     ("ebp","ecx");
+
+       &shr    ("ebx",1);
+       &shl    ("esi",31);
+       &mov    (&DWP(24,"edi"),"ebp");
+       &or     ("ebx","esi");                  # handle top-most carry bit
+       &mov    (&DWP(28,"edi"),"ebx");
+
+       &ret    ();
+&function_end_B("_ecp_nistz256_div_by_2");
+
+########################################################################
+# void ecp_nistz256_add(BN_ULONG edi[8],const BN_ULONG esi[8],
+#                                      const BN_ULONG ebp[8]);
+&function_begin("ecp_nistz256_add");
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_add");
+&function_end("ecp_nistz256_add");
+
+&function_begin_B("_ecp_nistz256_add");
+       &mov    ("eax",&DWP(0,"esi"));
+       &mov    ("ebx",&DWP(4,"esi"));
+       &mov    ("ecx",&DWP(8,"esi"));
+       &add    ("eax",&DWP(0,"ebp"));
+       &mov    ("edx",&DWP(12,"esi"));
+       &adc    ("ebx",&DWP(4,"ebp"));
+       &mov    (&DWP(0,"edi"),"eax");
+       &adc    ("ecx",&DWP(8,"ebp"));
+       &mov    (&DWP(4,"edi"),"ebx");
+       &adc    ("edx",&DWP(12,"ebp"));
+       &mov    (&DWP(8,"edi"),"ecx");
+       &mov    (&DWP(12,"edi"),"edx");
+
+       &mov    ("eax",&DWP(16,"esi"));
+       &mov    ("ebx",&DWP(20,"esi"));
+       &mov    ("ecx",&DWP(24,"esi"));
+       &adc    ("eax",&DWP(16,"ebp"));
+       &mov    ("edx",&DWP(28,"esi"));
+       &adc    ("ebx",&DWP(20,"ebp"));
+       &mov    (&DWP(16,"edi"),"eax");
+       &adc    ("ecx",&DWP(24,"ebp"));
+       &mov    (&DWP(20,"edi"),"ebx");
+       &adc    ("edx",&DWP(28,"ebp"));
+       &mov    (&DWP(24,"edi"),"ecx");
+       &sbb    ("esi","esi");                  # broadcast carry bit
+       &mov    (&DWP(28,"edi"),"edx");
+
+       # if a+b carries, subtract modulus.
+       #
+       # Note that because mod has special form, i.e. consists of
+       # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
+       # assigning carry bit to one register, %ebp, and its negative
+       # to another, %esi. But we started by calculating %esi...
+
+       &mov    ("eax",&DWP(0,"edi"));
+       &mov    ("ebp","esi");
+       &mov    ("ebx",&DWP(4,"edi"));
+       &shr    ("ebp",31);
+       &mov    ("ecx",&DWP(8,"edi"));
+       &sub    ("eax","esi");
+       &mov    ("edx",&DWP(12,"edi"));
+       &sbb    ("ebx","esi");
+       &mov    (&DWP(0,"edi"),"eax");
+       &sbb    ("ecx","esi");
+       &mov    (&DWP(4,"edi"),"ebx");
+       &sbb    ("edx",0);
+       &mov    (&DWP(8,"edi"),"ecx");
+       &mov    (&DWP(12,"edi"),"edx");
+
+       &mov    ("eax",&DWP(16,"edi"));
+       &mov    ("ebx",&DWP(20,"edi"));
+       &mov    ("ecx",&DWP(24,"edi"));
+       &sbb    ("eax",0);
+       &mov    ("edx",&DWP(28,"edi"));
+       &sbb    ("ebx",0);
+       &mov    (&DWP(16,"edi"),"eax");
+       &sbb    ("ecx","ebp");
+       &mov    (&DWP(20,"edi"),"ebx");
+       &sbb    ("edx","esi");
+       &mov    (&DWP(24,"edi"),"ecx");
+       &mov    (&DWP(28,"edi"),"edx");
+
+       &ret    ();
+&function_end_B("_ecp_nistz256_add");
+
+########################################################################
+# void ecp_nistz256_sub(BN_ULONG edi[8],const BN_ULONG esi[8],
+#                                      const BN_ULONG ebp[8]);
+&function_begin("ecp_nistz256_sub");
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_sub");
+&function_end("ecp_nistz256_sub");
+
+&function_begin_B("_ecp_nistz256_sub");
+       &mov    ("eax",&DWP(0,"esi"));
+       &mov    ("ebx",&DWP(4,"esi"));
+       &mov    ("ecx",&DWP(8,"esi"));
+       &sub    ("eax",&DWP(0,"ebp"));
+       &mov    ("edx",&DWP(12,"esi"));
+       &sbb    ("ebx",&DWP(4,"ebp"));
+       &mov    (&DWP(0,"edi"),"eax");
+       &sbb    ("ecx",&DWP(8,"ebp"));
+       &mov    (&DWP(4,"edi"),"ebx");
+       &sbb    ("edx",&DWP(12,"ebp"));
+       &mov    (&DWP(8,"edi"),"ecx");
+       &mov    (&DWP(12,"edi"),"edx");
+
+       &mov    ("eax",&DWP(16,"esi"));
+       &mov    ("ebx",&DWP(20,"esi"));
+       &mov    ("ecx",&DWP(24,"esi"));
+       &sbb    ("eax",&DWP(16,"ebp"));
+       &mov    ("edx",&DWP(28,"esi"));
+       &sbb    ("ebx",&DWP(20,"ebp"));
+       &sbb    ("ecx",&DWP(24,"ebp"));
+       &mov    (&DWP(16,"edi"),"eax");
+       &sbb    ("edx",&DWP(28,"ebp"));
+       &mov    (&DWP(20,"edi"),"ebx");
+       &sbb    ("esi","esi");                  # broadcast borrow bit
+       &mov    (&DWP(24,"edi"),"ecx");
+       &mov    (&DWP(28,"edi"),"edx");
+
+       # if a-b borrows, add modulus.
+       #
+       # Note that because mod has special form, i.e. consists of
+       # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
+       # assigning borrow bit to one register, %ebp, and its negative
+       # to another, %esi. But we started by calculating %esi...
+
+       &mov    ("eax",&DWP(0,"edi"));
+       &mov    ("ebp","esi");
+       &mov    ("ebx",&DWP(4,"edi"));
+       &shr    ("ebp",31);
+       &mov    ("ecx",&DWP(8,"edi"));
+       &add    ("eax","esi");
+       &mov    ("edx",&DWP(12,"edi"));
+       &adc    ("ebx","esi");
+       &mov    (&DWP(0,"edi"),"eax");
+       &adc    ("ecx","esi");
+       &mov    (&DWP(4,"edi"),"ebx");
+       &adc    ("edx",0);
+       &mov    (&DWP(8,"edi"),"ecx");
+       &mov    (&DWP(12,"edi"),"edx");
+
+       &mov    ("eax",&DWP(16,"edi"));
+       &mov    ("ebx",&DWP(20,"edi"));
+       &mov    ("ecx",&DWP(24,"edi"));
+       &adc    ("eax",0);
+       &mov    ("edx",&DWP(28,"edi"));
+       &adc    ("ebx",0);
+       &mov    (&DWP(16,"edi"),"eax");
+       &adc    ("ecx","ebp");
+       &mov    (&DWP(20,"edi"),"ebx");
+       &adc    ("edx","esi");
+       &mov    (&DWP(24,"edi"),"ecx");
+       &mov    (&DWP(28,"edi"),"edx");
+
+       &ret    ();
+&function_end_B("_ecp_nistz256_sub");
+
+########################################################################
+# void ecp_nistz256_neg(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_neg");
+       &mov    ("ebp",&wparam(1));
+       &mov    ("edi",&wparam(0));
+
+       &xor    ("eax","eax");
+       &stack_push(8);
+       &mov    (&DWP(0,"esp"),"eax");
+       &mov    ("esi","esp");
+       &mov    (&DWP(4,"esp"),"eax");
+       &mov    (&DWP(8,"esp"),"eax");
+       &mov    (&DWP(12,"esp"),"eax");
+       &mov    (&DWP(16,"esp"),"eax");
+       &mov    (&DWP(20,"esp"),"eax");
+       &mov    (&DWP(24,"esp"),"eax");
+       &mov    (&DWP(28,"esp"),"eax");
+       
+       &call   ("_ecp_nistz256_sub");
+
+       &stack_pop(8);
+&function_end("ecp_nistz256_neg");
+
+&function_begin_B("_picup_eax");
+       &mov    ("eax",&DWP(0,"esp"));
+       &ret    ();
+&function_end_B("_picup_eax");
+
+########################################################################
+# void ecp_nistz256_to_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_to_mont");
+       &mov    ("esi",&wparam(1));
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &lea    ("ebp",&DWP(&label("RR")."-".&label("pic"),"eax"));
+                                               if ($sse2) {
+       &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("eax",&DWP(0,"eax"));          }
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_mul_mont");
+&function_end("ecp_nistz256_to_mont");
+
+########################################################################
+# void ecp_nistz256_from_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_from_mont");
+       &mov    ("esi",&wparam(1));
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &lea    ("ebp",&DWP(&label("ONE")."-".&label("pic"),"eax"));
+                                               if ($sse2) {
+       &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("eax",&DWP(0,"eax"));          }
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_mul_mont");
+&function_end("ecp_nistz256_from_mont");
+
+########################################################################
+# void ecp_nistz256_mul_mont(BN_ULONG edi[8],const BN_ULONG esi[8],
+#                                           const BN_ULONG ebp[8]);
+&function_begin("ecp_nistz256_mul_mont");
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+                                               if ($sse2) {
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("eax",&DWP(0,"eax"));          }
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_mul_mont");
+&function_end("ecp_nistz256_mul_mont");
+
+########################################################################
+# void ecp_nistz256_sqr_mont(BN_ULONG edi[8],const BN_ULONG esi[8]);
+&function_begin("ecp_nistz256_sqr_mont");
+       &mov    ("esi",&wparam(1));
+                                               if ($sse2) {
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &picmeup("eax","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("eax",&DWP(0,"eax"));          }
+       &mov    ("edi",&wparam(0));
+       &mov    ("ebp","esi");
+       &call   ("_ecp_nistz256_mul_mont");
+&function_end("ecp_nistz256_sqr_mont");
+
+&function_begin_B("_ecp_nistz256_mul_mont");
+                                               if ($sse2) {
+       &and    ("eax",1<<24|1<<26);
+       &cmp    ("eax",1<<24|1<<26);            # see if XMM+SSE2 is on
+       &jne    (&label("mul_mont_ialu"));
+
+       ########################################
+       # SSE2 code path featuring 32x16-bit
+       # multiplications is ~2x faster than
+       # IALU counterpart (except on Atom)...
+       ########################################
+       # stack layout:
+       # +------------------------------------+< %esp
+       # | 7 16-byte temporary XMM words,     |
+       # | "sliding" toward lower address     |
+       # .                                    .
+       # +------------------------------------+
+       # | unused XMM word                    |
+       # +------------------------------------+< +128,%ebx
+       # | 8 16-byte XMM words holding copies |
+       # | of a[i]<<64|a[i]                   |
+       # .                                    .
+       # .                                    .
+       # +------------------------------------+< +256
+       &mov    ("edx","esp");
+       &sub    ("esp",0x100);
+
+       &movd   ("xmm7",&DWP(0,"ebp"));         # b[0] -> 0000.00xy
+       &lea    ("ebp",&DWP(4,"ebp"));
+       &pcmpeqd("xmm6","xmm6");
+       &psrlq  ("xmm6",48);                    # compose 0xffff<<64|0xffff
+
+       &pshuflw("xmm7","xmm7",0b11011100);     # 0000.00xy -> 0000.0x0y
+       &and    ("esp",-64);
+       &pshufd ("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
+       &lea    ("ebx",&DWP(0x80,"esp"));
+
+       &movd   ("xmm0",&DWP(4*0,"esi"));       # a[0] -> 0000.00xy
+       &pshufd ("xmm0","xmm0",0b11001100);     # 0000.00xy -> 00xy.00xy
+       &movd   ("xmm1",&DWP(4*1,"esi"));       # a[1] -> ...
+       &movdqa (&QWP(0x00,"ebx"),"xmm0");      # offload converted a[0]
+       &pmuludq("xmm0","xmm7");                # a[0]*b[0]
+
+       &movd   ("xmm2",&DWP(4*2,"esi"));
+       &pshufd ("xmm1","xmm1",0b11001100);
+       &movdqa (&QWP(0x10,"ebx"),"xmm1");
+       &pmuludq("xmm1","xmm7");                # a[1]*b[0]
+
+        &movq  ("xmm4","xmm0");                # clear upper 64 bits
+        &pslldq("xmm4",6);
+        &paddq ("xmm4","xmm0");
+        &movdqa("xmm5","xmm4");
+        &psrldq("xmm4",10);                    # upper 32 bits of a[0]*b[0]
+        &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[0]
+
+       # Upper half of a[0]*b[i] is carried into next multiplication
+       # iteration, while lower one "participates" in actual reduction.
+       # Normally latter is done by accumulating result of multiplication
+       # of modulus by "magic" digit, but thanks to special form of modulus
+       # and "magic" digit it can be performed only with additions and
+       # subtractions (see note in IALU section below). Note that we are
+       # not bothered with carry bits, they are accumulated in "flatten"
+       # phase after all multiplications and reductions.
+
+       &movd   ("xmm3",&DWP(4*3,"esi"));
+       &pshufd ("xmm2","xmm2",0b11001100);
+       &movdqa (&QWP(0x20,"ebx"),"xmm2");
+       &pmuludq("xmm2","xmm7");                # a[2]*b[0]
+        &paddq ("xmm1","xmm4");                # a[1]*b[0]+hw(a[0]*b[0]), carry
+       &movdqa (&QWP(0x00,"esp"),"xmm1");      # t[0]
+
+       &movd   ("xmm0",&DWP(4*4,"esi"));
+       &pshufd ("xmm3","xmm3",0b11001100);
+       &movdqa (&QWP(0x30,"ebx"),"xmm3");
+       &pmuludq("xmm3","xmm7");                # a[3]*b[0]
+       &movdqa (&QWP(0x10,"esp"),"xmm2");
+
+       &movd   ("xmm1",&DWP(4*5,"esi"));
+       &pshufd ("xmm0","xmm0",0b11001100);
+       &movdqa (&QWP(0x40,"ebx"),"xmm0");
+       &pmuludq("xmm0","xmm7");                # a[4]*b[0]
+        &paddq ("xmm3","xmm5");                # a[3]*b[0]+lw(a[0]*b[0]), reduction step
+       &movdqa (&QWP(0x20,"esp"),"xmm3");
+
+       &movd   ("xmm2",&DWP(4*6,"esi"));
+       &pshufd ("xmm1","xmm1",0b11001100);
+       &movdqa (&QWP(0x50,"ebx"),"xmm1");
+       &pmuludq("xmm1","xmm7");                # a[5]*b[0]
+       &movdqa (&QWP(0x30,"esp"),"xmm0");
+        &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
+
+       &movd   ("xmm3",&DWP(4*7,"esi"));
+       &pshufd ("xmm2","xmm2",0b11001100);
+       &movdqa (&QWP(0x60,"ebx"),"xmm2");
+       &pmuludq("xmm2","xmm7");                # a[6]*b[0]
+       &movdqa (&QWP(0x40,"esp"),"xmm1");
+        &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
+
+       &movd   ("xmm0",&DWP(0,"ebp"));         # b[1] -> 0000.00xy
+       &pshufd ("xmm3","xmm3",0b11001100);
+       &movdqa (&QWP(0x70,"ebx"),"xmm3");
+       &pmuludq("xmm3","xmm7");                # a[7]*b[0]
+
+       &pshuflw("xmm7","xmm0",0b11011100);     # 0000.00xy -> 0000.0x0y
+       &movdqa ("xmm0",&QWP(0x00,"ebx"));      # pre-load converted a[0]
+       &pshufd ("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
+
+       &mov    ("ecx",6);
+       &lea    ("ebp",&DWP(4,"ebp"));
+       &jmp    (&label("madd_sse2"));
+
+&set_label("madd_sse2",16);
+        &paddq ("xmm2","xmm5");                # a[6]*b[i-1]+lw(a[0]*b[i-1]), reduction step [modulo-scheduled]
+        &paddq ("xmm3","xmm4");                # a[7]*b[i-1]+lw(a[0]*b[i-1])*0xffffffff, reduction step [modulo-scheduled]
+       &movdqa ("xmm1",&QWP(0x10,"ebx"));
+       &pmuludq("xmm0","xmm7");                # a[0]*b[i]
+        &movdqa(&QWP(0x50,"esp"),"xmm2");
+
+       &movdqa ("xmm2",&QWP(0x20,"ebx"));
+       &pmuludq("xmm1","xmm7");                # a[1]*b[i]
+        &movdqa(&QWP(0x60,"esp"),"xmm3");
+       &paddq  ("xmm0",&QWP(0x00,"esp"));
+
+       &movdqa ("xmm3",&QWP(0x30,"ebx"));
+       &pmuludq("xmm2","xmm7");                # a[2]*b[i]
+        &movq  ("xmm4","xmm0");                # clear upper 64 bits
+        &pslldq("xmm4",6);
+       &paddq  ("xmm1",&QWP(0x10,"esp"));
+        &paddq ("xmm4","xmm0");
+        &movdqa("xmm5","xmm4");
+        &psrldq("xmm4",10);                    # upper 33 bits of a[0]*b[i]+t[0]
+
+       &movdqa ("xmm0",&QWP(0x40,"ebx"));
+       &pmuludq("xmm3","xmm7");                # a[3]*b[i]
+        &paddq ("xmm1","xmm4");                # a[1]*b[i]+hw(a[0]*b[i]), carry
+       &paddq  ("xmm2",&QWP(0x20,"esp"));
+       &movdqa (&QWP(0x00,"esp"),"xmm1");
+
+       &movdqa ("xmm1",&QWP(0x50,"ebx"));
+       &pmuludq("xmm0","xmm7");                # a[4]*b[i]
+       &paddq  ("xmm3",&QWP(0x30,"esp"));
+       &movdqa (&QWP(0x10,"esp"),"xmm2");
+        &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[i]
+
+       &movdqa ("xmm2",&DWP(0x60,"ebx"));
+       &pmuludq("xmm1","xmm7");                # a[5]*b[i]
+        &paddq ("xmm3","xmm5");                # a[3]*b[i]+lw(a[0]*b[i]), reduction step
+       &paddq  ("xmm0",&QWP(0x40,"esp"));
+       &movdqa (&QWP(0x20,"esp"),"xmm3");
+        &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
+
+       &movdqa ("xmm3","xmm7");
+       &pmuludq("xmm2","xmm7");                # a[6]*b[i]
+        &movd  ("xmm7",&DWP(0,"ebp"));         # b[i++] -> 0000.00xy
+        &lea   ("ebp",&DWP(4,"ebp"));
+       &paddq  ("xmm1",&QWP(0x50,"esp"));
+        &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
+       &movdqa (&QWP(0x30,"esp"),"xmm0");
+        &pshuflw("xmm7","xmm7",0b11011100);    # 0000.00xy -> 0000.0x0y
+
+       &pmuludq("xmm3",&QWP(0x70,"ebx"));      # a[7]*b[i]
+        &pshufd("xmm7","xmm7",0b11011100);     # 0000.0x0y -> 000x.000y
+        &movdqa("xmm0",&QWP(0x00,"ebx"));      # pre-load converted a[0]
+       &movdqa (&QWP(0x40,"esp"),"xmm1");
+       &paddq  ("xmm2",&QWP(0x60,"esp"));
+
+       &dec    ("ecx");
+       &jnz    (&label("madd_sse2"));
+
+        &paddq ("xmm2","xmm5");                # a[6]*b[6]+lw(a[0]*b[6]), reduction step [modulo-scheduled]
+        &paddq ("xmm3","xmm4");                # a[7]*b[6]+lw(a[0]*b[6])*0xffffffff, reduction step [modulo-scheduled]
+       &movdqa ("xmm1",&QWP(0x10,"ebx"));
+       &pmuludq("xmm0","xmm7");                # a[0]*b[7]
+        &movdqa(&QWP(0x50,"esp"),"xmm2");
+
+       &movdqa ("xmm2",&QWP(0x20,"ebx"));
+       &pmuludq("xmm1","xmm7");                # a[1]*b[7]
+        &movdqa(&QWP(0x60,"esp"),"xmm3");
+       &paddq  ("xmm0",&QWP(0x00,"esp"));
+
+       &movdqa ("xmm3",&QWP(0x30,"ebx"));
+       &pmuludq("xmm2","xmm7");                # a[2]*b[7]
+        &movq  ("xmm4","xmm0");                # clear upper 64 bits
+        &pslldq("xmm4",6);
+       &paddq  ("xmm1",&QWP(0x10,"esp"));
+        &paddq ("xmm4","xmm0");
+        &movdqa("xmm5","xmm4");
+        &psrldq("xmm4",10);                    # upper 33 bits of a[0]*b[i]+t[0]
+
+       &movdqa ("xmm0",&QWP(0x40,"ebx"));
+       &pmuludq("xmm3","xmm7");                # a[3]*b[7]
+        &paddq ("xmm1","xmm4");                # a[1]*b[7]+hw(a[0]*b[7]), carry
+       &paddq  ("xmm2",&QWP(0x20,"esp"));
+       &movdqa (&QWP(0x00,"esp"),"xmm1");
+
+       &movdqa ("xmm1",&QWP(0x50,"ebx"));
+       &pmuludq("xmm0","xmm7");                # a[4]*b[7]
+       &paddq  ("xmm3",&QWP(0x30,"esp"));
+       &movdqa (&QWP(0x10,"esp"),"xmm2");
+        &pand  ("xmm5","xmm6");                # lower 32 bits of a[0]*b[i]
+
+       &movdqa ("xmm2",&DWP(0x60,"ebx"));
+       &pmuludq("xmm1","xmm7");                # a[5]*b[7]
+        &paddq ("xmm3","xmm5");                # reduction step
+       &paddq  ("xmm0",&QWP(0x40,"esp"));
+       &movdqa (&QWP(0x20,"esp"),"xmm3");
+        &pshufd("xmm4","xmm5",0b10110001);     # xmm4 = xmm5<<32, reduction step
+
+       &movdqa ("xmm3",&QWP(0x70,"ebx"));
+       &pmuludq("xmm2","xmm7");                # a[6]*b[7]
+       &paddq  ("xmm1",&QWP(0x50,"esp"));
+        &psubq ("xmm4","xmm5");                # xmm4 = xmm5*0xffffffff, reduction step
+       &movdqa (&QWP(0x30,"esp"),"xmm0");
+
+       &pmuludq("xmm3","xmm7");                # a[7]*b[7]
+       &pcmpeqd("xmm7","xmm7");
+       &movdqa ("xmm0",&QWP(0x00,"esp"));
+       &pslldq ("xmm7",8);
+       &movdqa (&QWP(0x40,"esp"),"xmm1");
+       &paddq  ("xmm2",&QWP(0x60,"esp"));
+
+        &paddq ("xmm2","xmm5");                # a[6]*b[7]+lw(a[0]*b[7]), reduction step
+        &paddq ("xmm3","xmm4");                # a[6]*b[7]+lw(a[0]*b[7])*0xffffffff, reduction step
+        &movdqa(&QWP(0x50,"esp"),"xmm2");
+        &movdqa(&QWP(0x60,"esp"),"xmm3");
+
+       &movdqa ("xmm1",&QWP(0x10,"esp"));
+       &movdqa ("xmm2",&QWP(0x20,"esp"));
+       &movdqa ("xmm3",&QWP(0x30,"esp"));
+
+       &movq   ("xmm4","xmm0");                # "flatten"
+       &pand   ("xmm0","xmm7");
+       &xor    ("ebp","ebp");
+       &pslldq ("xmm4",6);
+        &movq  ("xmm5","xmm1");
+       &paddq  ("xmm0","xmm4");
+        &pand  ("xmm1","xmm7");
+       &psrldq ("xmm0",6);
+       &movd   ("eax","xmm0");
+       &psrldq ("xmm0",4);
+
+       &paddq  ("xmm5","xmm0");
+       &movdqa ("xmm0",&QWP(0x40,"esp"));
+       &sub    ("eax",-1);                     # start subtracting modulus,
+                                               # this is used to determine
+                                               # if result is larger/smaller
+                                               # than modulus (see below)
+       &pslldq ("xmm5",6);
+        &movq  ("xmm4","xmm2");
+       &paddq  ("xmm1","xmm5");
+        &pand  ("xmm2","xmm7");
+       &psrldq ("xmm1",6);
+       &mov    (&DWP(4*0,"edi"),"eax");
+       &movd   ("eax","xmm1");
+       &psrldq ("xmm1",4);
+
+       &paddq  ("xmm4","xmm1");
+       &movdqa ("xmm1",&QWP(0x50,"esp"));
+       &sbb    ("eax",-1);
+       &pslldq ("xmm4",6);
+        &movq  ("xmm5","xmm3");
+       &paddq  ("xmm2","xmm4");
+        &pand  ("xmm3","xmm7");
+       &psrldq ("xmm2",6);
+       &mov    (&DWP(4*1,"edi"),"eax");
+       &movd   ("eax","xmm2");
+       &psrldq ("xmm2",4);
+
+       &paddq  ("xmm5","xmm2");
+       &movdqa ("xmm2",&QWP(0x60,"esp"));
+       &sbb    ("eax",-1);
+       &pslldq ("xmm5",6);
+        &movq  ("xmm4","xmm0");
+       &paddq  ("xmm3","xmm5");
+        &pand  ("xmm0","xmm7");
+       &psrldq ("xmm3",6);
+       &mov    (&DWP(4*2,"edi"),"eax");
+       &movd   ("eax","xmm3");
+       &psrldq ("xmm3",4);
+
+       &paddq  ("xmm4","xmm3");
+       &sbb    ("eax",0);
+       &pslldq ("xmm4",6);
+        &movq  ("xmm5","xmm1");
+       &paddq  ("xmm0","xmm4");
+        &pand  ("xmm1","xmm7");
+       &psrldq ("xmm0",6);
+       &mov    (&DWP(4*3,"edi"),"eax");
+       &movd   ("eax","xmm0");
+       &psrldq ("xmm0",4);
+
+       &paddq  ("xmm5","xmm0");
+       &sbb    ("eax",0);
+       &pslldq ("xmm5",6);
+        &movq  ("xmm4","xmm2");
+       &paddq  ("xmm1","xmm5");
+        &pand  ("xmm2","xmm7");
+       &psrldq ("xmm1",6);
+       &movd   ("ebx","xmm1");
+       &psrldq ("xmm1",4);
+       &mov    ("esp","edx");
+
+       &paddq  ("xmm4","xmm1");
+       &pslldq ("xmm4",6);
+       &paddq  ("xmm2","xmm4");
+       &psrldq ("xmm2",6);
+       &movd   ("ecx","xmm2");
+       &psrldq ("xmm2",4);
+       &sbb    ("ebx",0);
+       &movd   ("edx","xmm2");
+       &pextrw ("esi","xmm2",2);               # top-most overflow bit
+       &sbb    ("ecx",1);
+       &sbb    ("edx",-1);
+       &sbb    ("esi",0);                      # borrow from subtraction
+
+       # Final step is "if result > mod, subtract mod", and at this point
+       # we have result - mod written to output buffer, as well as borrow
+       # bit from this subtraction, and if borrow bit is set, we add
+       # modulus back.
+       #
+       # Note that because mod has special form, i.e. consists of
+       # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
+       # assigning borrow bit to one register, %ebp, and its negative
+       # to another, %esi. But we started by calculating %esi...
+
+       &sub    ("ebp","esi");
+       &add    (&DWP(4*0,"edi"),"esi");        # add modulus or zero
+       &adc    (&DWP(4*1,"edi"),"esi");
+       &adc    (&DWP(4*2,"edi"),"esi");
+       &adc    (&DWP(4*3,"edi"),0);
+       &adc    ("eax",0);
+       &adc    ("ebx",0);
+       &mov    (&DWP(4*4,"edi"),"eax");
+       &adc    ("ecx","ebp");
+       &mov    (&DWP(4*5,"edi"),"ebx");
+       &adc    ("edx","esi");
+       &mov    (&DWP(4*6,"edi"),"ecx");
+       &mov    (&DWP(4*7,"edi"),"edx");
+
+       &ret    ();
+
+&set_label("mul_mont_ialu",16);                        }
+
+       ########################################
+       # IALU code path suitable for all CPUs.
+       ########################################
+       # stack layout:
+       # +------------------------------------+< %esp
+       # | 8 32-bit temporary words, accessed |
+       # | as circular buffer                 |
+       # .                                    .
+       # .                                    .
+       # +------------------------------------+< +32
+       # | offloaded destination pointer      |
+       # +------------------------------------+
+       # | unused                             |
+       # +------------------------------------+< +40
+       &sub    ("esp",10*4);
+
+       &mov    ("eax",&DWP(0*4,"esi"));                # a[0]
+       &mov    ("ebx",&DWP(0*4,"ebp"));                # b[0]
+       &mov    (&DWP(8*4,"esp"),"edi");                # off-load dst ptr
+
+       &mul    ("ebx");                                # a[0]*b[0]
+       &mov    (&DWP(0*4,"esp"),"eax");                # t[0]
+       &mov    ("eax",&DWP(1*4,"esi"));
+       &mov    ("ecx","edx")
+
+       &mul    ("ebx");                                # a[1]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(2*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(1*4,"esp"),"ecx");                # t[1]
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[2]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(3*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(2*4,"esp"),"ecx");                # t[2]
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[3]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(4*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(3*4,"esp"),"ecx");                # t[3]
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[4]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(5*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(4*4,"esp"),"ecx");                # t[4]
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[5]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(6*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(5*4,"esp"),"ecx");                # t[5]
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[6]*b[0]
+       &add    ("ecx","eax");
+       &mov    ("eax",&DWP(7*4,"esi"));
+       &adc    ("edx",0);
+       &mov    (&DWP(6*4,"esp"),"ecx");                # t[6]
+       &mov    ("ecx","edx");
+
+       &xor    ("edi","edi");                          # initial top-most carry
+       &mul    ("ebx");                                # a[7]*b[0]
+       &add    ("ecx","eax");                          # t[7]
+       &mov    ("eax",&DWP(0*4,"esp"));                # t[0]
+       &adc    ("edx",0);                              # t[8]
+
+for ($i=0;$i<7;$i++) {
+       my $j=$i+1;
+
+       # Reduction iteration is normally performed by accumulating
+       # result of multiplication of modulus by "magic" digit [and
+       # omitting least significant word, which is guaranteed to
+       # be 0], but thanks to special form of modulus and "magic"
+       # digit being equal to least significant word, it can be
+       # performed with additions and subtractions alone. Indeed:
+       #
+       #        ffff.0001.0000.0000.0000.ffff.ffff.ffff
+       # *                                         abcd
+       # + xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.abcd
+       #
+       # Now observing that ff..ff*x = (2^n-1)*x = 2^n*x-x, we
+       # rewrite above as:
+       #
+       #   xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.abcd
+       # + abcd.0000.abcd.0000.0000.abcd.0000.0000.0000
+       # -      abcd.0000.0000.0000.0000.0000.0000.abcd
+       #
+       # or marking redundant operations:
+       #
+       #   xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.xxxx.----
+       # + abcd.0000.abcd.0000.0000.abcd.----.----.----
+       # -      abcd.----.----.----.----.----.----.----
+
+       &add    (&DWP((($i+3)%8)*4,"esp"),"eax");       # t[3]+=t[0]
+       &adc    (&DWP((($i+4)%8)*4,"esp"),0);           # t[4]+=0
+       &adc    (&DWP((($i+5)%8)*4,"esp"),0);           # t[5]+=0
+       &adc    (&DWP((($i+6)%8)*4,"esp"),"eax");       # t[6]+=t[0]
+       &adc    ("ecx",0);                              # t[7]+=0
+       &adc    ("edx","eax");                          # t[8]+=t[0]
+       &adc    ("edi",0);                              # top-most carry
+        &mov   ("ebx",&DWP($j*4,"ebp"));               # b[i]
+       &sub    ("ecx","eax");                          # t[7]-=t[0]
+        &mov   ("eax",&DWP(0*4,"esi"));                # a[0]
+       &sbb    ("edx",0);                              # t[8]-=0
+       &mov    (&DWP((($i+7)%8)*4,"esp"),"ecx");
+       &sbb    ("edi",0);                              # top-most carry,
+                                                       # keep in mind that
+                                                       # netto result is
+                                                       # *addition* of value
+                                                       # with (abcd<<32)-abcd
+                                                       # on top, so that
+                                                       # underflow is
+                                                       # impossible, because
+                                                       # (abcd<<32)-abcd
+                                                       # doesn't underflow
+       &mov    (&DWP((($i+8)%8)*4,"esp"),"edx");
+
+       &mul    ("ebx");                                # a[0]*b[i]
+       &add    ("eax",&DWP((($j+0)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &mov    (&DWP((($j+0)%8)*4,"esp"),"eax");
+       &mov    ("eax",&DWP(1*4,"esi"));
+       &mov    ("ecx","edx")
+
+       &mul    ("ebx");                                # a[1]*b[i]
+       &add    ("ecx",&DWP((($j+1)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(2*4,"esi"));
+       &mov    (&DWP((($j+1)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[2]*b[i]
+       &add    ("ecx",&DWP((($j+2)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(3*4,"esi"));
+       &mov    (&DWP((($j+2)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[3]*b[i]
+       &add    ("ecx",&DWP((($j+3)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(4*4,"esi"));
+       &mov    (&DWP((($j+3)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[4]*b[i]
+       &add    ("ecx",&DWP((($j+4)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(5*4,"esi"));
+       &mov    (&DWP((($j+4)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[5]*b[i]
+       &add    ("ecx",&DWP((($j+5)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(6*4,"esi"));
+       &mov    (&DWP((($j+5)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[6]*b[i]
+       &add    ("ecx",&DWP((($j+6)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");
+       &adc    ("edx",0);
+       &mov    ("eax",&DWP(7*4,"esi"));
+       &mov    (&DWP((($j+6)%8)*4,"esp"),"ecx");
+       &mov    ("ecx","edx");
+
+       &mul    ("ebx");                                # a[7]*b[i]
+       &add    ("ecx",&DWP((($j+7)%8)*4,"esp"));
+       &adc    ("edx",0);
+       &add    ("ecx","eax");                          # t[7]
+       &mov    ("eax",&DWP((($j+0)%8)*4,"esp"));       # t[0]
+       &adc    ("edx","edi");                          # t[8]
+       &mov    ("edi",0);
+       &adc    ("edi",0);                              # top-most carry
+}
+       &mov    ("ebp",&DWP(8*4,"esp"));                # restore dst ptr
+       &xor    ("esi","esi");
+       my $j=$i+1;
+
+       # last multiplication-less reduction
+       &add    (&DWP((($i+3)%8)*4,"esp"),"eax");       # t[3]+=t[0]
+       &adc    (&DWP((($i+4)%8)*4,"esp"),0);           # t[4]+=0
+       &adc    (&DWP((($i+5)%8)*4,"esp"),0);           # t[5]+=0
+       &adc    (&DWP((($i+6)%8)*4,"esp"),"eax");       # t[6]+=t[0]
+       &adc    ("ecx",0);                              # t[7]+=0
+       &adc    ("edx","eax");                          # t[8]+=t[0]
+       &adc    ("edi",0);                              # top-most carry
+        &mov   ("ebx",&DWP((($j+1)%8)*4,"esp"));
+       &sub    ("ecx","eax");                          # t[7]-=t[0]
+        &mov   ("eax",&DWP((($j+0)%8)*4,"esp"));
+       &sbb    ("edx",0);                              # t[8]-=0
+       &mov    (&DWP((($i+7)%8)*4,"esp"),"ecx");
+       &sbb    ("edi",0);                              # top-most carry
+       &mov    (&DWP((($i+8)%8)*4,"esp"),"edx");
+
+       # Final step is "if result > mod, subtract mod", but we do it
+       # "other way around", namely write result - mod to output buffer
+       # and if subtraction borrowed, add modulus back.
+
+       &mov    ("ecx",&DWP((($j+2)%8)*4,"esp"));
+       &sub    ("eax",-1);
+       &mov    ("edx",&DWP((($j+3)%8)*4,"esp"));
+       &sbb    ("ebx",-1);
+       &mov    (&DWP(0*4,"ebp"),"eax");
+       &sbb    ("ecx",-1);
+       &mov    (&DWP(1*4,"ebp"),"ebx");
+       &sbb    ("edx",0);
+       &mov    (&DWP(2*4,"ebp"),"ecx");
+       &mov    (&DWP(3*4,"ebp"),"edx");
+
+       &mov    ("eax",&DWP((($j+4)%8)*4,"esp"));
+       &mov    ("ebx",&DWP((($j+5)%8)*4,"esp"));
+       &mov    ("ecx",&DWP((($j+6)%8)*4,"esp"));
+       &sbb    ("eax",0);
+       &mov    ("edx",&DWP((($j+7)%8)*4,"esp"));
+       &sbb    ("ebx",0);
+       &sbb    ("ecx",1);
+       &sbb    ("edx",-1);
+       &sbb    ("edi",0);
+
+       # Note that because mod has special form, i.e. consists of
+       # 0xffffffff, 1 and 0s, we can conditionally synthesize it by
+       # assigning borrow bit to one register, %ebp, and its negative
+       # to another, %esi. But we started by calculating %esi...
+
+       &sub    ("esi","edi");
+       &add    (&DWP(0*4,"ebp"),"edi");                # add modulus or zero
+       &adc    (&DWP(1*4,"ebp"),"edi");
+       &adc    (&DWP(2*4,"ebp"),"edi");
+       &adc    (&DWP(3*4,"ebp"),0);
+       &adc    ("eax",0);
+       &adc    ("ebx",0);
+       &mov    (&DWP(4*4,"ebp"),"eax");
+       &adc    ("ecx","esi");
+       &mov    (&DWP(5*4,"ebp"),"ebx");
+       &adc    ("edx","edi");
+       &mov    (&DWP(6*4,"ebp"),"ecx");
+       &mov    ("edi","ebp");                          # fulfill contract
+       &mov    (&DWP(7*4,"ebp"),"edx");
+
+       &add    ("esp",10*4);
+       &ret    ();
+&function_end_B("_ecp_nistz256_mul_mont");
+
+########################################################################
+# void ecp_nistz256_scatter_w5(void *edi,const P256_POINT *esi,
+#                                       int ebp);
+&function_begin("ecp_nistz256_scatter_w5");
+       &mov    ("edi",&wparam(0));
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+
+       &lea    ("edi",&DWP(128-4,"edi","ebp",4));
+       &mov    ("ebp",96/16);
+&set_label("scatter_w5_loop");
+       &mov    ("eax",&DWP(0,"esi"));
+       &mov    ("ebx",&DWP(4,"esi"));
+       &mov    ("ecx",&DWP(8,"esi"));
+       &mov    ("edx",&DWP(12,"esi"));
+       &lea    ("esi",&DWP(16,"esi"));
+       &mov    (&DWP(64*0-128,"edi"),"eax");
+       &mov    (&DWP(64*1-128,"edi"),"ebx");
+       &mov    (&DWP(64*2-128,"edi"),"ecx");
+       &mov    (&DWP(64*3-128,"edi"),"edx");
+       &lea    ("edi",&DWP(64*4,"edi"));
+       &dec    ("ebp");
+       &jnz    (&label("scatter_w5_loop"));
+&function_end("ecp_nistz256_scatter_w5");
+
+########################################################################
+# void ecp_nistz256_gather_w5(P256_POINT *edi,const void *esi,
+#                                            int ebp);
+&function_begin("ecp_nistz256_gather_w5");
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+
+       &lea    ("esi",&DWP(0,"esi","ebp",4));
+       &neg    ("ebp");
+       &sar    ("ebp",31);
+       &mov    ("edi",&wparam(0));
+       &lea    ("esi",&DWP(0,"esi","ebp",4));
+
+    for($i=0;$i<24;$i+=4) {
+       &mov    ("eax",&DWP(64*($i+0),"esi"));
+       &mov    ("ebx",&DWP(64*($i+1),"esi"));
+       &mov    ("ecx",&DWP(64*($i+2),"esi"));
+       &mov    ("edx",&DWP(64*($i+3),"esi"));
+       &and    ("eax","ebp");
+       &and    ("ebx","ebp");
+       &and    ("ecx","ebp");
+       &and    ("edx","ebp");
+       &mov    (&DWP(4*($i+0),"edi"),"eax");
+       &mov    (&DWP(4*($i+1),"edi"),"ebx");
+       &mov    (&DWP(4*($i+2),"edi"),"ecx");
+       &mov    (&DWP(4*($i+3),"edi"),"edx");
+    }
+&function_end("ecp_nistz256_gather_w5");
+
+########################################################################
+# void ecp_nistz256_scatter_w7(void *edi,const P256_POINT_AFFINE *esi,
+#                                       int ebp);
+&function_begin("ecp_nistz256_scatter_w7");
+       &mov    ("edi",&wparam(0));
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+
+       &lea    ("edi",&DWP(-1,"edi","ebp"));
+       &mov    ("ebp",64/4);
+&set_label("scatter_w7_loop");
+       &mov    ("eax",&DWP(0,"esi"));
+       &lea    ("esi",&DWP(4,"esi"));
+       &mov    (&BP(64*0,"edi"),"al");
+       &mov    (&BP(64*1,"edi"),"ah");
+       &shr    ("eax",16);
+       &mov    (&BP(64*2,"edi"),"al");
+       &mov    (&BP(64*3,"edi"),"ah");
+       &lea    ("edi",&DWP(64*4,"edi"));
+       &dec    ("ebp");
+       &jnz    (&label("scatter_w7_loop"));
+&function_end("ecp_nistz256_scatter_w7");
+
+########################################################################
+# void ecp_nistz256_gather_w7(P256_POINT_AFFINE *edi,const void *esi,
+#                                                   int ebp);
+&function_begin("ecp_nistz256_gather_w7");
+       &mov    ("esi",&wparam(1));
+       &mov    ("ebp",&wparam(2));
+
+       &add    ("esi","ebp");
+       &neg    ("ebp"),
+       &sar    ("ebp",31);
+       &mov    ("edi",&wparam(0));
+       &lea    ("esi",&DWP(0,"esi","ebp"));
+
+    for($i=0;$i<64;$i+=4) {
+       &movz   ("eax",&BP(64*($i+0),"esi"));
+       &movz   ("ebx",&BP(64*($i+1),"esi"));
+       &movz   ("ecx",&BP(64*($i+2),"esi"));
+       &and    ("eax","ebp");
+       &movz   ("edx",&BP(64*($i+3),"esi"));
+       &and    ("ebx","ebp");
+       &mov    (&BP($i+0,"edi"),"al");
+       &and    ("ecx","ebp");
+       &mov    (&BP($i+1,"edi"),"bl");
+       &and    ("edx","ebp");
+       &mov    (&BP($i+2,"edi"),"cl");
+       &mov    (&BP($i+3,"edi"),"dl");
+    }
+&function_end("ecp_nistz256_gather_w7");
+
+########################################################################
+# following subroutines are "literal" implementation of those found in
+# ecp_nistz256.c
+#
+########################################################################
+# void ecp_nistz256_point_double(P256_POINT *out,const P256_POINT *inp);
+#
+&function_begin("ecp_nistz256_point_double");
+{   my ($S,$M,$Zsqr,$in_x,$tmp0)=map(32*$_,(0..4));
+
+       &mov    ("esi",&wparam(1));
+
+       # above map() describes stack layout with 5 temporary
+       # 256-bit vectors on top, then we take extra word for
+       # OPENSS_ia32cap_P copy.
+       &stack_push(8*5+1);
+                                               if ($sse2) {
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("ebp",&DWP(0,"edx"));          }
+
+       &mov    ("eax",&DWP(0,"esi"));          # copy in_x
+       &mov    ("ebx",&DWP(4,"esi"));
+       &mov    ("ecx",&DWP(8,"esi"));
+       &mov    ("edx",&DWP(12,"esi"));
+       &mov    (&DWP($in_x+0,"esp"),"eax");
+       &mov    (&DWP($in_x+4,"esp"),"ebx");
+       &mov    (&DWP($in_x+8,"esp"),"ecx");
+       &mov    (&DWP($in_x+12,"esp"),"edx");
+       &mov    ("eax",&DWP(16,"esi"));
+       &mov    ("ebx",&DWP(20,"esi"));
+       &mov    ("ecx",&DWP(24,"esi"));
+       &mov    ("edx",&DWP(28,"esi"));
+       &mov    (&DWP($in_x+16,"esp"),"eax");
+       &mov    (&DWP($in_x+20,"esp"),"ebx");
+       &mov    (&DWP($in_x+24,"esp"),"ecx");
+       &mov    (&DWP($in_x+28,"esp"),"edx");
+       &mov    (&DWP(32*5,"esp"),"ebp");       # OPENSSL_ia32cap_P copy
+
+       &lea    ("ebp",&DWP(32,"esi"));
+       &lea    ("esi",&DWP(32,"esi"));
+       &lea    ("edi",&DWP($S,"esp"));
+       &call   ("_ecp_nistz256_add");          # p256_mul_by_2(S, in_y);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &mov    ("esi",64);
+       &add    ("esi",&wparam(1));
+       &lea    ("edi",&DWP($Zsqr,"esp"));
+       &mov    ("ebp","esi");
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Zsqr, in_z);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($S,"esp"));
+       &lea    ("ebp",&DWP($S,"esp"));
+       &lea    ("edi",&DWP($S,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(S, S);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &mov    ("ebp",&wparam(1));
+       &lea    ("esi",&DWP(32,"ebp"));
+       &lea    ("ebp",&DWP(64,"ebp"));
+       &lea    ("edi",&DWP($tmp0,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(tmp0, in_z, in_y);
+
+       &lea    ("esi",&DWP($in_x,"esp"));
+       &lea    ("ebp",&DWP($Zsqr,"esp"));
+       &lea    ("edi",&DWP($M,"esp"));
+       &call   ("_ecp_nistz256_add");          # p256_add(M, in_x, Zsqr);
+
+       &mov    ("edi",64);
+       &lea    ("esi",&DWP($tmp0,"esp"));
+       &lea    ("ebp",&DWP($tmp0,"esp"));
+       &add    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_add");          # p256_mul_by_2(res_z, tmp0);
+
+       &lea    ("esi",&DWP($in_x,"esp"));
+       &lea    ("ebp",&DWP($Zsqr,"esp"));
+       &lea    ("edi",&DWP($Zsqr,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(Zsqr, in_x, Zsqr);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($S,"esp"));
+       &lea    ("ebp",&DWP($S,"esp"));
+       &lea    ("edi",&DWP($tmp0,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(tmp0, S);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($M,"esp"));
+       &lea    ("ebp",&DWP($Zsqr,"esp"));
+       &lea    ("edi",&DWP($M,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(M, M, Zsqr);
+
+       &mov    ("edi",32);
+       &lea    ("esi",&DWP($tmp0,"esp"));
+       &add    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_div_by_2");     # p256_div_by_2(res_y, tmp0);
+
+       &lea    ("esi",&DWP($M,"esp"));
+       &lea    ("ebp",&DWP($M,"esp"));
+       &lea    ("edi",&DWP($tmp0,"esp"));
+       &call   ("_ecp_nistz256_add");          # 1/2 p256_mul_by_3(M, M);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in_x,"esp"));
+       &lea    ("ebp",&DWP($S,"esp"));
+       &lea    ("edi",&DWP($S,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S, S, in_x);
+
+       &lea    ("esi",&DWP($tmp0,"esp"));
+       &lea    ("ebp",&DWP($M,"esp"));
+       &lea    ("edi",&DWP($M,"esp"));
+       &call   ("_ecp_nistz256_add");          # 2/2 p256_mul_by_3(M, M);
+
+       &lea    ("esi",&DWP($S,"esp"));
+       &lea    ("ebp",&DWP($S,"esp"));
+       &lea    ("edi",&DWP($tmp0,"esp"));
+       &call   ("_ecp_nistz256_add");          # p256_mul_by_2(tmp0, S);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($M,"esp"));
+       &lea    ("ebp",&DWP($M,"esp"));
+       &mov    ("edi",&wparam(0));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(res_x, M);
+
+       &mov    ("esi","edi");                  # %edi is still res_x here
+       &lea    ("ebp",&DWP($tmp0,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, tmp0);
+
+       &lea    ("esi",&DWP($S,"esp"));
+       &mov    ("ebp","edi");                  # %edi is still res_x
+       &lea    ("edi",&DWP($S,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(S, S, res_x);
+
+       &mov    ("eax",&DWP(32*5,"esp"));       # OPENSSL_ia32cap_P copy
+       &mov    ("esi","edi");                  # %edi is still &S
+       &lea    ("ebp",&DWP($M,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S, S, M);
+
+       &mov    ("ebp",32);
+       &lea    ("esi",&DWP($S,"esp"));
+       &add    ("ebp",&wparam(0));
+       &mov    ("edi","ebp");
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, S, res_y);
+
+       &stack_pop(8*5+1);
+} &function_end("ecp_nistz256_point_double");
+
+########################################################################
+# void ecp_nistz256_point_add(P256_POINT *out,const P256_POINT *in1,
+#                                            const P256_POINT *in2);
+&function_begin("ecp_nistz256_point_add");
+{   my ($res_x,$res_y,$res_z,
+       $in1_x,$in1_y,$in1_z,
+       $in2_x,$in2_y,$in2_z,
+       $H,$Hsqr,$R,$Rsqr,$Hcub,
+       $U1,$U2,$S1,$S2)=map(32*$_,(0..17));
+    my ($Z1sqr, $Z2sqr) = ($Hsqr, $Rsqr);
+
+       &mov    ("esi",&wparam(2));
+
+       # above map() describes stack layout with 18 temporary
+       # 256-bit vectors on top, then we take extra words for
+       # !in1infty, !in2infty, result of check for zero and
+       # OPENSS_ia32cap_P copy. [one unused word for padding]
+       &stack_push(8*18+5);
+                                               if ($sse2) {
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("ebp",&DWP(0,"edx"));          }
+
+       &lea    ("edi",&DWP($in2_x,"esp"));
+    for($i=0;$i<96;$i+=16) {
+       &mov    ("eax",&DWP($i+0,"esi"));       # copy in2
+       &mov    ("ebx",&DWP($i+4,"esi"));
+       &mov    ("ecx",&DWP($i+8,"esi"));
+       &mov    ("edx",&DWP($i+12,"esi"));
+       &mov    (&DWP($i+0,"edi"),"eax");
+       &mov    (&DWP(32*18+12,"esp"),"ebp")    if ($i==0);
+       &mov    ("ebp","eax")                   if ($i==0);
+       &or     ("ebp","eax")                   if ($i!=0 && $i<64);
+       &mov    (&DWP($i+4,"edi"),"ebx");
+       &or     ("ebp","ebx")                   if ($i<64);
+       &mov    (&DWP($i+8,"edi"),"ecx");
+       &or     ("ebp","ecx")                   if ($i<64);
+       &mov    (&DWP($i+12,"edi"),"edx");
+       &or     ("ebp","edx")                   if ($i<64);
+    }
+       &xor    ("eax","eax");
+       &mov    ("esi",&wparam(1));
+       &sub    ("eax","ebp");
+       &or     ("ebp","eax");
+       &sar    ("ebp",31);
+       &mov    (&DWP(32*18+4,"esp"),"ebp");    # !in2infty
+
+       &lea    ("edi",&DWP($in1_x,"esp"));
+    for($i=0;$i<96;$i+=16) {
+       &mov    ("eax",&DWP($i+0,"esi"));       # copy in1
+       &mov    ("ebx",&DWP($i+4,"esi"));
+       &mov    ("ecx",&DWP($i+8,"esi"));
+       &mov    ("edx",&DWP($i+12,"esi"));
+       &mov    (&DWP($i+0,"edi"),"eax");
+       &mov    ("ebp","eax")                   if ($i==0);
+       &or     ("ebp","eax")                   if ($i!=0 && $i<64);
+       &mov    (&DWP($i+4,"edi"),"ebx");
+       &or     ("ebp","ebx")                   if ($i<64);
+       &mov    (&DWP($i+8,"edi"),"ecx");
+       &or     ("ebp","ecx")                   if ($i<64);
+       &mov    (&DWP($i+12,"edi"),"edx");
+       &or     ("ebp","edx")                   if ($i<64);
+    }
+       &xor    ("eax","eax");
+       &sub    ("eax","ebp");
+       &or     ("ebp","eax");
+       &sar    ("ebp",31);
+       &mov    (&DWP(32*18+0,"esp"),"ebp");    # !in1infty
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_z,"esp"));
+       &lea    ("ebp",&DWP($in2_z,"esp"));
+       &lea    ("edi",&DWP($Z2sqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z2sqr, in2_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in1_z,"esp"));
+       &lea    ("ebp",&DWP($in1_z,"esp"));
+       &lea    ("edi",&DWP($Z1sqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z1sqr, in1_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($Z2sqr,"esp"));
+       &lea    ("ebp",&DWP($in2_z,"esp"));
+       &lea    ("edi",&DWP($S1,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S1, Z2sqr, in2_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($Z1sqr,"esp"));
+       &lea    ("ebp",&DWP($in1_z,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Z1sqr, in1_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in1_y,"esp"));
+       &lea    ("ebp",&DWP($S1,"esp"));
+       &lea    ("edi",&DWP($S1,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S1, S1, in1_y);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_y,"esp"));
+       &lea    ("ebp",&DWP($S2,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S2, in2_y);
+
+       &lea    ("esi",&DWP($S2,"esp"));
+       &lea    ("ebp",&DWP($S1,"esp"));
+       &lea    ("edi",&DWP($R,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(R, S2, S1);
+
+       &or     ("ebx","eax");                  # see if result is zero
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &or     ("ebx","ecx");
+       &or     ("ebx","edx");
+       &or     ("ebx",&DWP(0,"edi"));
+       &or     ("ebx",&DWP(4,"edi"));
+        &lea   ("esi",&DWP($in1_x,"esp"));
+       &or     ("ebx",&DWP(8,"edi"));
+        &lea   ("ebp",&DWP($Z2sqr,"esp"));
+       &or     ("ebx",&DWP(12,"edi"));
+        &lea   ("edi",&DWP($U1,"esp"));
+       &mov    (&DWP(32*18+8,"esp"),"ebx");
+
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U1, in1_x, Z2sqr);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_x,"esp"));
+       &lea    ("ebp",&DWP($Z1sqr,"esp"));
+       &lea    ("edi",&DWP($U2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, in2_x, Z1sqr);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($U1,"esp"));
+       &lea    ("edi",&DWP($H,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(H, U2, U1);
+
+       &or     ("eax","ebx");                  # see if result is zero
+       &or     ("eax","ecx");
+       &or     ("eax","edx");
+       &or     ("eax",&DWP(0,"edi"));
+       &or     ("eax",&DWP(4,"edi"));
+       &or     ("eax",&DWP(8,"edi"));
+       &or     ("eax",&DWP(12,"edi"));
+
+       &data_byte(0x3e);                       # predict taken
+       &jnz    (&label("add_proceed"));        # is_equal(U1,U2)?
+
+       &mov    ("eax",&DWP(32*18+0,"esp"));
+       &and    ("eax",&DWP(32*18+4,"esp"));
+       &mov    ("ebx",&DWP(32*18+8,"esp"));
+       &jz     (&label("add_proceed"));        # (in1infty || in2infty)?
+       &test   ("ebx","ebx");
+       &jz     (&label("add_proceed"));        # is_equal(S1,S2)?
+
+       &mov    ("edi",&wparam(0));
+       &xor    ("eax","eax");
+       &mov    ("ecx",96/4);
+       &data_byte(0xfc,0xf3,0xab);             # cld; stosd
+       &jmp    (&label("add_done"));
+
+&set_label("add_proceed",16);
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($R,"esp"));
+       &lea    ("ebp",&DWP($R,"esp"));
+       &lea    ("edi",&DWP($Rsqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Rsqr, R);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($H,"esp"));
+       &lea    ("ebp",&DWP($in1_z,"esp"));
+       &lea    ("edi",&DWP($res_z,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, H, in1_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($H,"esp"));
+       &lea    ("ebp",&DWP($H,"esp"));
+       &lea    ("edi",&DWP($Hsqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Hsqr, H);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_z,"esp"));
+       &lea    ("ebp",&DWP($res_z,"esp"));
+       &lea    ("edi",&DWP($res_z,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, res_z, in2_z);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($Hsqr,"esp"));
+       &lea    ("ebp",&DWP($U1,"esp"));
+       &lea    ("edi",&DWP($U2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, U1, Hsqr);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($H,"esp"));
+       &lea    ("ebp",&DWP($Hsqr,"esp"));
+       &lea    ("edi",&DWP($Hcub,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(Hcub, Hsqr, H);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($U2,"esp"));
+       &lea    ("edi",&DWP($Hsqr,"esp"));
+       &call   ("_ecp_nistz256_add");          # p256_mul_by_2(Hsqr, U2);
+
+       &lea    ("esi",&DWP($Rsqr,"esp"));
+       &lea    ("ebp",&DWP($Hsqr,"esp"));
+       &lea    ("edi",&DWP($res_x,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, Rsqr, Hsqr);
+
+       &lea    ("esi",&DWP($res_x,"esp"));
+       &lea    ("ebp",&DWP($Hcub,"esp"));
+       &lea    ("edi",&DWP($res_x,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, Hcub);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($res_x,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, U2, res_x);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($Hcub,"esp"));
+       &lea    ("ebp",&DWP($S1,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S1, Hcub);
+
+       &mov    ("eax",&DWP(32*18+12,"esp"));   # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($R,"esp"));
+       &lea    ("ebp",&DWP($res_y,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_y, R, res_y);
+
+       &lea    ("esi",&DWP($res_y,"esp"));
+       &lea    ("ebp",&DWP($S2,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, res_y, S2);
+
+       &mov    ("ebp",&DWP(32*18+0,"esp"));    # !in1infty
+       &mov    ("esi",&DWP(32*18+4,"esp"));    # !in2infty
+       &mov    ("edi",&wparam(0));
+       &mov    ("edx","ebp");
+       &not    ("ebp");
+       &and    ("edx","esi");
+       &and    ("ebp","esi");
+       &not    ("esi");
+
+       ########################################
+       # conditional moves
+    for($i=64;$i<96;$i+=4) {
+       &mov    ("eax","edx");
+       &and    ("eax",&DWP($res_x+$i,"esp"));
+       &mov    ("ebx","ebp");
+       &and    ("ebx",&DWP($in2_x+$i,"esp"));
+       &mov    ("ecx","esi");
+       &and    ("ecx",&DWP($in1_x+$i,"esp"));
+       &or     ("eax","ebx");
+       &or     ("eax","ecx");
+       &mov    (&DWP($i,"edi"),"eax");
+    }
+    for($i=0;$i<64;$i+=4) {
+       &mov    ("eax","edx");
+       &and    ("eax",&DWP($res_x+$i,"esp"));
+       &mov    ("ebx","ebp");
+       &and    ("ebx",&DWP($in2_x+$i,"esp"));
+       &mov    ("ecx","esi");
+       &and    ("ecx",&DWP($in1_x+$i,"esp"));
+       &or     ("eax","ebx");
+       &or     ("eax","ecx");
+       &mov    (&DWP($i,"edi"),"eax");
+    }
+    &set_label("add_done");
+       &stack_pop(8*18+5);
+} &function_end("ecp_nistz256_point_add");
+
+########################################################################
+# void ecp_nistz256_point_add_affine(P256_POINT *out,
+#                                   const P256_POINT *in1,
+#                                   const P256_POINT_AFFINE *in2);
+&function_begin("ecp_nistz256_point_add_affine");
+{
+    my ($res_x,$res_y,$res_z,
+       $in1_x,$in1_y,$in1_z,
+       $in2_x,$in2_y,
+       $U2,$S2,$H,$R,$Hsqr,$Hcub,$Rsqr)=map(32*$_,(0..14));
+    my $Z1sqr = $S2;
+    my @ONE_mont=(1,0,0,-1,-1,-1,-2,0);
+
+       &mov    ("esi",&wparam(1));
+
+       # above map() describes stack layout with 15 temporary
+       # 256-bit vectors on top, then we take extra words for
+       # !in1infty, !in2infty, and OPENSS_ia32cap_P copy.
+       &stack_push(8*15+3);
+                                               if ($sse2) {
+       &call   ("_picup_eax");
+    &set_label("pic");
+       &picmeup("edx","OPENSSL_ia32cap_P","eax",&label("pic"));
+       &mov    ("ebp",&DWP(0,"edx"));          }
+
+       &lea    ("edi",&DWP($in1_x,"esp"));
+    for($i=0;$i<96;$i+=16) {
+       &mov    ("eax",&DWP($i+0,"esi"));       # copy in1
+       &mov    ("ebx",&DWP($i+4,"esi"));
+       &mov    ("ecx",&DWP($i+8,"esi"));
+       &mov    ("edx",&DWP($i+12,"esi"));
+       &mov    (&DWP($i+0,"edi"),"eax");
+       &mov    (&DWP(32*15+8,"esp"),"ebp")     if ($i==0);
+       &mov    ("ebp","eax")                   if ($i==0);
+       &or     ("ebp","eax")                   if ($i!=0 && $i<64);
+       &mov    (&DWP($i+4,"edi"),"ebx");
+       &or     ("ebp","ebx")                   if ($i<64);
+       &mov    (&DWP($i+8,"edi"),"ecx");
+       &or     ("ebp","ecx")                   if ($i<64);
+       &mov    (&DWP($i+12,"edi"),"edx");
+       &or     ("ebp","edx")                   if ($i<64);
+    }
+       &xor    ("eax","eax");
+       &mov    ("esi",&wparam(2));
+       &sub    ("eax","ebp");
+       &or     ("ebp","eax");
+       &sar    ("ebp",31);
+       &mov    (&DWP(32*15+0,"esp"),"ebp");    # !in1infty
+
+       &lea    ("edi",&DWP($in2_x,"esp"));
+    for($i=0;$i<64;$i+=16) {
+       &mov    ("eax",&DWP($i+0,"esi"));       # copy in2
+       &mov    ("ebx",&DWP($i+4,"esi"));
+       &mov    ("ecx",&DWP($i+8,"esi"));
+       &mov    ("edx",&DWP($i+12,"esi"));
+       &mov    (&DWP($i+0,"edi"),"eax");
+       &mov    ("ebp","eax")                   if ($i==0);
+       &or     ("ebp","eax")                   if ($i!=0);
+       &mov    (&DWP($i+4,"edi"),"ebx");
+       &or     ("ebp","ebx");
+       &mov    (&DWP($i+8,"edi"),"ecx");
+       &or     ("ebp","ecx");
+       &mov    (&DWP($i+12,"edi"),"edx");
+       &or     ("ebp","edx");
+    }
+       &xor    ("ebx","ebx");
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &sub    ("ebx","ebp");
+        &lea   ("esi",&DWP($in1_z,"esp"));
+       &or     ("ebx","ebp");
+        &lea   ("ebp",&DWP($in1_z,"esp"));
+       &sar    ("ebx",31);
+        &lea   ("edi",&DWP($Z1sqr,"esp"));
+       &mov    (&DWP(32*15+4,"esp"),"ebx");    # !in2infty
+
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Z1sqr, in1_z);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_x,"esp"));
+       &mov    ("ebp","edi");                  # %esi is stull &Z1sqr
+       &lea    ("edi",&DWP($U2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, Z1sqr, in2_x);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in1_z,"esp"));
+       &lea    ("ebp",&DWP($Z1sqr,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Z1sqr, in1_z);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($in1_x,"esp"));
+       &lea    ("edi",&DWP($H,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(H, U2, in1_x);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in2_y,"esp"));
+       &lea    ("ebp",&DWP($S2,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, S2, in2_y);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in1_z,"esp"));
+       &lea    ("ebp",&DWP($H,"esp"));
+       &lea    ("edi",&DWP($res_z,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_z, H, in1_z);
+
+       &lea    ("esi",&DWP($S2,"esp"));
+       &lea    ("ebp",&DWP($in1_y,"esp"));
+       &lea    ("edi",&DWP($R,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(R, S2, in1_y);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($H,"esp"));
+       &lea    ("ebp",&DWP($H,"esp"));
+       &lea    ("edi",&DWP($Hsqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Hsqr, H);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($R,"esp"));
+       &lea    ("ebp",&DWP($R,"esp"));
+       &lea    ("edi",&DWP($Rsqr,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_sqr_mont(Rsqr, R);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($in1_x,"esp"));
+       &lea    ("ebp",&DWP($Hsqr,"esp"));
+       &lea    ("edi",&DWP($U2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(U2, in1_x, Hsqr);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($H,"esp"));
+       &lea    ("ebp",&DWP($Hsqr,"esp"));
+       &lea    ("edi",&DWP($Hcub,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(Hcub, Hsqr, H);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($U2,"esp"));
+       &lea    ("edi",&DWP($Hsqr,"esp"));
+       &call   ("_ecp_nistz256_add");          # p256_mul_by_2(Hsqr, U2);
+
+       &lea    ("esi",&DWP($Rsqr,"esp"));
+       &lea    ("ebp",&DWP($Hsqr,"esp"));
+       &lea    ("edi",&DWP($res_x,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, Rsqr, Hsqr);
+
+       &lea    ("esi",&DWP($res_x,"esp"));
+       &lea    ("ebp",&DWP($Hcub,"esp"));
+       &lea    ("edi",&DWP($res_x,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_x, res_x, Hcub);
+
+       &lea    ("esi",&DWP($U2,"esp"));
+       &lea    ("ebp",&DWP($res_x,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, U2, res_x);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($Hcub,"esp"));
+       &lea    ("ebp",&DWP($in1_y,"esp"));
+       &lea    ("edi",&DWP($S2,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(S2, Hcub, in1_y);
+
+       &mov    ("eax",&DWP(32*15+8,"esp"));    # OPENSSL_ia32cap_P copy
+       &lea    ("esi",&DWP($R,"esp"));
+       &lea    ("ebp",&DWP($res_y,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_mul_mont");     # p256_mul_mont(res_y, res_y, R);
+
+       &lea    ("esi",&DWP($res_y,"esp"));
+       &lea    ("ebp",&DWP($S2,"esp"));
+       &lea    ("edi",&DWP($res_y,"esp"));
+       &call   ("_ecp_nistz256_sub");          # p256_sub(res_y, res_y, S2);
+
+       &mov    ("ebp",&DWP(32*15+0,"esp"));    # !in1infty
+       &mov    ("esi",&DWP(32*15+4,"esp"));    # !in2infty
+       &mov    ("edi",&wparam(0));
+       &mov    ("edx","ebp");
+       &not    ("ebp");
+       &and    ("edx","esi");
+       &and    ("ebp","esi");
+       &not    ("esi");
+
+       ########################################
+       # conditional moves
+    for($i=64;$i<96;$i+=4) {
+       my $one=@ONE_mont[($i-64)/4];
+
+       &mov    ("eax","edx");
+       &and    ("eax",&DWP($res_x+$i,"esp"));
+       &mov    ("ebx","ebp")                   if ($one && $one!=-1);
+       &and    ("ebx",$one)                    if ($one && $one!=-1);
+       &mov    ("ecx","esi");
+       &and    ("ecx",&DWP($in1_x+$i,"esp"));
+       &or     ("eax",$one==-1?"ebp":"ebx")    if ($one);
+       &or     ("eax","ecx");
+       &mov    (&DWP($i,"edi"),"eax");
+    }
+    for($i=0;$i<64;$i+=4) {
+       &mov    ("eax","edx");
+       &and    ("eax",&DWP($res_x+$i,"esp"));
+       &mov    ("ebx","ebp");
+       &and    ("ebx",&DWP($in2_x+$i,"esp"));
+       &mov    ("ecx","esi");
+       &and    ("ecx",&DWP($in1_x+$i,"esp"));
+       &or     ("eax","ebx");
+       &or     ("eax","ecx");
+       &mov    (&DWP($i,"edi"),"eax");
+    }
+       &stack_pop(8*15+3);
+} &function_end("ecp_nistz256_point_add_affine");
+
+&asm_finish();