Add AES-NI GCM stitch.

4e049c52 · Andy Polyakov · b4a9d5bf · 4e049c52 · 4e049c52 · 4e049c52
6 changed file
--- a/Configure
+++ b/Configure
@@ -128,7 +128,7 @@ my $x86_asm="x86cpuid.o:bn-586.o co-586.o x86-mont.o x86-gf2m.o:des-586.o crypt5

 my $x86_elf_asm="$x86_asm:elf";

-my $x86_64_asm="x86_64cpuid.o:x86_64-gcc.o x86_64-mont.o x86_64-mont5.o x86_64-gf2m.o modexp512-x86_64.o::aes-x86_64.o vpaes-x86_64.o bsaes-x86_64.o aesni-x86_64.o aesni-sha1-x86_64.o::md5-x86_64.o:sha1-x86_64.o sha256-x86_64.o sha512-x86_64.o::rc4-x86_64.o rc4-md5-x86_64.o:::wp-x86_64.o:cmll-x86_64.o cmll_misc.o:ghash-x86_64.o:e_padlock-x86_64.o";
+my $x86_64_asm="x86_64cpuid.o:x86_64-gcc.o x86_64-mont.o x86_64-mont5.o x86_64-gf2m.o modexp512-x86_64.o::aes-x86_64.o vpaes-x86_64.o bsaes-x86_64.o aesni-x86_64.o aesni-sha1-x86_64.o::md5-x86_64.o:sha1-x86_64.o sha256-x86_64.o sha512-x86_64.o::rc4-x86_64.o rc4-md5-x86_64.o:::wp-x86_64.o:cmll-x86_64.o cmll_misc.o:ghash-x86_64.o aesni-gcm-x86_64.o:e_padlock-x86_64.o";
 my $ia64_asm="ia64cpuid.o:bn-ia64.o ia64-mont.o::aes_core.o aes_cbc.o aes-ia64.o::md5-ia64.o:sha1-ia64.o sha256-ia64.o sha512-ia64.o::rc4-ia64.o rc4_skey.o:::::ghash-ia64.o::void";
 my $sparcv9_asm="sparcv9cap.o sparccpuid.o:bn-sparcv9.o sparcv9-mont.o sparcv9a-mont.o vis3-mont.o sparct4-mont.o sparcv9-gf2m.o:des_enc-sparc.o fcrypt_b.o:aes_core.o aes_cbc.o aes-sparcv9.o aest4-sparcv9.o::md5-sparcv9.o:sha1-sparcv9.o sha256-sparcv9.o sha512-sparcv9.o::::::camellia.o cmll_misc.o cmll_cbc.o cmllt4-sparcv9.o:ghash-sparcv9.o::void";
 my $sparcv8_asm=":sparcv8.o:des_enc-sparc.o fcrypt_b.o:::::::::::::void";

--- a/TABLE
+++ b/TABLE
@@ -318,7 +318,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -813,7 +813,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = auto
 $dso_scheme   = win32
@@ -1506,7 +1506,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = macosx
 $dso_scheme   = dlfcn
@@ -1671,7 +1671,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = auto
 $dso_scheme   = win32
@@ -1770,7 +1770,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = macosx
 $dso_scheme   = dlfcn
@@ -1836,7 +1836,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -2034,7 +2034,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -2562,7 +2562,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -2760,7 +2760,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -2826,7 +2826,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -4476,7 +4476,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -4509,7 +4509,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -4542,7 +4542,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -4740,7 +4740,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = mingw64
 $dso_scheme   = win32
@@ -5730,7 +5730,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn
@@ -5763,7 +5763,7 @@ $rmd160_obj   =
 $rc5_obj      = 
 $wp_obj       = wp-x86_64.o
 $cmll_obj     = cmll-x86_64.o cmll_misc.o
-$modes_obj    = ghash-x86_64.o
+$modes_obj    = ghash-x86_64.o aesni-gcm-x86_64.o
 $engines_obj  = e_padlock-x86_64.o
 $perlasm_scheme = elf
 $dso_scheme   = dlfcn

--- a/crypto/evp/e_aes.c
+++ b/crypto/evp/e_aes.c
@@ -227,6 +227,26 @@ void aesni_ccm64_decrypt_blocks (const unsigned char *in,
 			const unsigned char ivec[16],
 			unsigned char cmac[16]);

+#if defined(__x86_64) || defined(__x86_64__) || defined(_M_AMD64) || defined(_M_X64)
+size_t aesni_gcm_encrypt(const unsigned char *in,
+			unsigned char *out,
+			size_t len,
+			const void *key,
+			unsigned char ivec[16],
+			u64 *Xi);
+#define AES_gcm_encrypt	aesni_gcm_encrypt
+size_t aesni_gcm_decrypt(const unsigned char *in,
+			unsigned char *out,
+			size_t len,
+			const void *key,
+			unsigned char ivec[16],
+			u64 *Xi);
+#define AES_gcm_decrypt	aesni_gcm_decrypt
+void gcm_ghash_avx(u64 Xi[2],const u128 Htable[16],const u8 *in,size_t len);
+#define AES_GCM_ASM(gctx)	(gctx->ctr==aesni_ctr32_encrypt_blocks && \
+				 gctx->gcm.ghash==gcm_ghash_avx)
+#endif
+
 static int aesni_init_key(EVP_CIPHER_CTX *ctx, const unsigned char *key,
 		   const unsigned char *iv, int enc)
 	{
@@ -1297,8 +1317,24 @@ static int aes_gcm_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 		/* Encrypt payload */
 		if (gctx->ctr)
 			{
+			size_t bulk=0;
+#if defined(AES_GCM_ASM)
+			if (len>=32 && AES_GCM_ASM(gctx))
+				{
+				if (CRYPTO_gcm128_encrypt(&gctx->gcm,NULL,NULL,0))
+					return -1;
+
+				bulk = AES_gcm_encrypt(in,out,len,
+							gctx->gcm.key,
+							gctx->gcm.Yi.c,
+							gctx->gcm.Xi.u);
+				gctx->gcm.len.u[1] += bulk;
+				}
+#endif
 			if (CRYPTO_gcm128_encrypt_ctr32(&gctx->gcm,
-							in, out, len,
+							in +bulk,
+							out+bulk,
+							len-bulk,
 							gctx->ctr))
 				goto err;
 			}
@@ -1316,8 +1352,24 @@ static int aes_gcm_tls_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 		/* Decrypt */
 		if (gctx->ctr)
 			{
+			size_t bulk=0;
+#if defined(AES_GCM_ASM)
+			if (len>=16 && AES_GCM_ASM(gctx))
+				{
+				if (CRYPTO_gcm128_decrypt(&gctx->gcm,NULL,NULL,0))
+					return -1;
+
+				bulk = AES_gcm_decrypt(in,out,len,
+							gctx->gcm.key,
+							gctx->gcm.Yi.c,
+							gctx->gcm.Xi.u);
+				gctx->gcm.len.u[1] += bulk;
+				}
+#endif
 			if (CRYPTO_gcm128_decrypt_ctr32(&gctx->gcm,
-							in, out, len,
+							in +bulk,
+							out+bulk,
+							len-bulk,
 							gctx->ctr))
 				goto err;
 			}
@@ -1367,8 +1419,28 @@ static int aes_gcm_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 			{
 			if (gctx->ctr)
 				{
+				size_t bulk=0;
+#if defined(AES_GCM_ASM)
+				if (len>=32 && AES_GCM_ASM(gctx))
+					{
+					size_t res = (16-gctx->gcm.mres)%16;
+
+					if (CRYPTO_gcm128_encrypt(&gctx->gcm,
+							in,out,res))
+						return -1;
+
+					bulk = AES_gcm_encrypt(in+res,
+							out+res,len-res,								gctx->gcm.key,
+							gctx->gcm.Yi.c,
+							gctx->gcm.Xi.u);
+					gctx->gcm.len.u[1] += bulk;
+					bulk += res;
+					}
+#endif
 				if (CRYPTO_gcm128_encrypt_ctr32(&gctx->gcm,
-							in, out, len,
+							in +bulk,
+							out+bulk,
+							len-bulk,
 							gctx->ctr))
 					return -1;
 				}
@@ -1381,8 +1453,29 @@ static int aes_gcm_cipher(EVP_CIPHER_CTX *ctx, unsigned char *out,
 			{
 			if (gctx->ctr)
 				{
+				size_t bulk=0;
+#if defined(AES_GCM_ASM)
+				if (len>=16 && AES_GCM_ASM(gctx))
+					{
+					size_t res = (16-gctx->gcm.mres)%16;
+
+					if (CRYPTO_gcm128_decrypt(&gctx->gcm,
+							in,out,res))
+						return -1;
+
+					bulk = AES_gcm_decrypt(in+res,
+							out+res,len-res,
+							gctx->gcm.key,
+							gctx->gcm.Yi.c,
+							gctx->gcm.Xi.u);
+					gctx->gcm.len.u[1] += bulk;
+					bulk += res;
+					}
+#endif
 				if (CRYPTO_gcm128_decrypt_ctr32(&gctx->gcm,
-							in, out, len,
+							in +bulk,
+							out+bulk,
+							len-bulk,
 							gctx->ctr))
 					return -1;
 				}

--- a/crypto/evp/evptests.txt
+++ b/crypto/evp/evptests.txt
@@ -367,5 +367,7 @@ aes-128-gcm:00000000000000000000000000000000:000000000000000000000000:0000000000
 aes-128-gcm:00000000000000000000000000000000:000000000000000000000000:0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000:0388dace60b6a392f328c2b971b2fe78f795aaab494b5923f7fd89ff948bc1e0200211214e7394da2089b6acd093abe0c94da219118e297d7b7ebcbcc9c388f28ade7d85a8ee35616f7124a9d527029195b84d1b96c690ff2f2de30bf2ec89e00253786e126504f0dab90c48a30321de3345e6b0461e7c9e6c6b7afedde83f40::cac45f60e31efd3b5a43b98a22ce1aa1
 # 192 bytes plaintext, iv is chosen so that initial counter LSB is 0xFF
 aes-128-gcm:00000000000000000000000000000000:ffffffff000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000:000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000:56b3373ca9ef6e4a2b64fe1e9a17b61425f10d47a75a5fce13efc6bc784af24f4141bdd48cf7c770887afd573cca5418a9aeffcd7c5ceddfc6a78397b9a85b499da558257267caab2ad0b23ca476a53cb17fb41c4b8b475cb4f3f7165094c229c9e8c4dc0a2a5ff1903e501511221376a1cdb8364c5061a20cae74bc4acd76ceb0abc9fd3217ef9f8c90be402ddf6d8697f4f880dff15bfb7a6b28241ec8fe183c2d59e3f9dfff653c7126f0acb9e64211f42bae12af462b1070bef1ab5e3606::566f8ef683078bfdeeffa869d751a017
+# 288 bytes plaintext, iv is chosen so that initial counter LSB is 0xFF
+aes-128-gcm:00000000000000000000000000000000:ffffffff000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000:000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000: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::8b307f6b33286d0ab026a9ed3fe1e85f
 # 80 bytes plaintext, submitted by Intel
 aes-128-gcm:843ffcf5d2b72694d19ed01d01249412:dbcca32ebf9b804617c3aa9e:000102030405060708090a0b0c0d0e0f101112131415161718191a1b1c1d1e1f202122232425262728292a2b2c2d2e2f303132333435363738393a3b3c3d3e3f404142434445464748494a4b4c4d4e4f:6268c6fa2a80b2d137467f092f657ac04d89be2beaa623d61b5a868c8f03ff95d3dcee23ad2f1ab3a6c80eaf4b140eb05de3457f0fbc111a6b43d0763aa422a3013cf1dc37fe417d1fbfc449b75d4cc5:00000000000000000000000000000000101112131415161718191a1b1c1d1e1f:3b629ccfbc1119b7319e1dce2cd6fd6d
--- a/crypto/modes/Makefile
+++ b/crypto/modes/Makefile
@@ -50,6 +50,8 @@ ghash-x86.s:	asm/ghash-x86.pl
 	$(PERL) asm/ghash-x86.pl $(PERLASM_SCHEME) $(CFLAGS) $(PROCESSOR) > $@
 ghash-x86_64.s:	asm/ghash-x86_64.pl
 	$(PERL) asm/ghash-x86_64.pl $(PERLASM_SCHEME) > $@
+aesni-gcm-x86_64.s:	asm/aesni-gcm-x86_64.pl
+	$(PERL) asm/aesni-gcm-x86_64.pl $(PERLASM_SCHEME) > $@
 ghash-sparcv9.s:	asm/ghash-sparcv9.pl
 	$(PERL) asm/ghash-sparcv9.pl $@ $(CFLAGS)
 ghash-alpha.s:	asm/ghash-alpha.pl

--- a/crypto/modes/asm/aesni-gcm-x86_64.pl
+++ b/crypto/modes/asm/aesni-gcm-x86_64.pl
+#!/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/.
+# ====================================================================
+#
+#
+# AES-NI-CTR+GHASH stitch.
+#
+# February 2013
+#
+# OpenSSL GCM implementation is organized in such way that its
+# performance is rather close to the sum of its streamed components,
+# in the context parallelized AES-NI CTR and modulo-scheduled
+# PCLMULQDQ-enabled GHASH. Unfortunately, as no stitch implementation
+# was observed to perform significantly better than the sum of the
+# components on contemporary CPUs, the effort was deemed impossible to
+# justify. This module is based on combination of Intel submissions,
+# [1] and [2], with MOVBE twist suggested by Ilya Albrekht and Max
+# Locktyukhin of Intel Corp. who verified that it reduces shuffles
+# pressure with notable relative improvement on upcoming Haswell
+# processor. [Exact performance numbers to be added at launch.]
+#
+# [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
+# [2] http://www.intel.com/content/dam/www/public/us/en/documents/software-support/enabling-high-performance-gcm.pdf
+
+$flavour = shift;
+$output  = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
+		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
+	$avx = ($1>=2.19) + ($1>=2.22);
+}
+
+if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
+	    `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
+	$avx = ($1>=2.09) + ($1>=2.10);
+}
+
+if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
+	    `ml64 2>&1` =~ /Version ([0-9]+)\./) {
+	$avx = ($1>=10) + ($1>=11);
+}
+
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
+
+if ($avx>1) {{{
+
+($inp,$out,$len,$key,$ivp,$Xip)=("%rdi","%rsi","%rdx","%rcx","%r8","%r9");
+
+($Ii,$T1,$T2,$Hkey,
+ $Z0,$Z1,$Z2,$Z3,$Xi) = map("%xmm$_",(0..8));
+
+($inout0,$inout1,$inout2,$inout3,$inout4,$inout5,$rndkey) = map("%xmm$_",(9..15));
+
+($counter,$rounds,$ret,$const,$in0,$end0)=("%ebx","%ebp","%r10","%r11","%r14","%r15");
+
+$code=<<___;
+.text
+
+.type	_aesni_ctr32_ghash_6x,\@abi-omnipotent
+.align	32
+_aesni_ctr32_ghash_6x:
+	vmovdqu		0x20($const),$T2	# borrow $T2, .Lone_msb
+	sub		\$6,$len
+	vpxor		$Z0,$Z0,$Z0		# $Z0   = 0
+	vmovdqu		0x00-0x80($key),$rndkey
+	vpaddb		$T2,$T1,$inout1
+	vpaddb		$T2,$inout1,$inout2
+	vpaddb		$T2,$inout2,$inout3
+	vpaddb		$T2,$inout3,$inout4
+	vpaddb		$T2,$inout4,$inout5
+	vpxor		$rndkey,$T1,$inout0
+	vmovdqu		$Z0,16+8(%rsp)		# "$Z3" = 0
+	jmp		.Loop6x
+
+.align	32
+.Loop6x:
+	add		\$6<<24,$counter
+	jc		.Lhandle_ctr32		# discard $inout[1-5]?
+	vmovdqu		0x00-0x20($Xip),$Hkey	# $Hkey^1
+	  vpaddb	$T2,$inout5,$T1		# next counter value
+	  vpxor		$rndkey,$inout1,$inout1
+	  vpxor		$rndkey,$inout2,$inout2
+
+.Lresume_ctr32:
+	vmovdqu		$T1,($ivp)		# save next counter value
+	vpclmulqdq	\$0x10,$Hkey,$Z3,$Z1
+	  vpxor		$rndkey,$inout3,$inout3
+	  vmovups	0x10-0x80($key),$T2	# borrow $T2 for $rndkey
+	vpclmulqdq	\$0x01,$Hkey,$Z3,$Z2
+	xor		%r12,%r12
+	cmp		$in0,$end0
+
+	  vaesenc	$T2,$inout0,$inout0
+	vmovdqu		0x30+8(%rsp),$Ii	# I[4]
+	  vpxor		$rndkey,$inout4,$inout4
+	vpclmulqdq	\$0x00,$Hkey,$Z3,$T1
+	  vaesenc	$T2,$inout1,$inout1
+	  vpxor		$rndkey,$inout5,$inout5
+	setnc		%r12b
+	vpclmulqdq	\$0x11,$Hkey,$Z3,$Z3
+	  vaesenc	$T2,$inout2,$inout2
+	vmovdqu		0x10-0x20($Xip),$Hkey	# $Hkey^2
+	neg		%r12
+	  vaesenc	$T2,$inout3,$inout3
+	 vpxor		$Z1,$Z2,$Z2
+	vpclmulqdq	\$0x00,$Hkey,$Ii,$Z1
+	 vpxor		$Z0,$Xi,$Xi		# modulo-scheduled
+	  vaesenc	$T2,$inout4,$inout4
+	 vpxor		$Z1,$T1,$Z0
+	and		\$0x60,%r12
+	  vmovups	0x20-0x80($key),$rndkey
+	vpclmulqdq	\$0x10,$Hkey,$Ii,$T1
+	  vaesenc	$T2,$inout5,$inout5
+
+	vpclmulqdq	\$0x01,$Hkey,$Ii,$T2
+	lea		($in0,%r12),$in0
+	  vaesenc	$rndkey,$inout0,$inout0
+	 vpxor		16+8(%rsp),$Xi,$Xi	# modulo-scheduled [vpxor $Z3,$Xi,$Xi]
+	vpclmulqdq	\$0x11,$Hkey,$Ii,$Hkey
+	 vmovdqu	0x40+8(%rsp),$Ii	# I[3]
+	  vaesenc	$rndkey,$inout1,$inout1
+	movbe		0x58($in0),%r13
+	  vaesenc	$rndkey,$inout2,$inout2
+	movbe		0x50($in0),%r12
+	  vaesenc	$rndkey,$inout3,$inout3
+	mov		%r13,0x20+8(%rsp)
+	  vaesenc	$rndkey,$inout4,$inout4
+	mov		%r12,0x28+8(%rsp)
+	vmovdqu		0x30-0x20($Xip),$Z1	# borrow $Z1 for $Hkey^3
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vmovups	0x30-0x80($key),$rndkey
+	 vpxor		$T1,$Z2,$Z2
+	vpclmulqdq	\$0x00,$Z1,$Ii,$T1
+	  vaesenc	$rndkey,$inout0,$inout0
+	 vpxor		$T2,$Z2,$Z2
+	vpclmulqdq	\$0x10,$Z1,$Ii,$T2
+	  vaesenc	$rndkey,$inout1,$inout1
+	 vpxor		$Hkey,$Z3,$Z3
+	vpclmulqdq	\$0x01,$Z1,$Ii,$Hkey
+	  vaesenc	$rndkey,$inout2,$inout2
+	vpclmulqdq	\$0x11,$Z1,$Ii,$Z1
+	 vmovdqu	0x50+8(%rsp),$Ii	# I[2]
+	  vaesenc	$rndkey,$inout3,$inout3
+	  vaesenc	$rndkey,$inout4,$inout4
+	 vpxor		$T1,$Z0,$Z0
+	vmovdqu		0x40-0x20($Xip),$T1	# borrow $T1 for $Hkey^4
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vmovups	0x40-0x80($key),$rndkey
+	 vpxor		$T2,$Z2,$Z2
+	vpclmulqdq	\$0x00,$T1,$Ii,$T2
+	  vaesenc	$rndkey,$inout0,$inout0
+	 vpxor		$Hkey,$Z2,$Z2
+	vpclmulqdq	\$0x10,$T1,$Ii,$Hkey
+	  vaesenc	$rndkey,$inout1,$inout1
+	movbe		0x48($in0),%r13
+	 vpxor		$Z1,$Z3,$Z3
+	vpclmulqdq	\$0x01,$T1,$Ii,$Z1
+	  vaesenc	$rndkey,$inout2,$inout2
+	movbe		0x40($in0),%r12
+	vpclmulqdq	\$0x11,$T1,$Ii,$T1
+	 vmovdqu	0x60+8(%rsp),$Ii	# I[1]
+	  vaesenc	$rndkey,$inout3,$inout3
+	mov		%r13,0x30+8(%rsp)
+	  vaesenc	$rndkey,$inout4,$inout4
+	mov		%r12,0x38+8(%rsp)
+	 vpxor		$T2,$Z0,$Z0
+	vmovdqu		0x60-0x20($Xip),$T2	# borrow $T2 for $Hkey^5
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vmovups	0x50-0x80($key),$rndkey
+	 vpxor		$Hkey,$Z2,$Z2
+	vpclmulqdq	\$0x00,$T2,$Ii,$Hkey
+	  vaesenc	$rndkey,$inout0,$inout0
+	 vpxor		$Z1,$Z2,$Z2
+	vpclmulqdq	\$0x10,$T2,$Ii,$Z1
+	  vaesenc	$rndkey,$inout1,$inout1
+	movbe		0x38($in0),%r13
+	 vpxor		$T1,$Z3,$Z3
+	vpclmulqdq	\$0x01,$T2,$Ii,$T1
+	 vpxor		0x70+8(%rsp),$Xi,$Xi	# accumulate I[0]
+	  vaesenc	$rndkey,$inout2,$inout2
+	movbe		0x30($in0),%r12
+	vpclmulqdq	\$0x11,$T2,$Ii,$T2
+	  vaesenc	$rndkey,$inout3,$inout3
+	mov		%r13,0x40+8(%rsp)
+	  vaesenc	$rndkey,$inout4,$inout4
+	mov		%r12,0x48+8(%rsp)
+	 vpxor		$Hkey,$Z0,$Z0
+	 vmovdqu	0x70-0x20($Xip),$Hkey	# $Hkey^6
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vmovups	0x60-0x80($key),$rndkey
+	 vpxor		$Z1,$Z2,$Z2
+	vpclmulqdq	\$0x10,$Hkey,$Xi,$Z1
+	  vaesenc	$rndkey,$inout0,$inout0
+	 vpxor		$T1,$Z2,$Z2
+	vpclmulqdq	\$0x01,$Hkey,$Xi,$T1
+	  vaesenc	$rndkey,$inout1,$inout1
+	movbe		0x28($in0),%r13
+	 vpxor		$T2,$Z3,$Z3
+	vpclmulqdq	\$0x00,$Hkey,$Xi,$T2
+	  vaesenc	$rndkey,$inout2,$inout2
+	movbe		0x20($in0),%r12
+	vpclmulqdq	\$0x11,$Hkey,$Xi,$Xi
+	  vaesenc	$rndkey,$inout3,$inout3
+	mov		%r13,0x50+8(%rsp)
+	  vaesenc	$rndkey,$inout4,$inout4
+	mov		%r12,0x58+8(%rsp)
+	vpxor		$Z1,$Z2,$Z2
+	  vaesenc	$rndkey,$inout5,$inout5
+	vpxor		$T1,$Z2,$Z2
+
+	  vmovups	0x70-0x80($key),$rndkey
+	vpslldq		\$8,$Z2,$Z1
+	vpxor		$T2,$Z0,$Z0
+	vmovdqu		0x10($const),$Hkey	# .Lpoly
+
+	  vaesenc	$rndkey,$inout0,$inout0
+	vpxor		$Xi,$Z3,$Z3
+	  vaesenc	$rndkey,$inout1,$inout1
+	vpxor		$Z1,$Z0,$Z0
+	movbe		0x18($in0),%r13
+	  vaesenc	$rndkey,$inout2,$inout2
+	movbe		0x10($in0),%r12
+	vpalignr	\$8,$Z0,$Z0,$Ii		# 1st phase
+	vpclmulqdq	\$0x10,$Hkey,$Z0,$Z0
+	mov		%r13,0x60+8(%rsp)
+	  vaesenc	$rndkey,$inout3,$inout3
+	mov		%r12,0x68+8(%rsp)
+	  vaesenc	$rndkey,$inout4,$inout4
+	  vmovups	0x80-0x80($key),$T1	# borrow $T1 for $rndkey
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vaesenc	$T1,$inout0,$inout0
+	  vmovups	0x90-0x80($key),$rndkey
+	  vaesenc	$T1,$inout1,$inout1
+	vpsrldq		\$8,$Z2,$Z2
+	  vaesenc	$T1,$inout2,$inout2
+	vpxor		$Z2,$Z3,$Z3
+	  vaesenc	$T1,$inout3,$inout3
+	vpxor		$Ii,$Z0,$Z0
+	movbe		0x08($in0),%r13
+	  vaesenc	$T1,$inout4,$inout4
+	movbe		0x00($in0),%r12
+	  vaesenc	$T1,$inout5,$inout5
+	  vmovups	0xa0-0x80($key),$T1
+	  cmp		\$11,$rounds
+	  jb		.Lenc_tail		# 128-bit key
+
+	  vaesenc	$rndkey,$inout0,$inout0
+	  vaesenc	$rndkey,$inout1,$inout1
+	  vaesenc	$rndkey,$inout2,$inout2
+	  vaesenc	$rndkey,$inout3,$inout3
+	  vaesenc	$rndkey,$inout4,$inout4
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vaesenc	$T1,$inout0,$inout0
+	  vaesenc	$T1,$inout1,$inout1
+	  vaesenc	$T1,$inout2,$inout2
+	  vaesenc	$T1,$inout3,$inout3
+	  vaesenc	$T1,$inout4,$inout4
+	  vmovups	0xb0-0x80($key),$rndkey
+	  vaesenc	$T1,$inout5,$inout5
+	  vmovups	0xc0-0x80($key),$T1
+	  je		.Lenc_tail		# 192-bit key
+
+	  vaesenc	$rndkey,$inout0,$inout0
+	  vaesenc	$rndkey,$inout1,$inout1
+	  vaesenc	$rndkey,$inout2,$inout2
+	  vaesenc	$rndkey,$inout3,$inout3
+	  vaesenc	$rndkey,$inout4,$inout4
+	  vaesenc	$rndkey,$inout5,$inout5
+
+	  vaesenc	$T1,$inout0,$inout0
+	  vaesenc	$T1,$inout1,$inout1
+	  vaesenc	$T1,$inout2,$inout2
+	  vaesenc	$T1,$inout3,$inout3
+	  vaesenc	$T1,$inout4,$inout4
+	  vmovups	0xd0-0x80($key),$rndkey
+	  vaesenc	$T1,$inout5,$inout5
+	  vmovups	0xe0-0x80($key),$T1
+	  jmp		.Lenc_tail		# 256-bit key
+
+.align	32
+.Lhandle_ctr32:
+	vmovdqu		($const),$Ii		# borrow $Ii for .Lbswap_mask
+	  vpshufb	$Ii,$T1,$Z2		# byte-swap counter
+	  vmovdqu	0x30($const),$Z1	# borrow $Z1, .Ltwo_lsb
+	  vpaddd	0x40($const),$Z2,$inout1	# .Lone_lsb
+	  vpaddd	$Z1,$Z2,$inout2
+	vmovdqu		0x00-0x20($Xip),$Hkey	# $Hkey^1
+	  vpaddd	$Z1,$inout1,$inout3
+	  vpshufb	$Ii,$inout1,$inout1
+	  vpaddd	$Z1,$inout2,$inout4
+	  vpshufb	$Ii,$inout2,$inout2
+	  vpxor		$rndkey,$inout1,$inout1
+	  vpaddd	$Z1,$inout3,$inout5
+	  vpshufb	$Ii,$inout3,$inout3
+	  vpxor		$rndkey,$inout2,$inout2
+	  vpaddd	$Z1,$inout4,$T1		# byte-swapped next counter value
+	  vpshufb	$Ii,$inout4,$inout4
+	  vpshufb	$Ii,$inout5,$inout5
+	  vpshufb	$Ii,$T1,$T1		# next counter value
+	jmp		.Lresume_ctr32
+
+.align	32
+.Lenc_tail:
+	  vaesenc	$rndkey,$inout0,$inout0
+	vmovdqu		$Z3,16+8(%rsp)		# postpone vpxor $Z3,$Xi,$Xi
+	vpalignr	\$8,$Z0,$Z0,$Xi		# 2nd phase
+	  vaesenc	$rndkey,$inout1,$inout1
+	vpclmulqdq	\$0x10,$Hkey,$Z0,$Z0
+	  vpxor		0x00($inp),$T1,$T2
+	  vaesenc	$rndkey,$inout2,$inout2
+	  vpxor		0x10($inp),$T1,$Ii
+	  vaesenc	$rndkey,$inout3,$inout3
+	  vpxor		0x20($inp),$T1,$Z1
+	  vaesenc	$rndkey,$inout4,$inout4
+	  vpxor		0x30($inp),$T1,$Z2
+	  vaesenc	$rndkey,$inout5,$inout5
+	  vpxor		0x40($inp),$T1,$Z3
+	  vpxor		0x50($inp),$T1,$Hkey
+	  vmovdqu	($ivp),$T1		# load next counter value
+
+	  vaesenclast	$T2,$inout0,$inout0
+	  vmovdqu	0x20($const),$T2	# borrow $T2, .Lone_msb
+	  vaesenclast	$Ii,$inout1,$inout1
+	 vpaddb		$T2,$T1,$Ii
+	mov		%r13,0x70+8(%rsp)
+	lea		0x60($inp),$inp
+	  vaesenclast	$Z1,$inout2,$inout2
+	 vpaddb		$T2,$Ii,$Z1
+	mov		%r12,0x78+8(%rsp)
+	lea		0x60($out),$out
+	  vmovdqu	0x00-0x80($key),$rndkey
+	  vaesenclast	$Z2,$inout3,$inout3
+	 vpaddb		$T2,$Z1,$Z2
+	  vaesenclast	$Z3, $inout4,$inout4
+	 vpaddb		$T2,$Z2,$Z3
+	  vaesenclast	$Hkey,$inout5,$inout5
+	 vpaddb		$T2,$Z3,$Hkey
+
+	add		\$0x60,$ret
+	sub		\$0x6,$len
+	jc		.L6x_done
+
+	  vmovups	$inout0,-0x60($out)	# save output
+	 vpxor		$rndkey,$T1,$inout0
+	  vmovups	$inout1,-0x50($out)
+	 vmovdqa	$Ii,$inout1		# 0 latency
+	  vmovups	$inout2,-0x40($out)
+	 vmovdqa	$Z1,$inout2		# 0 latency
+	  vmovups	$inout3,-0x30($out)
+	 vmovdqa	$Z2,$inout3		# 0 latency
+	  vmovups	$inout4,-0x20($out)
+	 vmovdqa	$Z3,$inout4		# 0 latency
+	  vmovups	$inout5,-0x10($out)
+	 vmovdqa	$Hkey,$inout5		# 0 latency
+	vmovdqu		0x20+8(%rsp),$Z3	# I[5]
+	jmp		.Loop6x
+
+.L6x_done:
+	vpxor		16+8(%rsp),$Xi,$Xi	# modulo-scheduled
+	vpxor		$Z0,$Xi,$Xi		# modulo-scheduled
+
+	ret
+.size	_aesni_ctr32_ghash_6x,.-_aesni_ctr32_ghash_6x
+___
+######################################################################
+#
+# size_t aesni_gcm_[en|de]crypt(const void *inp, void *out, size_t len,
+#		const AES_KEY *key, unsigned char iv[16],
+#		struct { u128 Xi,H,Htbl[9]; } *Xip);
+$code.=<<___;
+.globl	aesni_gcm_decrypt
+.type	aesni_gcm_decrypt,\@function,6
+.align	32
+aesni_gcm_decrypt:
+	xor	$ret,$ret
+	cmp	\$0x60,$len			# minimal accepted length
+	jb	.Lgcm_dec_abort
+
+	lea	(%rsp),%rax			# save stack pointer
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+___
+$code.=<<___ if ($win64);
+	lea	-0xa8(%rsp),%rsp
+	movaps	%xmm6,-0xd8(%rax)
+	movaps	%xmm7,-0xc8(%rax)
+	movaps	%xmm8,-0xb8(%rax)
+	movaps	%xmm9,-0xa8(%rax)
+	movaps	%xmm10,-0x98(%rax)
+	movaps	%xmm11,-0x88(%rax)
+	movaps	%xmm12,-0x78(%rax)
+	movaps	%xmm13,-0x68(%rax)
+	movaps	%xmm14,-0x58(%rax)
+	movaps	%xmm15,-0x48(%rax)
+.Lgcm_dec_body:
+___
+$code.=<<___;
+	vzeroupper
+
+	vmovdqu		($ivp),$T1		# input counter value
+	sub		\$128,%rsp
+	mov		12($ivp),$counter
+	lea		.Lbswap_mask(%rip),$const
+	vmovdqu		($Xip),$Xi		# load Xi
+	and		\$-64,%rsp		# ensure stack alignment
+	vmovdqu		($const),$Ii		# borrow $Ii for .Lbswap_mask
+	lea		0x80($key),$key		# size optimization
+	lea		0x20+0x20($Xip),$Xip	# size optimization
+	mov		0xf0-0x80($key),$rounds
+	vpshufb		$Ii,$Xi,$Xi
+
+	vmovdqu		0x50($inp),$Z3		# I[5]
+	lea		($inp),$in0
+	vmovdqu		0x40($inp),$Z0
+	lea		-0xc0($inp,$len),$end0
+	vmovdqu		0x30($inp),$Z1
+	shr		\$4,$len
+	xor		$ret,$ret
+	vmovdqu		0x20($inp),$Z2
+	 vpshufb	$Ii,$Z3,$Z3		# passed to _aesni_ctr32_ghash_6x
+	vmovdqu		0x10($inp),$T2
+	 vpshufb	$Ii,$Z0,$Z0
+	vmovdqu		($inp),$Hkey
+	 vpshufb	$Ii,$Z1,$Z1
+	vmovdqu		$Z0,0x30(%rsp)
+	 vpshufb	$Ii,$Z2,$Z2
+	vmovdqu		$Z1,0x40(%rsp)
+	 vpshufb	$Ii,$T2,$T2
+	vmovdqu		$Z2,0x50(%rsp)
+	 vpshufb	$Ii,$Hkey,$Hkey
+	vmovdqu		$T2,0x60(%rsp)
+	vmovdqu		$Hkey,0x70(%rsp)
+
+	call		_aesni_ctr32_ghash_6x
+
+	vmovups		$inout0,-0x60($out)	# save output
+	vmovups		$inout1,-0x50($out)
+	vmovups		$inout2,-0x40($out)
+	vmovups		$inout3,-0x30($out)
+	vmovups		$inout4,-0x20($out)
+	vmovups		$inout5,-0x10($out)
+
+	vpshufb		($const),$Xi,$Xi	# .Lbswap_mask
+	vmovdqu		$Xi,-0x40($Xip)		# output Xi
+
+	vzeroupper
+___
+$code.=<<___ if ($win64);
+	movaps	-0xd8(%rax),%xmm6
+	movaps	-0xd8(%rax),%xmm7
+	movaps	-0xb8(%rax),%xmm8
+	movaps	-0xa8(%rax),%xmm9
+	movaps	-0x98(%rax),%xmm10
+	movaps	-0x88(%rax),%xmm11
+	movaps	-0x78(%rax),%xmm12
+	movaps	-0x68(%rax),%xmm13
+	movaps	-0x58(%rax),%xmm14
+	movaps	-0x48(%rax),%xmm15
+___
+$code.=<<___;
+	mov	-48(%rax),%r15
+	mov	-40(%rax),%r14
+	mov	-32(%rax),%r13
+	mov	-24(%rax),%r12
+	mov	-16(%rax),%rbp
+	mov	-8(%rax),%rbx
+	lea	(%rax),%rsp		# restore %rsp
+.Lgcm_dec_abort:
+	mov	$ret,%rax		# return value
+	ret
+.size	aesni_gcm_decrypt,.-aesni_gcm_decrypt
+___
+
+$code.=<<___;
+.type	_aesni_ctr32_6x,\@abi-omnipotent
+.align	32
+_aesni_ctr32_6x:
+	vmovdqu		0x00-0x80($key),$Z0	# borrow $Z0 for $rndkey
+	vmovdqu		0x20($const),$T2	# borrow $T2, .Lone_msb
+	lea		-1($rounds),%r13
+	vmovups		0x10-0x80($key),$rndkey
+	lea		0x20-0x80($key),%r12
+	vpxor		$Z0,$T1,$inout0
+	add		\$6<<24,$counter
+	jc		.Lhandle_ctr32_2
+	vpaddb		$T2,$T1,$inout1
+	vpaddb		$T2,$inout1,$inout2
+	vpxor		$Z0,$inout1,$inout1
+	vpaddb		$T2,$inout2,$inout3
+	vpxor		$Z0,$inout2,$inout2
+	vpaddb		$T2,$inout3,$inout4
+	vpxor		$Z0,$inout3,$inout3
+	vpaddb		$T2,$inout4,$inout5
+	vpxor		$Z0,$inout4,$inout4
+	vpaddb		$T2,$inout5,$T1
+	vpxor		$Z0,$inout5,$inout5
+	jmp		.Loop_ctr32
+
+.align	16
+.Loop_ctr32:
+	vaesenc		$rndkey,$inout0,$inout0
+	vaesenc		$rndkey,$inout1,$inout1
+	vaesenc		$rndkey,$inout2,$inout2
+	vaesenc		$rndkey,$inout3,$inout3
+	vaesenc		$rndkey,$inout4,$inout4
+	vaesenc		$rndkey,$inout5,$inout5
+	vmovups		(%r12),$rndkey
+	lea		0x10(%r12),%r12
+	dec		%r13d
+	jnz		.Loop_ctr32
+
+	vmovdqu		(%r12),$Hkey		# last round key
+	vaesenc		$rndkey,$inout0,$inout0
+	vpxor		0x00($inp),$Hkey,$Z0
+	vaesenc		$rndkey,$inout1,$inout1
+	vpxor		0x10($inp),$Hkey,$Z1
+	vaesenc		$rndkey,$inout2,$inout2
+	vpxor		0x20($inp),$Hkey,$Z2
+	vaesenc		$rndkey,$inout3,$inout3
+	vpxor		0x30($inp),$Hkey,$Xi
+	vaesenc		$rndkey,$inout4,$inout4
+	vpxor		0x40($inp),$Hkey,$T2
+	vaesenc		$rndkey,$inout5,$inout5
+	vpxor		0x50($inp),$Hkey,$Hkey
+	lea		0x60($inp),$inp
+
+	vaesenclast	$Z0,$inout0,$inout0
+	vaesenclast	$Z1,$inout1,$inout1
+	vaesenclast	$Z2,$inout2,$inout2
+	vaesenclast	$Xi,$inout3,$inout3
+	vaesenclast	$T2,$inout4,$inout4
+	vaesenclast	$Hkey,$inout5,$inout5
+	vmovups		$inout0,0x00($out)
+	vmovups		$inout1,0x10($out)
+	vmovups		$inout2,0x20($out)
+	vmovups		$inout3,0x30($out)
+	vmovups		$inout4,0x40($out)
+	vmovups		$inout5,0x50($out)
+	lea		0x60($out),$out
+
+	ret
+.align	32
+.Lhandle_ctr32_2:
+	vpshufb		$Ii,$T1,$Z2		# byte-swap counter
+	vmovdqu		0x30($const),$Z1	# borrow $Z1, .Ltwo_lsb
+	vpaddd		0x40($const),$Z2,$inout1	# .Lone_lsb
+	vpaddd		$Z1,$Z2,$inout2
+	vpaddd		$Z1,$inout1,$inout3
+	vpshufb		$Ii,$inout1,$inout1
+	vpaddd		$Z1,$inout2,$inout4
+	vpshufb		$Ii,$inout2,$inout2
+	vpxor		$Z0,$inout1,$inout1
+	vpaddd		$Z1,$inout3,$inout5
+	vpshufb		$Ii,$inout3,$inout3
+	vpxor		$Z0,$inout2,$inout2
+	vpaddd		$Z1,$inout4,$T1		# byte-swapped next counter value
+	vpshufb		$Ii,$inout4,$inout4
+	vpxor		$Z0,$inout3,$inout3
+	vpshufb		$Ii,$inout5,$inout5
+	vpxor		$Z0,$inout4,$inout4
+	vpshufb		$Ii,$T1,$T1		# next counter value
+	vpxor		$Z0,$inout5,$inout5
+	jmp	.Loop_ctr32
+.size	_aesni_ctr32_6x,.-_aesni_ctr32_6x
+
+.globl	aesni_gcm_encrypt
+.type	aesni_gcm_encrypt,\@function,6
+.align	32
+aesni_gcm_encrypt:
+	xor	$ret,$ret
+	cmp	\$0x60*3,$len			# minimal accepted length
+	jb	.Lgcm_enc_abort
+
+	lea	(%rsp),%rax			# save stack pointer
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+___
+$code.=<<___ if ($win64);
+	lea	-0xa8(%rsp),%rsp
+	movaps	%xmm6,-0xd8(%rax)
+	movaps	%xmm7,-0xc8(%rax)
+	movaps	%xmm8,-0xb8(%rax)
+	movaps	%xmm9,-0xa8(%rax)
+	movaps	%xmm10,-0x98(%rax)
+	movaps	%xmm11,-0x88(%rax)
+	movaps	%xmm12,-0x78(%rax)
+	movaps	%xmm13,-0x68(%rax)
+	movaps	%xmm14,-0x58(%rax)
+	movaps	%xmm15,-0x48(%rax)
+.Lgcm_enc_body:
+___
+$code.=<<___;
+	vzeroupper
+
+	vmovdqu		($ivp),$T1		# input counter value
+	sub		\$128,%rsp
+	mov		12($ivp),$counter
+	lea		.Lbswap_mask(%rip),$const
+	lea		0x80($key),$key		# size optimization
+	vmovdqu		($const),$Ii		# borrow $Ii for .Lbswap_mask
+	and		\$-64,%rsp		# ensure stack alignment
+	mov		0xf0-0x80($key),$rounds
+
+	lea		($out),$in0
+	lea		-0xc0($out,$len),$end0
+	shr		\$4,$len
+
+	call		_aesni_ctr32_6x
+	vpshufb		$Ii,$inout0,$Xi		# save bswapped output on stack
+	vpshufb		$Ii,$inout1,$T2
+	vmovdqu		$Xi,0x70(%rsp)
+	vpshufb		$Ii,$inout2,$Z0
+	vmovdqu		$T2,0x60(%rsp)
+	vpshufb		$Ii,$inout3,$Z1
+	vmovdqu		$Z0,0x50(%rsp)
+	vpshufb		$Ii,$inout4,$Z2
+	vmovdqu		$Z1,0x40(%rsp)
+	vpshufb		$Ii,$inout5,$Z3		# passed to _aesni_ctr32_ghash_6x
+	vmovdqu		$Z2,0x30(%rsp)
+
+	call		_aesni_ctr32_6x
+
+	vmovdqu		($Xip),$Xi		# load Xi
+	lea		0x20+0x20($Xip),$Xip	# size optimization
+	sub		\$12,$len
+	mov		\$0x60*2,$ret
+	vpshufb		$Ii,$Xi,$Xi
+
+	call		_aesni_ctr32_ghash_6x
+	vmovdqu		0x20(%rsp),$Z3		# I[5]
+	 vmovdqu	($const),$Ii		# borrow $Ii for .Lbswap_mask
+	vmovdqu		0x00-0x20($Xip),$Hkey	# $Hkey^1
+	vpunpckhqdq	$Z3,$Z3,$T1
+	vmovdqu		0x20-0x20($Xip),$rndkey	# borrow $rndkey for $HK
+	 vmovups	$inout0,-0x60($out)	# save output
+	 vpshufb	$Ii,$inout0,$inout0	# but keep bswapped copy
+	vpxor		$Z3,$T1,$T1
+	 vmovups	$inout1,-0x50($out)
+	 vpshufb	$Ii,$inout1,$inout1
+	 vmovups	$inout2,-0x40($out)
+	 vpshufb	$Ii,$inout2,$inout2
+	 vmovups	$inout3,-0x30($out)
+	 vpshufb	$Ii,$inout3,$inout3
+	 vmovups	$inout4,-0x20($out)
+	 vpshufb	$Ii,$inout4,$inout4
+	 vmovups	$inout5,-0x10($out)
+	 vpshufb	$Ii,$inout5,$inout5
+	 vmovdqu	$inout0,0x10(%rsp)	# free $inout0
+___
+{ my ($HK,$T3)=($rndkey,$inout0);
+
+$code.=<<___;
+	 vmovdqu	0x30(%rsp),$Z2		# I[4]
+	 vmovdqu	0x10-0x20($Xip),$Ii	# borrow $Ii for $Hkey^2
+	 vpunpckhqdq	$Z2,$Z2,$T2
+	vpclmulqdq	\$0x00,$Hkey,$Z3,$Z1
+	 vpxor		$Z2,$T2,$T2
+	vpclmulqdq	\$0x11,$Hkey,$Z3,$Z3
+	vpclmulqdq	\$0x00,$HK,$T1,$T1
+
+	 vmovdqu	0x40(%rsp),$T3		# I[3]
+	vpclmulqdq	\$0x00,$Ii,$Z2,$Z0
+	 vmovdqu	0x30-0x20($Xip),$Hkey	# $Hkey^3
+	vpxor		$Z1,$Z0,$Z0
+	 vpunpckhqdq	$T3,$T3,$Z1
+	vpclmulqdq	\$0x11,$Ii,$Z2,$Z2
+	 vpxor		$T3,$Z1,$Z1
+	vpxor		$Z3,$Z2,$Z2
+	vpclmulqdq	\$0x10,$HK,$T2,$T2
+	 vmovdqu	0x50-0x20($Xip),$HK
+	vpxor		$T1,$T2,$T2
+
+	 vmovdqu	0x50(%rsp),$T1		# I[2]
+	vpclmulqdq	\$0x00,$Hkey,$T3,$Z3
+	 vmovdqu	0x40-0x20($Xip),$Ii	# borrow $Ii for $Hkey^4
+	vpxor		$Z0,$Z3,$Z3
+	 vpunpckhqdq	$T1,$T1,$Z0
+	vpclmulqdq	\$0x11,$Hkey,$T3,$T3
+	 vpxor		$T1,$Z0,$Z0
+	vpxor		$Z2,$T3,$T3
+	vpclmulqdq	\$0x00,$HK,$Z1,$Z1
+	vpxor		$T2,$Z1,$Z1
+
+	 vmovdqu	0x60(%rsp),$T2		# I[1]
+	vpclmulqdq	\$0x00,$Ii,$T1,$Z2
+	 vmovdqu	0x60-0x20($Xip),$Hkey	# $Hkey^5
+	vpxor		$Z3,$Z2,$Z2
+	 vpunpckhqdq	$T2,$T2,$Z3
+	vpclmulqdq	\$0x11,$Ii,$T1,$T1
+	 vpxor		$T2,$Z3,$Z3
+	vpxor		$T3,$T1,$T1
+	vpclmulqdq	\$0x10,$HK,$Z0,$Z0
+	 vmovdqu	0x80-0x20($Xip),$HK
+	vpxor		$Z1,$Z0,$Z0
+
+	 vpxor		0x70(%rsp),$Xi,$Xi	# accumulate I[0]
+	vpclmulqdq	\$0x00,$Hkey,$T2,$Z1
+	 vmovdqu	0x70-0x20($Xip),$Ii	# borrow $Ii for $Hkey^6
+	 vpunpckhqdq	$Xi,$Xi,$T3
+	vpxor		$Z2,$Z1,$Z1
+	vpclmulqdq	\$0x11,$Hkey,$T2,$T2
+	 vpxor		$Xi,$T3,$T3
+	vpxor		$T1,$T2,$T2
+	vpclmulqdq	\$0x00,$HK,$Z3,$Z3
+	vpxor		$Z0,$Z3,$Z0
+
+	vpclmulqdq	\$0x00,$Ii,$Xi,$Z2
+	 vmovdqu	0x00-0x20($Xip),$Hkey	# $Hkey^1
+	 vpunpckhqdq	$inout5,$inout5,$T1
+	vpclmulqdq	\$0x11,$Ii,$Xi,$Xi
+	 vpxor		$inout5,$T1,$T1
+	vpxor		$Z1,$Z2,$Z1
+	vpclmulqdq	\$0x10,$HK,$T3,$T3
+	 vmovdqu	0x20-0x20($Xip),$HK
+	vpxor		$T2,$Xi,$Z3
+	vpxor		$Z0,$T3,$Z2
+
+	 vmovdqu	0x10-0x20($Xip),$Ii	# borrow $Ii for $Hkey^2
+	  vpxor		$Z1,$Z3,$T3		# aggregated Karatsuba post-processing
+	vpclmulqdq	\$0x00,$Hkey,$inout5,$Z0
+	  vpxor		$T3,$Z2,$Z2
+	 vpunpckhqdq	$inout4,$inout4,$T2
+	vpclmulqdq	\$0x11,$Hkey,$inout5,$inout5
+	 vpxor		$inout4,$T2,$T2
+	  vpslldq	\$8,$Z2,$T3
+	vpclmulqdq	\$0x00,$HK,$T1,$T1
+	  vpxor		$T3,$Z1,$Xi
+	  vpsrldq	\$8,$Z2,$Z2
+	  vpxor		$Z2,$Z3,$Z3
+
+	vpclmulqdq	\$0x00,$Ii,$inout4,$Z1
+	 vmovdqu	0x30-0x20($Xip),$Hkey	# $Hkey^3
+	vpxor		$Z0,$Z1,$Z1
+	 vpunpckhqdq	$inout3,$inout3,$T3
+	vpclmulqdq	\$0x11,$Ii,$inout4,$inout4
+	 vpxor		$inout3,$T3,$T3
+	vpxor		$inout5,$inout4,$inout4
+	  vpalignr	\$8,$Xi,$Xi,$inout5	# 1st phase
+	vpclmulqdq	\$0x10,$HK,$T2,$T2
+	 vmovdqu	0x50-0x20($Xip),$HK
+	vpxor		$T1,$T2,$T2
+
+	vpclmulqdq	\$0x00,$Hkey,$inout3,$Z0
+	 vmovdqu	0x40-0x20($Xip),$Ii	# borrow $Ii for $Hkey^4
+	vpxor		$Z1,$Z0,$Z0
+	 vpunpckhqdq	$inout2,$inout2,$T1
+	vpclmulqdq	\$0x11,$Hkey,$inout3,$inout3
+	 vpxor		$inout2,$T1,$T1
+	vpxor		$inout4,$inout3,$inout3
+	  vxorps	0x10(%rsp),$Z3,$Z3	# accumulate $inout0
+	vpclmulqdq	\$0x00,$HK,$T3,$T3
+	vpxor		$T2,$T3,$T3
+
+	  vpclmulqdq	\$0x10,0x10($const),$Xi,$Xi
+	  vxorps	$inout5,$Xi,$Xi
+
+	vpclmulqdq	\$0x00,$Ii,$inout2,$Z1
+	 vmovdqu	0x60-0x20($Xip),$Hkey	# $Hkey^5
+	vpxor		$Z0,$Z1,$Z1
+	 vpunpckhqdq	$inout1,$inout1,$T2
+	vpclmulqdq	\$0x11,$Ii,$inout2,$inout2
+	 vpxor		$inout1,$T2,$T2
+	  vpalignr	\$8,$Xi,$Xi,$inout5	# 2nd phase
+	vpxor		$inout3,$inout2,$inout2
+	vpclmulqdq	\$0x10,$HK,$T1,$T1
+	 vmovdqu	0x80-0x20($Xip),$HK
+	vpxor		$T3,$T1,$T1
+
+	  vxorps	$Z3,$inout5,$inout5
+	  vpclmulqdq	\$0x10,0x10($const),$Xi,$Xi
+	  vxorps	$inout5,$Xi,$Xi
+
+	vpclmulqdq	\$0x00,$Hkey,$inout1,$Z0
+	 vmovdqu	0x70-0x20($Xip),$Ii	# borrow $Ii for $Hkey^6
+	vpxor		$Z1,$Z0,$Z0
+	 vpunpckhqdq	$Xi,$Xi,$T3
+	vpclmulqdq	\$0x11,$Hkey,$inout1,$inout1
+	 vpxor		$Xi,$T3,$T3
+	vpxor		$inout2,$inout1,$inout1
+	vpclmulqdq	\$0x00,$HK,$T2,$T2
+	vpxor		$T1,$T2,$T2
+
+	vpclmulqdq	\$0x00,$Ii,$Xi,$Z1
+	vpclmulqdq	\$0x11,$Ii,$Xi,$Z3
+	vpxor		$Z0,$Z1,$Z1
+	vpclmulqdq	\$0x10,$HK,$T3,$Z2
+	vpxor		$inout1,$Z3,$Z3
+	vpxor		$T2,$Z2,$Z2
+
+	vpxor		$Z1,$Z3,$Z0		# aggregated Karatsuba post-processing
+	vpxor		$Z0,$Z2,$Z2
+	vpslldq		\$8,$Z2,$T1
+	vmovdqu		0x10($const),$Hkey	# .Lpoly
+	vpsrldq		\$8,$Z2,$Z2
+	vpxor		$T1,$Z1,$Xi
+	vpxor		$Z2,$Z3,$Z3
+
+	vpalignr	\$8,$Xi,$Xi,$T2		# 1st phase
+	vpclmulqdq	\$0x10,$Hkey,$Xi,$Xi
+	vpxor		$T2,$Xi,$Xi
+
+	vpalignr	\$8,$Xi,$Xi,$T2		# 2nd phase
+	vpclmulqdq	\$0x10,$Hkey,$Xi,$Xi
+	vpxor		$Z3,$T2,$T2
+	vpxor		$T2,$Xi,$Xi
+___
+}
+$code.=<<___;
+	vpshufb		($const),$Xi,$Xi	# .Lbswap_mask
+	vmovdqu		$Xi,-0x40($Xip)		# output Xi
+
+	vzeroupper
+___
+$code.=<<___ if ($win64);
+	movaps	-0xd8(%rax),%xmm6
+	movaps	-0xc8(%rax),%xmm7
+	movaps	-0xb8(%rax),%xmm8
+	movaps	-0xa8(%rax),%xmm9
+	movaps	-0x98(%rax),%xmm10
+	movaps	-0x88(%rax),%xmm11
+	movaps	-0x78(%rax),%xmm12
+	movaps	-0x68(%rax),%xmm13
+	movaps	-0x58(%rax),%xmm14
+	movaps	-0x48(%rax),%xmm15
+___
+$code.=<<___;
+	mov	-48(%rax),%r15
+	mov	-40(%rax),%r14
+	mov	-32(%rax),%r13
+	mov	-24(%rax),%r12
+	mov	-16(%rax),%rbp
+	mov	-8(%rax),%rbx
+	lea	(%rax),%rsp		# restore %rsp
+.Lgcm_enc_abort:
+	mov	$ret,%rax		# return value
+	ret
+.size	aesni_gcm_encrypt,.-aesni_gcm_encrypt
+___
+
+$code.=<<___;
+.align	64
+.Lbswap_mask:
+	.byte	15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
+.Lpoly:
+	.byte	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
+.Lone_msb:
+	.byte	0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,1
+.Ltwo_lsb:
+	.byte	2,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+.Lone_lsb:
+	.byte	1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+.asciz	"AES-NI GCM module for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align	64
+___
+if ($win64) {
+$rec="%rcx";
+$frame="%rdx";
+$context="%r8";
+$disp="%r9";
+
+$code.=<<___
+.extern	__imp_RtlVirtualUnwind
+.type	gcm_se_handler,\@abi-omnipotent
+.align	16
+gcm_se_handler:
+	push	%rsi
+	push	%rdi
+	push	%rbx
+	push	%rbp
+	push	%r12
+	push	%r13
+	push	%r14
+	push	%r15
+	pushfq
+	sub	\$64,%rsp
+
+	mov	120($context),%rax	# pull context->Rax
+	mov	248($context),%rbx	# pull context->Rip
+
+	mov	8($disp),%rsi		# disp->ImageBase
+	mov	56($disp),%r11		# disp->HandlerData
+
+	mov	0(%r11),%r10d		# HandlerData[0]
+	lea	(%rsi,%r10),%r10	# prologue label
+	cmp	%r10,%rbx		# context->Rip<prologue label
+	jb	.Lcommon_seh_tail
+
+	mov	152($context),%rax	# pull context->Rsp
+
+	mov	4(%r11),%r10d		# HandlerData[1]
+	lea	(%rsi,%r10),%r10	# epilogue label
+	cmp	%r10,%rbx		# context->Rip>=epilogue label
+	jae	.Lcommon_seh_tail
+
+	mov	120($context),%rax	# pull context->Rax
+
+	mov	-48(%rax),%r15
+	mov	-40(%rax),%r14
+	mov	-32(%rax),%r13
+	mov	-24(%rax),%r12
+	mov	-16(%rax),%rbp
+	mov	-8(%rax),%rbx
+	mov	%r15,240($context)
+	mov	%r14,232($context)
+	mov	%r13,224($context)
+	mov	%r12,216($context)
+	mov	%rbp,160($context)
+	mov	%rbx,144($context)
+
+	lea	-0xd8(%rax),%rsi	# %xmm save area
+	lea	512($context),%rdi	# & context.Xmm6
+	mov	\$20,%ecx		# 10*sizeof(%xmm0)/sizeof(%rax)
+	.long	0xa548f3fc		# cld; rep movsq
+
+.Lcommon_seh_tail:
+	mov	8(%rax),%rdi
+	mov	16(%rax),%rsi
+	mov	%rax,152($context)	# restore context->Rsp
+	mov	%rsi,168($context)	# restore context->Rsi
+	mov	%rdi,176($context)	# restore context->Rdi
+
+	mov	40($disp),%rdi		# disp->ContextRecord
+	mov	$context,%rsi		# context
+	mov	\$154,%ecx		# sizeof(CONTEXT)
+	.long	0xa548f3fc		# cld; rep movsq
+
+	mov	$disp,%rsi
+	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
+	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
+	mov	0(%rsi),%r8		# arg3, disp->ControlPc
+	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
+	mov	40(%rsi),%r10		# disp->ContextRecord
+	lea	56(%rsi),%r11		# &disp->HandlerData
+	lea	24(%rsi),%r12		# &disp->EstablisherFrame
+	mov	%r10,32(%rsp)		# arg5
+	mov	%r11,40(%rsp)		# arg6
+	mov	%r12,48(%rsp)		# arg7
+	mov	%rcx,56(%rsp)		# arg8, (NULL)
+	call	*__imp_RtlVirtualUnwind(%rip)
+
+	mov	\$1,%eax		# ExceptionContinueSearch
+	add	\$64,%rsp
+	popfq
+	pop	%r15
+	pop	%r14
+	pop	%r13
+	pop	%r12
+	pop	%rbp
+	pop	%rbx
+	pop	%rdi
+	pop	%rsi
+	ret
+.size	gcm_se_handler,.-gcm_se_handler
+
+.section	.pdata
+.align	4
+	.rva	.LSEH_begin_aesni_gcm_decrypt
+	.rva	.LSEH_end_aesni_gcm_decrypt
+	.rva	.LSEH_gcm_dec_info
+
+	.rva	.LSEH_begin_aesni_gcm_encrypt
+	.rva	.LSEH_end_aesni_gcm_encrypt
+	.rva	.LSEH_gcm_enc_info
+.section	.xdata
+.align	8
+.LSEH_gcm_dec_info:
+	.byte	9,0,0,0
+	.rva	gcm_se_handler
+	.rva	.Lgcm_dec_body,.Lgcm_dec_abort
+.LSEH_gcm_enc_info:
+	.byte	9,0,0,0
+	.rva	gcm_se_handler
+	.rva	.Lgcm_enc_body,.Lgcm_enc_abort
+___
+}
+}}} else {{{
+$code=<<___;	# assembler is too old
+.text
+
+.globl	aesni_gcm_encrypt
+.type	aesni_gcm_encrypt,\@abi-omnipotent
+aesni_gcm_encrypt:
+	xor	%eax,%eax
+	ret
+.size	aesni_gcm_encrypt,.-aesni_gcm_encrypt
+
+.globl	aesni_gcm_decrypt
+.type	aesni_gcm_decrypt,\@abi-omnipotent
+aesni_gcm_decrypt:
+	xor	%eax,%eax
+	ret
+.size	aesni_gcm_decrypt,.-aesni_gcm_decrypt
+___
+}}}
+
+$code =~ s/\`([^\`]*)\`/eval($1)/gem;
+
+print $code;
+
+close STDOUT;