sha/asm/keccak1600-avx2.pl: optimized remodelled version.

New register usage pattern allows to achieve sligtly better
performance. Not as much as I hoped for. Performance is believed
to be limited by irreconcilable write-back conflicts, rather than
lack of computational resources or data dependencies.
Reviewed-by: NRich Salz <rsalz@openssl.org>

crypto/sha/asm/keccak1600-avx2.pl

@@ -111,10 +111,16 @@ my @A_jagged = ([0,0], [1,0], [1,1], [1,2], [1,3],	# [0][0..4]
 #
 #			r=1088(*)
 #
-# Haswell		9.5
-# Skylake		8.8
+# Haswell		8.7/+10%
+# Skylake		7.8/+20%
+# Ryzen			17(**)
 #
-# (*)	Corresponds to SHA3-256.
+# (*)	Corresponds to SHA3-256. Percentage after slash is improvement
+#	coefficient in comparison to scalar keccak1600-x86_64.pl.
+# (**)	It's expected that Ryzen performs poorly, because instruction
+#	issue rate is limited to two AVX2 instructions per cycle and
+#	in addition vpblendd is reportedly bound to specific port.
+#	Obviously this code path should not be executed on Ryzen.
 
 my @T = map("%ymm$_",(7..15));
 my ($C14,$C00,$D00,$D14) = @T[5..8];
@@ -134,137 +140,133 @@ __KeccakF1600:
 .align	32
 .Loop_avx2:
 	######################################### Theta
-	vpxor		$A01,$A31,$C14
-	vpxor		$A21,$C14,$C14
-	vpxor		$A41,$C14,$C14
-	vpxor		$A11,$C14,$C14		# C[1..4]
-	vpermq		\$0b10110001,$A20,$C00
+	vpshufd		\$0b01001110,$A20,$C00
+	vpxor		$A31,$A41,$C14
+	vpxor		$A11,$A21,@T[2]
+	vpxor		$A01,$C14,$C14
+	vpxor		@T[2],$C14,$C14		# C[1..4]
+
+	vpermq		\$0b10010011,$C14,@T[4]
 	vpxor		$A20,$C00,$C00
 	vpermq		\$0b01001110,$C00,@T[0]
-	vpxor		$A00,$C00,$C00
-	vpxor		@T[0],$C00,$C00		# C[0..0]
 
 	vpsrlq		\$63,$C14,@T[1]
-	vpaddq		$C14,$C14,@T[3]
-	vpor		@T[3],@T[1],@T[1]	# ROL64(C[1..4],1)
+	vpaddq		$C14,$C14,@T[2]
+	vpor		@T[2],@T[1],@T[1]	# ROL64(C[1..4],1)
 
-	vpsrlq		\$63,$C00,@T[0]
-	vpaddq		$C00,$C00,@T[2]
-	vpor		@T[2],@T[0],@T[0]	# ROL64(C[0..0],1)
+	vpermq		\$0b00111001,@T[1],$D14
+	vpxor		@T[4],@T[1],$D00
+	vpermq		\$0b00000000,$D00,$D00	# D[0..0] = ROL64(C[1],1) ^ C[4]
 
-	vpermq		\$0b00000000,@T[1],$D00	
-	vpermq		\$0b11111111,$C14,@T[3]
-	vpxor		@T[3],$D00,$D00		# D[0..0] = ROL64(C[1],1) ^ C[4]
+	vpxor		$A00,$C00,$C00
+	vpxor		@T[0],$C00,$C00		# C[0..0]
 
-	vpermq		\$0b00111001,@T[1],$D14
-	vpblendd	\$0b11000000,@T[0],$D14,$D14
-	vpermq		\$0b10010011,$C14,@T[2]
-	vpblendd	\$0b00000011,$C00,@T[2],@T[2]
-	vpxor		@T[2],$D14,$D14		# D[1..4] = ROL64(C[2..4,0),1) ^ C[0..3]
+	vpsrlq		\$63,$C00,@T[0]
+	vpaddq		$C00,$C00,@T[1]
+	vpor		@T[0],@T[1],@T[1]	# ROL64(C[0..0],1)
 
-	vpxor		$D00,$A00,$A00		# ^= D[0..0]
 	vpxor		$D00,$A20,$A20		# ^= D[0..0]
-	vpxor		$D14,$A01,$A01		# ^= D[1..4]
-	vpxor		$D14,$A31,$A31		# ^= D[1..4]
-	vpxor		$D14,$A21,$A21		# ^= D[1..4]
-	vpxor		$D14,$A41,$A41		# ^= D[1..4]
-	vpxor		$D14,$A11,$A11		# ^= D[1..4]
-
-	######################################### Rho
-	vpsllvq		0*32-96(%r8),$A20,@T[0]
-	vpsrlvq		0*32-96(%r9),$A20,$A20
-	vpor		@T[0],$A20,$A20
+	vpxor		$D00,$A00,$A00		# ^= D[0..0]
+
+	vpblendd	\$0b11000000,@T[1],$D14,$D14
+	vpblendd	\$0b00000011,$C00,@T[4],@T[4]
+	vpxor		@T[4],$D14,$D14		# D[1..4] = ROL64(C[2..4,0),1) ^ C[0..3]
 
-	vpsllvq		1*32-96(%r8),$A01,@T[1]
-	vpsrlvq		1*32-96(%r9),$A01,$A01
-	vpor		@T[1],$A01,$A01
+	######################################### Rho + Pi + pre-Chi shuffle
+	vpsllvq		0*32-96(%r8),$A20,@T[3]
+	vpsrlvq		0*32-96(%r9),$A20,$A20
+	vpor		@T[3],$A20,$A20
 
-	vpsllvq		2*32-96(%r8),$A31,@T[2]
+	 vpxor		$D14,$A31,$A31		# ^= D[1..4] from Theta
+	vpsllvq		2*32-96(%r8),$A31,@T[4]
 	vpsrlvq		2*32-96(%r9),$A31,$A31
-	vpor		@T[2],$A31,$A31
+	vpor		@T[4],$A31,$A31
 
-	vpsllvq		3*32-96(%r8),$A21,@T[3]
+	 vpxor		$D14,$A21,$A21		# ^= D[1..4] from Theta
+	vpsllvq		3*32-96(%r8),$A21,@T[5]
 	vpsrlvq		3*32-96(%r9),$A21,$A21
-	vpor		@T[3],$A21,$A21
+	vpor		@T[5],$A21,$A21
 
-	vpsllvq		4*32-96(%r8),$A41,@T[4]
+	 vpxor		$D14,$A41,$A41		# ^= D[1..4] from Theta
+	vpsllvq		4*32-96(%r8),$A41,@T[6]
 	vpsrlvq		4*32-96(%r9),$A41,$A41
-	vpor		@T[4],$A41,$A41
+	vpor		@T[6],$A41,$A41
+
+	 vpxor		$D14,$A11,$A11		# ^= D[1..4] from Theta
+	 vpermq		\$0b10001101,$A20,@T[3]	# $A20 -> future $A31
+	 vpermq		\$0b10001101,$A31,@T[4]	# $A31 -> future $A21
+	vpsllvq		5*32-96(%r8),$A11,@T[7]
+	vpsrlvq		5*32-96(%r9),$A11,@T[1]
+	vpor		@T[7],@T[1],@T[1]	# $A11 -> future $A01
+
+	 vpxor		$D14,$A01,$A01		# ^= D[1..4] from Theta
+	 vpermq		\$0b00011011,$A21,@T[5]	# $A21 -> future $A41
+	 vpermq		\$0b01110010,$A41,@T[6]	# $A41 -> future $A11
+	vpsllvq		1*32-96(%r8),$A01,@T[8]
+	vpsrlvq		1*32-96(%r9),$A01,@T[2]
+	vpor		@T[8],@T[2],@T[2]	# $A01 -> future $A20
 
-	vpsllvq		5*32-96(%r8),$A11,@T[5]
-	vpsrlvq		5*32-96(%r9),$A11,$A11
-	vpor		@T[5],$A11,$A11
+	######################################### Chi
+	vpsrldq		\$8,@T[1],@T[7]
+	vpandn		@T[7],@T[1],@T[0]	# tgting  [0][0] [0][0] [0][0] [0][0]
 
-	######################################### Pi + pre-Chi shuffle
-	vpermq		\$0b01110010,$A41,@T[6]	# vpermq \$0b00011011,$A41,$A11
-	vpermq		\$0b00011011,$A21,@T[5]	# vpermq \$0b01110010,$A21,$A41
-	vpermq		\$0b10001101,$A31,@T[4]	# vpermq \$0b10001101,$A31,$A21
-	vpermq		\$0b10001101,$A20,@T[3]	# vpermq \$0b01110010,$A20,$A31
-	vmovdqa		$A01,@T[2]
-	vmovdqa		$A11,@T[1]
+	vpblendd	\$0b00001100,@T[6],@T[2],$A31	#               [4][4] [2][0]
+	vpblendd	\$0b00001100,@T[2],@T[4],@T[8]	#               [4][0] [2][1]
+	 vpblendd	\$0b00001100,@T[4],@T[3],$A41	#               [4][2] [2][4]
+	 vpblendd	\$0b00001100,@T[3],@T[2],@T[7]	#               [4][3] [2][0]
+	vpblendd	\$0b00110000,@T[4],$A31,$A31	#        [1][3] [4][4] [2][0]
+	vpblendd	\$0b00110000,@T[5],@T[8],@T[8]	#        [1][4] [4][0] [2][1]
+	 vpblendd	\$0b00110000,@T[2],$A41,$A41	#        [1][0] [4][2] [2][4]
+	 vpblendd	\$0b00110000,@T[6],@T[7],@T[7]	#        [1][1] [4][3] [2][0]
+	vpblendd	\$0b11000000,@T[5],$A31,$A31	# [3][2] [1][3] [4][4] [2][0]
+	vpblendd	\$0b11000000,@T[6],@T[8],@T[8]	# [3][3] [1][4] [4][0] [2][1]
+	 vpblendd	\$0b11000000,@T[6],$A41,$A41	# [3][3] [1][0] [4][2] [2][4]
+	 vpblendd	\$0b11000000,@T[4],@T[7],@T[7]	# [3][4] [1][1] [4][3] [2][0]
+	vpandn		@T[8],$A31,$A31		# tgting  [3][1] [1][2] [4][3] [2][4]
+	 vpandn		@T[7],$A41,$A41		# tgting  [3][2] [1][4] [4][1] [2][3]
 
-	######################################### Chi
-	vpermq		\$0b00000000,@T[1],@T[0]	# [0][1] [0][1] [0][1] [0][1]
-	vpermq		\$0b01010101,@T[1],@T[7]	# [0][2] [0][2] [0][2] [0][2]
-	vpandn		@T[7],@T[0],@T[0]	# tgting  [0][0] [0][0] [0][0] [0][0]
+	vpblendd	\$0b00001100,@T[2],@T[5],$A11	#               [4][0] [2][3]
+	vpblendd	\$0b00001100,@T[5],@T[3],@T[8]	#               [4][1] [2][4]
+	 vpxor		@T[3],$A31,$A31
+	vpblendd	\$0b00110000,@T[3],$A11,$A11	#        [1][2] [4][0] [2][3]
+	vpblendd	\$0b00110000,@T[4],@T[8],@T[8]	#        [1][3] [4][1] [2][4]
+	 vpxor		@T[5],$A41,$A41
+	vpblendd	\$0b11000000,@T[4],$A11,$A11	# [3][4] [1][2] [4][0] [2][3]
+	vpblendd	\$0b11000000,@T[2],@T[8],@T[8]	# [3][0] [1][3] [4][1] [2][4]
+	vpandn		@T[8],$A11,$A11		# tgting  [3][3] [1][1] [4][4] [2][2]
+	vpxor		@T[6],$A11,$A11
 
-	vpermq		\$0b00111001,@T[1],$A01		# [0][1] [0][4] [0][3] [0][2]
-	vpermq		\$0b00011110,@T[1],@T[8]	# [0][1] [0][2] [0][4] [0][3]
-	vpblendd	\$0b11000000,$A00,$A01,$A01	# [0][0] [0][4] [0][3] [0][2]
-	vpblendd	\$0b00110000,$A00,@T[8],@T[8]	# [0][1] [0][0] [0][4] [0][3]
-	vpandn		@T[8],$A01,$A01		# tgting  [0][4] [0][3] [0][2] [0][1]
+	  vpermq	\$0b00011110,@T[1],$A21		# [0][1] [0][2] [0][4] [0][3]
+	  vpblendd	\$0b00110000,$A00,$A21,@T[8]	# [0][1] [0][0] [0][4] [0][3]
+	  vpermq	\$0b00111001,@T[1],$A01		# [0][1] [0][4] [0][3] [0][2]
+	  vpblendd	\$0b11000000,$A00,$A01,$A01	# [0][0] [0][4] [0][3] [0][2]
+	  vpandn	@T[8],$A01,$A01		# tgting  [0][4] [0][3] [0][2] [0][1]
 
 	vpblendd	\$0b00001100,@T[5],@T[4],$A20	#               [4][1] [2][1]
-	vpblendd	\$0b00110000,@T[6],$A20,$A20	#        [1][1] [4][1] [2][1]
-	vpblendd	\$0b11000000,@T[3],$A20,$A20	# [3][1] [1][1] [4][1] [2][1]
 	vpblendd	\$0b00001100,@T[4],@T[6],@T[7]	#               [4][2] [2][2]
+	vpblendd	\$0b00110000,@T[6],$A20,$A20	#        [1][1] [4][1] [2][1]
 	vpblendd	\$0b00110000,@T[3],@T[7],@T[7]	#        [1][2] [4][2] [2][2]
+	vpblendd	\$0b11000000,@T[3],$A20,$A20	# [3][1] [1][1] [4][1] [2][1]
 	vpblendd	\$0b11000000,@T[5],@T[7],@T[7]	# [3][2] [1][2] [4][2] [2][2]
 	vpandn		@T[7],$A20,$A20		# tgting  [3][0] [1][0] [4][0] [2][0]
+	vpxor		@T[2],$A20,$A20
 
-	vpblendd	\$0b00001100,@T[6],@T[2],$A31	#               [4][4] [2][0]
-	vpblendd	\$0b00110000,@T[4],$A31,$A31	#        [1][3] [4][4] [2][0]
-	vpblendd	\$0b11000000,@T[5],$A31,$A31	# [3][2] [1][3] [4][4] [2][0]
-	vpblendd	\$0b00001100,@T[2],@T[4],@T[8]	#               [4][0] [2][1]
-	vpblendd	\$0b00110000,@T[5],@T[8],@T[8]	#        [1][4] [4][0] [2][1]
-	vpblendd	\$0b11000000,@T[6],@T[8],@T[8]	# [3][3] [1][4] [4][0] [2][1]
-	vpandn		@T[8],$A31,$A31		# tgting  [3][1] [1][2] [4][3] [2][4]
+	 vpermq		\$0b00000000,@T[0],@T[0]	# [0][0] [0][0] [0][0] [0][0]
+	 vpermq		\$0b00011011,$A31,$A31	# post-Chi shuffle
+	 vpermq		\$0b10001101,$A41,$A41
+	 vpermq		\$0b01110010,$A11,$A11
 
 	vpblendd	\$0b00001100,@T[3],@T[6],$A21	#               [4][3] [2][2]
-	vpblendd	\$0b00110000,@T[5],$A21,$A21	#        [1][4] [4][3] [2][2]
-	vpblendd	\$0b11000000,@T[2],$A21,$A21	# [3][0] [1][4] [4][3] [2][2]
 	vpblendd	\$0b00001100,@T[6],@T[5],@T[7]	#               [4][4] [2][3]
+	vpblendd	\$0b00110000,@T[5],$A21,$A21	#        [1][4] [4][3] [2][2]
 	vpblendd	\$0b00110000,@T[2],@T[7],@T[7]	#        [1][0] [4][4] [2][3]
+	vpblendd	\$0b11000000,@T[2],$A21,$A21	# [3][0] [1][4] [4][3] [2][2]
 	vpblendd	\$0b11000000,@T[3],@T[7],@T[7]	# [3][1] [1][0] [4][4] [2][3]
 	vpandn		@T[7],$A21,$A21		# tgting  [3][4] [1][3] [4][2] [2][1]
 
-	vpblendd	\$0b00001100,@T[4],@T[3],$A41	#               [4][2] [2][4]
-	vpblendd	\$0b00110000,@T[2],$A41,$A41	#        [1][0] [4][2] [2][4]
-	vpblendd	\$0b11000000,@T[6],$A41,$A41	# [3][3] [1][0] [4][2] [2][4]
-	vpblendd	\$0b00001100,@T[3],@T[2],@T[8]	#               [4][3] [2][0]
-	vpblendd	\$0b00110000,@T[6],@T[8],@T[8]	#        [1][1] [4][3] [2][0]
-	vpblendd	\$0b11000000,@T[4],@T[8],@T[8]	# [3][4] [1][1] [4][3] [2][0]
-	vpandn		@T[8],$A41,$A41		# tgting  [3][2] [1][4] [4][1] [2][3]
-
-	vpblendd	\$0b00001100,@T[2],@T[5],$A11	#               [4][0] [2][3]
-	vpblendd	\$0b00110000,@T[3],$A11,$A11	#        [1][2] [4][0] [2][3]
-	vpblendd	\$0b11000000,@T[4],$A11,$A11	# [3][4] [1][2] [4][0] [2][3]
-	vpblendd	\$0b00001100,@T[5],@T[3],@T[7]	#               [4][1] [2][4]
-	vpblendd	\$0b00110000,@T[4],@T[7],@T[7]	#        [1][3] [4][1] [2][4]
-	vpblendd	\$0b11000000,@T[2],@T[7],@T[7]	# [3][0] [1][3] [4][1] [2][4]
-	vpandn		@T[7],$A11,$A11		# tgting  [3][3] [1][1] [4][4] [2][2]
-
 	vpxor		@T[0],$A00,$A00
 	vpxor		@T[1],$A01,$A01
-	vpxor		@T[2],$A20,$A20
-	vpxor		@T[3],$A31,$A31
 	vpxor		@T[4],$A21,$A21
-	vpxor		@T[5],$A41,$A41
-	vpxor		@T[6],$A11,$A11
-
-	vpermq		\$0b00011011,$A31,$A31	# post-Chi shuffle
-	vpermq		\$0b10001101,$A41,$A41
-	vpermq		\$0b01110010,$A11,$A11
 
 	######################################### Iota
 	vpxor		(%r10),$A00,$A00