perlasm/sparcv9_modes.pl: "cooperative" optimizations based on suggestions

from David Miller.

1 files changed, 309 insertions, 11 deletions

diff --git a/crypto/perlasm/sparcv9_modes.pl b/crypto/perlasm/sparcv9_modes.pl
index e6b8a24cc2..d372586bbc 100644
--- a/crypto/perlasm/sparcv9_modes.pl
+++ b/crypto/perlasm/sparcv9_modes.pl
@@ -1,9 +1,28 @@
 #!/usr/bin/env perl
 
-# Specific mode implementations for SPARC T4 modules.
+# Specific modes implementations for SPARC Architecture 2011. There
+# is T4 dependency though, an ASI value that is not specified in the
+# Architecture Manual. But as SPARC universe is rather monocultural,
+# we imply that processor capable of executing crypto instructions
+# can handle the ASI in question as well. This means that we ought to
+# keep eyes open when new processors emerge...
+#
+# As for above mentioned ASI. It's so called "block initializing
+# store" which cancels "read" in "read-update-write" on cache lines.
+# This is "cooperative" optimization, as it reduces overall pressure
+# on memory interface. Benefits can't be observed/quantified with
+# usual benchmarks, on the contrary you can notice that single-thread
+# performance for parallelizable modes is ~1.5% worse. Special note
+# about commented 'membar' instructions, otherwise recommended by
+# manual. Rationale is following. Memory view is consistent from
+# viewpoint of processor executing the code even when ASI in question
+# is used. If thread on another processor has to access the result,
+# its availability would have to be mediated and it can be done only
+# through a syncronization operation which would requre ... 'membar'.
+# All this based on suggestions from David Miller.
 
 my ($inp,$out,$len,$key,$ivec,$enc)=map("%i$_",(0..5));
-my ($ileft,$iright,$ooff,$omask,$ivoff)=map("%l$_",(1..7));
+my ($ileft,$iright,$ooff,$omask,$ivoff,$blk_init)=map("%l$_",(0..7));
 
 sub alg_cbc_encrypt_implement {
 my ($alg,$bits) = @_;
@@ -13,6 +32,7 @@ $::code.=<<___;
 .align	32
 ${alg}${bits}_t4_cbc_encrypt:
 	save		%sp, -$::frame, %sp
+	sub		$inp, $out, $blk_init	! $inp!=$out
 ___
 $::code.=<<___ if (!$::evp);
 	andcc		$ivec, 7, $ivoff
@@ -33,8 +53,9 @@ $::code.=<<___ if ($::evp);
 	ld		[$ivec + 12], %f3
 ___
 $::code.=<<___;
+	prefetch	[$inp], 20
+	prefetch	[$inp + 63], 20
 	call		_${alg}${bits}_load_enckey
-	srlx		$len, 4, $len
 	and		$inp, 7, $ileft
 	andn		$inp, 7, $inp
 	sll		$ileft, 3, $ileft
@@ -42,9 +63,16 @@ $::code.=<<___;
 	mov		0xff, $omask
 	sub		$iright, $ileft, $iright
 	and		$out, 7, $ooff
-	alignaddrl	$out, %g0, $out
+	cmp		$len, 127
+	movrnz		$ooff, 0, $blk_init		! if (	$out&7 ||
+	movleu		$::size_t_cc, 0, $blk_init	!	$len<128 ||
+	brnz,pn		$blk_init, .L${bits}cbc_enc_blk	!	$inp==$out)
 	srl		$omask, $ooff, $omask
 
+	alignaddrl	$out, %g0, $out
+	srlx		$len, 4, $len
+	prefetch	[$out], 22
+
 .L${bits}_cbc_enc_loop:
 	ldx		[$inp + 0], %o0
 	brz,pt		$ileft, 4f
@@ -65,6 +93,8 @@ $::code.=<<___;
 
 	fxor		%f12, %f0, %f0		! ^= ivec
 	fxor		%f14, %f2, %f2
+	prefetch	[$out + 63], 22
+	prefetch	[$inp + 16+63], 20
 	call		_${alg}${bits}_encrypt_1x
 	add		$inp, 16, $inp
 
@@ -141,6 +171,68 @@ ___
 $::code.=<<___;
 	ret
 	restore
+
+!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+.align	32
+.L${bits}cbc_enc_blk:
+	add	$out, $len, $blk_init
+	and	$blk_init, 63, $blk_init	! tail
+	sub	$len, $blk_init, $len
+	add	$blk_init, 15, $blk_init	! round up to 16n
+	srlx	$len, 4, $len
+	srl	$blk_init, 4, $blk_init
+
+.L${bits}_cbc_enc_blk_loop:
+	ldx		[$inp + 0], %o0
+	brz,pt		$ileft, 5f
+	ldx		[$inp + 8], %o1
+
+	ldx		[$inp + 16], %o2
+	sllx		%o0, $ileft, %o0
+	srlx		%o1, $iright, %g1
+	sllx		%o1, $ileft, %o1
+	or		%g1, %o0, %o0
+	srlx		%o2, $iright, %o2
+	or		%o2, %o1, %o1
+5:
+	xor		%g4, %o0, %o0		! ^= rk[0]
+	xor		%g5, %o1, %o1
+	movxtod		%o0, %f12
+	movxtod		%o1, %f14
+
+	fxor		%f12, %f0, %f0		! ^= ivec
+	fxor		%f14, %f2, %f2
+	prefetch	[$inp + 16+63], 20
+	call		_${alg}${bits}_encrypt_1x
+	add		$inp, 16, $inp
+	sub		$len, 1, $len
+		
+	stda		%f0, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f2, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	brnz,pt		$len, .L${bits}_cbc_enc_blk_loop
+	add		$out, 8, $out
+
+	!membar		0x0f
+	brnz,pt		$blk_init, .L${bits}_cbc_enc_loop
+	mov		$blk_init, $len
+___
+$::code.=<<___ if ($::evp);
+	st		%f0, [$ivec + 0]
+	st		%f1, [$ivec + 4]
+	st		%f2, [$ivec + 8]
+	st		%f3, [$ivec + 12]
+___
+$::code.=<<___ if (!$::evp);
+	brnz,pn		$ivoff, 3b
+	nop
+
+	std		%f0, [$ivec + 0]	! write out ivec
+	std		%f2, [$ivec + 8]
+___
+$::code.=<<___;
+	ret
+	restore
 .type	${alg}${bits}_t4_cbc_encrypt,#function
 .size	${alg}${bits}_t4_cbc_encrypt,.-${alg}${bits}_t4_cbc_encrypt
 ___
@@ -154,6 +246,7 @@ $::code.=<<___;
 .align	32
 ${alg}${bits}_t4_cbc_decrypt:
 	save		%sp, -$::frame, %sp
+	sub		$inp, $out, $blk_init	! $inp!=$out
 ___
 $::code.=<<___ if (!$::evp);
 	andcc		$ivec, 7, $ivoff
@@ -174,9 +267,9 @@ $::code.=<<___ if ($::evp);
 	ld		[$ivec + 12], %f15
 ___
 $::code.=<<___;
+	prefetch	[$inp], 20
+	prefetch	[$inp + 63], 20
 	call		_${alg}${bits}_load_deckey
-	srlx		$len, 4, $len
-	andcc		$len, 1, %g0		! is number of blocks even?
 	and		$inp, 7, $ileft
 	andn		$inp, 7, $inp
 	sll		$ileft, 3, $ileft
@@ -184,9 +277,17 @@ $::code.=<<___;
 	mov		0xff, $omask
 	sub		$iright, $ileft, $iright
 	and		$out, 7, $ooff
+	cmp		$len, 127
+	movrnz		$ooff, 0, $blk_init		! if (	$out&7 ||
+	movleu		$::size_t_cc, 0, $blk_init	!	$len<128 ||
+	brnz,pn		$blk_init, .L${bits}cbc_dec_blk	!	$inp==$out)
+	srl		$omask, $ooff, $omask
+
+	andcc		$len, 16, %g0		! is number of blocks even?
+	srlx		$len, 4, $len
 	alignaddrl	$out, %g0, $out
 	bz		%icc, .L${bits}_cbc_dec_loop2x
-	srl		$omask, $ooff, $omask
+	prefetch	[$out], 22
 .L${bits}_cbc_dec_loop:
 	ldx		[$inp + 0], %o0
 	brz,pt		$ileft, 4f
@@ -205,6 +306,8 @@ $::code.=<<___;
 	movxtod		%o2, %f0
 	movxtod		%o3, %f2
 
+	prefetch	[$out + 63], 22
+	prefetch	[$inp + 16+63], 20
 	call		_${alg}${bits}_decrypt_1x
 	add		$inp, 16, $inp
 
@@ -304,6 +407,8 @@ $::code.=<<___;
 	movxtod		%o4, %f4
 	movxtod		%o5, %f6
 
+	prefetch	[$out + 63], 22
+	prefetch	[$inp + 32+63], 20
 	call		_${alg}${bits}_decrypt_2x
 	add		$inp, 32, $inp
 
@@ -395,6 +500,97 @@ ___
 $::code.=<<___;
 	ret
 	restore
+
+!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+.align	32
+.L${bits}cbc_dec_blk:
+	add	$out, $len, $blk_init
+	and	$blk_init, 63, $blk_init	! tail
+	sub	$len, $blk_init, $len
+	add	$blk_init, 15, $blk_init	! round up to 16n
+	srlx	$len, 4, $len
+	srl	$blk_init, 4, $blk_init
+	sub	$len, 1, $len
+	add	$blk_init, 1, $blk_init
+
+.L${bits}_cbc_dec_blk_loop2x:
+	ldx		[$inp + 0], %o0
+	ldx		[$inp + 8], %o1
+	ldx		[$inp + 16], %o2
+	brz,pt		$ileft, 5f
+	ldx		[$inp + 24], %o3
+
+	ldx		[$inp + 32], %o4
+	sllx		%o0, $ileft, %o0
+	srlx		%o1, $iright, %g1
+	or		%g1, %o0, %o0
+	sllx		%o1, $ileft, %o1
+	srlx		%o2, $iright, %g1
+	or		%g1, %o1, %o1
+	sllx		%o2, $ileft, %o2
+	srlx		%o3, $iright, %g1
+	or		%g1, %o2, %o2
+	sllx		%o3, $ileft, %o3
+	srlx		%o4, $iright, %o4
+	or		%o4, %o3, %o3
+5:
+	xor		%g4, %o0, %o4		! ^= rk[0]
+	xor		%g5, %o1, %o5
+	movxtod		%o4, %f0
+	movxtod		%o5, %f2
+	xor		%g4, %o2, %o4
+	xor		%g5, %o3, %o5
+	movxtod		%o4, %f4
+	movxtod		%o5, %f6
+
+	prefetch	[$inp + 32+63], 20
+	call		_${alg}${bits}_decrypt_2x
+	add		$inp, 32, $inp
+	subcc		$len, 2, $len
+
+	movxtod		%o0, %f8
+	movxtod		%o1, %f10
+	fxor		%f12, %f0, %f0		! ^= ivec
+	fxor		%f14, %f2, %f2
+	movxtod		%o2, %f12
+	movxtod		%o3, %f14
+	fxor		%f8, %f4, %f4
+	fxor		%f10, %f6, %f6
+
+	stda		%f0, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f2, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f4, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f6, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	bgu,pt		$::size_t_cc, .L${bits}_cbc_dec_blk_loop2x
+	add		$out, 8, $out
+
+	add		$blk_init, $len, $len
+	andcc		$len, 1, %g0		! is number of blocks even?
+	!membar		0x0f
+	bnz,pt		%icc, .L${bits}_cbc_dec_loop
+	srl		$len, 0, $len
+	brnz,pn		$len, .L${bits}_cbc_dec_loop2x
+	nop
+___
+$::code.=<<___ if ($::evp);
+	st		%f0, [$ivec + 0]
+	st		%f1, [$ivec + 4]
+	st		%f2, [$ivec + 8]
+	st		%f3, [$ivec + 12]
+___
+$::code.=<<___ if (!$::evp);
+	brnz,pn		$ivoff, 3b
+	nop
+
+	std		%f0, [$ivec + 0]	! write out ivec
+	std		%f2, [$ivec + 8]
+___
+$::code.=<<___;
+	ret
+	restore
 .type	${alg}${bits}_t4_cbc_decrypt,#function
 .size	${alg}${bits}_t4_cbc_decrypt,.-${alg}${bits}_t4_cbc_decrypt
 ___
@@ -409,8 +605,10 @@ $::code.=<<___;
 ${alg}${bits}_t4_ctr32_encrypt:
 	save		%sp, -$::frame, %sp
 
+	prefetch	[$inp], 20
+	prefetch	[$inp + 63], 20
 	call		_${alg}${bits}_load_enckey
-	nop
+	sllx		$len, 4, $len
 
 	ld		[$ivec + 0], %l4	! counter
 	ld		[$ivec + 4], %l5
@@ -424,7 +622,7 @@ ${alg}${bits}_t4_ctr32_encrypt:
 	xor		%g1, %g5, %g5
 	movxtod		%g4, %f14		! most significant 64 bits
 
-	andcc		$len, 1, %g0		! is number of blocks even?
+	sub		$inp, $out, $blk_init	! $inp!=$out
 	and		$inp, 7, $ileft
 	andn		$inp, 7, $inp
 	sll		$ileft, 3, $ileft
@@ -432,9 +630,16 @@ ${alg}${bits}_t4_ctr32_encrypt:
 	mov		0xff, $omask
 	sub		$iright, $ileft, $iright
 	and		$out, 7, $ooff
+	cmp		$len, 127
+	movrnz		$ooff, 0, $blk_init		! if (	$out&7 ||
+	movleu		$::size_t_cc, 0, $blk_init	!	$len<128 ||
+	brnz,pn		$blk_init, .L${bits}_ctr32_blk	!	$inp==$out)
+	srl		$omask, $ooff, $omask
+
+	andcc		$len, 16, %g0		! is number of blocks even?
 	alignaddrl	$out, %g0, $out
 	bz		%icc, .L${bits}_ctr32_loop2x
-	srl		$omask, $ooff, $omask
+	srlx		$len, 4, $len
 .L${bits}_ctr32_loop:
 	ldx		[$inp + 0], %o0
 	brz,pt		$ileft, 4f
@@ -452,6 +657,8 @@ ${alg}${bits}_t4_ctr32_encrypt:
 	add		%l7, 1, %l7
 	movxtod		%g1, %f2
 	srl		%l7, 0, %l7		! clruw
+	prefetch	[$out + 63], 22
+	prefetch	[$inp + 16+63], 20
 ___
 $::code.=<<___ if ($alg eq "aes");
 	aes_eround01	%f16, %f14, %f2, %f4
@@ -531,7 +738,8 @@ $::code.=<<___;
 	add		%l7, 1, %l7
 	movxtod		%g1, %f6
 	srl		%l7, 0, %l7		! clruw
-
+	prefetch	[$out + 63], 22
+	prefetch	[$inp + 32+63], 20
 ___
 $::code.=<<___ if ($alg eq "aes");
 	aes_eround01	%f16, %f14, %f2, %f8
@@ -594,6 +802,96 @@ $::code.=<<___;
 
 	ret
 	restore
+
+!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+.align	32
+.L${bits}_ctr32_blk:
+	add	$out, $len, $blk_init
+	and	$blk_init, 63, $blk_init	! tail
+	sub	$len, $blk_init, $len
+	add	$blk_init, 15, $blk_init	! round up to 16n
+	srlx	$len, 4, $len
+	srl	$blk_init, 4, $blk_init
+	sub	$len, 1, $len
+	add	$blk_init, 1, $blk_init
+
+.L${bits}_ctr32_blk_loop2x:
+	ldx		[$inp + 0], %o0
+	ldx		[$inp + 8], %o1
+	ldx		[$inp + 16], %o2
+	brz,pt		$ileft, 5f
+	ldx		[$inp + 24], %o3
+
+	ldx		[$inp + 32], %o4
+	sllx		%o0, $ileft, %o0
+	srlx		%o1, $iright, %g1
+	or		%g1, %o0, %o0
+	sllx		%o1, $ileft, %o1
+	srlx		%o2, $iright, %g1
+	or		%g1, %o1, %o1
+	sllx		%o2, $ileft, %o2
+	srlx		%o3, $iright, %g1
+	or		%g1, %o2, %o2
+	sllx		%o3, $ileft, %o3
+	srlx		%o4, $iright, %o4
+	or		%o4, %o3, %o3
+5:
+	xor		%g5, %l7, %g1		! ^= rk[0]
+	add		%l7, 1, %l7
+	movxtod		%g1, %f2
+	srl		%l7, 0, %l7		! clruw
+	xor		%g5, %l7, %g1
+	add		%l7, 1, %l7
+	movxtod		%g1, %f6
+	srl		%l7, 0, %l7		! clruw
+	prefetch	[$inp + 32+63], 20
+___
+$::code.=<<___ if ($alg eq "aes");
+	aes_eround01	%f16, %f14, %f2, %f8
+	aes_eround23	%f18, %f14, %f2, %f2
+	aes_eround01	%f16, %f14, %f6, %f10
+	aes_eround23	%f18, %f14, %f6, %f6
+___
+$::code.=<<___ if ($alg eq "cmll");
+	camellia_f	%f16, %f2, %f14, %f2
+	camellia_f	%f16, %f6, %f14, %f6
+	camellia_f	%f18, %f14, %f2, %f0
+	camellia_f	%f18, %f14, %f6, %f4
+___
+$::code.=<<___;
+	call		_${alg}${bits}_encrypt_2x+16
+	add		$inp, 32, $inp
+	subcc		$len, 2, $len
+
+	movxtod		%o0, %f8
+	movxtod		%o1, %f10
+	movxtod		%o2, %f12
+	fxor		%f8, %f0, %f0		! ^= inp
+	movxtod		%o3, %f8
+	fxor		%f10, %f2, %f2
+	fxor		%f12, %f4, %f4
+	fxor		%f8, %f6, %f6
+
+	stda		%f0, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f2, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f4, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	add		$out, 8, $out
+	stda		%f6, [$out]0xf2		! ASI_BLK_INIT, T4-specific
+	bgu,pt		$::size_t_cc, .L${bits}_ctr32_blk_loop2x
+	add		$out, 8, $out
+
+	add		$blk_init, $len, $len
+	andcc		$len, 1, %g0		! is number of blocks even?
+	!membar		0x0f
+	bnz,pt		%icc, .L${bits}_ctr32_loop
+	srl		$len, 0, $len
+	brnz,pn		$len, .L${bits}_ctr32_loop2x
+	nop
+
+	ret
+	restore
 .type	${alg}${bits}_t4_ctr32_encrypt,#function
 .size	${alg}${bits}_t4_ctr32_encrypt,.-${alg}${bits}_t4_ctr32_encrypt
 ___