ppc64-mont.pl: adapt for 32-bit and engage for all builds.

Configure

@@ -135,8 +135,8 @@ my $alpha_asm="alphacpuid.o:bn_asm.o alpha-mont.o::::::::::::void";
 my $mips3_asm=":bn-mips3.o::::::::::::void";
 my $s390x_asm="s390xcpuid.o:bn-s390x.o s390x-mont.o::aes-s390x.o:::sha1-s390x.o sha256-s390x.o sha512-s390x.o::rc4-s390x.o:::::void";
 my $armv4_asm=":bn_asm.o armv4-mont.o::aes_cbc.o aes-armv4.o:::sha1-armv4-large.o sha256-armv4.o sha512-armv4.o:::::::void";
-my $ppc32_asm="ppccpuid.o:bn-ppc.o::aes_core.o aes_cbc.o aes-ppc.o:::sha1-ppc.o sha256-ppc.o::::::";
-my $ppc64_asm="ppccpuid.o:bn-ppc.o ppc-mont.o::aes_core.o aes_cbc.o aes-ppc.o:::sha1-ppc.o sha256-ppc.o sha512-ppc.o::::::";
+my $ppc32_asm="ppccpuid.o ppccap.o:bn-ppc.o ppc-mont.o ppc64-mont.o::aes_core.o aes_cbc.o aes-ppc.o:::sha1-ppc.o sha256-ppc.o::::::";
+my $ppc64_asm="ppccpuid.o ppccap.o:bn-ppc.o ppc-mont.o ppc64-mont.o::aes_core.o aes_cbc.o aes-ppc.o:::sha1-ppc.o sha256-ppc.o sha512-ppc.o::::::";
 my $no_asm=":::::::::::::void";
 
 # As for $BSDthreads. Idea is to maintain "collective" set of flags,
@@ -547,7 +547,7 @@ my %table=(
 
 ##### MacOS X (a.k.a. Rhapsody or Darwin) setup
 "rhapsody-ppc-cc","cc:-O3 -DB_ENDIAN::(unknown):MACOSX_RHAPSODY::BN_LLONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR:${no_asm}::",
-"darwin-ppc-cc","cc:-arch ppc -O3 -DB_ENDIAN::-D_REENTRANT:MACOSX:-Wl,-search_paths_first%:BN_LLONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR:${ppc32_asm}:osx32:dlfcn:darwin-shared:-fPIC -fno-common:-arch ppc -dynamiclib:.\$(SHLIB_MAJOR).\$(SHLIB_MINOR).dylib",
+"darwin-ppc-cc","cc:-arch ppc -O3 -DB_ENDIAN -Wa,-force_cpusubtype_ALL::-D_REENTRANT:MACOSX:-Wl,-search_paths_first%:BN_LLONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR:${ppc32_asm}:osx32:dlfcn:darwin-shared:-fPIC -fno-common:-arch ppc -dynamiclib:.\$(SHLIB_MAJOR).\$(SHLIB_MINOR).dylib",
 "darwin64-ppc-cc","cc:-arch ppc64 -O3 -DB_ENDIAN::-D_REENTRANT:MACOSX:-Wl,-search_paths_first%:SIXTY_FOUR_BIT_LONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR:${ppc64_asm}:osx64:dlfcn:darwin-shared:-fPIC -fno-common:-arch ppc64 -dynamiclib:.\$(SHLIB_MAJOR).\$(SHLIB_MINOR).dylib",
 "darwin-i386-cc","cc:-arch i386 -O3 -fomit-frame-pointer -DL_ENDIAN::-D_REENTRANT:MACOSX:-Wl,-search_paths_first%:BN_LLONG RC4_INT RC4_CHUNK DES_UNROLL BF_PTR:${x86_asm}:macosx:dlfcn:darwin-shared:-fPIC -fno-common:-arch i386 -dynamiclib:.\$(SHLIB_MAJOR).\$(SHLIB_MINOR).dylib",
 "debug-darwin-i386-cc","cc:-arch i386 -g3 -DL_ENDIAN::-D_REENTRANT:MACOSX:-Wl,-search_paths_first%:BN_LLONG RC4_INT RC4_CHUNK DES_UNROLL BF_PTR:${x86_asm}:macosx:dlfcn:darwin-shared:-fPIC -fno-common:-arch i386 -dynamiclib:.\$(SHLIB_MAJOR).\$(SHLIB_MINOR).dylib",

TABLE

@@ -814,8 +814,8 @@ $thread_cflag = -qthreaded
 $sys_id       = AIX
 $lflags       = 
 $bn_ops       = BN_LLONG RC4_CHAR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -845,8 +845,8 @@ $thread_cflag = -pthread
 $sys_id       = AIX
 $lflags       = 
 $bn_ops       = BN_LLONG RC4_CHAR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -907,8 +907,8 @@ $thread_cflag = -qthreaded
 $sys_id       = AIX
 $lflags       = 
 $bn_ops       = SIXTY_FOUR_BIT_LONG RC4_CHAR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o ppc-mont.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -938,8 +938,8 @@ $thread_cflag = -pthread
 $sys_id       = AIX
 $lflags       = 
 $bn_ops       = SIXTY_FOUR_BIT_LONG RC4_CHAR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o ppc-mont.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -1211,14 +1211,14 @@ $multilib     =
 
 *** darwin-ppc-cc
 $cc           = cc
-$cflags       = -arch ppc -O3 -DB_ENDIAN
+$cflags       = -arch ppc -O3 -DB_ENDIAN -Wa,-force_cpusubtype_ALL
 $unistd       = 
 $thread_cflag = -D_REENTRANT
 $sys_id       = MACOSX
 $lflags       = -Wl,-search_paths_first%
 $bn_ops       = BN_LLONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -1248,8 +1248,8 @@ $thread_cflag = -D_REENTRANT
 $sys_id       = MACOSX
 $lflags       = -Wl,-search_paths_first%
 $bn_ops       = SIXTY_FOUR_BIT_LONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o ppc-mont.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -1682,8 +1682,8 @@ $thread_cflag = -D_REENTRANT
 $sys_id       = MACOSX
 $lflags       = 
 $bn_ops       = BN_LLONG RC4_CHAR RC4_CHUNK DES_UNROLL BF_PTR
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -2327,7 +2327,7 @@ $multilib     =
 
 *** debug-steve32
 $cc           = gcc
-$cflags       = -Wall -pedantic -DPEDANTIC -Wno-long-long -Wsign-compare -Wmissing-prototypes -Wshadow -Wformat -Werror -DCRYPTO_MDEBUG_ALL -DCRYPTO_MDEBUG_ABORT -DREF_CHECK -DOPENSSL_NO_DEPRECATED -m32 -DL_ENDIAN -DCONF_DEBUG -DDEBUG_SAFESTACK -DDEBUG_UNUSED -g -pipe
+$cflags       = -Wall -pedantic -DPEDANTIC -Wno-long-long -Wsign-compare -Wmissing-prototypes -Wshadow -Wformat -Werror -DCRYPTO_MDEBUG_ALL -DCRYPTO_MDEBUG_ABORT -DREF_CHECK -DOPENSSL_NO_DEPRECATED -m32 -DL_ENDIAN -DCONF_DEBUG -DDEBUG_SAFESTACK -g -pipe
 $unistd       = 
 $thread_cflag = -D_REENTRANT
 $sys_id       = 
@@ -2358,7 +2358,7 @@ $multilib     =
 
 *** debug-steve64
 $cc           = gcc
-$cflags       = -Wall -pedantic -DPEDANTIC -Wno-long-long -Wsign-compare -Wmissing-prototypes -Wshadow -Wformat -Werror -DCRYPTO_MDEBUG_ALL -DCRYPTO_MDEBUG_ABORT -DREF_CHECK -DOPENSSL_NO_DEPRECATED -m64 -DL_ENDIAN -DTERMIO -DCONF_DEBUG -DDEBUG_SAFESTACK -DDEBUG_UNUSED -g -DMD32_REG_T=int
+$cflags       = -Wall -pedantic -DPEDANTIC -Wno-long-long -Wsign-compare -Wmissing-prototypes -Wshadow -Wformat -Werror -DCRYPTO_MDEBUG_ALL -DCRYPTO_MDEBUG_ABORT -DREF_CHECK -DOPENSSL_NO_DEPRECATED -m64 -DL_ENDIAN -DTERMIO -DCONF_DEBUG -DDEBUG_SAFESTACK -g -DMD32_REG_T=int
 $unistd       = 
 $thread_cflag = -D_REENTRANT
 $sys_id       = 
@@ -3666,8 +3666,8 @@ $thread_cflag = -D_REENTRANT
 $sys_id       = 
 $lflags       = -ldl
 $bn_ops       = BN_LLONG RC4_CHAR RC4_CHUNK DES_RISC1 DES_UNROLL
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 
@@ -3697,8 +3697,8 @@ $thread_cflag = -D_REENTRANT
 $sys_id       = 
 $lflags       = -ldl
 $bn_ops       = SIXTY_FOUR_BIT_LONG RC4_CHAR RC4_CHUNK DES_RISC1 DES_UNROLL
-$cpuid_obj    = ppccpuid.o
-$bn_obj       = bn-ppc.o ppc-mont.o
+$cpuid_obj    = ppccpuid.o ppccap.o
+$bn_obj       = bn-ppc.o ppc-mont.o ppc64-mont.o
 $des_obj      = 
 $aes_obj      = aes_core.o aes_cbc.o aes-ppc.o
 $bf_obj       = 

crypto/bn/Makefile

@@ -103,6 +103,7 @@ pa-risc2.o: asm/pa-risc2.s
 # ppc - AIX, Linux, MacOS X...
 bn-ppc.s:	asm/ppc.pl;	$(PERL) asm/ppc.pl $(PERLASM_SCHEME) $@
 ppc-mont.s:	asm/ppc-mont.pl;$(PERL) asm/ppc-mont.pl $(PERLASM_SCHEME) $@
+ppc64-mont.s:	asm/ppc64-mont.pl;$(PERL) asm/ppc64-mont.pl $(PERLASM_SCHEME) $@
 
 alpha-mont.s:	asm/alpha-mont.pl
 	$(PERL) $< | $(CC) -E - | tee $@ > /dev/null

crypto/bn/asm/ppc-mont.pl

@@ -108,14 +108,19 @@ $code=<<___;
 .machine "any"
 .text
 
-.globl	.bn_mul_mont
+.globl	.bn_mul_mont_int
 .align	4
-.bn_mul_mont:
+.bn_mul_mont_int:
 	cmpwi	$num,4
 	mr	$rp,r3		; $rp is reassigned
 	li	r3,0
 	bltlr
-
+___
+$code.=<<___ if ($BNSZ==4);
+	cmpwi	$num,32		; longer key performance is not better
+	bgelr
+___
+$code.=<<___;
 	slwi	$num,$num,`log($BNSZ)/log(2)`
 	li	$tj,-4096
 	addi	$ovf,$num,`$FRAME+$RZONE`

crypto/bn/asm/ppc64-mont.pl

@@ -45,23 +45,41 @@
 # on 1.8GHz PPC970, it's only 5-55% faster. Still far from impressive
 # in absolute terms, but it's apparently the way Power 6 is...
 
+# December 2009
+
+# Adapted for 32-bit build this module delivers 25-120%, more for
+# longer keys, performance improvement on 1.8GHz PPC970. However!
+# This implementation utilizes even 64-bit integer operations and
+# trouble is that most PPC operating systems don't preserve upper
+# halves of general purpose registers upong signal delivery. They do
+# preserve them upon context switch, but not signalling:-( This means
+# that asynchronous signals have to be blocked upon entry to this
+# subroutine. Signal masking (and complementary unmasking) has quite
+# an impact on performance, naturally larger for shorter keys. It's
+# so severe that shorter key performance as low as 1/3 of expected
+# one. This is why this routine should be engaged for longer key
+# operations only, see crypto/ppccap.c for further details.
+# Alternative is to break dependance on upper halves on GPRs...
+# MacOS X is an exception from this and doesn't require signal
+# masking, and that's where above improvement coefficients were
+# collected.
+
 $flavour = shift;
 
 if ($flavour =~ /32/) {
 	$SIZE_T=4;
 	$RZONE=	224;
 	$FRAME=	$SIZE_T*12+8*12;
-	$fname=	"bn_mul_mont_ppc64";
+	$fname=	"bn_mul_mont_fpu64";
 
 	$STUX=	"stwux";	# store indexed and update
 	$PUSH=	"stw";
 	$POP=	"lwz";
-	die "not implemented yet";
 } elsif ($flavour =~ /64/) {
 	$SIZE_T=8;
 	$RZONE=	288;
 	$FRAME=	$SIZE_T*12+8*12;
-	$fname=	"bn_mul_mont";
+	$fname=	"bn_mul_mont_fpu64";
 
 	# same as above, but 64-bit mnemonics...
 	$STUX=	"stdux";	# store indexed and update
@@ -181,14 +199,14 @@ $code=<<___;
 .globl	.$fname
 .align	5
 .$fname:
-	cmpwi	$num,4
+	cmpwi	$num,`3*8/$SIZE_T`
 	mr	$rp,r3		; $rp is reassigned
 	li	r3,0		; possible "not handled" return code
 	bltlr-
-	andi.	r0,$num,1	; $num has to be even
+	andi.	r0,$num,`16/$SIZE_T-1`		; $num has to be "even"
 	bnelr-
 
-	slwi	$num,$num,3	; num*=8
+	slwi	$num,$num,`log($SIZE_T)/log(2)`	; num*=sizeof(BN_LONG)
 	li	$i,-4096
 	slwi	$tp,$num,2	; place for {an}p_{lh}[num], i.e. 4*num
 	add	$tp,$tp,$num	; place for tp[num+1]
@@ -220,11 +238,25 @@ $code=<<___;
 	stfd	f23,`12*$SIZE_T+72`($sp)
 	stfd	f24,`12*$SIZE_T+80`($sp)
 	stfd	f25,`12*$SIZE_T+88`($sp)
-
+___
+$code.=<<___ if ($SIZE_T==8);
 	ld	$a0,0($ap)	; pull ap[0] value
 	ld	$n0,0($n0)	; pull n0[0] value
 	ld	$t3,0($bp)	; bp[0]
-
+___
+$code.=<<___ if ($SIZE_T==4);
+	mr	$t1,$n0
+	lwz	$a0,0($ap)	; pull ap[0,1] value
+	lwz	$t0,4($ap)
+	lwz	$n0,0($t1)	; pull n0[0,1] value
+	lwz	$t1,4($t1)
+	lwz	$t3,0($bp)	; bp[0,1]
+	lwz	$t2,4($bp)
+	insrdi	$a0,$t0,32,0
+	insrdi	$n0,$t1,32,0
+	insrdi	$t3,$t2,32,0
+___
+$code.=<<___;
 	addi	$tp,$sp,`$FRAME+$TRANSFER+8+64`
 	li	$i,-64
 	add	$nap_d,$tp,$num
@@ -258,6 +290,8 @@ $code=<<___;
 	std	$t5,`$FRAME+40`($sp)
 	std	$t6,`$FRAME+48`($sp)
 	std	$t7,`$FRAME+56`($sp)
+___
+$code.=<<___ if ($SIZE_T==8);
 	lwz	$t0,4($ap)		; load a[j] as 32-bit word pair
 	lwz	$t1,0($ap)
 	lwz	$t2,12($ap)		; load a[j+1] as 32-bit word pair
@@ -266,6 +300,18 @@ $code=<<___;
 	lwz	$t5,0($np)
 	lwz	$t6,12($np)		; load n[j+1] as 32-bit word pair
 	lwz	$t7,8($np)
+___
+$code.=<<___ if ($SIZE_T==4);
+	lwz	$t0,0($ap)		; load a[j..j+3] as 32-bit word pairs
+	lwz	$t1,4($ap)
+	lwz	$t2,8($ap)
+	lwz	$t3,12($ap)
+	lwz	$t4,0($np)		; load n[j..j+3] as 32-bit word pairs
+	lwz	$t5,4($np)
+	lwz	$t6,8($np)
+	lwz	$t7,12($np)
+___
+$code.=<<___;
 	lfd	$ba,`$FRAME+0`($sp)
 	lfd	$bb,`$FRAME+8`($sp)
 	lfd	$bc,`$FRAME+16`($sp)
@@ -374,6 +420,8 @@ $code=<<___;
 
 .align	5
 L1st:
+___
+$code.=<<___ if ($SIZE_T==8);
 	lwz	$t0,4($ap)		; load a[j] as 32-bit word pair
 	lwz	$t1,0($ap)
 	lwz	$t2,12($ap)		; load a[j+1] as 32-bit word pair
@@ -382,6 +430,18 @@ L1st:
 	lwz	$t5,0($np)
 	lwz	$t6,12($np)		; load n[j+1] as 32-bit word pair
 	lwz	$t7,8($np)
+___
+$code.=<<___ if ($SIZE_T==4);
+	lwz	$t0,0($ap)		; load a[j..j+3] as 32-bit word pairs
+	lwz	$t1,4($ap)
+	lwz	$t2,8($ap)
+	lwz	$t3,12($ap)
+	lwz	$t4,0($np)		; load n[j..j+3] as 32-bit word pairs
+	lwz	$t5,4($np)
+	lwz	$t6,8($np)
+	lwz	$t7,12($np)
+___
+$code.=<<___;
 	std	$t0,`$FRAME+64`($sp)
 	std	$t1,`$FRAME+72`($sp)
 	std	$t2,`$FRAME+80`($sp)
@@ -559,7 +619,17 @@ L1st:
 	li	$i,8			; i=1
 .align	5
 Louter:
+___
+$code.=<<___ if ($SIZE_T==8);
 	ldx	$t3,$bp,$i	; bp[i]
+___
+$code.=<<___ if ($SIZE_T==4);
+	add	$t0,$bp,$i
+	lwz	$t3,0($t0)		; bp[i,i+1]
+	lwz	$t0,4($t0)
+	insrdi	$t3,$t0,32,0
+___
+$code.=<<___;
 	ld	$t6,`$FRAME+$TRANSFER+8`($sp)	; tp[0]
 	mulld	$t7,$a0,$t3	; ap[0]*bp[i]
 
@@ -761,6 +831,13 @@ Linner:
 	stfd	$T0b,`$FRAME+8`($sp)
 	 add	$t7,$t7,$carry
 	 addc	$t3,$t0,$t1
+___
+$code.=<<___ if ($SIZE_T==4);		# adjust XER[CA]
+	extrdi	$t0,$t0,32,0
+	extrdi	$t1,$t1,32,0
+	adde	$t0,$t0,$t1
+___
+$code.=<<___;
 	stfd	$T1a,`$FRAME+16`($sp)
 	stfd	$T1b,`$FRAME+24`($sp)
 	 insrdi	$t4,$t7,16,0		; 64..127 bits
@@ -768,6 +845,13 @@ Linner:
 	stfd	$T2a,`$FRAME+32`($sp)
 	stfd	$T2b,`$FRAME+40`($sp)
 	 adde	$t5,$t4,$t2
+___
+$code.=<<___ if ($SIZE_T==4);		# adjust XER[CA]
+	extrdi	$t4,$t4,32,0
+	extrdi	$t2,$t2,32,0
+	adde	$t4,$t4,$t2
+___
+$code.=<<___;
 	stfd	$T3a,`$FRAME+48`($sp)
 	stfd	$T3b,`$FRAME+56`($sp)
 	 addze	$carry,$carry
@@ -816,7 +900,21 @@ Linner:
 	ld	$t7,`$FRAME+72`($sp)
 
 	addc	$t3,$t0,$t1
+___
+$code.=<<___ if ($SIZE_T==4);		# adjust XER[CA]
+	extrdi	$t0,$t0,32,0
+	extrdi	$t1,$t1,32,0
+	adde	$t0,$t0,$t1
+___
+$code.=<<___;
 	adde	$t5,$t4,$t2
+___
+$code.=<<___ if ($SIZE_T==4);		# adjust XER[CA]
+	extrdi	$t4,$t4,32,0
+	extrdi	$t2,$t2,32,0
+	adde	$t4,$t4,$t2
+___
+$code.=<<___;
 	addze	$carry,$carry
 
 	std	$t3,-16($tp)		; tp[j-1]
@@ -835,7 +933,9 @@ Linner:
 	subf	$nap_d,$t7,$nap_d	; rewind pointer
 	cmpw	$i,$num
 	blt-	Louter
+___
 
+$code.=<<___ if ($SIZE_T==8);
 	subf	$np,$num,$np	; rewind np
 	addi	$j,$j,1		; restore counter
 	subfc	$i,$i,$i	; j=0 and "clear" XER[CA]
@@ -883,7 +983,74 @@ Lcopy:				; copy or in-place refresh
 	stdx	$i,$t4,$i
 	addi	$i,$i,16
 	bdnz-	Lcopy
+___
+$code.=<<___ if ($SIZE_T==4);
+	subf	$np,$num,$np	; rewind np
+	addi	$j,$j,1		; restore counter
+	subfc	$i,$i,$i	; j=0 and "clear" XER[CA]
+	addi	$tp,$sp,`$FRAME+$TRANSFER`
+	addi	$np,$np,-4
+	addi	$rp,$rp,-4
+	addi	$ap,$sp,`$FRAME+$TRANSFER+4`
+	mtctr	$j
+
+.align	4
+Lsub:	ld	$t0,8($tp)	; load tp[j..j+3] in 64-bit word order
+	ldu	$t2,16($tp)
+	lwz	$t4,4($np)	; load np[j..j+3] in 32-bit word order
+	lwz	$t5,8($np)
+	lwz	$t6,12($np)
+	lwzu	$t7,16($np)
+	extrdi	$t1,$t0,32,0
+	extrdi	$t3,$t2,32,0
+	subfe	$t4,$t4,$t0	; tp[j]-np[j]
+	 stw	$t0,4($ap)	; save tp[j..j+3] in 32-bit word order
+	subfe	$t5,$t5,$t1	; tp[j+1]-np[j+1]
+	 stw	$t1,8($ap)
+	subfe	$t6,$t6,$t2	; tp[j+2]-np[j+2]
+	 stw	$t2,12($ap)
+	subfe	$t7,$t7,$t3	; tp[j+3]-np[j+3]
+	 stwu	$t3,16($ap)
+	stw	$t4,4($rp)
+	stw	$t5,8($rp)
+	stw	$t6,12($rp)
+	stwu	$t7,16($rp)
+	bdnz-	Lsub
+
+	li	$i,0
+	subfe	$ovf,$i,$ovf	; handle upmost overflow bit
+	addi	$tp,$sp,`$FRAME+$TRANSFER+4`
+	subf	$rp,$num,$rp	; rewind rp
+	and	$ap,$tp,$ovf
+	andc	$np,$rp,$ovf
+	or	$ap,$ap,$np	; ap=borrow?tp:rp
+	addi	$tp,$sp,`$FRAME+$TRANSFER`
+	mtctr	$j
+
+.align	4
+Lcopy:				; copy or in-place refresh
+	lwz	$t0,4($ap)
+	lwz	$t1,8($ap)
+	lwz	$t2,12($ap)
+	lwzu	$t3,16($ap)
+	std	$i,8($nap_d)	; zap nap_d
+	std	$i,16($nap_d)
+	std	$i,24($nap_d)
+	std	$i,32($nap_d)
+	std	$i,40($nap_d)
+	std	$i,48($nap_d)
+	std	$i,56($nap_d)
+	stdu	$i,64($nap_d)
+	stw	$t0,4($rp)
+	stw	$t1,8($rp)
+	stw	$t2,12($rp)
+	stwu	$t3,16($rp)
+	std	$i,8($tp)	; zap tp at once
+	stdu	$i,16($tp)
+	bdnz-	Lcopy
+___
 
+$code.=<<___;
 	$POP	r14,`2*$SIZE_T`($sp)
 	$POP	r15,`3*$SIZE_T`($sp)
 	$POP	r16,`4*$SIZE_T`($sp)

crypto/ppccap.c

@@ -0,0 +1,88 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <setjmp.h>
+#include <signal.h>
+#include <openssl/bn.h>
+
+#define PPC_FPU64	(1<<0)
+
+static int OPENSSL_ppccap_P = 0;
+
+static sigset_t all_masked;
+
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np, const BN_ULONG *n0, int num)
+	{
+	int bn_mul_mont_fpu64(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np, const BN_ULONG *n0, int num);
+	int bn_mul_mont_int(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np, const BN_ULONG *n0, int num);
+
+	if (sizeof(size_t)==4)
+		{
+#if (defined(__APPLE__) && defined(__MACH__))
+		if ((OPENSSL_ppccap_P&PPC_FPU64))
+			return bn_mul_mont_fpu64(rp,ap,bp,np,n0,num);
+#else
+		/* boundary of 32 was experimentally determined on
+		   Linux 2.6.22, might have to be adjusted on AIX... */
+		if ((num>=32) && (OPENSSL_ppccap_P&PPC_FPU64))
+			{
+			sigset_t oset;
+			int ret;
+
+			sigprocmask(SIG_SETMASK,&all_masked,&oset);
+			ret=bn_mul_mont_fpu64(rp,ap,bp,np,n0,num);
+			sigprocmask(SIG_SETMASK,&oset,NULL);
+
+			return ret;
+			}
+#endif
+		}
+	else if ((OPENSSL_ppccap_P&PPC_FPU64))
+		/* this is a "must" on Power 6, but run-time detection
+		 * is not implemented yet... */
+		return bn_mul_mont_fpu64(rp,ap,bp,np,n0,num);
+
+	return bn_mul_mont_int(rp,ap,bp,np,n0,num);
+	}
+
+static sigjmp_buf ill_jmp;
+static void ill_handler (int sig) { siglongjmp(ill_jmp,sig); }
+
+void OPENSSL_cpuid_setup(void)
+	{
+	char *e;
+
+	sigfillset(&all_masked);
+	sigdelset(&all_masked,SIGSEGV);
+	sigdelset(&all_masked,SIGILL);
+
+	if ((e=getenv("OPENSSL_ppccap")))
+		{
+		OPENSSL_ppccap_P=strtoul(e,NULL,0);
+		return;
+		}
+
+	if (sizeof(size_t)==4)
+		{
+		struct sigaction	ill_oact,ill_act;
+		sigset_t		oset;
+
+		memset(&ill_act,0,sizeof(ill_act));
+		ill_act.sa_handler = ill_handler;
+		sigfillset(&ill_act.sa_mask);
+		sigdelset(&ill_act.sa_mask,SIGILL);
+		sigprocmask(SIG_SETMASK,&ill_act.sa_mask,&oset);
+		sigaction (SIGILL,&ill_act,&ill_oact);
+		if (sigsetjmp(ill_jmp,0) == 0)
+			{
+			OPENSSL_ppc64_probe();
+			OPENSSL_ppccap_P |= PPC_FPU64;
+			}
+		else
+			{
+			OPENSSL_ppccap_P &= ~PPC_FPU64;
+			}
+		sigaction (SIGILL,&ill_oact,NULL);
+		sigprocmask(SIG_SETMASK,&oset,NULL);
+		}
+	}

crypto/ppccpuid.pl

@@ -23,9 +23,11 @@ $code=<<___;
 .machine	"any"
 .text
 
-.globl	.OPENSSL_cpuid_setup
+.globl	.OPENSSL_ppc64_probe
 .align	4
-.OPENSSL_cpuid_setup:
+.OPENSSL_ppc64_probe:
+	fcfid	f1,f1
+	extrdi	r0,r0,32,0
 	blr
 
 .globl	.OPENSSL_wipe_cpu