sparcv9cap.c: update from master.

crypto/sparc_arch.h

@@ -0,0 +1,101 @@
+#ifndef __SPARC_ARCH_H__
+#define __SPARC_ARCH_H__
+
+#define SPARCV9_TICK_PRIVILEGED	(1<<0)
+#define SPARCV9_PREFER_FPU	(1<<1)
+#define SPARCV9_VIS1		(1<<2)
+#define SPARCV9_VIS2		(1<<3)	/* reserved */
+#define SPARCV9_FMADD		(1<<4)	/* reserved for SPARC64 V */
+#define SPARCV9_BLK		(1<<5)	/* VIS1 block copy */
+#define SPARCV9_VIS3		(1<<6)
+#define SPARCV9_RANDOM		(1<<7)
+#define SPARCV9_64BIT_STACK	(1<<8)
+
+/*
+ * OPENSSL_sparcv9cap_P[1] is copy of Compatibility Feature Register,
+ * %asr26, SPARC-T4 and later. There is no SPARCV9_CFR bit in
+ * OPENSSL_sparcv9cap_P[0], as %cfr copy is sufficient...
+ */
+#define CFR_AES		0x00000001 /* Supports AES opcodes     */
+#define CFR_DES		0x00000002 /* Supports DES opcodes     */
+#define CFR_KASUMI	0x00000004 /* Supports KASUMI opcodes  */
+#define CFR_CAMELLIA	0x00000008 /* Supports CAMELLIA opcodes*/
+#define CFR_MD5		0x00000010 /* Supports MD5 opcodes     */
+#define CFR_SHA1	0x00000020 /* Supports SHA1 opcodes    */
+#define CFR_SHA256	0x00000040 /* Supports SHA256 opcodes  */
+#define CFR_SHA512	0x00000080 /* Supports SHA512 opcodes  */
+#define CFR_MPMUL	0x00000100 /* Supports MPMUL opcodes   */
+#define CFR_MONTMUL	0x00000200 /* Supports MONTMUL opcodes */
+#define CFR_MONTSQR	0x00000400 /* Supports MONTSQR opcodes */
+#define CFR_CRC32C	0x00000800 /* Supports CRC32C opcodes  */
+
+#if defined(OPENSSL_PIC) && !defined(__PIC__)
+# define __PIC__
+#endif
+
+#if defined(__SUNPRO_C) && defined(__sparcv9) && !defined(__arch64__)
+# define __arch64__
+#endif
+
+#define SPARC_PIC_THUNK(reg)	\
+	.align	32;		\
+.Lpic_thunk:			\
+	jmp	%o7 + 8;	\
+	 add	%o7, reg, reg;
+
+#define SPARC_PIC_THUNK_CALL(reg)			\
+	sethi	%hi(_GLOBAL_OFFSET_TABLE_-4), reg;	\
+	call	.Lpic_thunk;				\
+	 or	reg, %lo(_GLOBAL_OFFSET_TABLE_+4), reg;
+
+#if 1
+# define SPARC_SETUP_GOT_REG(reg)	SPARC_PIC_THUNK_CALL(reg)
+#else
+# define SPARC_SETUP_GOT_REG(reg)	\
+	sethi	%hi(_GLOBAL_OFFSET_TABLE_-4), reg;	\
+	call	.+8;					\
+	or	reg,%lo(_GLOBAL_OFFSET_TABLE_+4), reg;	\
+	add	%o7, reg, reg
+#endif
+
+#if defined(__arch64__)
+
+# define SPARC_LOAD_ADDRESS(SYM, reg)	\
+	setx	SYM, %o7, reg;
+# define LDPTR		ldx
+# define SIZE_T_CC	%xcc
+# define STACK_FRAME	192
+# define STACK_BIAS	2047
+# define STACK_7thARG	(STACK_BIAS+176)
+
+#else
+
+# define SPARC_LOAD_ADDRESS(SYM, reg)	\
+	set	SYM, reg;
+# define LDPTR		ld
+# define SIZE_T_CC	%icc
+# define STACK_FRAME	112
+# define STACK_BIAS	0
+# define STACK_7thARG	92
+# define SPARC_LOAD_ADDRESS_LEAF(SYM,reg,tmp) SPARC_LOAD_ADDRESS(SYM,reg)
+
+#endif
+
+#ifdef __PIC__
+# undef	SPARC_LOAD_ADDRESS
+# undef SPARC_LOAD_ADDRESS_LEAF
+# define SPARC_LOAD_ADDRESS(SYM, reg)	\
+	SPARC_SETUP_GOT_REG(reg);	\
+	sethi	%hi(SYM), %o7;		\
+	or	%o7, %lo(SYM), %o7;	\
+	LDPTR	[reg + %o7], reg;
+#endif
+
+#ifndef SPARC_LOAD_ADDRESS_LEAF
+# define SPARC_LOAD_ADDRESS_LEAF(SYM, reg, tmp)	\
+	mov	%o7, tmp;			\
+	SPARC_LOAD_ADDRESS(SYM, reg)		\
+	mov	tmp, %o7;
+#endif
+
+#endif	/* __SPARC_ARCH_H__ */

crypto/sparccpuid.S

@@ -251,6 +251,11 @@ _sparcv9_vis1_probe:
 !	UltraSPARC IIe		7
 !	UltraSPARC III		7
 !	UltraSPARC T1		24
+!	SPARC T4		65(*)
+!
+! (*)	result has lesser to do with VIS instruction latencies, rdtick
+!	appears that slow, but it does the trick in sense that FP and
+!	VIS code paths are still slower than integer-only ones.
 !
 ! Numbers for T2 and SPARC64 V-VII are more than welcomed.
 !
@@ -260,6 +265,8 @@ _sparcv9_vis1_probe:
 .global	_sparcv9_vis1_instrument
 .align	8
 _sparcv9_vis1_instrument:
+	.word	0x81b00d80	!fxor	%f0,%f0,%f0
+	.word	0x85b08d82	!fxor	%f2,%f2,%f2
 	.word	0x91410000	!rd	%tick,%o0
 	.word	0x81b00d80	!fxor	%f0,%f0,%f0
 	.word	0x85b08d82	!fxor	%f2,%f2,%f2
@@ -314,6 +321,30 @@ _sparcv9_fmadd_probe:
 .type	_sparcv9_fmadd_probe,#function
 .size	_sparcv9_fmadd_probe,.-_sparcv9_fmadd_probe
 
+.global	_sparcv9_rdcfr
+.align	8
+_sparcv9_rdcfr:
+	retl
+	.word	0x91468000	!rd	%asr26,%o0
+.type	_sparcv9_rdcfr,#function
+.size	_sparcv9_rdcfr,.-_sparcv9_rdcfr
+
+.global	_sparcv9_vis3_probe
+.align	8
+_sparcv9_vis3_probe:
+	retl
+	.word	0x81b022a0	!xmulx	%g0,%g0,%g0
+.type	_sparcv9_vis3_probe,#function
+.size	_sparcv9_vis3_probe,.-_sparcv9_vis3_probe
+
+.global	_sparcv9_random
+.align	8
+_sparcv9_random:
+	retl
+	.word	0x91b002a0	!random	%o0
+.type	_sparcv9_random,#function
+.size	_sparcv9_random,.-_sparcv9_vis3_probe
+
 .global	OPENSSL_cleanse
 .align	32
 OPENSSL_cleanse:
@@ -397,6 +428,102 @@ OPENSSL_cleanse:
 .type	OPENSSL_cleanse,#function
 .size	OPENSSL_cleanse,.-OPENSSL_cleanse
 
+.global	_sparcv9_vis1_instrument_bus
+.align	8
+_sparcv9_vis1_instrument_bus:
+	mov	%o1,%o3					! save cnt
+	.word	0x99410000	!rd	%tick,%o4	! tick
+	mov	%o4,%o5					! lasttick = tick
+	set	0,%g4					! diff
+
+	andn	%o0,63,%g1
+	.word	0xc1985e00	!ldda	[%g1]0xf0,%f0	! block load
+	.word	0x8143e040	!membar	#Sync
+	.word	0xc1b85c00	!stda	%f0,[%g1]0xe0	! block store and commit
+	.word	0x8143e040	!membar	#Sync
+	ld	[%o0],%o4
+	add	%o4,%g4,%g4
+	.word	0xc9e2100c	!cas	[%o0],%o4,%g4
+
+.Loop:	.word	0x99410000	!rd	%tick,%o4
+	sub	%o4,%o5,%g4				! diff=tick-lasttick
+	mov	%o4,%o5					! lasttick=tick
+
+	andn	%o0,63,%g1
+	.word	0xc1985e00	!ldda	[%g1]0xf0,%f0	! block load
+	.word	0x8143e040	!membar	#Sync
+	.word	0xc1b85c00	!stda	%f0,[%g1]0xe0	! block store and commit
+	.word	0x8143e040	!membar	#Sync
+	ld	[%o0],%o4
+	add	%o4,%g4,%g4
+	.word	0xc9e2100c	!cas	[%o0],%o4,%g4
+	subcc	%o1,1,%o1				! --$cnt
+	bnz	.Loop
+	add	%o0,4,%o0				! ++$out
+
+	retl
+	mov	%o3,%o0
+.type	_sparcv9_vis1_instrument_bus,#function
+.size	_sparcv9_vis1_instrument_bus,.-_sparcv9_vis1_instrument_bus
+
+.global	_sparcv9_vis1_instrument_bus2
+.align	8
+_sparcv9_vis1_instrument_bus2:
+	mov	%o1,%o3					! save cnt
+	sll	%o1,2,%o1				! cnt*=4
+
+	.word	0x99410000	!rd	%tick,%o4	! tick
+	mov	%o4,%o5					! lasttick = tick
+	set	0,%g4					! diff
+
+	andn	%o0,63,%g1
+	.word	0xc1985e00	!ldda	[%g1]0xf0,%f0	! block load
+	.word	0x8143e040	!membar	#Sync
+	.word	0xc1b85c00	!stda	%f0,[%g1]0xe0	! block store and commit
+	.word	0x8143e040	!membar	#Sync
+	ld	[%o0],%o4
+	add	%o4,%g4,%g4
+	.word	0xc9e2100c	!cas	[%o0],%o4,%g4
+
+	.word	0x99410000	!rd	%tick,%o4	! tick
+	sub	%o4,%o5,%g4				! diff=tick-lasttick
+	mov	%o4,%o5					! lasttick=tick
+	mov	%g4,%g5					! lastdiff=diff
+.Loop2:
+	andn	%o0,63,%g1
+	.word	0xc1985e00	!ldda	[%g1]0xf0,%f0	! block load
+	.word	0x8143e040	!membar	#Sync
+	.word	0xc1b85c00	!stda	%f0,[%g1]0xe0	! block store and commit
+	.word	0x8143e040	!membar	#Sync
+	ld	[%o0],%o4
+	add	%o4,%g4,%g4
+	.word	0xc9e2100c	!cas	[%o0],%o4,%g4
+
+	subcc	%o2,1,%o2				! --max
+	bz	.Ldone2
+	nop
+
+	.word	0x99410000	!rd	%tick,%o4	! tick
+	sub	%o4,%o5,%g4				! diff=tick-lasttick
+	mov	%o4,%o5					! lasttick=tick
+	cmp	%g4,%g5
+	mov	%g4,%g5					! lastdiff=diff
+
+	.word	0x83408000	!rd	%ccr,%g1
+	and	%g1,4,%g1				! isolate zero flag
+	xor	%g1,4,%g1				! flip zero flag
+
+	subcc	%o1,%g1,%o1				! conditional --$cnt
+	bnz	.Loop2
+	add	%o0,%g1,%o0				! conditional ++$out
+
+.Ldone2:
+	srl	%o1,2,%o1
+	retl
+	sub	%o3,%o1,%o0
+.type	_sparcv9_vis1_instrument_bus2,#function
+.size	_sparcv9_vis1_instrument_bus2,.-_sparcv9_vis1_instrument_bus2
+
 .section	".init",#alloc,#execinstr
 	call	OPENSSL_cpuid_setup
 	nop

crypto/sparcv9cap.c

@@ -4,15 +4,15 @@
 #include <setjmp.h>
 #include <signal.h>
 #include <sys/time.h>
+#include <unistd.h>
 #include <openssl/bn.h>
 
-#define SPARCV9_TICK_PRIVILEGED	(1<<0)
+#include "sparc_arch.h"
-#define SPARCV9_PREFER_FPU	(1<<1)
-#define SPARCV9_VIS1		(1<<2)
-#define SPARCV9_VIS2		(1<<3)	/* reserved */
-#define SPARCV9_FMADD		(1<<4)	/* reserved for SPARC64 V */
 
-static int OPENSSL_sparcv9cap_P=SPARCV9_TICK_PRIVILEGED;
+#if defined(__GNUC__) && defined(__linux)
+__attribute__((visibility("hidden")))
+#endif
+unsigned int OPENSSL_sparcv9cap_P[2]={SPARCV9_TICK_PRIVILEGED,0};
 
 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)
 	{
@@ -20,7 +20,7 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_U
 	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 (num>=8 && !(num&1) &&
-	    (OPENSSL_sparcv9cap_P&(SPARCV9_PREFER_FPU|SPARCV9_VIS1)) ==
+	    (OPENSSL_sparcv9cap_P[0]&(SPARCV9_PREFER_FPU|SPARCV9_VIS1)) ==
 		(SPARCV9_PREFER_FPU|SPARCV9_VIS1))
 		return bn_mul_mont_fpu(rp,ap,bp,np,n0,num);
 	else
@@ -32,10 +32,15 @@ void		_sparcv9_vis1_probe(void);
 unsigned long	_sparcv9_vis1_instrument(void);
 void		_sparcv9_vis2_probe(void);
 void		_sparcv9_fmadd_probe(void);
+unsigned long	_sparcv9_rdcfr(void);
+void		_sparcv9_vis3_probe(void);
+unsigned long	_sparcv9_random(void);
+size_t 		_sparcv9_vis1_instrument_bus(unsigned int *,size_t);
+size_t		_sparcv9_vis1_instrument_bus2(unsigned int *,size_t,size_t);
 
 unsigned long OPENSSL_rdtsc(void)
 	{
-	if (OPENSSL_sparcv9cap_P&SPARCV9_TICK_PRIVILEGED)
+	if (OPENSSL_sparcv9cap_P[0]&SPARCV9_TICK_PRIVILEGED)
 #if defined(__sun) && defined(__SVR4)
 		return gethrtime();
 #else
@@ -45,6 +50,24 @@ unsigned long OPENSSL_rdtsc(void)
 		return _sparcv9_rdtick();
 	}
 
+size_t OPENSSL_instrument_bus(unsigned int *out,size_t cnt)
+	{
+	if ((OPENSSL_sparcv9cap_P[0]&(SPARCV9_TICK_PRIVILEGED|SPARCV9_BLK)) ==
+			SPARCV9_BLK)
+		return _sparcv9_vis1_instrument_bus(out,cnt);
+	else
+		return 0;
+	}
+
+size_t OPENSSL_instrument_bus2(unsigned int *out,size_t cnt,size_t max)
+	{
+	if ((OPENSSL_sparcv9cap_P[0]&(SPARCV9_TICK_PRIVILEGED|SPARCV9_BLK)) ==
+			SPARCV9_BLK)
+		return _sparcv9_vis1_instrument_bus2(out,cnt,max);
+	else
+		return 0;
+	}
+
 #if 0 && defined(__sun) && defined(__SVR4)
 /* This code path is disabled, because of incompatibility of
  * libdevinfo.so.1 and libmalloc.so.1 (see below for details)
@@ -69,18 +92,18 @@ static int walk_nodename(di_node_t node, di_node_name_t di_node_name)
 	if (!strcmp (name,"SUNW,UltraSPARC") ||
 	    !strncmp(name,"SUNW,UltraSPARC-I",17))  /* covers II,III,IV */
 		{
-		OPENSSL_sparcv9cap_P |= SPARCV9_PREFER_FPU|SPARCV9_VIS1;
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU|SPARCV9_VIS1;
 
 		/* %tick is privileged only on UltraSPARC-I/II, but not IIe */
 		if (name[14]!='\0' && name[17]!='\0' && name[18]!='\0')
-			OPENSSL_sparcv9cap_P &= ~SPARCV9_TICK_PRIVILEGED;
+			OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
 
 		return DI_WALK_TERMINATE;
 		}
 	/* This is expected to catch remaining UltraSPARCs, such as T1 */
 	else if (!strncmp(name,"SUNW,UltraSPARC",15))
 		{
-		OPENSSL_sparcv9cap_P &= ~SPARCV9_TICK_PRIVILEGED;
+		OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
 
 		return DI_WALK_TERMINATE;
 		}
@@ -99,7 +122,7 @@ void OPENSSL_cpuid_setup(void)
 
 	if ((e=getenv("OPENSSL_sparcv9cap")))
 		{
-		OPENSSL_sparcv9cap_P=strtoul(e,NULL,0);
+		OPENSSL_sparcv9cap_P[0]=strtoul(e,NULL,0);
 		return;
 		}
 
@@ -107,17 +130,17 @@ void OPENSSL_cpuid_setup(void)
 		{
 		if (strcmp(si,"sun4v"))
 			/* FPU is preferred for all CPUs, but US-T1/2 */
-			OPENSSL_sparcv9cap_P |= SPARCV9_PREFER_FPU;
+			OPENSSL_sparcv9cap_P[0] |= SPARCV9_PREFER_FPU;
 		}
 
 	if (sysinfo(SI_ISALIST,si,sizeof(si))>0)
 		{
 		if (strstr(si,"+vis"))
-			OPENSSL_sparcv9cap_P |= SPARCV9_VIS1;
+			OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1|SPARCV9_BLK;
 		if (strstr(si,"+vis2"))
 			{
-			OPENSSL_sparcv9cap_P |= SPARCV9_VIS2;
+			OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2;
-			OPENSSL_sparcv9cap_P &= ~SPARCV9_TICK_PRIVILEGED;
+			OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
 			return;
 			}
 		}
@@ -177,12 +200,14 @@ void OPENSSL_cpuid_setup(void)
  
 	if ((e=getenv("OPENSSL_sparcv9cap")))
 		{
-		OPENSSL_sparcv9cap_P=strtoul(e,NULL,0);
+		OPENSSL_sparcv9cap_P[0]=strtoul(e,NULL,0);
+		if ((e=strchr(e,':')))
+			OPENSSL_sparcv9cap_P[1]=strtoul(e+1,NULL,0);
 		return;
 		}
 
 	/* Initial value, fits UltraSPARC-I&II... */
-	OPENSSL_sparcv9cap_P = SPARCV9_PREFER_FPU|SPARCV9_TICK_PRIVILEGED;
+	OPENSSL_sparcv9cap_P[0] = SPARCV9_PREFER_FPU|SPARCV9_TICK_PRIVILEGED;
 
 	sigfillset(&all_masked);
 	sigdelset(&all_masked,SIGILL);
@@ -205,33 +230,73 @@ void OPENSSL_cpuid_setup(void)
 	if (sigsetjmp(common_jmp,1) == 0)
 		{
 		_sparcv9_rdtick();
-		OPENSSL_sparcv9cap_P &= ~SPARCV9_TICK_PRIVILEGED;
+		OPENSSL_sparcv9cap_P[0] &= ~SPARCV9_TICK_PRIVILEGED;
 		}
 
 	if (sigsetjmp(common_jmp,1) == 0)
 		{
 		_sparcv9_vis1_probe();
-		OPENSSL_sparcv9cap_P |= SPARCV9_VIS1;
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS1|SPARCV9_BLK;
 		/* detect UltraSPARC-Tx, see sparccpud.S for details... */
 		if (_sparcv9_vis1_instrument() >= 12)
-			OPENSSL_sparcv9cap_P &= ~(SPARCV9_VIS1|SPARCV9_PREFER_FPU);
+			OPENSSL_sparcv9cap_P[0] &= ~(SPARCV9_VIS1|SPARCV9_PREFER_FPU);
 		else
 			{
 			_sparcv9_vis2_probe();
-			OPENSSL_sparcv9cap_P |= SPARCV9_VIS2;
+			OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS2;
 			}
 		}
 
 	if (sigsetjmp(common_jmp,1) == 0)
 		{
 		_sparcv9_fmadd_probe();
-		OPENSSL_sparcv9cap_P |= SPARCV9_FMADD;
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_FMADD;
+		}
+
+	/*
+	 * VIS3 flag is tested independently from VIS1, unlike VIS2 that is,
+	 * because VIS3 defines even integer instructions.
+	 */
+	if (sigsetjmp(common_jmp,1) == 0)
+		{
+		_sparcv9_vis3_probe();
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_VIS3;
+		}
+
+	if (sigsetjmp(common_jmp,1) == 0)
+		{
+		(void)_sparcv9_random();
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_RANDOM;
+		}
+
+	/*
+	 * In wait for better solution _sparcv9_rdcfr is masked by
+	 * VIS3 flag, because it goes to uninterruptable endless
+	 * loop on UltraSPARC II running Solaris. Things might be
+	 * different on Linux...
+	 */
+	if ((OPENSSL_sparcv9cap_P[0]&SPARCV9_VIS3) &&
+	    sigsetjmp(common_jmp,1) == 0)
+		{
+		OPENSSL_sparcv9cap_P[1] = (unsigned int)_sparcv9_rdcfr();
 		}
 
 	sigaction(SIGBUS,&bus_oact,NULL);
 	sigaction(SIGILL,&ill_oact,NULL);
 
 	sigprocmask(SIG_SETMASK,&oset,NULL);
+
+	if (sizeof(size_t)==8)
+		OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK;
+#ifdef __linux
+	else
+		{
+		int ret = syscall(340);
+
+		if (ret>=0 && ret&1)
+			OPENSSL_sparcv9cap_P[0] |= SPARCV9_64BIT_STACK;
+		}
+#endif
 	}
 
 #endif