From 4b2603e46ca98be6036ca2ebf83f31e793047ed9 Mon Sep 17 00:00:00 2001 From: Andy Polyakov Date: Sun, 5 Sep 2010 19:41:41 +0000 Subject: [PATCH] sparcv9cap.c: disengange Solaris-specific CPU detection routine in favour of unified procedure relying on SIGILL. PR: 2321 --- crypto/sparccpuid.S | 42 ++++++++++++++++++++----- crypto/sparcv9cap.c | 74 ++++++++++++++++++++++++++++++--------------- 2 files changed, 83 insertions(+), 33 deletions(-) diff --git a/crypto/sparccpuid.S b/crypto/sparccpuid.S index 1e0756a173..ae61f7f5ce 100644 --- a/crypto/sparccpuid.S +++ b/crypto/sparccpuid.S @@ -232,13 +232,23 @@ _sparcv9_rdtick: .type _sparcv9_rdtick,#function .size _sparcv9_rdtick,.-_sparcv9_rdtick +.global _sparcv9_vis1_probe +.align 8 +_sparcv9_vis1_probe: + .word 0x81b00d80 !fxor %f0,%f0,%f0 + add %sp,BIAS+2,%o1 + retl + .word 0xc19a5a40 !ldda [%o1]ASI_FP16_P,%f0 +.type _sparcv9_vis1_probe,#function +.size _sparcv9_vis1_probe,.-_sparcv9_vis1_probe + ! Probe and instrument VIS1 instruction. Output is number of cycles it ! takes to execute rdtick and pair of VIS1 instructions. US-Tx VIS unit ! is slow (documented to be 6 cycles on T2) and the core is in-order ! single-issue, it should be possible to distinguish Tx reliably... ! Observed return values are: ! -! UltraSPARC IIi 7 +! UltraSPARC IIe 7 ! UltraSPARC III 7 ! UltraSPARC T1 24 ! @@ -247,9 +257,9 @@ _sparcv9_rdtick: ! It would be possible to detect specifically US-T1 by instrumenting ! fmul8ulx16, which is emulated on T1 and as such accounts for quite ! a lot of %tick-s, couple of thousand on Linux... -.global _sparcv9_vis1_probe +.global _sparcv9_vis1_instrument .align 8 -_sparcv9_vis1_probe: +_sparcv9_vis1_instrument: .word 0x91410000 !rd %tick,%o0 .word 0x81b00d80 !fxor %f0,%f0,%f0 .word 0x85b08d82 !fxor %f2,%f2,%f2 @@ -281,12 +291,28 @@ _sparcv9_vis1_probe: .word 0x38680002 !bgu,a %xcc,.+8 mov %o3,%o0 - ! check for ASI_FP16_P is redundant... - add %sp,BIAS+2,%o1 retl - .word 0xc19a5a40 !ldda [%o1]ASI_FP16_P,%f0 -.type _sparcv9_vis1_probe,#function -.size _sparcv9_vis1_probe,.-_sparcv9_vis1_probe + nop +.type _sparcv9_vis1_instrument,#function +.size _sparcv9_vis1_instrument,.-_sparcv9_vis1_instrument + +.global _sparcv9_vis2_probe +.align 8 +_sparcv9_vis2_probe: + retl + .word 0x81b00980 !bshuffle %f0,%f0,%f0 +.type _sparcv9_vis2_probe,#function +.size _sparcv9_vis2_probe,.-_sparcv9_vis2_probe + +.global _sparcv9_fmadd_probe +.align 8 +_sparcv9_fmadd_probe: + .word 0x81b00d80 !fxor %f0,%f0,%f0 + .word 0x85b08d82 !fxor %f2,%f2,%f2 + retl + .word 0x81b80440 !fmaddd %f0,%f0,%f2,%f0 +.type _sparcv9_fmadd_probe,#function +.size _sparcv9_fmadd_probe,.-_sparcv9_fmadd_probe .global OPENSSL_cleanse .align 32 diff --git a/crypto/sparcv9cap.c b/crypto/sparcv9cap.c index bb0c8e73a9..ed195ab402 100644 --- a/crypto/sparcv9cap.c +++ b/crypto/sparcv9cap.c @@ -27,7 +27,10 @@ int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_U } unsigned long _sparcv9_rdtick(void); -unsigned long _sparcv9_vis1_probe(void); +void _sparcv9_vis1_probe(void); +unsigned long _sparcv9_vis1_instrument(void); +void _sparcv9_vis2_probe(void); +void _sparcv9_fmadd_probe(void); unsigned long OPENSSL_rdtsc(void) { @@ -41,8 +44,11 @@ unsigned long OPENSSL_rdtsc(void) return _sparcv9_rdtick(); } -#if defined(__sun) && defined(__SVR4) - +#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) + */ +#include #include #include #include @@ -114,7 +120,21 @@ void OPENSSL_cpuid_setup(void) return; } } - +#ifdef M_KEEP + /* + * Solaris libdevinfo.so.1 is effectively incomatible with + * libmalloc.so.1. Specifically, if application is linked with + * -lmalloc, it crashes upon startup with SIGSEGV in + * free(3LIBMALLOC) called by di_fini. Prior call to + * mallopt(M_KEEP,0) somehow helps... But not always... + */ + if ((h = dlopen(NULL,RTLD_LAZY))) + { + union { void *p; int (*f)(int,int); } sym; + if ((sym.p = dlsym(h,"mallopt"))) (*sym.f)(M_KEEP,0); + dlclose(h); + } +#endif if ((h = dlopen("libdevinfo.so.1",RTLD_LAZY))) do { di_init_t di_init; @@ -150,6 +170,10 @@ void OPENSSL_cpuid_setup(void) struct sigaction common_act,ill_oact,bus_oact; sigset_t all_masked,oset; int sig; + static int trigger=0; + + if (trigger) return; + trigger=1; if ((e=getenv("OPENSSL_sparcv9cap"))) { @@ -157,8 +181,8 @@ void OPENSSL_cpuid_setup(void) return; } - /* For now we assume that the rest supports UltraSPARC-I* only */ - OPENSSL_sparcv9cap_P |= SPARCV9_PREFER_FPU|SPARCV9_VIS1; + /* Initial value, fits UltraSPARC-I&II... */ + OPENSSL_sparcv9cap_P = SPARCV9_PREFER_FPU|SPARCV9_TICK_PRIVILEGED; sigfillset(&all_masked); sigdelset(&all_masked,SIGILL); @@ -176,34 +200,34 @@ void OPENSSL_cpuid_setup(void) common_act.sa_mask = all_masked; sigaction(SIGILL,&common_act,&ill_oact); - if (sigsetjmp(common_jmp,0) == 0) + sigaction(SIGBUS,&common_act,&bus_oact);/* T1 fails 16-bit ldda [on Linux] */ + + if (sigsetjmp(common_jmp,1) == 0) { _sparcv9_rdtick(); OPENSSL_sparcv9cap_P &= ~SPARCV9_TICK_PRIVILEGED; } - else - { - /* This happens on US-I&II, which have working VIS1 - * and fast FPU... In other words we are done... */ - OPENSSL_sparcv9cap_P |= SPARCV9_TICK_PRIVILEGED; - sigaction(SIGILL,&ill_oact,NULL); - sigprocmask(SIG_SETMASK,&oset,NULL); - return; - } - sigaction(SIGILL,&ill_oact,NULL); - sigaction(SIGILL,&common_act,&ill_oact); - sigaction(SIGBUS,&common_act,&bus_oact);/* T1 fails 16-bit ldda */ - if ((sig=sigsetjmp(common_jmp,0)) == 0) + if (sigsetjmp(common_jmp,1) == 0) { - /* see sparccpud.S for details... */ - if (_sparcv9_vis1_probe() >= 12) - OPENSSL_sparcv9cap_P &= ~SPARCV9_VIS1; + _sparcv9_vis1_probe(); + OPENSSL_sparcv9cap_P |= SPARCV9_VIS1; + /* detect UltraSPARC-Tx, see sparccpud.S for details... */ + if (_sparcv9_vis1_instrument() >= 12) + OPENSSL_sparcv9cap_P &= ~(SPARCV9_VIS1|SPARCV9_PREFER_FPU); + else + { + _sparcv9_vis2_probe(); + OPENSSL_sparcv9cap_P |= SPARCV9_VIS2; + } } - else + + if (sigsetjmp(common_jmp,1) == 0) { - OPENSSL_sparcv9cap_P &= ~SPARCV9_VIS1; + _sparcv9_fmadd_probe(); + OPENSSL_sparcv9cap_P |= SPARCV9_FMADD; } + sigaction(SIGBUS,&bus_oact,NULL); sigaction(SIGILL,&ill_oact,NULL); -- 2.34.1