/*--------------------------------------------------------------------*/
/*--- Machine-related stuff.                           m_machine.c ---*/
/*--------------------------------------------------------------------*/

/*
   This file is part of Valgrind, a dynamic binary instrumentation
   framework.

   Copyright (C) 2000-2010 Julian Seward 
      jseward@acm.org

   This program is free software; you can redistribute it and/or
   modify it under the terms of the GNU General Public License as
   published by the Free Software Foundation; either version 2 of the
   License, or (at your option) any later version.

   This program is distributed in the hope that it will be useful, but
   WITHOUT ANY WARRANTY; without even the implied warranty of
   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
   General Public License for more details.

   You should have received a copy of the GNU General Public License
   along with this program; if not, write to the Free Software
   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
   02111-1307, USA.

   The GNU General Public License is contained in the file COPYING.
*/

#include "pub_core_basics.h"
#include "pub_core_vki.h"
#include "pub_core_libcsetjmp.h"   // setjmp facilities
#include "pub_core_threadstate.h"
#include "pub_core_libcassert.h"
#include "pub_core_libcbase.h"
#include "pub_core_libcfile.h"
#include "pub_core_mallocfree.h"
#include "pub_core_machine.h"
#include "pub_core_cpuid.h"
#include "pub_core_libcsignal.h"   // for ppc32 messing with SIGILL and SIGFPE
#include "pub_core_debuglog.h"


#define INSTR_PTR(regs)    ((regs).vex.VG_INSTR_PTR)
#define STACK_PTR(regs)    ((regs).vex.VG_STACK_PTR)
#define FRAME_PTR(regs)    ((regs).vex.VG_FRAME_PTR)

Addr VG_(get_IP) ( ThreadId tid ) {
   return INSTR_PTR( VG_(threads)[tid].arch );
}
Addr VG_(get_SP) ( ThreadId tid ) {
   return STACK_PTR( VG_(threads)[tid].arch );
}
Addr VG_(get_FP) ( ThreadId tid ) {
   return FRAME_PTR( VG_(threads)[tid].arch );
}

void VG_(set_IP) ( ThreadId tid, Addr ip ) {
   INSTR_PTR( VG_(threads)[tid].arch ) = ip;
}
void VG_(set_SP) ( ThreadId tid, Addr sp ) {
   STACK_PTR( VG_(threads)[tid].arch ) = sp;
}

void VG_(get_UnwindStartRegs) ( /*OUT*/UnwindStartRegs* regs,
                                ThreadId tid )
{
#  if defined(VGA_x86)
   regs->r_pc = (ULong)VG_(threads)[tid].arch.vex.guest_EIP;
   regs->r_sp = (ULong)VG_(threads)[tid].arch.vex.guest_ESP;
   regs->misc.X86.r_ebp
      = VG_(threads)[tid].arch.vex.guest_EBP;
#  elif defined(VGA_amd64)
   regs->r_pc = VG_(threads)[tid].arch.vex.guest_RIP;
   regs->r_sp = VG_(threads)[tid].arch.vex.guest_RSP;
   regs->misc.AMD64.r_rbp
      = VG_(threads)[tid].arch.vex.guest_RBP;
#  elif defined(VGA_ppc32)
   regs->r_pc = (ULong)VG_(threads)[tid].arch.vex.guest_CIA;
   regs->r_sp = (ULong)VG_(threads)[tid].arch.vex.guest_GPR1;
   regs->misc.PPC32.r_lr
      = VG_(threads)[tid].arch.vex.guest_LR;
#  elif defined(VGA_ppc64)
   regs->r_pc = VG_(threads)[tid].arch.vex.guest_CIA;
   regs->r_sp = VG_(threads)[tid].arch.vex.guest_GPR1;
   regs->misc.PPC64.r_lr
      = VG_(threads)[tid].arch.vex.guest_LR;
#  elif defined(VGA_arm)
   regs->r_pc = (ULong)VG_(threads)[tid].arch.vex.guest_R15T;
   regs->r_sp = (ULong)VG_(threads)[tid].arch.vex.guest_R13;
   regs->misc.ARM.r14
      = VG_(threads)[tid].arch.vex.guest_R14;
   regs->misc.ARM.r12
      = VG_(threads)[tid].arch.vex.guest_R12;
   regs->misc.ARM.r11
      = VG_(threads)[tid].arch.vex.guest_R11;
   regs->misc.ARM.r7
      = VG_(threads)[tid].arch.vex.guest_R7;
#  elif defined(VGA_s390x)
   regs->r_pc = (ULong)VG_(threads)[tid].arch.vex.guest_IA;
   regs->r_sp = (ULong)VG_(threads)[tid].arch.vex.guest_SP;
   regs->misc.S390X.r_fp
      = VG_(threads)[tid].arch.vex.guest_r11;
   regs->misc.S390X.r_lr
      = VG_(threads)[tid].arch.vex.guest_r14;
#  else
#    error "Unknown arch"
#  endif
}


void VG_(set_syscall_return_shadows) ( ThreadId tid,
                                       /* shadow vals for the result */
                                       UWord s1res, UWord s2res,
                                       /* shadow vals for the error val */
                                       UWord s1err, UWord s2err )
{
#  if defined(VGP_x86_linux)
   VG_(threads)[tid].arch.vex_shadow1.guest_EAX = s1res;
   VG_(threads)[tid].arch.vex_shadow2.guest_EAX = s2res;
#  elif defined(VGP_amd64_linux)
   VG_(threads)[tid].arch.vex_shadow1.guest_RAX = s1res;
   VG_(threads)[tid].arch.vex_shadow2.guest_RAX = s2res;
#  elif defined(VGP_ppc32_linux) || defined(VGP_ppc64_linux)
   VG_(threads)[tid].arch.vex_shadow1.guest_GPR3 = s1res;
   VG_(threads)[tid].arch.vex_shadow2.guest_GPR3 = s2res;
#  elif defined(VGP_arm_linux)
   VG_(threads)[tid].arch.vex_shadow1.guest_R0 = s1res;
   VG_(threads)[tid].arch.vex_shadow2.guest_R0 = s2res;
#  elif defined(VGO_darwin)
   // GrP fixme darwin syscalls may return more values (2 registers plus error)
#  elif defined(VGP_s390x_linux)
   VG_(threads)[tid].arch.vex_shadow1.guest_r2 = s1res;
   VG_(threads)[tid].arch.vex_shadow2.guest_r2 = s2res;
#  else
#    error "Unknown plat"
#  endif
}

void
VG_(get_shadow_regs_area) ( ThreadId tid, 
                            /*DST*/UChar* dst,
                            /*SRC*/Int shadowNo, PtrdiffT offset, SizeT size )
{
   void*        src;
   ThreadState* tst;
   vg_assert(shadowNo == 0 || shadowNo == 1 || shadowNo == 2);
   vg_assert(VG_(is_valid_tid)(tid));
   // Bounds check
   vg_assert(0 <= offset && offset < sizeof(VexGuestArchState));
   vg_assert(offset + size <= sizeof(VexGuestArchState));
   // Copy
   tst = & VG_(threads)[tid];
   src = NULL;
   switch (shadowNo) {
      case 0: src = (void*)(((Addr)&(tst->arch.vex)) + offset); break;
      case 1: src = (void*)(((Addr)&(tst->arch.vex_shadow1)) + offset); break;
      case 2: src = (void*)(((Addr)&(tst->arch.vex_shadow2)) + offset); break;
   }
   tl_assert(src != NULL);
   VG_(memcpy)( dst, src, size);
}

void
VG_(set_shadow_regs_area) ( ThreadId tid, 
                            /*DST*/Int shadowNo, PtrdiffT offset, SizeT size,
                            /*SRC*/const UChar* src )
{
   void*        dst;
   ThreadState* tst;
   vg_assert(shadowNo == 0 || shadowNo == 1 || shadowNo == 2);
   vg_assert(VG_(is_valid_tid)(tid));
   // Bounds check
   vg_assert(0 <= offset && offset < sizeof(VexGuestArchState));
   vg_assert(offset + size <= sizeof(VexGuestArchState));
   // Copy
   tst = & VG_(threads)[tid];
   dst = NULL;
   switch (shadowNo) {
      case 0: dst = (void*)(((Addr)&(tst->arch.vex)) + offset); break;
      case 1: dst = (void*)(((Addr)&(tst->arch.vex_shadow1)) + offset); break;
      case 2: dst = (void*)(((Addr)&(tst->arch.vex_shadow2)) + offset); break;
   }
   tl_assert(dst != NULL);
   VG_(memcpy)( dst, src, size);
}


static void apply_to_GPs_of_tid(VexGuestArchState* vex, void (*f)(Addr))
{
#if defined(VGA_x86)
   (*f)(vex->guest_EAX);
   (*f)(vex->guest_ECX);
   (*f)(vex->guest_EDX);
   (*f)(vex->guest_EBX);
   (*f)(vex->guest_ESI);
   (*f)(vex->guest_EDI);
   (*f)(vex->guest_ESP);
   (*f)(vex->guest_EBP);
#elif defined(VGA_amd64)
   (*f)(vex->guest_RAX);
   (*f)(vex->guest_RCX);
   (*f)(vex->guest_RDX);
   (*f)(vex->guest_RBX);
   (*f)(vex->guest_RSI);
   (*f)(vex->guest_RDI);
   (*f)(vex->guest_RSP);
   (*f)(vex->guest_RBP);
   (*f)(vex->guest_R8);
   (*f)(vex->guest_R9);
   (*f)(vex->guest_R10);
   (*f)(vex->guest_R11);
   (*f)(vex->guest_R12);
   (*f)(vex->guest_R13);
   (*f)(vex->guest_R14);
   (*f)(vex->guest_R15);
#elif defined(VGA_ppc32) || defined(VGA_ppc64)
   (*f)(vex->guest_GPR0);
   (*f)(vex->guest_GPR1);
   (*f)(vex->guest_GPR2);
   (*f)(vex->guest_GPR3);
   (*f)(vex->guest_GPR4);
   (*f)(vex->guest_GPR5);
   (*f)(vex->guest_GPR6);
   (*f)(vex->guest_GPR7);
   (*f)(vex->guest_GPR8);
   (*f)(vex->guest_GPR9);
   (*f)(vex->guest_GPR10);
   (*f)(vex->guest_GPR11);
   (*f)(vex->guest_GPR12);
   (*f)(vex->guest_GPR13);
   (*f)(vex->guest_GPR14);
   (*f)(vex->guest_GPR15);
   (*f)(vex->guest_GPR16);
   (*f)(vex->guest_GPR17);
   (*f)(vex->guest_GPR18);
   (*f)(vex->guest_GPR19);
   (*f)(vex->guest_GPR20);
   (*f)(vex->guest_GPR21);
   (*f)(vex->guest_GPR22);
   (*f)(vex->guest_GPR23);
   (*f)(vex->guest_GPR24);
   (*f)(vex->guest_GPR25);
   (*f)(vex->guest_GPR26);
   (*f)(vex->guest_GPR27);
   (*f)(vex->guest_GPR28);
   (*f)(vex->guest_GPR29);
   (*f)(vex->guest_GPR30);
   (*f)(vex->guest_GPR31);
   (*f)(vex->guest_CTR);
   (*f)(vex->guest_LR);
#elif defined(VGA_arm)
   (*f)(vex->guest_R0);
   (*f)(vex->guest_R1);
   (*f)(vex->guest_R2);
   (*f)(vex->guest_R3);
   (*f)(vex->guest_R4);
   (*f)(vex->guest_R5);
   (*f)(vex->guest_R6);
   (*f)(vex->guest_R8);
   (*f)(vex->guest_R9);
   (*f)(vex->guest_R10);
   (*f)(vex->guest_R11);
   (*f)(vex->guest_R12);
   (*f)(vex->guest_R13);
   (*f)(vex->guest_R14);
#elif defined(VGA_s390x)
   (*f)(vex->guest_r0);
   (*f)(vex->guest_r1);
   (*f)(vex->guest_r2);
   (*f)(vex->guest_r3);
   (*f)(vex->guest_r4);
   (*f)(vex->guest_r5);
   (*f)(vex->guest_r6);
   (*f)(vex->guest_r7);
   (*f)(vex->guest_r8);
   (*f)(vex->guest_r9);
   (*f)(vex->guest_r10);
   (*f)(vex->guest_r11);
   (*f)(vex->guest_r12);
   (*f)(vex->guest_r13);
   (*f)(vex->guest_r14);
   (*f)(vex->guest_r15);
#else
#  error Unknown arch
#endif
}


void VG_(apply_to_GP_regs)(void (*f)(UWord))
{
   ThreadId tid;

   for (tid = 1; tid < VG_N_THREADS; tid++) {
      if (VG_(is_valid_tid)(tid)) {
         ThreadState* tst = VG_(get_ThreadState)(tid);
         apply_to_GPs_of_tid(&(tst->arch.vex), f);
      }
   }
}

void VG_(thread_stack_reset_iter)(/*OUT*/ThreadId* tid)
{
   *tid = (ThreadId)(-1);
}

Bool VG_(thread_stack_next)(/*MOD*/ThreadId* tid,
                            /*OUT*/Addr* stack_min, 
                            /*OUT*/Addr* stack_max)
{
   ThreadId i;
   for (i = (*tid)+1; i < VG_N_THREADS; i++) {
      if (i == VG_INVALID_THREADID)
         continue;
      if (VG_(threads)[i].status != VgTs_Empty) {
         *tid       = i;
         *stack_min = VG_(get_SP)(i);
         *stack_max = VG_(threads)[i].client_stack_highest_word;
         return True;
      }
   }
   return False;
}

Addr VG_(thread_get_stack_max)(ThreadId tid)
{
   vg_assert(0 <= tid && tid < VG_N_THREADS && tid != VG_INVALID_THREADID);
   vg_assert(VG_(threads)[tid].status != VgTs_Empty);
   return VG_(threads)[tid].client_stack_highest_word;
}

SizeT VG_(thread_get_stack_size)(ThreadId tid)
{
   vg_assert(0 <= tid && tid < VG_N_THREADS && tid != VG_INVALID_THREADID);
   vg_assert(VG_(threads)[tid].status != VgTs_Empty);
   return VG_(threads)[tid].client_stack_szB;
}

Addr VG_(thread_get_altstack_min)(ThreadId tid)
{
   vg_assert(0 <= tid && tid < VG_N_THREADS && tid != VG_INVALID_THREADID);
   vg_assert(VG_(threads)[tid].status != VgTs_Empty);
   return (Addr)VG_(threads)[tid].altstack.ss_sp;
}

SizeT VG_(thread_get_altstack_size)(ThreadId tid)
{
   vg_assert(0 <= tid && tid < VG_N_THREADS && tid != VG_INVALID_THREADID);
   vg_assert(VG_(threads)[tid].status != VgTs_Empty);
   return VG_(threads)[tid].altstack.ss_size;
}

//-------------------------------------------------------------
/* Details about the capabilities of the underlying (host) CPU.  These
   details are acquired by (1) enquiring with the CPU at startup, or
   (2) from the AT_SYSINFO entries the kernel gave us (ppc32 cache
   line size).  It's a bit nasty in the sense that there's no obvious
   way to stop uses of some of this info before it's ready to go.

   Current dependencies are:

   x86:   initially:  call VG_(machine_get_hwcaps)

          then safe to use VG_(machine_get_VexArchInfo) 
                       and VG_(machine_x86_have_mxcsr)
   -------------
   amd64: initially:  call VG_(machine_get_hwcaps)

          then safe to use VG_(machine_get_VexArchInfo) 
   -------------
   ppc32: initially:  call VG_(machine_get_hwcaps)
                      call VG_(machine_ppc32_set_clszB)

          then safe to use VG_(machine_get_VexArchInfo) 
                       and VG_(machine_ppc32_has_FP)
                       and VG_(machine_ppc32_has_VMX)
   -------------
   ppc64: initially:  call VG_(machine_get_hwcaps)
                      call VG_(machine_ppc64_set_clszB)

          then safe to use VG_(machine_get_VexArchInfo) 
                       and VG_(machine_ppc64_has_VMX)

   -------------
   s390x: initially:  call VG_(machine_get_hwcaps)

          then safe to use VG_(machine_get_VexArchInfo)

   VG_(machine_get_hwcaps) may use signals (although it attempts to
   leave signal state unchanged) and therefore should only be
   called before m_main sets up the client's signal state.
*/

/* --------- State --------- */
static Bool hwcaps_done = False;

/* --- all archs --- */
static VexArch     va;
static VexArchInfo vai;

#if defined(VGA_x86)
UInt VG_(machine_x86_have_mxcsr) = 0;
#endif
#if defined(VGA_ppc32)
UInt VG_(machine_ppc32_has_FP)  = 0;
UInt VG_(machine_ppc32_has_VMX) = 0;
#endif
#if defined(VGA_ppc64)
ULong VG_(machine_ppc64_has_VMX) = 0;
#endif
#if defined(VGA_arm)
Int VG_(machine_arm_archlevel) = 4;
#endif

/* fixs390: anything for s390x here ? */

/* For hwcaps detection on ppc32/64, s390x, and arm we'll need to do SIGILL
   testing, so we need a VG_MINIMAL_JMP_BUF. */
#if defined(VGA_ppc32) || defined(VGA_ppc64) \
    || defined(VGA_arm) || defined(VGA_s390x)
#include "pub_tool_libcsetjmp.h"
static VG_MINIMAL_JMP_BUF(env_unsup_insn);
static void handler_unsup_insn ( Int x ) {
   VG_MINIMAL_LONGJMP(env_unsup_insn);
}
#endif


/* Helper function for VG_(machine_get_hwcaps), assumes the SIGILL/etc
 * handlers are installed.  Determines the the sizes affected by dcbz
 * and dcbzl instructions and updates the given VexArchInfo structure
 * accordingly.
 *
 * Not very defensive: assumes that as long as the dcbz/dcbzl
 * instructions don't raise a SIGILL, that they will zero an aligned,
 * contiguous block of memory of a sensible size. */
#if defined(VGA_ppc32) || defined(VGA_ppc64)
static void find_ppc_dcbz_sz(VexArchInfo *arch_info)
{
   Int dcbz_szB = 0;
   Int dcbzl_szB;
#  define MAX_DCBZL_SZB (128) /* largest known effect of dcbzl */
   char test_block[4*MAX_DCBZL_SZB];
   char *aligned = test_block;
   Int i;

   /* round up to next max block size, assumes MAX_DCBZL_SZB is pof2 */
   aligned = (char *)(((HWord)aligned + MAX_DCBZL_SZB) & ~(MAX_DCBZL_SZB - 1));
   vg_assert((aligned + MAX_DCBZL_SZB) <= &test_block[sizeof(test_block)]);

   /* dcbz often clears 32B, although sometimes whatever the native cache
    * block size is */
   VG_(memset)(test_block, 0xff, sizeof(test_block));
   __asm__ __volatile__("dcbz 0,%0"
                        : /*out*/
                        : "r" (aligned) /*in*/
                        : "memory" /*clobber*/);
   for (dcbz_szB = 0, i = 0; i < sizeof(test_block); ++i) {
      if (!test_block[i])
         ++dcbz_szB;
   }
   vg_assert(dcbz_szB == 32 || dcbz_szB == 64 || dcbz_szB == 128);

   /* dcbzl clears 128B on G5/PPC970, and usually 32B on other platforms */
   if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
      dcbzl_szB = 0; /* indicates unsupported */
   }
   else {
      VG_(memset)(test_block, 0xff, sizeof(test_block));
      /* some older assemblers won't understand the dcbzl instruction
       * variant, so we directly emit the instruction ourselves */
      __asm__ __volatile__("mr 9, %0 ; .long 0x7C204FEC" /*dcbzl 0,9*/
                           : /*out*/
                           : "r" (aligned) /*in*/
                           : "memory", "r9" /*clobber*/);
      for (dcbzl_szB = 0, i = 0; i < sizeof(test_block); ++i) {
         if (!test_block[i])
            ++dcbzl_szB;
      }
      vg_assert(dcbzl_szB == 32 || dcbzl_szB == 64 || dcbzl_szB == 128);
   }

   arch_info->ppc_dcbz_szB  = dcbz_szB;
   arch_info->ppc_dcbzl_szB = dcbzl_szB;

   VG_(debugLog)(1, "machine", "dcbz_szB=%d dcbzl_szB=%d\n",
                 dcbz_szB, dcbzl_szB);
#  undef MAX_DCBZL_SZB
}
#endif /* defined(VGA_ppc32) || defined(VGA_ppc64) */

#ifdef VGA_s390x

/* Read /proc/cpuinfo. Look for lines like these

   processor 0: version = FF,  identification = 0117C9,  machine = 2064

   and return the machine model or VEX_S390X_MODEL_INVALID on error. */

static UInt VG_(get_machine_model)(void)
{
   static struct model_map {
      HChar name[5];
      UInt  id;
   } model_map[] = {
      { "2064", VEX_S390X_MODEL_Z900 },
      { "2066", VEX_S390X_MODEL_Z800 },
      { "2084", VEX_S390X_MODEL_Z990 },
      { "2086", VEX_S390X_MODEL_Z890 },
      { "2094", VEX_S390X_MODEL_Z9_EC },
      { "2096", VEX_S390X_MODEL_Z9_BC },
      { "2097", VEX_S390X_MODEL_Z10_EC },
      { "2098", VEX_S390X_MODEL_Z10_BC },
      { "2817", VEX_S390X_MODEL_Z196 },
   };

   Int    model, n, fh;
   SysRes fd;
   SizeT  num_bytes, file_buf_size;
   HChar *p, *m, *model_name, *file_buf;

   /* Slurp contents of /proc/cpuinfo into FILE_BUF */
   fd = VG_(open)( "/proc/cpuinfo", 0, VKI_S_IRUSR );
   if ( sr_isError(fd) ) return VEX_S390X_MODEL_INVALID;

   fh  = sr_Res(fd);

   /* Determine the size of /proc/cpuinfo.
      Work around broken-ness in /proc file system implementation.
      fstat returns a zero size for /proc/cpuinfo although it is
      claimed to be a regular file. */
   num_bytes = 0;
   file_buf_size = 1000;
   file_buf = VG_(malloc)("cpuinfo", file_buf_size + 1);
   while (42) {
      n = VG_(read)(fh, file_buf, file_buf_size);
      if (n < 0) break;

      num_bytes += n;
      if (n < file_buf_size) break;  /* reached EOF */
   }

   if (n < 0) num_bytes = 0;   /* read error; ignore contents */

   if (num_bytes > file_buf_size) {
      VG_(free)( file_buf );
      VG_(lseek)( fh, 0, VKI_SEEK_SET );
      file_buf = VG_(malloc)( "cpuinfo", num_bytes + 1 );
      n = VG_(read)( fh, file_buf, num_bytes );
      if (n < 0) num_bytes = 0;
   }

   file_buf[num_bytes] = '\0';
   VG_(close)(fh);

   /* Parse file */
   model = VEX_S390X_MODEL_INVALID;
   for (p = file_buf; *p; ++p) {
      /* Beginning of line */
     if (VG_(strncmp)( p, "processor", sizeof "processor" - 1 ) != 0) continue;

     m = VG_(strstr)( p, "machine" );
     if (m == NULL) continue;

     p = m + sizeof "machine" - 1;
     while ( VG_(isspace)( *p ) || *p == '=') {
       if (*p == '\n') goto next_line;
       ++p;
     }

     model_name = p;
     for (n = 0; n < sizeof model_map / sizeof model_map[0]; ++n) {
       struct model_map *mm = model_map + n;
       SizeT len = VG_(strlen)( mm->name );
       if ( VG_(strncmp)( mm->name, model_name, len ) == 0 &&
            VG_(isspace)( model_name[len] )) {
         if (mm->id < model) model = mm->id;
         p = model_name + len;
         break;
       }
     }
     /* Skip until end-of-line */
     while (*p != '\n')
       ++p;
   next_line: ;
   }

   VG_(free)( file_buf );
   VG_(debugLog)(1, "machine", "model = %s\n", model_map[model].name);

   return model;
}

#endif /* VGA_s390x */

/* Determine what insn set and insn set variant the host has, and
   record it.  To be called once at system startup.  Returns False if
   this a CPU incapable of running Valgrind. */

Bool VG_(machine_get_hwcaps)( void )
{
   vg_assert(hwcaps_done == False);
   hwcaps_done = True;

   // Whack default settings into vai, so that we only need to fill in
   // any interesting bits.
   LibVEX_default_VexArchInfo(&vai);

#if defined(VGA_x86)
   { Bool have_sse1, have_sse2, have_cx8, have_lzcnt;
     UInt eax, ebx, ecx, edx, max_extended;
     UChar vstr[13];
     vstr[0] = 0;

     if (!VG_(has_cpuid)())
        /* we can't do cpuid at all.  Give up. */
        return False;

     VG_(cpuid)(0, 0, &eax, &ebx, &ecx, &edx);
     if (eax < 1)
        /* we can't ask for cpuid(x) for x > 0.  Give up. */
        return False;

     /* Get processor ID string, and max basic/extended index
        values. */
     VG_(memcpy)(&vstr[0], &ebx, 4);
     VG_(memcpy)(&vstr[4], &edx, 4);
     VG_(memcpy)(&vstr[8], &ecx, 4);
     vstr[12] = 0;

     VG_(cpuid)(0x80000000, 0, &eax, &ebx, &ecx, &edx);
     max_extended = eax;

     /* get capabilities bits into edx */
     VG_(cpuid)(1, 0, &eax, &ebx, &ecx, &edx);

     have_sse1 = (edx & (1<<25)) != 0; /* True => have sse insns */
     have_sse2 = (edx & (1<<26)) != 0; /* True => have sse2 insns */

     /* cmpxchg8b is a minimum requirement now; if we don't have it we
        must simply give up.  But all CPUs since Pentium-I have it, so
        that doesn't seem like much of a restriction. */
     have_cx8 = (edx & (1<<8)) != 0; /* True => have cmpxchg8b */
     if (!have_cx8)
        return False;

     /* Figure out if this is an AMD that can do LZCNT. */
     have_lzcnt = False;
     if (0 == VG_(strcmp)(vstr, "AuthenticAMD")
         && max_extended >= 0x80000001) {
        VG_(cpuid)(0x80000001, 0, &eax, &ebx, &ecx, &edx);
        have_lzcnt = (ecx & (1<<5)) != 0; /* True => have LZCNT */
     }

     if (have_sse2 && have_sse1) {
        va          = VexArchX86;
        vai.hwcaps  = VEX_HWCAPS_X86_SSE1;
        vai.hwcaps |= VEX_HWCAPS_X86_SSE2;
        if (have_lzcnt)
           vai.hwcaps |= VEX_HWCAPS_X86_LZCNT;
        VG_(machine_x86_have_mxcsr) = 1;
        return True;
     }

     if (have_sse1) {
        va          = VexArchX86;
        vai.hwcaps  = VEX_HWCAPS_X86_SSE1;
        VG_(machine_x86_have_mxcsr) = 1;
        return True;
     }

     va         = VexArchX86;
     vai.hwcaps = 0; /*baseline - no sse at all*/
     VG_(machine_x86_have_mxcsr) = 0;
     return True;
   }

#elif defined(VGA_amd64)
   { Bool have_sse3, have_cx8, have_cx16;
     Bool have_lzcnt;
     UInt eax, ebx, ecx, edx, max_extended;
     UChar vstr[13];
     vstr[0] = 0;

     if (!VG_(has_cpuid)())
        /* we can't do cpuid at all.  Give up. */
        return False;

     VG_(cpuid)(0, 0, &eax, &ebx, &ecx, &edx);
     if (eax < 1)
        /* we can't ask for cpuid(x) for x > 0.  Give up. */
        return False;

     /* Get processor ID string, and max basic/extended index
        values. */
     VG_(memcpy)(&vstr[0], &ebx, 4);
     VG_(memcpy)(&vstr[4], &edx, 4);
     VG_(memcpy)(&vstr[8], &ecx, 4);
     vstr[12] = 0;

     VG_(cpuid)(0x80000000, 0, &eax, &ebx, &ecx, &edx);
     max_extended = eax;

     /* get capabilities bits into edx */
     VG_(cpuid)(1, 0, &eax, &ebx, &ecx, &edx);

     // we assume that SSE1 and SSE2 are available by default
     have_sse3 = (ecx & (1<<0)) != 0;  /* True => have sse3 insns */
     // ssse3  is ecx:9
     // sse41  is ecx:19
     // sse42  is ecx:20

     /* cmpxchg8b is a minimum requirement now; if we don't have it we
        must simply give up.  But all CPUs since Pentium-I have it, so
        that doesn't seem like much of a restriction. */
     have_cx8 = (edx & (1<<8)) != 0; /* True => have cmpxchg8b */
     if (!have_cx8)
        return False;

     /* on amd64 we tolerate older cpus, which don't have cmpxchg16b */
     have_cx16 = (ecx & (1<<13)) != 0; /* True => have cmpxchg16b */

     /* Figure out if this is an AMD that can do LZCNT. */
     have_lzcnt = False;
     if (0 == VG_(strcmp)(vstr, "AuthenticAMD")
         && max_extended >= 0x80000001) {
        VG_(cpuid)(0x80000001, 0, &eax, &ebx, &ecx, &edx);
        have_lzcnt = (ecx & (1<<5)) != 0; /* True => have LZCNT */
     }

     va         = VexArchAMD64;
     vai.hwcaps = (have_sse3 ? VEX_HWCAPS_AMD64_SSE3 : 0)
                  | (have_cx16 ? VEX_HWCAPS_AMD64_CX16 : 0)
                  | (have_lzcnt ? VEX_HWCAPS_AMD64_LZCNT : 0);
     return True;
   }

#elif defined(VGA_ppc32)
   {
     /* Find out which subset of the ppc32 instruction set is supported by
        verifying whether various ppc32 instructions generate a SIGILL
        or a SIGFPE. An alternative approach is to check the AT_HWCAP and
        AT_PLATFORM entries in the ELF auxiliary table -- see also
        the_iifii.client_auxv in m_main.c.
      */
     vki_sigset_t          saved_set, tmp_set;
     vki_sigaction_fromK_t saved_sigill_act, saved_sigfpe_act;
     vki_sigaction_toK_t     tmp_sigill_act,   tmp_sigfpe_act;

     volatile Bool have_F, have_V, have_FX, have_GX, have_VX;
     Int r;

     /* This is a kludge.  Really we ought to back-convert saved_act
        into a toK_t using VG_(convert_sigaction_fromK_to_toK), but
        since that's a no-op on all ppc32 platforms so far supported,
        it's not worth the typing effort.  At least include most basic
        sanity check: */
     vg_assert(sizeof(vki_sigaction_fromK_t) == sizeof(vki_sigaction_toK_t));

     VG_(sigemptyset)(&tmp_set);
     VG_(sigaddset)(&tmp_set, VKI_SIGILL);
     VG_(sigaddset)(&tmp_set, VKI_SIGFPE);

     r = VG_(sigprocmask)(VKI_SIG_UNBLOCK, &tmp_set, &saved_set);
     vg_assert(r == 0);

     r = VG_(sigaction)(VKI_SIGILL, NULL, &saved_sigill_act);
     vg_assert(r == 0);
     tmp_sigill_act = saved_sigill_act;

     r = VG_(sigaction)(VKI_SIGFPE, NULL, &saved_sigfpe_act);
     vg_assert(r == 0);
     tmp_sigfpe_act = saved_sigfpe_act;

     /* NODEFER: signal handler does not return (from the kernel's point of
        view), hence if it is to successfully catch a signal more than once,
        we need the NODEFER flag. */
     tmp_sigill_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigill_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigill_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigill_act.ksa_handler = handler_unsup_insn;
     r = VG_(sigaction)(VKI_SIGILL, &tmp_sigill_act, NULL);
     vg_assert(r == 0);

     tmp_sigfpe_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigfpe_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigfpe_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigfpe_act.ksa_handler = handler_unsup_insn;
     r = VG_(sigaction)(VKI_SIGFPE, &tmp_sigfpe_act, NULL);
     vg_assert(r == 0);

     /* standard FP insns */
     have_F = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_F = False;
     } else {
        __asm__ __volatile__(".long 0xFC000090"); /*fmr 0,0 */
     }

     /* Altivec insns */
     have_V = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_V = False;
     } else {
        /* Unfortunately some older assemblers don't speak Altivec (or
           choose not to), so to be safe we directly emit the 32-bit
           word corresponding to "vor 0,0,0".  This fixes a build
           problem that happens on Debian 3.1 (ppc32), and probably
           various other places. */
        __asm__ __volatile__(".long 0x10000484"); /*vor 0,0,0*/
     }

     /* General-Purpose optional (fsqrt, fsqrts) */
     have_FX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_FX = False;
     } else {
        __asm__ __volatile__(".long 0xFC00002C"); /*fsqrt 0,0 */
     }

     /* Graphics optional (stfiwx, fres, frsqrte, fsel) */
     have_GX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_GX = False;
     } else {
        __asm__ __volatile__(".long 0xFC000034"); /* frsqrte 0,0 */
     }

     /* VSX support implies Power ISA 2.06 */
     have_VX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_VX = False;
     } else {
        __asm__ __volatile__(".long 0xf0000564"); /* xsabsdp XT,XB */
     }


     /* determine dcbz/dcbzl sizes while we still have the signal
      * handlers registered */
     find_ppc_dcbz_sz(&vai);

     r = VG_(sigaction)(VKI_SIGILL, &saved_sigill_act, NULL);
     vg_assert(r == 0);
     r = VG_(sigaction)(VKI_SIGFPE, &saved_sigfpe_act, NULL);
     vg_assert(r == 0);
     r = VG_(sigprocmask)(VKI_SIG_SETMASK, &saved_set, NULL);
     vg_assert(r == 0);
     VG_(debugLog)(1, "machine", "F %d V %d FX %d GX %d VX %d\n", 
                    (Int)have_F, (Int)have_V, (Int)have_FX,
                    (Int)have_GX, (Int)have_VX);
     /* Make FP a prerequisite for VMX (bogusly so), and for FX and GX. */
     if (have_V && !have_F)
        have_V = False;
     if (have_FX && !have_F)
        have_FX = False;
     if (have_GX && !have_F)
        have_GX = False;

     VG_(machine_ppc32_has_FP)  = have_F ? 1 : 0;
     VG_(machine_ppc32_has_VMX) = have_V ? 1 : 0;

     va = VexArchPPC32;

     vai.hwcaps = 0;
     if (have_F)  vai.hwcaps |= VEX_HWCAPS_PPC32_F;
     if (have_V)  vai.hwcaps |= VEX_HWCAPS_PPC32_V;
     if (have_FX) vai.hwcaps |= VEX_HWCAPS_PPC32_FX;
     if (have_GX) vai.hwcaps |= VEX_HWCAPS_PPC32_GX;
     if (have_VX) vai.hwcaps |= VEX_HWCAPS_PPC32_VX;

     /* But we're not done yet: VG_(machine_ppc32_set_clszB) must be
        called before we're ready to go. */
     return True;
   }

#elif defined(VGA_ppc64)
   {
     /* Same instruction set detection algorithm as for ppc32. */
     vki_sigset_t          saved_set, tmp_set;
     vki_sigaction_fromK_t saved_sigill_act, saved_sigfpe_act;
     vki_sigaction_toK_t     tmp_sigill_act,   tmp_sigfpe_act;

     volatile Bool have_F, have_V, have_FX, have_GX, have_VX;
     Int r;

     /* This is a kludge.  Really we ought to back-convert saved_act
        into a toK_t using VG_(convert_sigaction_fromK_to_toK), but
        since that's a no-op on all ppc64 platforms so far supported,
        it's not worth the typing effort.  At least include most basic
        sanity check: */
     vg_assert(sizeof(vki_sigaction_fromK_t) == sizeof(vki_sigaction_toK_t));

     VG_(sigemptyset)(&tmp_set);
     VG_(sigaddset)(&tmp_set, VKI_SIGILL);
     VG_(sigaddset)(&tmp_set, VKI_SIGFPE);

     r = VG_(sigprocmask)(VKI_SIG_UNBLOCK, &tmp_set, &saved_set);
     vg_assert(r == 0);

     r = VG_(sigaction)(VKI_SIGILL, NULL, &saved_sigill_act);
     vg_assert(r == 0);
     tmp_sigill_act = saved_sigill_act;

     VG_(sigaction)(VKI_SIGFPE, NULL, &saved_sigfpe_act);
     tmp_sigfpe_act = saved_sigfpe_act;

     /* NODEFER: signal handler does not return (from the kernel's point of
        view), hence if it is to successfully catch a signal more than once,
        we need the NODEFER flag. */
     tmp_sigill_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigill_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigill_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigill_act.ksa_handler = handler_unsup_insn;
     VG_(sigaction)(VKI_SIGILL, &tmp_sigill_act, NULL);

     tmp_sigfpe_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigfpe_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigfpe_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigfpe_act.ksa_handler = handler_unsup_insn;
     VG_(sigaction)(VKI_SIGFPE, &tmp_sigfpe_act, NULL);

     /* standard FP insns */
     have_F = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_F = False;
     } else {
        __asm__ __volatile__("fmr 0,0");
     }

     /* Altivec insns */
     have_V = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_V = False;
     } else {
        __asm__ __volatile__(".long 0x10000484"); /*vor 0,0,0*/
     }

     /* General-Purpose optional (fsqrt, fsqrts) */
     have_FX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_FX = False;
     } else {
        __asm__ __volatile__(".long 0xFC00002C"); /*fsqrt 0,0*/
     }

     /* Graphics optional (stfiwx, fres, frsqrte, fsel) */
     have_GX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_GX = False;
     } else {
        __asm__ __volatile__(".long 0xFC000034"); /*frsqrte 0,0*/
     }

     /* VSX support implies Power ISA 2.06 */
     have_VX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_VX = False;
     } else {
        __asm__ __volatile__(".long 0xf0000564"); /* xsabsdp XT,XB */
     }

     /* determine dcbz/dcbzl sizes while we still have the signal
      * handlers registered */
     find_ppc_dcbz_sz(&vai);

     VG_(sigaction)(VKI_SIGILL, &saved_sigill_act, NULL);
     VG_(sigaction)(VKI_SIGFPE, &saved_sigfpe_act, NULL);
     VG_(sigprocmask)(VKI_SIG_SETMASK, &saved_set, NULL);
     VG_(debugLog)(1, "machine", "F %d V %d FX %d GX %d VX %d\n", 
                    (Int)have_F, (Int)have_V, (Int)have_FX,
                    (Int)have_GX, (Int)have_VX);
     /* on ppc64, if we don't even have FP, just give up. */
     if (!have_F)
        return False;

     VG_(machine_ppc64_has_VMX) = have_V ? 1 : 0;

     va = VexArchPPC64;

     vai.hwcaps = 0;
     if (have_V)  vai.hwcaps |= VEX_HWCAPS_PPC64_V;
     if (have_FX) vai.hwcaps |= VEX_HWCAPS_PPC64_FX;
     if (have_GX) vai.hwcaps |= VEX_HWCAPS_PPC64_GX;
     if (have_VX) vai.hwcaps |= VEX_HWCAPS_PPC64_VX;

     /* But we're not done yet: VG_(machine_ppc64_set_clszB) must be
        called before we're ready to go. */
     return True;
   }

#elif defined(VGA_s390x)
   {
     /* Instruction set detection code borrowed from ppc above. */
     vki_sigset_t          saved_set, tmp_set;
     vki_sigaction_fromK_t saved_sigill_act;
     vki_sigaction_toK_t     tmp_sigill_act;

     volatile Bool have_LDISP, have_EIMM, have_GIE, have_DFP, have_FGX;
     Int r, model;

     /* Unblock SIGILL and stash away the old action for that signal */
     VG_(sigemptyset)(&tmp_set);
     VG_(sigaddset)(&tmp_set, VKI_SIGILL);

     r = VG_(sigprocmask)(VKI_SIG_UNBLOCK, &tmp_set, &saved_set);
     vg_assert(r == 0);

     r = VG_(sigaction)(VKI_SIGILL, NULL, &saved_sigill_act);
     vg_assert(r == 0);
     tmp_sigill_act = saved_sigill_act;

     /* NODEFER: signal handler does not return (from the kernel's point of
        view), hence if it is to successfully catch a signal more than once,
        we need the NODEFER flag. */
     tmp_sigill_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigill_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigill_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigill_act.ksa_handler = handler_unsup_insn;
     VG_(sigaction)(VKI_SIGILL, &tmp_sigill_act, NULL);

     /* Determine hwcaps. Note, we cannot use the stfle insn because it
        is not supported on z900. */

     have_LDISP = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_LDISP = False;
     } else {
       /* BASR loads the address of the next insn into r1. Needed to avoid
          a segfault in XY. */
        __asm__ __volatile__("basr %%r1,%%r0\n\t"
                             ".long  0xe3001000\n\t"  /* XY  0,0(%r1) */
                             ".short 0x0057" : : : "r0", "r1", "cc", "memory");
     }

     have_EIMM = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_EIMM = False;
     } else {
        __asm__ __volatile__(".long  0xc0090000\n\t"  /* iilf r0,0 */
                             ".short 0x0000" : : : "r0", "memory");
     }

     have_GIE = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_GIE = False;
     } else {
        __asm__ __volatile__(".long  0xc2010000\n\t"  /* msfi r0,0 */
                             ".short 0x0000" : : : "r0", "memory");
     }

     have_DFP = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_DFP = False;
     } else {
        __asm__ __volatile__(".long 0xb3d20000"
                               : : : "r0", "cc", "memory");  /* adtr r0,r0,r0 */
     }

     have_FGX = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_FGX = False;
     } else {
        __asm__ __volatile__(".long 0xb3cd0000" : : : "r0");  /* lgdr r0,f0 */
     }

     /* Restore signals */
     r = VG_(sigaction)(VKI_SIGILL, &saved_sigill_act, NULL);
     vg_assert(r == 0);
     r = VG_(sigprocmask)(VKI_SIG_SETMASK, &saved_set, NULL);
     vg_assert(r == 0);
     va = VexArchS390X;

     model = VG_(get_machine_model)();

     VG_(debugLog)(1, "machine", "machine %d  LDISP %d EIMM %d GIE %d DFP %d "
                   "FGX %d\n", model, have_LDISP, have_EIMM, have_GIE,
                   have_DFP, have_FGX);

     if (model == VEX_S390X_MODEL_INVALID) return False;

     vai.hwcaps = model;
     if (have_LDISP) {
        /* Use long displacement only on machines >= z990. For all other machines
           it is millicoded and therefore slow. */
        if (model >= VEX_S390X_MODEL_Z990)
           vai.hwcaps |= VEX_HWCAPS_S390X_LDISP;
     }
     if (have_EIMM)  vai.hwcaps |= VEX_HWCAPS_S390X_EIMM;
     if (have_GIE)   vai.hwcaps |= VEX_HWCAPS_S390X_GIE;
     if (have_DFP)   vai.hwcaps |= VEX_HWCAPS_S390X_DFP;
     if (have_FGX)   vai.hwcaps |= VEX_HWCAPS_S390X_FGX;

     VG_(debugLog)(1, "machine", "hwcaps = 0x%x\n", vai.hwcaps);

     return True;
   }

#elif defined(VGA_arm)
   {
     /* Same instruction set detection algorithm as for ppc32. */
     vki_sigset_t          saved_set, tmp_set;
     vki_sigaction_fromK_t saved_sigill_act, saved_sigfpe_act;
     vki_sigaction_toK_t     tmp_sigill_act,   tmp_sigfpe_act;

     volatile Bool have_VFP, have_VFP2, have_VFP3, have_NEON;
     volatile Int archlevel;
     Int r;

     /* This is a kludge.  Really we ought to back-convert saved_act
        into a toK_t using VG_(convert_sigaction_fromK_to_toK), but
        since that's a no-op on all ppc64 platforms so far supported,
        it's not worth the typing effort.  At least include most basic
        sanity check: */
     vg_assert(sizeof(vki_sigaction_fromK_t) == sizeof(vki_sigaction_toK_t));

     VG_(sigemptyset)(&tmp_set);
     VG_(sigaddset)(&tmp_set, VKI_SIGILL);
     VG_(sigaddset)(&tmp_set, VKI_SIGFPE);

     r = VG_(sigprocmask)(VKI_SIG_UNBLOCK, &tmp_set, &saved_set);
     vg_assert(r == 0);

     r = VG_(sigaction)(VKI_SIGILL, NULL, &saved_sigill_act);
     vg_assert(r == 0);
     tmp_sigill_act = saved_sigill_act;

     VG_(sigaction)(VKI_SIGFPE, NULL, &saved_sigfpe_act);
     tmp_sigfpe_act = saved_sigfpe_act;

     /* NODEFER: signal handler does not return (from the kernel's point of
        view), hence if it is to successfully catch a signal more than once,
        we need the NODEFER flag. */
     tmp_sigill_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigill_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigill_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigill_act.ksa_handler = handler_unsup_insn;
     VG_(sigaction)(VKI_SIGILL, &tmp_sigill_act, NULL);

     tmp_sigfpe_act.sa_flags &= ~VKI_SA_RESETHAND;
     tmp_sigfpe_act.sa_flags &= ~VKI_SA_SIGINFO;
     tmp_sigfpe_act.sa_flags |=  VKI_SA_NODEFER;
     tmp_sigfpe_act.ksa_handler = handler_unsup_insn;
     VG_(sigaction)(VKI_SIGFPE, &tmp_sigfpe_act, NULL);

     /* VFP insns */
     have_VFP = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_VFP = False;
     } else {
        __asm__ __volatile__(".word 0xEEB02B42"); /* VMOV.F64 d2, d2 */
     }
     /* There are several generation of VFP extension but they differs very
        little so for now we will not distinguish them. */
     have_VFP2 = have_VFP;
     have_VFP3 = have_VFP;

     /* NEON insns */
     have_NEON = True;
     if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
        have_NEON = False;
     } else {
        __asm__ __volatile__(".word 0xF2244154"); /* VMOV q2, q2 */
     }

     /* ARM architecture level */
     archlevel = 5; /* v5 will be base level */
     if (archlevel < 7) {
        archlevel = 7;
        if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
           archlevel = 5;
        } else {
           __asm__ __volatile__(".word 0xF45FF000"); /* PLI [PC,#-0] */
        }
     }
     if (archlevel < 6) {
        archlevel = 6;
        if (VG_MINIMAL_SETJMP(env_unsup_insn)) {
           archlevel = 5;
        } else {
           __asm__ __volatile__(".word 0xE6822012"); /* PKHBT r2, r2, r2 */
        }
     }

     VG_(convert_sigaction_fromK_to_toK)(&saved_sigill_act, &tmp_sigill_act);
     VG_(convert_sigaction_fromK_to_toK)(&saved_sigfpe_act, &tmp_sigfpe_act);
     VG_(sigaction)(VKI_SIGILL, &tmp_sigill_act, NULL);
     VG_(sigaction)(VKI_SIGFPE, &tmp_sigfpe_act, NULL);
     VG_(sigprocmask)(VKI_SIG_SETMASK, &saved_set, NULL);

     VG_(debugLog)(1, "machine", "ARMv%d VFP %d VFP2 %d VFP3 %d NEON %d\n",
           archlevel, (Int)have_VFP, (Int)have_VFP2, (Int)have_VFP3,
           (Int)have_NEON);

     VG_(machine_arm_archlevel) = archlevel;

     va = VexArchARM;

     vai.hwcaps = VEX_ARM_ARCHLEVEL(archlevel);
     if (have_VFP3) vai.hwcaps |= VEX_HWCAPS_ARM_VFP3;
     if (have_VFP2) vai.hwcaps |= VEX_HWCAPS_ARM_VFP2;
     if (have_VFP)  vai.hwcaps |= VEX_HWCAPS_ARM_VFP;
     if (have_NEON) vai.hwcaps |= VEX_HWCAPS_ARM_NEON;

     return True;
   }

#else
#  error "Unknown arch"
#endif
}

/* Notify host cpu cache line size. */
#if defined(VGA_ppc32)
void VG_(machine_ppc32_set_clszB)( Int szB )
{
   vg_assert(hwcaps_done);

   /* Either the value must not have been set yet (zero) or we can
      tolerate it being set to the same value multiple times, as the
      stack scanning logic in m_main is a bit stupid. */
   vg_assert(vai.ppc_cache_line_szB == 0
             || vai.ppc_cache_line_szB == szB);

   vg_assert(szB == 32 || szB == 64 || szB == 128);
   vai.ppc_cache_line_szB = szB;
}
#endif


/* Notify host cpu cache line size. */
#if defined(VGA_ppc64)
void VG_(machine_ppc64_set_clszB)( Int szB )
{
   vg_assert(hwcaps_done);

   /* Either the value must not have been set yet (zero) or we can
      tolerate it being set to the same value multiple times, as the
      stack scanning logic in m_main is a bit stupid. */
   vg_assert(vai.ppc_cache_line_szB == 0
             || vai.ppc_cache_line_szB == szB);

   vg_assert(szB == 32 || szB == 64 || szB == 128);
   vai.ppc_cache_line_szB = szB;
}
#endif


/* Notify host's ability to handle NEON instructions. */
#if defined(VGA_arm)
void VG_(machine_arm_set_has_NEON)( Bool has_neon )
{
   vg_assert(hwcaps_done);
   /* There's nothing else we can sanity check. */

   if (has_neon) {
      vai.hwcaps |= VEX_HWCAPS_ARM_NEON;
   } else {
      vai.hwcaps &= ~VEX_HWCAPS_ARM_NEON;
   }
}
#endif


/* Fetch host cpu info, once established. */
void VG_(machine_get_VexArchInfo)( /*OUT*/VexArch* pVa,
                                   /*OUT*/VexArchInfo* pVai )
{
   vg_assert(hwcaps_done);
   if (pVa)  *pVa  = va;
   if (pVai) *pVai = vai;
}


// Given a pointer to a function as obtained by "& functionname" in C,
// produce a pointer to the actual entry point for the function.
void* VG_(fnptr_to_fnentry)( void* f )
{
#  if defined(VGP_x86_linux) || defined(VGP_amd64_linux)  \
      || defined(VGP_arm_linux)                           \
      || defined(VGP_ppc32_linux) || defined(VGO_darwin)  \
      || defined(VGP_s390x_linux)
   return f;
#  elif defined(VGP_ppc64_linux)
   /* ppc64-linux uses the AIX scheme, in which f is a pointer to a
      3-word function descriptor, of which the first word is the entry
      address. */
   UWord* descr = (UWord*)f;
   return (void*)(descr[0]);
#  else
#    error "Unknown platform"
#  endif
}

/*--------------------------------------------------------------------*/
/*--- end                                                          ---*/
/*--------------------------------------------------------------------*/