debian/patches/kfreebsd/local-dosavesse.diff

   1  unfortunately, it does not work
   2  in mean time use local-nosavesse.diff
   3
   4 --- a/linuxthreads/descr.h
   5 +++ b/linuxthreads/descr.h
   6 @@ -114,6 +114,7 @@
   7    /* This overlaps tcbhead_t (see tls.h), as used for TLS without threads.  */
   8    union
   9    {
  10 +    tcbhead_t tcbheader;
  11      struct
  12      {
  13        void *tcb;               /* Pointer to the TCB.  This is not always
  14 --- a/linuxthreads/sysdeps/x86_64/tcb-offsets.sym
  15 +++ b/linuxthreads/sysdeps/x86_64/tcb-offsets.sym
  16 @@ -3,3 +3,4 @@
  17
  18  MULTIPLE_THREADS_OFFSET                offsetof (tcbhead_t, multiple_threads)
  19  POINTER_GUARD                  offsetof (tcbhead_t, pointer_guard)
  20 +RTLD_SAVESPACE_SSE             offsetof (tcbhead_t, rtld_savespace_sse)
  21 --- a/linuxthreads/sysdeps/x86_64/tls.h
  22 +++ b/linuxthreads/sysdeps/x86_64/tls.h
  23 @@ -26,6 +26,7 @@
  24  # include <stdbool.h>
  25  # include <stddef.h>
  26  # include <stdint.h>
  27 +# include <xmmintrin.h>
  28
  29  /* Type for the dtv.  */
  30  typedef union dtv
  31 @@ -46,6 +47,7 @@
  32    dtv_t *dtv;
  33    void *self;          /* Pointer to the thread descriptor.  */
  34    int multiple_threads;
  35 +  int __unused1;
  36    uintptr_t sysinfo;
  37    uintptr_t stack_guard;
  38    uintptr_t pointer_guard;
  39 @@ -53,7 +55,13 @@
  40    long gscope_flag;
  41  # else
  42    int gscope_flag;
  43 +  int __unused2;
  44  # endif
  45 +  void *__padding1[7];
  46 +  int rtld_must_xmm_save;
  47 +  int __unused3;
  48 +   /* Have space for the post-AVX register size.  */
  49 +  __m128 rtld_savespace_sse[8][4];
  50  } tcbhead_t;
  51
  52  #else /* __ASSEMBLER__ */
  53 @@ -77,13 +85,19 @@
  54  #  define TLS_INIT_TCB_SIZE sizeof (tcbhead_t)
  55
  56  /* Alignment requirements for the initial TCB.  */
  57 -#  define TLS_INIT_TCB_ALIGN __alignof__ (tcbhead_t)
  58 +//#  define TLS_INIT_TCB_ALIGN __alignof__ (tcbhead_t)
  59 +// Normally the above would be correct  But we have to store post-AVX
  60 +// vector registers in the TCB and we want the storage to be aligned.
  61 +// unfortunately there isn't yet a type for these values and hence no
  62 +// 32-byte alignment requirement.  Make this explicit, for now.
  63 +#  define TLS_INIT_TCB_ALIGN 32
  64
  65  /* This is the size of the TCB.  */
  66  #  define TLS_TCB_SIZE sizeof (struct _pthread_descr_struct)
  67
  68  /* Alignment requirements for the TCB.  */
  69  #  define TLS_TCB_ALIGN __alignof__ (struct _pthread_descr_struct)
  70 +// It is already 32B aligned
  71
  72  /* The TCB can have any size and the memory following the address the
  73     thread pointer points to is unspecified.  Allocate the TCB there.  */
  74 @@ -189,7 +203,42 @@
  75  #define THREAD_GSCOPE_WAIT() \
  76    do { /* GL(dl_wait_lookup_done) () */ } while (0)
  77
  78 +
  79 +# ifdef SHARED
  80 +/* Defined in dl-trampoline.S.  */
  81 +extern void _dl_x86_64_save_sse (void);
  82 +extern void _dl_x86_64_restore_sse (void);
  83 +
  84 +# define RTLD_CHECK_FOREIGN_CALL \
  85 +  (THREAD_GETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save) != 0)
  86 +
  87 +/* NB: Don't use the xchg operation because that would imply a lock
  88 +   prefix which is expensive and unnecessary.  The cache line is also
  89 +   not contested at all.  */
  90 +#  define RTLD_ENABLE_FOREIGN_CALL \
  91 +  int old_rtld_must_xmm_save = THREAD_GETMEM (THREAD_SELF,                   \
  92 +                                             p_header.tcbheader.rtld_must_xmm_save); \
  93 +  THREAD_SETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save, 1)
  94 +
  95 +#  define RTLD_PREPARE_FOREIGN_CALL \
  96 +  do if (THREAD_GETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save))  \
  97 +    {                                                                        \
  98 +      _dl_x86_64_save_sse ();                                                \
  99 +      THREAD_SETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save, 0);  \
 100 +    }                                                                        \
 101 +  while (0)
 102 +
 103 +#  define RTLD_FINALIZE_FOREIGN_CALL \
 104 +  do {                                                                       \
 105 +    if (THREAD_GETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save) == 0) \
 106 +      _dl_x86_64_restore_sse ();                                             \
 107 +    THREAD_SETMEM (THREAD_SELF, p_header.tcbheader.rtld_must_xmm_save,       \
 108 +                  old_rtld_must_xmm_save);                                   \
 109 +  } while (0)
 110 +# endif
 111 +
 112  # endif        /* HAVE_TLS_SUPPORT */
 113 +
 114  #endif /* __ASSEMBLER__ */
 115
 116  #endif /* tls.h */