Preserve SSE registers in runtime relocations on x86-64.

SSE registers are used for passing parameters and must be preserved
in runtime relocations.  This is inside ld.so enforced through the
tests in tst-xmmymm.sh.  But the malloc routines used after startup
come from libc.so and can be arbitrarily complex.  It's overkill
to save the SSE registers all the time because of that.  These calls
are rare.  Instead we save them on demand.  The new infrastructure
put in place in this patch makes this possible and efficient.

ChangeLog

@@ -1,3 +1,18 @@
+2009-07-29  Ulrich Drepper  <drepper@redhat.com>
+
+	* elf/dl-runtime.c (_dl_fixup): Indicate before _dl_lookup_symbol_x
+	call that registers used in calling conventions need to be preserved.
+	* elf/dl-lookup.c (do_lookup_x): Use RTLD_*_FOREIGN_CALL macros
+	to preserve register content if necessary.
+	* sysdeps/x86_64/dl-trampoline.S (_dl_x86_64_save_sse): New function.
+	(_dl_x86_64_restore_sse): New function.
+	* sysdeps/x86_64/tst-xmmymm.sh: There is now one more function that
+	is allowed to modify xmm/ymm registers.
+
+	* stdio-common/scanf15.c: Undefine _LIBC.  We want to test from an
+	application's perspective.
+	* stdio-common/scanf17.c: Likewise.
+
 2009-07-28  Ulrich Drepper  <drepper@redhat.com>
 
 	* csu/libc-tls.c (__libc_setup_tls) [TLS_TCB_AT_TP]: Don't add TCB

elf/dl-lookup.c

@@ -380,6 +380,10 @@ do_lookup_x (const char *undef_name, uint_fast32_t new_hash,
 		  if (size * 3 <= tab->n_elements * 4)
 		    {
 		      /* Expand the table.  */
+#ifdef RTLD_CHECK_FOREIGN_CALL
+		      /* This must not happen during runtime relocations.  */
+		      assert (!RTLD_CHECK_FOREIGN_CALL);
+#endif
 		      size_t newsize = _dl_higher_prime_number (size + 1);
 		      struct unique_sym *newentries
 			= calloc (sizeof (struct unique_sym), newsize);
@@ -405,6 +409,11 @@ do_lookup_x (const char *undef_name, uint_fast32_t new_hash,
 		}
 	      else
 		{
+#ifdef RTLD_CHECK_FOREIGN_CALL
+		  /* This must not happen during runtime relocations.  */
+		  assert (!RTLD_CHECK_FOREIGN_CALL);
+#endif
+
 #define INITIAL_NUNIQUE_SYM_TABLE 31
 		  size = INITIAL_NUNIQUE_SYM_TABLE;
 		  entries = calloc (sizeof (struct unique_sym), size);
@@ -600,6 +609,10 @@ add_dependency (struct link_map *undef_map, struct link_map *map, int flags)
 	  unsigned int max
 	    = undef_map->l_reldepsmax ? undef_map->l_reldepsmax * 2 : 10;
 
+#ifdef RTLD_PREPARE_FOREIGN_CALL
+	  RTLD_PREPARE_FOREIGN_CALL;
+#endif
+
 	  newp = malloc (sizeof (*newp) + max * sizeof (struct link_map *));
 	  if (newp == NULL)
 	    {

elf/dl-runtime.c

@@ -111,13 +111,21 @@ _dl_fixup (
 	  flags |= DL_LOOKUP_GSCOPE_LOCK;
 	}
 
+#ifdef RTLD_ENABLE_FOREIGN_CALL
+      RTLD_ENABLE_FOREIGN_CALL;
+#endif
+
       result = _dl_lookup_symbol_x (strtab + sym->st_name, l, &sym, l->l_scope,
 				    version, ELF_RTYPE_CLASS_PLT, flags, NULL);
 
       /* We are done with the global scope.  */
       if (!RTLD_SINGLE_THREAD_P)
 	THREAD_GSCOPE_RESET_FLAG ();
 
+#ifdef RTLD_FINALIZE_FOREIGN_CALL
+      RTLD_FINALIZE_FOREIGN_CALL;
+#endif
+
       /* Currently result contains the base load address (or link map)
 	 of the object that defines sym.  Now add in the symbol
 	 offset.  */

nptl/ChangeLog

@@ -1,3 +1,11 @@
+2009-07-29  Ulrich Drepper  <drepper@redhat.com>
+
+	* sysdeps/x86_64/tls.h (tcbhead_t): Add room for SSE registers the
+	dynamic linker might have to save.  Define RTLD_CHECK_FOREIGN_CALL,
+	RTLD_ENABLE_FOREIGN_CALL, RTLD_PREPARE_FOREIGN_CALL, and
+	RTLD_FINALIZE_FOREIGN_CALL.  Pretty printing.
+	* sysdeps/x86_64/tcb-offsets.sym: Add RTLD_SAVESPACE_SSE.
+
 2009-07-28  Ulrich Drepper  <drepper@redhat.com>
 
 	* pthread_mutex_lock.c [NO_INCR] (__pthread_mutex_cond_lock_adjust):

nptl/sysdeps/x86_64/tcb-offsets.sym

@@ -15,3 +15,4 @@ VGETCPU_CACHE_OFFSET	offsetof (tcbhead_t, vgetcpu_cache)
 #ifndef __ASSUME_PRIVATE_FUTEX
 PRIVATE_FUTEX		offsetof (tcbhead_t, private_futex)
 #endif
+RTLD_SAVESPACE_SSE	offsetof (tcbhead_t, rtld_savespace_sse)

nptl/sysdeps/x86_64/tls.h

@@ -29,6 +29,7 @@
 # include <sysdep.h>
 # include <kernel-features.h>
 # include <bits/wordsize.h>
+# include <xmmintrin.h>
 
 
 /* Type for the dtv.  */
@@ -55,16 +56,23 @@ typedef struct
   uintptr_t stack_guard;
   uintptr_t pointer_guard;
   unsigned long int vgetcpu_cache[2];
-#ifndef __ASSUME_PRIVATE_FUTEX
+# ifndef __ASSUME_PRIVATE_FUTEX
   int private_futex;
-#else
+# else
   int __unused1;
-#endif
-#if __WORDSIZE == 64
-  int __pad1;
-#endif
+# endif
+# if __WORDSIZE == 64
+  int rtld_must_xmm_save;
+# endif
   /* Reservation of some values for the TM ABI.  */
   void *__private_tm[5];
+# if __WORDSIZE == 64
+  long int __unused2;
+  /* Have space for the post-AVX register size.  */
+  __m128 rtld_savespace_sse[8][4];
+
+  void *__padding[8];
+# endif
 } tcbhead_t;
 
 #else /* __ASSEMBLER__ */
@@ -298,7 +306,7 @@ typedef struct
 
 
 /* Atomic compare and exchange on TLS, returning old value.  */
-#define THREAD_ATOMIC_CMPXCHG_VAL(descr, member, newval, oldval) \
+# define THREAD_ATOMIC_CMPXCHG_VAL(descr, member, newval, oldval) \
   ({ __typeof (descr->member) __ret;					      \
      __typeof (oldval) __old = (oldval);				      \
      if (sizeof (descr->member) == 4)					      \
@@ -313,7 +321,7 @@ typedef struct
 
 
 /* Atomic logical and.  */
-#define THREAD_ATOMIC_AND(descr, member, val) \
+# define THREAD_ATOMIC_AND(descr, member, val) \
   (void) ({ if (sizeof ((descr)->member) == 4)				      \
 	      asm volatile (LOCK_PREFIX "andl %1, %%fs:%P0"		      \
 			    :: "i" (offsetof (struct pthread, member)),	      \
@@ -324,7 +332,7 @@ typedef struct
 
 
 /* Atomic set bit.  */
-#define THREAD_ATOMIC_BIT_SET(descr, member, bit) \
+# define THREAD_ATOMIC_BIT_SET(descr, member, bit) \
   (void) ({ if (sizeof ((descr)->member) == 4)				      \
 	      asm volatile (LOCK_PREFIX "orl %1, %%fs:%P0"		      \
 			    :: "i" (offsetof (struct pthread, member)),	      \
@@ -334,7 +342,7 @@ typedef struct
 	      abort (); })
 
 
-#define CALL_THREAD_FCT(descr) \
+# define CALL_THREAD_FCT(descr) \
   ({ void *__res;							      \
      asm volatile ("movq %%fs:%P2, %%rdi\n\t"				      \
 		   "callq *%%fs:%P1"					      \
@@ -355,18 +363,18 @@ typedef struct
 
 
 /* Set the pointer guard field in the TCB head.  */
-#define THREAD_SET_POINTER_GUARD(value) \
+# define THREAD_SET_POINTER_GUARD(value) \
   THREAD_SETMEM (THREAD_SELF, header.pointer_guard, value)
-#define THREAD_COPY_POINTER_GUARD(descr) \
+# define THREAD_COPY_POINTER_GUARD(descr) \
   ((descr)->header.pointer_guard					      \
    = THREAD_GETMEM (THREAD_SELF, header.pointer_guard))
 
 
 /* Get and set the global scope generation counter in the TCB head.  */
-#define THREAD_GSCOPE_FLAG_UNUSED 0
-#define THREAD_GSCOPE_FLAG_USED   1
-#define THREAD_GSCOPE_FLAG_WAIT   2
-#define THREAD_GSCOPE_RESET_FLAG() \
+# define THREAD_GSCOPE_FLAG_UNUSED 0
+# define THREAD_GSCOPE_FLAG_USED   1
+# define THREAD_GSCOPE_FLAG_WAIT   2
+# define THREAD_GSCOPE_RESET_FLAG() \
   do									      \
     { int __res;							      \
       asm volatile ("xchgl %0, %%fs:%P1"				      \
@@ -377,11 +385,40 @@ typedef struct
 	lll_futex_wake (&THREAD_SELF->header.gscope_flag, 1, LLL_PRIVATE);    \
     }									      \
   while (0)
-#define THREAD_GSCOPE_SET_FLAG() \
+# define THREAD_GSCOPE_SET_FLAG() \
   THREAD_SETMEM (THREAD_SELF, header.gscope_flag, THREAD_GSCOPE_FLAG_USED)
-#define THREAD_GSCOPE_WAIT() \
+# define THREAD_GSCOPE_WAIT() \
   GL(dl_wait_lookup_done) ()
 
+
+# ifdef SHARED
+/* Defined in dl-trampoline.S.  */
+extern void _dl_x86_64_save_sse (void);
+extern void _dl_x86_64_restore_sse (void);
+
+# define RTLD_CHECK_FOREIGN_CALL \
+  (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) != 0)
+
+#  define RTLD_ENABLE_FOREIGN_CALL \
+  THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 1)
+
+#  define RTLD_PREPARE_FOREIGN_CALL \
+  do if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save))	      \
+    {									      \
+      _dl_x86_64_save_sse ();						      \
+      THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 0);	      \
+    }									      \
+  while (0)
+
+#  define RTLD_FINALIZE_FOREIGN_CALL \
+  do {									      \
+    if (THREAD_GETMEM (THREAD_SELF, header.rtld_must_xmm_save) == 0)	      \
+      _dl_x86_64_restore_sse ();					      \
+    THREAD_SETMEM (THREAD_SELF, header.rtld_must_xmm_save, 0);		      \
+  } while (0)
+# endif
+
+
 #endif /* __ASSEMBLER__ */
 
 #endif	/* tls.h */

stdio-common/scanf15.c

@@ -1,5 +1,6 @@
 #undef _GNU_SOURCE
 #define _XOPEN_SOURCE 600
+#undef _LIBC
 /* The following macro definitions are a hack.  They word around disabling
    the GNU extension while still using a few internal headers.  */
 #define u_char unsigned char

stdio-common/scanf17.c

@@ -1,5 +1,6 @@
 #undef _GNU_SOURCE
 #define _XOPEN_SOURCE 600
+#undef _LIBC
 /* The following macro definitions are a hack.  They word around disabling
    the GNU extension while still using a few internal headers.  */
 #define u_char unsigned char

sysdeps/x86_64/dl-trampoline.S

@@ -390,3 +390,85 @@ L(no_avx4):
 	cfi_endproc
 	.size _dl_runtime_profile, .-_dl_runtime_profile
 #endif
+
+
+#ifdef SHARED
+	.globl _dl_x86_64_save_sse
+	.type _dl_x86_64_save_sse, @function
+	.align 16
+	cfi_startproc
+_dl_x86_64_save_sse:
+# ifdef HAVE_AVX_SUPPORT
+	cmpl	$0, L(have_avx)(%rip)
+	jne	1f
+	movq	%rbx, %r11		# Save rbx
+	movl	$1, %eax
+	cpuid
+	movq	%r11,%rbx		# Restore rbx
+	movl	$1, %eax
+	testl	$(1 << 28), %ecx
+	jne	2f
+	negl	%eax
+2:	movl	%eax, L(have_avx)(%rip)
+	cmpl	$0, %eax
+
+1:	js	L(no_avx5)
+
+#  define YMM_SIZE 32
+	vmovdqa %ymm0, %fs:RTLD_SAVESPACE_SSE+0*YMM_SIZE
+	vmovdqa %ymm1, %fs:RTLD_SAVESPACE_SSE+1*YMM_SIZE
+	vmovdqa %ymm2, %fs:RTLD_SAVESPACE_SSE+2*YMM_SIZE
+	vmovdqa %ymm3, %fs:RTLD_SAVESPACE_SSE+3*YMM_SIZE
+	vmovdqa %ymm4, %fs:RTLD_SAVESPACE_SSE+4*YMM_SIZE
+	vmovdqa %ymm5, %fs:RTLD_SAVESPACE_SSE+5*YMM_SIZE
+	vmovdqa %ymm6, %fs:RTLD_SAVESPACE_SSE+6*YMM_SIZE
+	vmovdqa %ymm7, %fs:RTLD_SAVESPACE_SSE+7*YMM_SIZE
+	ret
+L(no_avx5):
+# endif
+# define YMM_SIZE 16
+	movdqa	%xmm0, %fs:RTLD_SAVESPACE_SSE+0*XMM_SIZE
+	movdqa	%xmm1, %fs:RTLD_SAVESPACE_SSE+1*XMM_SIZE
+	movdqa	%xmm2, %fs:RTLD_SAVESPACE_SSE+2*XMM_SIZE
+	movdqa	%xmm3, %fs:RTLD_SAVESPACE_SSE+3*XMM_SIZE
+	movdqa	%xmm4, %fs:RTLD_SAVESPACE_SSE+4*XMM_SIZE
+	movdqa	%xmm5, %fs:RTLD_SAVESPACE_SSE+5*XMM_SIZE
+	movdqa	%xmm6, %fs:RTLD_SAVESPACE_SSE+6*XMM_SIZE
+	movdqa	%xmm7, %fs:RTLD_SAVESPACE_SSE+7*XMM_SIZE
+	ret
+	cfi_endproc
+	.size _dl_x86_64_save_sse, .-_dl_x86_64_save_sse
+
+
+	.globl _dl_x86_64_restore_sse
+	.type _dl_x86_64_restore_sse, @function
+	.align 16
+	cfi_startproc
+_dl_x86_64_restore_sse:
+# ifdef HAVE_AVX_SUPPORT
+	cmpl	$0, L(have_avx)(%rip)
+	js	L(no_avx6)
+
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+0*YMM_SIZE, %ymm0
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+1*YMM_SIZE, %ymm1
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+2*YMM_SIZE, %ymm2
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+3*YMM_SIZE, %ymm3
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+4*YMM_SIZE, %ymm4
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+5*YMM_SIZE, %ymm5
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+6*YMM_SIZE, %ymm6
+	vmovdqa %fs:RTLD_SAVESPACE_SSE+7*YMM_SIZE, %ymm7
+	ret
+L(no_avx6):
+# endif
+	movdqa	%fs:RTLD_SAVESPACE_SSE+0*XMM_SIZE, %xmm0
+	movdqa	%fs:RTLD_SAVESPACE_SSE+1*XMM_SIZE, %xmm1
+	movdqa	%fs:RTLD_SAVESPACE_SSE+2*XMM_SIZE, %xmm2
+	movdqa	%fs:RTLD_SAVESPACE_SSE+3*XMM_SIZE, %xmm3
+	movdqa	%fs:RTLD_SAVESPACE_SSE+4*XMM_SIZE, %xmm4
+	movdqa	%fs:RTLD_SAVESPACE_SSE+5*XMM_SIZE, %xmm5
+	movdqa	%fs:RTLD_SAVESPACE_SSE+6*XMM_SIZE, %xmm6
+	movdqa	%fs:RTLD_SAVESPACE_SSE+7*XMM_SIZE, %xmm7
+	ret
+	cfi_endproc
+	.size _dl_x86_64_restore_sse, .-_dl_x86_64_restore_sse
+#endif

sysdeps/x86_64/tst-xmmymm.sh

@@ -59,10 +59,11 @@ for f in $tocheck; do
   objdump -d "$objpfx"../*/"$f" |
   awk 'BEGIN { last="" } /^[[:xdigit:]]* <[_[:alnum:]]*>:$/ { fct=substr($2, 2, length($2)-3) } /,%[xy]mm[[:digit:]]*$/ { if (last != fct) { print fct; last=fct} }' |
   while read fct; do
-    if test "$fct" != "_dl_runtime_profile"; then
-      echo "function $fct in $f modifies xmm/ymm" >> "$tmp"
-      result=1
+    if test "$fct" = "_dl_runtime_profile" -o "$fct" = "_dl_x86_64_restore_sse"; then
+      continue;
     fi
+    echo "function $fct in $f modifies xmm/ymm" >> "$tmp"
+    result=1
   done
 done