[multiple changes]

2004-02-12  Olivier Hainque  <hainque@act-europe.fr>

	* decl.c (components_to_record): Don't claim that the internal fields
	we make to hold the variant parts are semantically addressable, because
	they are not.

	* exp_pakd.adb (Create_Packed_Array_Type): Rename Esiz into PASize and
	adjust the comment describing the modular type form when we can use it.
	(Install_PAT): Account for the Esiz renaming.

	* init.c (__gnat_error_handler for alpha-tru64): Arrange to clear the
	sc_onstack context indication before raising the exception to which
	the signal is mapped. Allows better handling of later signals possibly
	triggered by the resumed user code if the exception is handled.

2004-02-12  Arnaud Charlet  <charlet@act-europe.fr>

	* 5zinit.adb: Removed, no longer used.

2004-02-12  Robert Dewar  <dewar@gnat.com>

	* ali.adb: Remove separating space between parameters on R line. Makes
	format consistent with format used by the binder for Set_Globals call.

	* atree.ads, atree.adb: Minor reformatting (new function header format)

	* bindgen.adb: Add Run-Time Globals documentation section containing
	detailed documentation of the globals passed from the binder file to
	the run time.

	* gnatls.adb: Minor reformatting

	* init.c (__gnat_set_globals): Add note pointing to documentation in
	bindgen.

	* lib-writ.ads, lib-writ.adb: Remove separating space between
	parameters on R line.
	Makes format consistent with format used by the binder for Set_Globals
	call.

	* osint.ads: Add 2004 to copyright notice
	Minor reformatting

	* snames.ads: Correct capitalization of FIFO_Within_Priorities
	Noticed during code reading, documentation issue only

	* usage.adb: Remove junk line for obsolete C switch
	Noticed during code reading

2004-02-12  Vincent Celier  <celier@gnat.com>

	* bld.adb (Process_Declarative_Items): For Source_Dirs call gprcmd
	extend for each directory, so that multiple /** directories are
	extended individually.
	(Recursive_Process): Set the default for LANGUAGES to ada

	* gprcmd.adb: Define new command "ignore", to do nothing.
	Implement new comment "path".

	* Makefile.generic: Suppress output when SILENT is set
	Make sure that when compiler for C/C++ is gcc, the correct -x switch is
	used, so that the correct compiler is invoked.
	When compiler is gcc/g++, put search path in env vars C_INCLUDE_PATH/
	CXX_INCLUDE_PATH, to avoid failure with too long command lines.

2004-02-12  Jerome Guitton  <guitton@act-europe.fr>

	* Makefile.in: Clean ups and remove obsolete targets.

2004-02-12  Ed Schonberg  <schonberg@gnat.com>

	* exp_ch5.adb: Remove Possible_Unligned_Slice, in favor of the similar
	predicate declared in exp_util.

	* exp_util.adb: Add comments.

	* sem_ch10.adb (Analyze_Subunit): Remove ultimate parent unit from
	visibility before compiling context of the subunit.

	* sem_res.adb (Check_Parameterless_Call): If the context expects a
	value but the name is a procedure, do not attempt to analyze as a call,
	in order to obtain more telling diagnostics.

	* sem_util.adb (Wrong_Type): Further enhancement to diagnose missing
	'Access on parameterless function calls.
	(Normalize_Actuals): For a parameterless function call with missing
	actuals, defer diagnostic until resolution of enclosing call.

	* sem_util.adb (Wrong_Type): If the context type is an access to
	subprogram and the expression is a procedure name, suggest a missing
	'attribute.

From-SVN: r77704

1 files changed, 53 insertions, 3 deletions

diff --git a/gcc/ada/init.c b/gcc/ada/init.c
index 4e4400f..7db7f1f 100644
--- a/gcc/ada/init.c
+++ b/gcc/ada/init.c
@@ -39,6 +39,10 @@
     installed by this file are used to handle resulting signals that come
     from these probes failing (i.e. touching protected pages) */
 
+/* This file should be kept synchronized with 2sinit.ads, 2sinit.adb, and
+   5zinit.adb. All these files implement the required functionality for
+   different targets. */
+
 /* The following include is here to meet the published VxWorks requirement
    that the __vxworks header appear before any other include. */
 #ifdef __vxworks
@@ -154,6 +158,9 @@ __gnat_get_interrupt_state (int intrup)
    binder file is not in the shared library. Global references across library
    boundaries like this are not handled correctly in all systems.  */
 
+/* For detailed description of the parameters to this routine, see the
+   section titled Run-Time Globals in package Bindgen (bindgen.adb) */
+
 void
 __gnat_set_globals (int main_priority,
                     int time_slice_val,
@@ -363,6 +370,7 @@ __gnat_initialize (void)
    exclude this case in the above test.  */
 
 #include <signal.h>
+#include <setjmp.h>
 #include <sys/siginfo.h>
 
 static void __gnat_error_handler (int, siginfo_t *, struct sigcontext *);
@@ -440,7 +448,48 @@ __gnat_error_handler (int sig, siginfo_t *sip, struct sigcontext *context)
   if (mstate != 0)
     *mstate = *context;
 
-  Raise_From_Signal_Handler (exception, (char *) msg);
+  /* We are now going to raise the exception corresponding to the signal we
+     caught, which may eventually end up resuming the application code if the
+     exception is handled.
+
+     When the exception is handled, merely arranging for the *exception*
+     handler's context (stack pointer, program counter, other registers, ...)
+     to be installed is *not* enough to let the kernel think we've left the
+     *signal* handler.  This has annoying implications if an alternate stack
+     has been setup for this *signal* handler, because the kernel thinks we
+     are still running on that alternate stack even after the jump, which
+     causes trouble at least as soon as another signal is raised.
+
+     We deal with this by forcing a "local" longjmp within the signal handler
+     below, forcing the "on alternate stack" indication to be reset (kernel
+     wise) on the way.  If no alternate stack has been setup, this should be a
+     neutral operation. Otherwise, we will be in a delicate situation for a
+     short while because we are going to run the exception propagation code
+     within the alternate stack area (that is, with the stack pointer inside
+     the alternate stack bounds), but with the corresponding flag off from the
+     kernel's standpoint.  We expect this to be ok as long as the propagation
+     code does not trigger a signal itself, which is expected.
+
+     ??? A better approach would be to at least delay this operation until the
+     last second, that is, until just before we jump to the exception handler,
+     if any.  */
+  {
+    jmp_buf handler_jmpbuf;
+
+    if (setjmp (handler_jmpbuf) != 0)
+      Raise_From_Signal_Handler (exception, (char *) msg);
+    else
+      {
+	/* Arrange for the "on alternate stack" flag to be reset.  See the
+	   comments around "jmp_buf offsets" in /usr/include/setjmp.h.  */
+	struct sigcontext * handler_context
+	  = (struct sigcontext *) & handler_jmpbuf;
+
+	handler_context->sc_onstack = 0;
+	
+	longjmp (handler_jmpbuf, 1);
+      }
+  }
 }
 
 void
@@ -461,11 +510,12 @@ __gnat_install_handler (void)
      we want this to happen for tasks also.  */
 
   static char sig_stack [8*1024];
-  /* 8K allocated here because 4K is not enough for the GCC/ZCX scheme.  */
+  /* 8K is a mininum to be able to propagate an exception using the GCC/ZCX
+     scheme.  */
 
   struct sigaltstack ss;
 
-  ss.ss_sp = (void *) & sig_stack;
+  ss.ss_sp = (void *) sig_stack;
   ss.ss_size = sizeof (sig_stack);
   ss.ss_flags = 0;