Target-specific thread switching backend

Target-specific thread switching backend

[Maciej W. Rozycki]

Hello,

 In preparation for submitting support for the MDI target I propose the 
following enhancement to target_switch_to_thread().  The MDI implements a 
multi-threaded target which remembers the currently selected thread within 
itself.  All the operations that have thread-local scope, such as 
accessing of the CPU registers, use that thread identifier.  Therefore I 
think it is reasonable to propagate the thread-switch event down to the 
target, which is what the following change implements.

 Tested using the mipsisa32-sde-elf target, with the mips-sim-sde32/-EB 
and mips-sim-sde32/-EL boards with no regressions.

2007-12-05  Maciej W. Rozycki  <macro@mips.com>

	* target.c (update_current_target): Initialize
	to_switch_to_thread.
	* target.h (target_switch_to_thread): New macro.
	* thread.c (switch_to_thread): Call target_switch_to_thread().

 OK to apply?

  Maciej

13327-1.diff
Index: binutils-quilt/src/gdb/target.h
===================================================================
--- binutils-quilt.orig/src/gdb/target.h	2007-12-03 14:28:21.000000000 +0000
+++ binutils-quilt/src/gdb/target.h	2007-12-03 14:28:24.000000000 +0000
@@ -393,6 +393,7 @@
     void (*to_mourn_inferior) (void);
     int (*to_can_run) (void);
     void (*to_notice_signals) (ptid_t ptid);
+    void (*to_switch_to_thread) (void);
     int (*to_thread_alive) (ptid_t ptid);
     void (*to_find_new_threads) (void);
     char *(*to_pid_to_str) (ptid_t);
@@ -878,6 +879,14 @@
 #define target_notice_signals(ptid) \
      (*current_target.to_notice_signals) (ptid)
 
+/* Switch to a selected thread.  */
+
+#define target_switch_to_thread()					\
+  do									\
+    if (current_target.to_switch_to_thread)				\
+      (*current_target.to_switch_to_thread) ();				\
+  while (0)
+
 /* Check to see if a thread is still alive.  */
 
 #define target_thread_alive(ptid) \
Index: binutils-quilt/src/gdb/thread.c
===================================================================
--- binutils-quilt.orig/src/gdb/thread.c	2007-12-03 14:20:57.000000000 +0000
+++ binutils-quilt/src/gdb/thread.c	2007-12-03 14:28:24.000000000 +0000
@@ -462,6 +462,7 @@
   inferior_ptid = ptid;
   reinit_frame_cache ();
   registers_changed ();
+  target_switch_to_thread ();
   stop_pc = read_pc ();
 }
 
Index: binutils-quilt/src/gdb/target.c
===================================================================
--- binutils-quilt.orig/src/gdb/target.c	2007-12-03 14:20:57.000000000 +0000
+++ binutils-quilt/src/gdb/target.c	2007-12-03 14:28:24.000000000 +0000
@@ -438,6 +438,7 @@
       INHERIT (to_mourn_inferior, t);
       INHERIT (to_can_run, t);
       INHERIT (to_notice_signals, t);
+      INHERIT (to_switch_to_thread, t);
       INHERIT (to_thread_alive, t);
       INHERIT (to_find_new_threads, t);
       INHERIT (to_pid_to_str, t);

Re: Target-specific thread switching backend

[Daniel Jacobowitz]

On Wed, Dec 05, 2007 at 02:13:18PM +0000, Maciej W. Rozycki wrote:
> Hello,
> 
>  In preparation for submitting support for the MDI target I propose the 
> following enhancement to target_switch_to_thread().  The MDI implements a 
> multi-threaded target which remembers the currently selected thread within 
> itself.  All the operations that have thread-local scope, such as 
> accessing of the CPU registers, use that thread identifier.  Therefore I 
> think it is reasonable to propagate the thread-switch event down to the 
> target, which is what the following change implements.

I don't see why this is necessary, or how it is safe.  At every
operation, the requested thread should be indicated by inferior_ptid
(or a ptid_t argument in some cases).  And switch_to_thread is usually
but not always called when inferior_ptid is changed, e.g.
regcache_raw_read.

-- 
Daniel Jacobowitz
CodeSourcery

Re: Target-specific thread switching backend

[Maciej W. Rozycki]

On Wed, 5 Dec 2007, Daniel Jacobowitz wrote:

> I don't see why this is necessary, or how it is safe.  At every
> operation, the requested thread should be indicated by inferior_ptid
> (or a ptid_t argument in some cases).  And switch_to_thread is usually
> but not always called when inferior_ptid is changed, e.g.
> regcache_raw_read.

 Well, I found it cleaner for the hardware to follow what GDB thinks that 
the current thread is, but I suppose I can make the switch be done in the 
calls that access registers and hardware breakpoints as appropriate only 
and cache the last seen value of inferior_ptid within the target code to 
make it reasonably fast.

 It sounds worrying though, that the notion of the current thread within 
GDB is changed freely somewhere when there is a dedicated call to switch 
threads, hmm...

  Maciej

Re: Target-specific thread switching backend

[Daniel Jacobowitz]

On Wed, Dec 05, 2007 at 03:22:34PM +0000, Maciej W. Rozycki wrote:
>  It sounds worrying though, that the notion of the current thread within 
> GDB is changed freely somewhere when there is a dedicated call to switch 
> threads, hmm...

Yes, especially when all the thread IDs should be really be arguments
to these functions.  For instance, the global current_regcache is
silly; we should support more than one regcache live at a time.

-- 
Daniel Jacobowitz
CodeSourcery