Re: [PATCH] add gdbarch_in_function_epilogue_p hook for sparc64

[jose dot marchesi at oracle dot com (Jose E. Marchesi)]

Hi Joel.  Thanks for the review!

    >         2013-10-16  Jose E. Marchesi  <jose.marchesi@oracle.com>
    >         
    >         	* sparc-tdep.c (sparc_in_function_epilogue_p): New function.
    >         	(X_RETTURN): New macro.
    >         	* sparc-tdep.h: sparc_in_function_epilogue_p prototype.
    >         
    >         	* sparc64-tdep.c (sparc64_init_abi): Hook
    >         	sparc_in_function_epilogue_p.
    
    Regarding testing this function on sparc32:
    
    Can you tell us which testcase in our testsuite this patch fixes?
    Although I am allowed to run the testsuite, I can still run individual
    testcases by hand (if not too complex, of course). Otherwise, would
    you have a small reproducer I could use to test on sparc32?

As mentioned in the original submission, the failing test was
gdb.base/watch-cond.exp.
    
    Comments about the patch below.
    
    > +/* Macros to identify some instructions.  */
    > +#define X_RETTURN(i) ((X_OP (i) == 0x2) && (X_OP3 (i) == 0x39))
    
    Can your comment say a little more precisely what instruction it
    identifies? I think the parens around the equality operators are
    superfluous and should be removed.

Fixed in the amended patch below.
   
    > +/* Return true if we are in a function's epilogue, i.e. after an
    > +   instruction that destroyed a function's stack frame.  */
    > +
    > +int
    > +sparc_in_function_epilogue_p (struct gdbarch *gdbarch, CORE_ADDR pc)
    > +{
    
    The general principle for function meant to be used as gdbarch callbacks
    is just to say:
    
    /* Implement "in_function_epilogue_p".  */
    
    That way, if we ever change the function's prototype, we don't have
    to update the documentation everywhere.  This callback should only
    be documented in gdbarch.sh (and repeated in gdbarch.h, which is
    generated from gdbarch.sh).

Fixed in the amended patch below.
    
    In your case, since there are 32bit and 64bit versions, you can add
    something like, for instance:
    
       This implementation works on both sparc32 and sparc64.

That fact is implicitly true for all functions defined in sparc-tdep.c
with prefix sparc_.
    
    > +  /* This function must return true if we are one instruction after an
    > +     instruction that destroyed the stack frame of the current
    > +     function.  The SPARC instructions used to restore the callers
    > +     stack frame are RESTORE and RETURN/RETT.
    > +
    > +     Of these RETURN/RETT is a branch instruction and thus we return
    > +     true if we are in its delay slot.
    > +
    > +     RESTORE is almost always found in the delay slot of a branch
    > +     instruction that transfers control to the caller, such as JMPL.
    > +     Thus the next instruction is in the caller frame and we don't
    > +     need to do anything about it.  */
    > +
    > +  unsigned int insn = sparc_fetch_instruction (pc - 4);  
    > +  return X_RETTURN (insn);
    
    Small quirk of the GDB Coding Style: We require an empty line between
    local variable declarations and the statements after. Also, I notice
    there are trailing spaces.

Fixed in the amended patch below.
    
    >    set_gdbarch_skip_prologue (gdbarch, sparc64_skip_prologue);
    >  
    > +  /* Detect whether PC is in function epilogue.  */
    > +  set_gdbarch_in_function_epilogue_p (gdbarch, sparc_in_function_epilogue_p);
    > +
    
    I would normally not comment on this, but since I've made other
    comments, I'll ask that the comment be revmoved, and that you
    avoid the empty line between the call to set_gdbarch_skip_prologue
    just above and the call to set_gdbarch_in_function_epilogue_p that
    you are adding.

Fixed in the amended patch below.    
    
2013-10-16  Jose E. Marchesi  <jose.marchesi@oracle.com>

	* sparc-tdep.c (sparc_in_function_epilogue_p): New function.
	(X_RETTURN): New macro.
	* sparc-tdep.h: sparc_in_function_epilogue_p prototype.

	* sparc64-tdep.c (sparc64_init_abi): Hook
	sparc_in_function_epilogue_p.


diff --git a/gdb/sparc-tdep.c b/gdb/sparc-tdep.c
index 38b345b..311a156 100644
--- a/gdb/sparc-tdep.c
+++ b/gdb/sparc-tdep.c
@@ -88,6 +88,9 @@ struct regset;
 #define X_DISP19(i) ((((i) & 0x7ffff) ^ 0x40000) - 0x40000)
 #define X_DISP10(i) ((((((i) >> 11) && 0x300) | (((i) >> 5) & 0xff)) ^ 0x200) - 0x200)
 #define X_SIMM13(i) ((((i) & 0x1fff) ^ 0x1000) - 0x1000)
+/* Macros to identify some instructions.  */
+/* RETURN (RETT in V8) */
+#define X_RETTURN(i) ((X_OP (i) == 0x2) && (X_OP3 (i) == 0x39))
 
 /* Fetch the instruction at PC.  Instructions are always big-endian
    even if the processor operates in little-endian mode.  */
@@ -452,6 +455,29 @@ sparc32_pseudo_register_write (struct gdbarch *gdbarch,
   regcache_raw_write (regcache, regnum + 1, buf + 4);
 }
 
+/* Implement "in_function_epilogue_p".  */
+
+int
+sparc_in_function_epilogue_p (struct gdbarch *gdbarch, CORE_ADDR pc)
+{
+  /* This function must return true if we are one instruction after an
+     instruction that destroyed the stack frame of the current
+     function.  The SPARC instructions used to restore the callers
+     stack frame are RESTORE and RETURN/RETT.
+
+     Of these RETURN/RETT is a branch instruction and thus we return
+     true if we are in its delay slot.
+
+     RESTORE is almost always found in the delay slot of a branch
+     instruction that transfers control to the caller, such as JMPL.
+     Thus the next instruction is in the caller frame and we don't
+     need to do anything about it.  */
+
+  unsigned int insn = sparc_fetch_instruction (pc - 4);
+
+  return X_RETTURN (insn);
+}
+
 
 static CORE_ADDR
 sparc32_frame_align (struct gdbarch *gdbarch, CORE_ADDR address)
diff --git a/gdb/sparc-tdep.h b/gdb/sparc-tdep.h
index b83d711..a065ebe 100644
--- a/gdb/sparc-tdep.h
+++ b/gdb/sparc-tdep.h
@@ -193,6 +193,9 @@ extern struct sparc_frame_cache *
 extern struct sparc_frame_cache *
   sparc32_frame_cache (struct frame_info *this_frame, void **this_cache);
 
+extern int
+  sparc_in_function_epilogue_p (struct gdbarch *gdbarch, CORE_ADDR pc);
+
 
 
 extern int sparc_software_single_step (struct frame_info *frame);
diff --git a/gdb/sparc64-tdep.c b/gdb/sparc64-tdep.c
index 52958df..9e4db3a 100644
--- a/gdb/sparc64-tdep.c
+++ b/gdb/sparc64-tdep.c
@@ -1196,6 +1196,7 @@ sparc64_init_abi (struct gdbarch_info info, struct gdbarch *gdbarch)
     (gdbarch, default_stabs_argument_has_addr);
 
   set_gdbarch_skip_prologue (gdbarch, sparc64_skip_prologue);
+  set_gdbarch_in_function_epilogue_p (gdbarch, sparc_in_function_epilogue_p);
 
   /* Hook in the DWARF CFI frame unwinder.  */
   dwarf2_frame_set_init_reg (gdbarch, sparc64_dwarf2_frame_init_reg);