Improve arm_skip_prologue by using arm_analyze_prologue
authorYao Qi <yao@codesourcery.com>
Thu, 27 Nov 2014 03:37:54 +0000 (11:37 +0800)
committerYao Qi <yao@codesourcery.com>
Fri, 12 Dec 2014 00:46:34 +0000 (08:46 +0800)
commit21daaaaffcbda47b724858dd99ee2082043ef2da
tree01b636ad807269afc219b6d4dc3f9f60416c0412
parentf303bc3e6ca29f0413376e38164dc5cdc0893d4b
Improve arm_skip_prologue by using arm_analyze_prologue

Hi,
I see many fails in dw2-dir-file-name.exp on arm target when test
case is compiled with -marm, however, these fails are disappeared when
test case is compiled with -mthumb.

The difference of pass and fail shown below is that "0x000085d4 in" isn't
printed out, but test case expects to see it.

-Breakpoint 2, compdir_missing__ldir_missing__file_basename () at tmp-dw2-dir-file-name.c:999^M
-(gdb) FAIL: gdb.dwarf2/dw2-dir-file-name.exp: compdir_missing__ldir_missing__file_basename: continue to breakpoint: compdir_missing__ldir_missing__file_basename
+Breakpoint 2, 0x000085d4 in compdir_missing__ldir_missing__file_basename () at tmp-dw2-dir-file-name.c:999^M
+(gdb) PASS: gdb.dwarf2/dw2-dir-file-name.exp: compdir_missing__ldir_missing__file_basename: continue to breakpoint: compdir_missing__ldir_missing__file_basename

This difference is caused by setting breakpoint at the first instruction
in the function (actually, the first instruction in prologue, at [1]),
so that frame_show_address returns false, and print_frame doesn't print the
address.

   0x00008620 <+0>:     push    {r11}           ; (str r11, [sp, #-4]!)  <--[1]
   0x00008624 <+4>:     add     r11, sp, #0
   0x00008628 <+8>:     ldr     r3, [pc, #24]   ; 0x8648 <compdir_missing__ldir_missing__file_basename+40>
   0x0000862c <+12>:    ldr     r3, [r3]
   0x00008630 <+16>:    add     r3, r3, #1
   0x00008634 <+20>:    ldr     r2, [pc, #12]   ; 0x8648 <compdir_missing__ldir_missing__file_basename+40>

Then, it must be the arm_skip_prologue's fault that unable to skip
instructions in prologue.  At the end of arm_skip_prologue, it matches
several instructions, such as "str  r(0123),[r11,#-nn]" and
"str  r(0123),[sp,#nn]", but "push {r11}" isn't handled.

These instruction matching code in arm_skip_prologue, which can be regarded
as leftover of development for many years, should be merged to
arm_analyze_prologue and use arm_analyze_prologue in arm_skip_prologue.
Here is the something like the history of arm_{skip,scan,analyze}_prologue.
Around 2002, there are arm_skip_prologue and arm_scan_prologue, but code are
duplicated to some extent.  When match an instruction, both functions should
be modified, for example in Michael Snyder's patch
https://sourceware.org/ml/gdb-patches/2002-05/msg00205.html and Michael
expressed the willingness to merge both into one.  Daniel added code call
thumb_analyze_prologue in arm_skip_prologue in 2006, but didn't handle its
counterpart arm_analyze_prologue, which is added in 2010
<https://sourceware.org/ml/gdb-patches/2010-03/msg00820.html>
however, the instructions matching at the bottom of arm_skip_prologue wasn't
cleaned up.  This patch is to merge them into arm_analyze_prologue.

gdb:

2014-12-12  Yao Qi  <yao@codesourcery.com>

PR tdep/14261
* arm-tdep.c (arm_skip_prologue): Remove unused local variable
'skip_pc'.  Remove code skipping prologue instructions, use
arm_analyze_prologue instead.
(arm_analyze_prologue): Stop the scanning for unrecognized
instruction when skipping prologue.
gdb/ChangeLog
gdb/arm-tdep.c