PLEASE APPROVE: patch to fix PR 5879, a serious gcc 3.1 regression

[Jason Merrill]

OK.  The problem seems to be a clash between the sibcall code and the EH
code.  With the patch, looking at the 00.rtl dump, the end of
initializeClass looks like

-----
(call_insn 835 825 836 (call_placeholder 833 826 0 429 (call_insn 834 833 0 (parallel[ 
                (call (mem:SI (symbol_ref:SI ("_Jv_MonitorExit")) [0 S4
                A32])
                    (const_int 0 [0x0]))
                (use (const_int 0 [0x0]))
                (clobber (reg:SI 14 lr))
            ] ) -1 (nil)
        (nil)
        (expr_list (use (reg:SI 0 r0))
            (nil)))) -1 (nil)
    (nil)
    (nil))

(note 836 835 837 0x408202c0 NOTE_INSN_BLOCK_END)

(jump_insn 837 836 838 (set (pc)
        (label_ref 841)) -1 (nil)
    (nil))

(barrier 838 837 839)

(jump_insn 839 838 840 (set (pc)
        (label_ref 844)) -1 (nil)
    (nil))

(barrier 840 839 841)

(code_label 841 840 842 414 "" "" [0 uses])

(note 842 841 844 NOTE_INSN_FUNCTION_END)

(code_label 844 842 0 387 "" "" [0 uses])
-----

without the patch, it looks like

-----
(call_insn 835 825 836 (call_placeholder 833 826 0 429 (call_insn 834 833 0 (parallel[ 
                (call (mem:SI (symbol_ref:SI ("_Jv_MonitorExit")) [0 S4 A32])
                    (const_int 0 [0x0]))
                (use (const_int 0 [0x0]))
                (clobber (reg:SI 14 lr))
            ] ) -1 (nil)
        (nil)
        (expr_list (use (reg:SI 0 r0))
            (nil)))) -1 (nil)
    (nil)
    (nil))

(note 836 835 837 0x40806ec0 NOTE_INSN_BLOCK_END)

(note 837 836 839 NOTE_INSN_FUNCTION_END)

(code_label 839 837 0 360 "" "" [0 uses])
-----

note that these two are equivalent; the first just has an extra jump; in
both cases we fall through to the end of the function.  However, that extra
jump seems to be enough to prevent the sibcall code from turning the call
to _Jv_MonitorExit into a sibcall; without the jump, it does.  After the
sibcall pass, the EH pass adds the call to _Unwind_SjLj_Unregister; but by
this time it's too late, as we never return from _Jv_MonitorExit.

Seems to me that the solution is to perform EH processing before the
sibcall pass; this also allows us to sibcallize the call to
_Unwind_SjLj_Unregister.  Richard, what do you think?

Adam, Anthony, does this fix things for you?

2002-03-19  Jason Merrill  <jason@redhat.com>

	* toplev.c (rest_of_compilation): Do EH processing before
	sibcalls.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: text/x-patch
Size: 1974 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20020319/5154584f/attachment.bin>