On Mon, 7 Nov 2011, Alan Modra wrote:
> On Mon, Nov 07, 2011 at 10:53:51AM +0100, Olivier Hainque wrote:
> > Another bug we're running into here is an unwarranted move of the sp
> > restore prior to register fetches despite an attempt at preventing that
> > with a stack_tie instruction (VxWorks target).
> >
> > http://gcc.gnu.org/ml/gcc/2011-03/msg00123.html
> >
> > The failure can still be exposed on mainline with a minor adjustment
> > to the C testcase quoted in the msg.
>
> Even after revision 181056?
FWIW, when I encountered such a problem in 4.4-based tools I found I
needed the following change to stack_tie patterns to fix it (applied to
and rediffed against trunk, but not tested there), in addition to a
DEFAULT_ABI == ABI_V4 check such as you added. Given all the confusion
about this class of bugs and whether they should be addressed in each back
end or in a target-independent way, I didn't try to send this upstream at
the time.
2011-04-11 Joseph Myers <jos...@codesourcery.com>
* config/rs6000/rs6000.md (stack_tie): Use (mem:BLK (scratch)) as
output of SET.
(restore_stack_block, restore_stack_nonlocal): Update UNSPEC_TIE
patterns.
Index: gcc/config/rs6000/rs6000.md
===================================================================
--- gcc/config/rs6000/rs6000.md (revision 181085)
+++ gcc/config/rs6000/rs6000.md (working copy)
@@ -11989,7 +11989,7 @@
(define_expand "restore_stack_block"
[(set (match_dup 2) (match_dup 3))
(set (match_dup 4) (match_dup 2))
- (set (match_dup 5) (unspec:BLK [(match_dup 5)] UNSPEC_TIE))
+ (set (mem:BLK (scratch)) (unspec:BLK [(match_dup 5)] UNSPEC_TIE))
(set (match_operand 0 "register_operand" "")
(match_operand 1 "register_operand" ""))]
""
@@ -12022,7 +12022,7 @@
[(set (match_dup 2) (match_operand 1 "memory_operand" ""))
(set (match_dup 3) (match_dup 4))
(set (match_dup 5) (match_dup 2))
- (set (match_dup 6) (unspec:BLK [(match_dup 6)] UNSPEC_TIE))
+ (set (mem:BLK (scratch)) (unspec:BLK [(match_dup 6)] UNSPEC_TIE))
(set (match_operand 0 "register_operand" "") (match_dup 3))]
""
"
@@ -15902,8 +15902,8 @@
; These are to explain that changes to the stack pointer should
; not be moved over stores to stack memory.
(define_insn "stack_tie"
- [(set (match_operand:BLK 0 "memory_operand" "+m")
- (unspec:BLK [(match_dup 0)] UNSPEC_TIE))]
+ [(set (mem:BLK (scratch))
+ (unspec:BLK [(match_operand:BLK 0 "memory_operand" "+m")] UNSPEC_TIE))]
""
""
[(set_attr "length" "0")])
--
Joseph S. Myers
jos...@codesourcery.com
