Fix xstormy16's eqbranchsi pattern
Nick Clifton
nickc@redhat.com
Wed Feb 6 10:36:00 GMT 2008
Hi Guys,
The xstormy16 port's eqbranchsi pattern has had problems because it
was using a match_dup inside a clobber and frequently an
optimization would change the register that was being duplicated
without changing the match_dup. So I am applying the patch below in
order to fix this in the same way that the problem was fixed for the
ineqbranchsi pattern.
Cheers
Nick
gcc/ChangeLog
2008-02-06 Nick Clifton <nickc@redhat.com>
* config/stormy16/stormy16.md (eqbranchsi): Replace a match_dup
inside the clobber with a match_operand and duplicated operand
number in the constraint.
(ineqbranchsi): Delete redundant comment.
Index: gcc/config/stormy16/stormy16.md
===================================================================
--- gcc/config/stormy16/stormy16.md (revision 132112)
+++ gcc/config/stormy16/stormy16.md (working copy)
@@ -880,11 +880,7 @@
(const_int 0)])
(label_ref (match_operand 0 "" ""))
(pc)))
-;; Although I would greatly like the 'match_dup' in the following line
-;; to actually be a register constraint, there is (at the time of writing) no
-;; way for reload to insert an output reload on the edges out of a branch.
-;; If reload is fixed to use insert_insn_on_edge, this can be changed.
- (clobber (match_dup 2))]
+ (clobber (match_operand:SI 3 "register_operand" "=2"))]
""
"*
{
@@ -902,10 +898,6 @@
"ri")])
(label_ref (match_operand 0 "" ""))
(pc)))
- ;; This clobber is problematic. Too many gcc optimizations will change
- ;; operand 2 without changing the clobber. At the time of writing there
- ;; is no way around this. :-( For an example try compiling:
- ;; gcc.c-torture/compile/20000403-1.c -O3 -fomit-frame-pointer -funroll-loops
(clobber (match_operand:SI 5 "register_operand" "=2"))
(clobber (match_operand:BI 4 "" "=&y"))]
""
More information about the Gcc-patches
mailing list