------- Comment #4 from ramana at gcc dot gnu dot org 2010-04-09 07:38 -------
It appears as though there's a latent bug in arm_ccfsm_state_machine. If you
mark this correctly as being "predicable" which the rev insn is - the bug goes
away.
If arm_final_prescan_insn sees something of the following form, it goes ahead
and assumes it can convert the rev in this pattern into a condexec and removes
the branch.
(jump_insn:TI 12 15 19 /tmp/bitswap.c:23 (set (pc)
(if_then_else (eq (reg:CC 24 cc)
(const_int 0 [0x0]))
(label_ref:SI 23)
(pc))) 228 {*arm_cond_branch} (expr_list:REG_DEAD (reg:CC 24 cc)
(expr_list:REG_BR_PROB (const_int 3900 [0xf3c])
(nil)))
-> 23)
(note 19 12 21 [bb 3] NOTE_INSN_BASIC_BLOCK)
(insn:TI 21 19 22 /tmp/bitswap.c:23 (set (reg:SI 2 r2 [orig:141 D.4313 ] [141])
(bswap:SI (reg:SI 2 r2 [152]))) 347 {arm_rev} (nil))
(insn:TI 22 21 23 /tmp/bitswap.c:23 (set (reg/v:DI 2 r2 [orig:133 __res ]
[133])
(zero_extend:DI (reg:SI 2 r2 [orig:141 D.4313 ] [141]))) 138
{*arm_zero_extendsidi2} (expr_list:REG_UNUSED (reg:SI 3 r3)
(nil)))
The patch below which I'm testing, fixes the immediate problem with rev as
below. Longer term we ought to remove the ccfsm_state_machine once we fix the
problem with conditional calls.
Index: arm.md
===================================================================
--- arm.md (revision 158138)
+++ arm.md (working copy)
@@ -11201,8 +11201,9 @@
[(set (match_operand:SI 0 "s_register_operand" "=r")
(bswap:SI (match_operand:SI 1 "s_register_operand" "r")))]
"TARGET_EITHER && arm_arch6"
- "rev\t%0, %1"
- [(set (attr "length")
+ "rev%?\t%0, %1"
+ [(set_attr "predicable" "yes")
+ (set (attr "length")
(if_then_else (eq_attr "is_thumb" "yes")
(const_int 2)
(const_int 4)))]
Ramana
--
ramana at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |ramana at gcc dot gnu dot
|dot org |org
Status|NEW |ASSIGNED
Last reconfirmed|2010-04-09 07:02:26 |2010-04-09 07:38:25
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43698
