http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54762
Bug #: 54762
Summary: [SH] Utilize zero-displacement branches for
conditional far branches
Classification: Unclassified
Product: gcc
Version: 4.8.0
Status: UNCONFIRMED
Severity: enhancement
Priority: P3
Component: target
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: olege...@gcc.gnu.org
Target: sh*-*-*
When the displacement of a conditional branch is too small the branch is
'widened', which results in code like this (a random snippet from CSiBE):
mov.l r8,@-r15
sts.l pr,@-r15
mov.l @(32,r4),r8
tst r8,r8
bt .L293
mov.l @r8,r1
cmp/eq r4,r1
bt .L295 <<<
.L293:
bra .L297 <<<
mov #-2,r6 <<<
.L295:
mov.l @(4,r8),r2
< lots of code >
.L297:
In such cases zero-displacement branches could be utilized to branch around the
far branch. For example:
mov.l r8,@-r15
sts.l pr,@-r15
mov.l @(32,r4),r8
tst r8,r8
bf 0f
bra .L297
0: mov #-2,r6
mov.l @r8,r1
cmp/eq r4,r1
bt 0f
bra .L297
0: mov #-2,r6
mov.l @(4,r8),r2
< lots of code >
.L297:
Maybe this could be achieved by adding variations to the existing "branch_true"
and "branch_false" patterns like:
(define_insn "branch_true_far"
[(set (pc) (if_then_else (ne (match_operand 1 "t_reg_operand" "")
(const_int 0))
(label_ref (match_operand 0 "" ""))
(pc)))]
"TARGET_SH1"
{
return "bf 0f" "\n"
" bra %l0" "\n"
"0: %#";
}
[(set_attr "type" "cbranch")
(set_attr "needs_delay_slot" "yes")
(set_attr "length" "4")]) ;; not sure whether length should be 4 or 6
;; in this case...
