https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64761
Bug ID: 64761
Summary: [4.9/5 Regression] -freorder-blocks-and-partition
causes some failures on SH
Product: gcc
Version: 5.0
Status: UNCONFIRMED
Keywords: ice-on-valid-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kkojima at gcc dot gnu.org
Target: sh*-*-*
For SH, we have two failures on testsuite
FAIL: gcc.dg/tree-prof/cold_partition_label.c compilation, -fprofile-use
-D_PROFILE_USE (internal compiler error)
FAIL: gcc.dg/tree-prof/va-arg-pack-1.c compilation, -fprofile-use
-D_PROFILE_USE (internal compiler error)
with -freorder-blocks-and-partition. Although both test cases
are added at 4.9, they don't fail with 4.8.4 compiler. So
this is a 4.9/5 regression from 4.8.
In both cases, (set (pc) (label_ref ...)) can't be recognized
because CROSSING_JUMP_P (insn) is set. I've added a new insn
(define_insn "*jump_compact_crossing"
[(set (pc)
(label_ref (match_operand 0 "" "")))]
"TARGET_SH1
&& flag_reorder_blocks_and_partition
&& CROSSING_JUMP_P (insn)"
{
/* The length is 16 if the delay slot is unfilled. */
return output_far_jump (insn, operands[0]);
}
[(set_attr "type" "jump")
(set_attr "length" "16")])
to sh.md. I guess that it's enough ATM, though it isn't optimal.
Then we get another failures for both tests like:
Error: displacement to defined symbol .L10 overflows 8-bit field
from assembler.
It looks almost wrong branches came from follow jump optimization.
Giving SH specific TARGET_CAN_FOLLOW_JUMP target macro as
static bool
sh_can_follow_jump (const rtx_insn *, const rtx_insn *)
{
if (TARGET_SH1
&& flag_reorder_blocks_and_partition
&& simplejump_p (branch2)
&& CROSSING_JUMP_P (branch2))
return false;
return true;
}
could fix them. Joern had already proposed
https://gcc.gnu.org/ml/gcc-patches/2013-04/msg01332.html
which replaces MD_REDIRECT_BRANCH with TARGET_CAN_FOLLOW_JUMP
and these changes could be merged. I'll attach the patches
for it.
