------- Comment #7 from kkojima at gcc dot gnu dot org 2010-04-16 22:34 -------
Created an attachment (id=20404)
--> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=20404&action=view)
A bit reduced test case
It seems that we see a yet another corner case for find_barrier.
Also it looks latent on the trunk and 4.5. Here is what is going
on, I guess.
For PIC, sh.c:fixup_mova replaces casesi_worker_1 insn with
casesi_worker_2 insn and a constant table entry is added to
a constant pool. In the problematic case, find_barrier inserts
this constant pool in the middle of insns for casesi_worker_2.
Then the constant pool including the above constant table entry
is put before some of those insns.
This results a code sequence like
mov.l .L94,r0 ! insn for casesi_worker_2
...
CP:
.long bar
...
.L94:
.long .L24-.L86
.align 5
.L87:
add r0,r1 ! insn for casesi_worker_2
mova .L86,r0 ! insn for casesi_worker_2
mov.b @(r0,r1),r1 ! insn for casesi_worker_2
...
.L24:
.byte .L22-.L25
.byte .L23-.L25
...
otherCP:
.L86:
which looks unexpected to the range computation in find_barrier.
It seems that the constant entry .long .L24-.L86 was assumed to
be put after the label .L86. The patch below is an experimental
one to avoid the issue, though there would be better ways.
--- ORIG/gcc-4_4-branch/gcc/config/sh/sh.c 2010-04-15 19:59:21.000000000
+0900
+++ gcc-4_4-branch/gcc/config/sh/sh.c 2010-04-16 19:46:26.000000000 +0900
@@ -3884,6 +3884,7 @@ find_barrier (int num_mova, rtx mova, rt
int si_limit;
int hi_limit;
rtx orig = from;
+ rtx last_symoff = NULL_RTX;
/* For HImode: range is 510, add 4 because pc counts from address of
second instruction after this one, subtract 2 for the jump instruction
@@ -4013,8 +4014,17 @@ find_barrier (int num_mova, rtx mova, rt
if (mova_p (from))
{
+ rtx src;
+
switch (untangle_mova (&num_mova, &mova, from))
{
+ case 1:
+ src = SET_SRC (PATTERN (from));
+ if (GET_CODE (src) == CONST
+ && GET_CODE (XEXP (src, 0)) == UNSPEC
+ && XINT (XEXP (src, 0), 1) == UNSPEC_SYMOFF)
+ last_symoff = from;
+ break;
case 0: return find_barrier (0, 0, mova);
case 2:
{
@@ -4120,6 +4130,9 @@ find_barrier (int num_mova, rtx mova, rt
so we'll make one. */
rtx label = gen_label_rtx ();
+ if (last_symoff)
+ from = last_symoff;
+
/* If we exceeded the range, then we must back up over the last
instruction we looked at. Otherwise, we just need to undo the
NEXT_INSN at the end of the loop. */
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43744
