http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47918
Summary: [4.6 regression] noreturn discovery broke non local
gotos on m68k
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: mi...@it.uu.se
gcc.dg/torture/stackalign/non-local-goto-4.c and several similar test cases
fail with current 4.6 on m68k-linux. Bisection identified r160124 as the
cause:
Author: hubicka
Date: Tue Jun 1 21:55:49 2010
New Revision: 160124
URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=160124
Log:
* ipa-pure-const.c (local_pure_const): Do NORETURN discovery.
The essence of non-local-goto-4.c is a function x() with a local self-recursive
function y() and a non-local goto from y() to x():
x(a)
{
__label__ xlab;
void y(a)
{
if (a==0)
goto xlab;
y (a-1);
}
y (a);
xlab:;
return a;
}
With r160123, gcc -O3 -fomit-frame-pointer -S generates this for x() and y():
.text
.align 2
.type y.1201, @function
y.1201:
link.w %fp,#0
move.l #.L2,%a1
move.l (%a0),%fp
move.l 4(%a0),%sp
jmp (%a1)
.size y.1201, .-y.1201
.align 2
.globl x
.type x, @function
x:
lea (-12,%sp),%sp
fmovem #252,-(%sp)
movem.l #16190,-(%sp)
lea (128,%sp),%a0
move.l %a0,116(%sp)
move.l %sp,120(%sp)
move.l 132(%sp),-(%sp)
lea (120,%sp),%a0
jsr y.1201
addq.l #4,%sp
.L4:
move.l 132(%sp),%d0
movem.l (%sp)+,#31996
fmovem (%sp)+,#63
lea (12,%sp),%sp
rts
.L2:
move.l 132(%sp),%d0
movem.l (%sp)+,#31996
fmovem (%sp)+,#63
lea (12,%sp),%sp
rts
.size x, .-x
This works. But r160124 changes the code to:
.align 2
.type y.1201, @function
y.1201:
link.w %fp,#0
move.l #.L2,%a1
move.l (%a0),%fp
move.l 4(%a0),%sp
jmp (%a1)
.size y.1201, .-y.1201
.align 2
.globl x
.type x, @function
x:
lea (-12,%sp),%sp
fmovem #252,-(%sp)
movem.l #16190,-(%sp)
lea (128,%sp),%a0
move.l %a0,116(%sp)
move.l %sp,120(%sp)
move.l 132(%sp),-(%sp)
lea (120,%sp),%a0
jsr y.1201
.L2:
.L4:
move.l 136(%sp),%d0
movem.l (%sp)+,#31996
fmovem (%sp)+,#63
lea (12,%sp),%sp
rts
.size x, .-x
A diff shows that L2 and L4 have been merged, a stack increment before L4 has
been removed and the stack offset for loading %d0 in L4 has been adjusted for
that. However, the path leading up to L2 is the same as before, but L2 is now
the same as L4 so uses an incorrect (off-by-4) stack offset when fetching %d2.
--- non-local-goto-4.s-r160123 2011-02-27 23:44:28.000000000 +0100
+++ non-local-goto-4.s-r160124 2011-02-27 23:36:27.000000000 +0100
@@ -23,15 +23,9 @@
move.l 132(%sp),-(%sp)
lea (120,%sp),%a0
jsr y.1201
- addq.l #4,%sp
-.L4:
- move.l 132(%sp),%d0
- movem.l (%sp)+,#31996
- fmovem (%sp)+,#63
- lea (12,%sp),%sp
- rts
.L2:
- move.l 132(%sp),%d0
+.L4:
+ move.l 136(%sp),%d0
movem.l (%sp)+,#31996
fmovem (%sp)+,#63
lea (12,%sp),%sp
