For this code, at -O2:
extern char *x;
extern void foo (void);
void f (char *s, char *se, char *mp, char *y)
{
while (s != se)
{
char *p;
foo ();
p = s + *mp;
*y++ = *p;
s = p;
}
x = s;
}
I get with LAST_UPDATED "Tue Aug 16 12:30:26 UTC 2005":
/tmp/b.c:15: error: insn does not satisfy its constraints:
(insn 74 26 29 2 (parallel [
(set (reg:QI 13 r13)
(mem:QI (plus:SI (sign_extend:SI (mem:QI (reg/v/f:SI 3 r3
[orig:31 mp ] [31]) [0 S1 A8]))
(reg/v/f:SI 0 r0 [orig:29 s ] [29])) [0 S1 A8]))
(set (reg/v/f:SI 9 r9 [orig:28 p ] [28])
(plus:SI (sign_extend:SI (mem:QI (reg/v/f:SI 3 r3 [orig:31 mp ]
[31]) [0 S1 A8]))
(reg/v/f:SI 0 r0 [orig:29 s ] [29])))
]) 18 {*mov_sideqi} (nil)
(expr_list:REG_DEAD (reg/v/f:SI 0 r0 [orig:29 s ] [29])
(nil)))
/tmp/b.c:15: internal compiler error: in copyprop_hardreg_forward_1, at
regrename.c:1578
At first sight, there's a cris.md bug that has been there forever:
the insn is supposed to match "*mov_side<mode>", but that can only happen
with help of the "%" commutative marker. That marker isn't handled after
reload, and the insn is generated by the moverside peephole2 pattern.
All patterns *have* to be able to match whatever (canonical) RTL is generated
by the target after reload, without the commutativity marker.
So I'd say it's a target bug (the generated RTL is canonical) and IMHO arguably
also a documentation bug. But there's more...
--
Summary: cris.md rtl canonicalization bug
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P2
Component: target
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: hp at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: cris-axis-elf
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23424
