[Bug testsuite/47013] FAIL: gcc.dg/sms-*.c scan-rtl-dump-times sms "SMS succeeded" *
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47013 --- Comment #7 from revital eres 2010-12-27 14:06:43 UTC --- Created attachment 22853 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22853 SMS dump created with -r168082 on powerpc.
[Bug testsuite/47013] FAIL: gcc.dg/sms-*.c scan-rtl-dump-times sms "SMS succeeded" *
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47013 --- Comment #6 from revital eres 2010-12-27 14:00:54 UTC --- Looking more closely at sms-6.c it seems there is another issue which causes SMS to FAIL in the loop in foo function: Although the restrict keyword is attached to the pointers dependence edges are created between the write and read. (Please see the SMS dump file) The edges were not created in -r146278. I used -O2 -fmodulo-sched -fdump-rtl-sms -fno-auto-inc-dec -mno-update to compile on PowerPC.
[Bug testsuite/47013] FAIL: gcc.dg/sms-*.c scan-rtl-dump-times sms "SMS succeeded" *
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47013
--- Comment #4 from revital eres 2010-12-20 10:55:40
UTC ---
(In reply to comment #3)
> Thanks for looking at the problem. Do you understand why the tests pass on
> *86*-*-* and not on powerpc*-*-*?
Yes, the dump file is checked only for powerpc*-*-* and spu-*-*:
/* { dg-final { scan-rtl-dump-times "SMS succeeded" 1 "sms" { target
powerpc*-*-* spu-*-* } } } */
[Bug testsuite/47013] FAIL: gcc.dg/sms-*.c scan-rtl-dump-times sms "SMS succeeded" *
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47013 --- Comment #2 from revital eres 2010-12-20 09:26:21 UTC --- The testcases fail (even when adding the flags mentioned before) due to the decision to apply SMS when stage_count greater than 1. If stage_count is 1 it means the new SMSed kernel includes instructions from a single iteration of the original loop -- in that case SMS fails and leaves the scheduling task to the scheduler. In r146278 SMS has been applied also when SC equals 1 so the testcases passed with that version.
[Bug testsuite/47013] FAIL: gcc.dg/sms-*.c scan-rtl-dump-times sms "SMS succeeded" *
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47013 --- Comment #1 from revital eres 2010-12-20 09:07:53 UTC --- In some of the testcases additional flags are needed to make SMS apply: (though it does not mean it succeeds) sms-5.c: add -mno-update -fno-auto-inc-dec sms-4.c: -funsafe-loop-optimizations -fno-auto-inc-dec (-funsafe-loop-optimizations is related to PR32283 http://gcc.gnu.org/ml/gcc/2010-07/msg00379.html) sms-7.c: add -fno-auto-inc-dec
[Bug c/45834] New: Redundant inter-loop edges in DDG
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45834
Summary: Redundant inter-loop edges in DDG
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: e...@il.ibm.com
CC: b...@broadcom.com
Despite the patch to check alias sets in add_inter_loop_mem_dep function
(http://gcc.gnu.org/ml/gcc-patches/2009-01/msg00789.html); inter-loop edges are
created in the ddg of the loop below.
When changing the data types to int the edges are avoided.
Tested on PowerPC using GCC -r164646 with the flags:
-O3 -fmodulo-sched -fdump-rtl-sms -fmodulo-sched-allow-regmoves -c
-fno-auto-inc-dec
void
foo(unsigned char * __restrict__ dst,
unsigned char * __restrict__ src1,
unsigned char * __restrict__ src2)
{
int x, y;
for( x = 0; x < 100; x+=1)
{
dst[x] = ( src1[x] * src2[x]);
}
}
Snippet from SMS dump file which shows the edges:
SMS loop num: 1, file: test.c, line: 7
Node num: 0
(insn 62 61 63 3 (set (reg:QI 189 [ MEM[base: src2_11(D), index: ivtmp.15_29,
offset: 0B] ])
(mem:QI (plus:DI (reg/v/f:DI 188 [ src2 ])
(reg:DI 185 [ ivtmp.15 ])) [0 MEM[base: src2_11(D), index:
ivtmp.15_29, offset: 0B]+0 S1 A8])) test.c:10 395 {*movqi_internal}
(nil))
OUT ARCS: [62 -(T,2,0)-> 64]
IN ARCS: [66 -(T,3,1)-> 62] [65 -(T,14,1)-> 62]
Node num: 1
(insn 63 62 64 3 (set (reg:QI 190 [ MEM[base: src1_7(D), index: ivtmp.15_29,
offset: 0B] ])
(mem:QI (plus:DI (reg/v/f:DI 187 [ src1 ])
(reg:DI 185 [ ivtmp.15 ])) [0 MEM[base: src1_7(D), index:
ivtmp.15_29, offset: 0B]+0 S1 A8])) test.c:10 395 {*movqi_internal}
(nil))
OUT ARCS: [63 -(T,2,0)-> 64]
IN ARCS: [66 -(T,3,1)-> 63] [65 -(T,14,1)-> 63]
...
Node num: 3
(insn 65 64 66 3 (set (mem:QI (plus:DI (reg/v/f:DI 186 [ dst ])
(reg:DI 185 [ ivtmp.15 ])) [0 MEM[base: dst_4(D), index:
ivtmp.15_29, offset: 0B]+0 S1 A8])
(subreg:QI (reg:SI 191) 3)) test.c:10 395 {*movqi_internal}
(expr_list:REG_DEAD (reg:SI 191)
(nil)))
OUT ARCS: [65 -(O,0,0)-> 80] [65 -(T,14,1)-> 63] [65 -(T,14,1)-> 62]
IN ARCS: [66 -(T,3,1)-> 65] [64 -(T,9,0)-> 65]
