Sorry, the original patch has problems. Because the
Update_loops_for_mainopt() is also called by
Update_loops_for_preopt(). My patch will reset these fields in preopt,
which will cause a segmentation fault in preopt since the field is
NULL.
I made a new patch for this problem. The original
Update_loops_for_mainopt() is renamed to Update_loops_for_allopt() and
called by both Update_loops_for_mainopt() and
Update_loops_for_preopt(). A new function Update_loops_for_mainopt()
is created to call Update_loops_for_allopt() and reset the high-level
SCF fields.
Could a gatekeeper review this again? Thank you very much.
Index: opt_cfg.cxx
===================================================================
--- opt_cfg.cxx (revision 3687)
+++ opt_cfg.cxx (working copy)
@@ -5227,7 +5227,7 @@
// the trip count exit is at the first block of the loop or the last block
// of the loop. Also compute the loop may exit early conservatively.
//
-void Update_loops_for_mainopt(BB_LOOP *loop)
+static void Update_loops_for_allopt(BB_LOOP *loop)
{
// loop->Header() must be defined
@@ -5303,7 +5303,7 @@
//
void Update_loops_for_preopt(BB_LOOP *loop)
{
- Update_loops_for_mainopt(loop);
+ Update_loops_for_allopt(loop);
// Fix up the loop attributes if it wasn't a natural loop.
//
@@ -5318,6 +5318,22 @@
! loop->Dotail()->Postdominates_strictly(loop->Body()));
}
+// If we are in mainopt, reset the high-level SCF related field:
+// 1. these fields are not used in mainopt and later phases
+// 2. the mainopt CFG transformation will break high-level SCF assumptions
+//
+void Update_loops_for_mainopt(BB_LOOP *loop)
+{
+ Update_loops_for_allopt(loop);
+
+ // reset high SCF fields if we are in WOPT
+ loop->Set_start(NULL);
+ loop->Set_end(NULL);
+ loop->Set_body(NULL);
+ loop->Set_step(NULL);
+ loop->Set_merge(NULL);
+}
+
// A BB is a loop header block if it dominates the head of one of
// its incoming edge.
static bool
2011/7/5 Jian-Xin Lai <laij...@gmail.com>:
> Hi,
>
> Can someone review the patch for #829:
> https://bugs.open64.net/show_bug.cgi?id=829
>
> There is a WHILE_DO loop in the case. In CFG construction, a BB_LOOP
> will be created and attached to the following BB_NODE:
> _start: BB6 (empty)
> _body: BB7 (label for loopback of the WHILE_DO)
> _cond: BB10 (cond and TRUEBR to BB7)
>
> The CFG is:
> BB1
> |
> BB2
> / \
> +-----> BB3 BB4
> | \ /
> | BB5 ---> BB12
> | |
> | BB6
> | |
> | BB7 <-------+
> | | |
> +--------- BB8 |
> | |
> BB9 |
> | |
> BB10 ------+
> |
> BB11
> |
> BB12 <-- BB5
> |
> BB13 (RETURN)
>
> After MainOPT CFG transformation, BB3 and BB5 is cloned and new CFG is
> BB1
> |
> BB2
> / \
> BB14 BB4
> | |
> BB12 <-- BB15 BB17 --> BB12
> | |
> BB18 BB19
> \ /
> +-----------> BB6
> | |
> | BB7 <-------+
> | | |
> BB5 <-BB3 <- BB8 |
> | |
> BB9 |
> | |
> BB10 ------+
> |
> BB11
> |
> BB12 <-- BB5, BB15, BB17
> |
> BB13 (RETURN)
>
> There will be two loops (6, 7, 8, 3, 5, 9, 10) and (7, 8, 9, 10).
>
> In CFG::Analyze_loops() in opt_cfg.cxx, when call
> Fix_SCF_for_mainopt() on BB7, it will set its start/end/step/merge
> with the same BB_LOOP*:
> 5466 BB_LOOP *loop = bb->Loop();
> 5467 if (loop != NULL && loop->Header() == bb) {
> 5468 if (loop->End() != NULL) {
> 5469 if(loop->Start())
> 5470 loop->Start()->Set_loop(loop);
> 5471 else {
> 5472 Is_True(loop->Is_flag_set(LOOP_PRE_WHILE),("wrong
> non bottom test loop lowering\n"));
> 5473 }
> 5474 loop->End()->Set_loop(loop);
> 5475 loop->Body()->Set_loop(loop);
> 5476 if (loop->Step())
> 5477 loop->Step()->Set_loop(loop);
>
> After calling this function, BB6 and BB7 will share the same BB_LOOP
> info. In Find_real_loops(), Collect_loop_body() will be called on both
> BB6 and BB7 with the same BB_LOOP*. BB10 will be added to the
> BB_LOOP's True_body_set twice. An assertion will happen in the second
> time.
>
> In existing code, we already cleaned some of the fields, like the
> child, parent, preheader, tail, loopback, etc
> 5397 static BB_LOOP *
> 5398 Allocate_loop(BB_NODE *bb, BB_LOOP *parent, CFG *cfg)
> 5399 {
> 5400 BB_LOOP *loop = bb->Loop();
> 5401 if (loop == NULL)
> 5402 loop = CXX_NEW( BB_LOOP(NULL, NULL, NULL, NULL, NULL, NULL),
> 5403 cfg->Mem_pool() );
> 5404 loop->Set_true_body_set(CXX_NEW(BB_NODE_SET(cfg->Total_bb_count(),
> cfg,
> 5405 cfg->Mem_pool(),
> BBNS_EMPTY),
> 5406 cfg->Mem_pool()));
> 5407 loop->Set_body_set(NULL); // body_set not supported in mainopt
> 5408 loop->Set_header(bb);
> 5409 loop->Set_child(NULL);
> 5410 loop->Set_parent(NULL);
> 5411 loop->Set_depth(0);
> 5412 loop->Set_max_depth(0);
> 5413 Update_loops_for_mainopt(loop);
> 5414 return loop;
> 5415 }
>
> 5230 void Update_loops_for_mainopt(BB_LOOP *loop)
> 5231 {
> 5232 // loop->Header() must be defined
> 5233
> 5234 loop->Set_loopback(NULL);
> 5235 loop->Set_preheader(NULL);
> 5236 loop->Set_tail(NULL);
> 5237 loop->Reset_well_formed();
> 5238
> 5239 loop->Set_test_at_entry(false);
> 5240 loop->Set_test_at_exit(false);
> 5241 loop->Set_exit_early(true);
> 5242
>
> To fix this issue, I think we need to clear all necessary field in
> BB_LOOP when we call Update_loops_for_mainopt():
> Index: ../osprey/be/opt/opt_cfg.cxx
> ===================================================================
> --- ../osprey/be/opt/opt_cfg.cxx (revision 3648)
> +++ ../osprey/be/opt/opt_cfg.cxx (working copy)
> @@ -5231,6 +5231,12 @@
> {
> // loop->Header() must be defined
>
> + loop->Set_start(NULL);
> + loop->Set_end(NULL);
> + loop->Set_body(NULL);
> + loop->Set_step(NULL);
> + loop->Set_merge(NULL);
> +
> loop->Set_loopback(NULL);
> loop->Set_preheader(NULL);
> loop->Set_tail(NULL);
>
>
> --
> Regards,
> Lai Jian-Xin
>
--
Regards,
Lai Jian-Xin
------------------------------------------------------------------------------
All of the data generated in your IT infrastructure is seriously valuable.
Why? It contains a definitive record of application performance, security
threats, fraudulent activity, and more. Splunk takes this data and makes
sense of it. IT sense. And common sense.
http://p.sf.net/sfu/splunk-d2d-c2
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
