[PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[Feng Xue OS]

[-- Attachment #1: Type: text/plain, Size: 529 bytes --]

When checking a self-recursively generated value for aggregate jump
function, wrong aggregate lattice was used, which will cause infinite
constant propagation. This patch is composed to fix this issue. 

2020-01-03  Feng Xue  <fxue@os.amperecomputing.com>

        PR ipa/93084
        * ipa-cp.c (self_recursively_generated_p): Find matched aggregate
        lattice for a value to check.
        (propagate_vals_across_arith_jfunc): Add an assertion to ensure
        finite propagation in self-recursive scc.

Feng

[-- Attachment #2: 0001-Fix-pr93084.patch --]
[-- Type: application/octet-stream, Size: 2545 bytes --]

From e04cd6f51e2745337c57b6a2fabe0c56d8be4aa1 Mon Sep 17 00:00:00 2001
From: Feng Xue <fxue@os.amperecomputing.com>
Date: Fri, 3 Jan 2020 22:00:55 +0800
Subject: [PATCH] Fix pr93084

---
 gcc/ipa-cp.c                           | 21 +++++++++++--
 gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c | 42 ++++++++++++++++++++++++++
 2 files changed, 61 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c

diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
index 4381b35a809..11631986293 100644
--- a/gcc/ipa-cp.c
+++ b/gcc/ipa-cp.c
@@ -1917,10 +1917,25 @@ self_recursively_generated_p (ipcp_value<tree> *val)
 
       class ipcp_param_lattices *plats = ipa_get_parm_lattices (info,
 								src->index);
-      ipcp_lattice<tree> *src_lat = src->offset == -1 ? &plats->itself
-						      : plats->aggs;
+      ipcp_lattice<tree> *src_lat;
       ipcp_value<tree> *src_val;
 
+      if (src->offset == -1)
+	src_lat = &plats->itself;
+      else
+	{
+	  struct ipcp_agg_lattice *src_aglat;
+
+	  for (src_aglat = plats->aggs; src_aglat; src_aglat = src_aglat->next)
+	    if (src_aglat->offset == src->offset)
+	      break;
+
+	  if (!src_aglat)
+	    return false;
+
+	  src_lat = src_aglat;
+	}
+
       for (src_val = src_lat->values; src_val; src_val = src_val->next)
 	if (src_val == val)
 	  break;
@@ -2017,6 +2032,8 @@ propagate_vals_across_arith_jfunc (cgraph_edge *cs,
 	    val_seeds.safe_push (src_val);
 	}
 
+      gcc_assert ((int) val_seeds.length () <= param_ipa_cp_value_list_size);
+
       /* Recursively generate lattice values with a limited count.  */
       FOR_EACH_VEC_ELT (val_seeds, i, src_val)
 	{
diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c b/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c
new file mode 100644
index 00000000000..18d29bdd0b3
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c
@@ -0,0 +1,42 @@
+/* { dg-do compile } */
+/* { dg-options "-O3 -fdump-ipa-cp-details -fno-early-inlining --param ipa-cp-max-recursive-depth=8 --param ipa-cp-eval-threshold=1" } */
+
+struct V {
+  int f0;
+  int f1;
+};
+
+int data[100];
+
+int fn ();
+
+int recur_fn (struct V * __restrict v)
+{
+  int i = v->f0;
+  int j = v->f1;
+  struct V t;
+
+  if (j > 100)
+    {
+      fn ();
+      return 1;
+    }
+
+  data[i] = i;
+
+  t.f0 = i - 2;
+  t.f1 = j + 1;
+
+  recur_fn (&t);
+
+  return i * j;
+}
+
+int main ()
+{
+  struct V v = {1, 3};
+
+  return recur_fn (&v);
+}
+
+/* { dg-final { scan-ipa-dump-times "Creating a specialized node of recur_fn/\[0-9\]*\\." 8 "cp" } } */
-- 
2.17.1

Re: [PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[Martin Jambor]

Hi,

On Fri, Jan 03 2020, Feng Xue OS wrote:
> When checking a self-recursively generated value for aggregate jump
> function, wrong aggregate lattice was used, which will cause infinite
> constant propagation. This patch is composed to fix this issue. 
>
> 2020-01-03  Feng Xue  <fxue@os.amperecomputing.com>
>
>         PR ipa/93084
>         * ipa-cp.c (self_recursively_generated_p): Find matched aggregate
>         lattice for a value to check.
>         (propagate_vals_across_arith_jfunc): Add an assertion to ensure
>         finite propagation in self-recursive scc.

as far as I am concerned, the patch looks good.

Thanks,

Martin

Re: [PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[Christophe Lyon]

On Tue, 7 Jan 2020 at 14:18, Martin Jambor <mjambor@suse.cz> wrote:
>
> Hi,
>
> On Fri, Jan 03 2020, Feng Xue OS wrote:
> > When checking a self-recursively generated value for aggregate jump
> > function, wrong aggregate lattice was used, which will cause infinite
> > constant propagation. This patch is composed to fix this issue.
> >
> > 2020-01-03  Feng Xue  <fxue@os.amperecomputing.com>
> >
> >         PR ipa/93084
> >         * ipa-cp.c (self_recursively_generated_p): Find matched aggregate
> >         lattice for a value to check.
> >         (propagate_vals_across_arith_jfunc): Add an assertion to ensure
> >         finite propagation in self-recursive scc.
>
> as far as I am concerned, the patch looks good.

Hi,
The new test introduced by this patch (gcc.dg/ipa/ipa-clone-3.c) fails
on arm, powerpc and mips according to gcc-testresults.
FAIL: gcc.dg/ipa/ipa-clone-3.c scan-ipa-dump-times cp "Creating a
specialized node of recur_fn/[0-9]*\\." 8

Can you have a look?

Thanks,

Christophe

>
> Thanks,
>
> Martin

Re: [PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[Martin Jambor]

Hi,

On Thu, Jan 09 2020, Christophe Lyon wrote:
> On Tue, 7 Jan 2020 at 14:18, Martin Jambor <mjambor@suse.cz> wrote:
>>
>> Hi,
>>
>> On Fri, Jan 03 2020, Feng Xue OS wrote:
>> > When checking a self-recursively generated value for aggregate jump
>> > function, wrong aggregate lattice was used, which will cause infinite
>> > constant propagation. This patch is composed to fix this issue.
>> >
>> > 2020-01-03  Feng Xue  <fxue@os.amperecomputing.com>
>> >
>> >         PR ipa/93084
>> >         * ipa-cp.c (self_recursively_generated_p): Find matched aggregate
>> >         lattice for a value to check.
>> >         (propagate_vals_across_arith_jfunc): Add an assertion to ensure
>> >         finite propagation in self-recursive scc.
>>
>> as far as I am concerned, the patch looks good.
>
> Hi,
> The new test introduced by this patch (gcc.dg/ipa/ipa-clone-3.c) fails
> on arm, powerpc and mips according to gcc-testresults.
> FAIL: gcc.dg/ipa/ipa-clone-3.c scan-ipa-dump-times cp "Creating a
> specialized node of recur_fn/[0-9]*\\." 8
>
> Can you have a look?
>

thank you, next time please also consider filing a bug with the exact
architecture triplets/quadruplets where it fails.  I started with a look
at aarch64-suse-linux and then at ppc64le-redhat-linux but apparently
those were not the arms and powerpcs I was looking for.

After building a cross-compiler for armv8l-unknown-linux-gnueabihf I can
see the failure, aggregate jump functions do not get built for:

  main ()
  { 
    struct V v;
  
    <bb 2> [local count: 1073741824]:
    v = *.LC0;
    _4 = recur_fn (&v);
    return _4;
  }

which is how the release_ssa dump looks like for this target, even
though it might actually be simpler than for the supported and more
common form:

  main ()
  { 
    struct V v;
  
    <bb 2> [local count: 1073741824]:
    v.f0 = 1;
    v.f1 = 3;
    _5 = recur_fn (&v);
    return _5;
  }

The different form stems from the fact that can_move_by_pieces returns
false in gimplify_init_constructor.  I'll put this on my TODO list.

Meanwhile, I' about to commit this which should fix the testcase for now
after lightly testing it with make -k check-gcc RUNTESTFLAGS="ipa.exp".

Thanks,

Martin


2020-01-10  Martin Jambor  <mjambor@suse.cz>

	* gcc.dg/ipa/ipa-clone-3.c: Replace struct initializer with
	piecemeal initialization.
---
 gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c b/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c
index 18d29bdd0b3..a73cb8b63fc 100644
--- a/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c
+++ b/gcc/testsuite/gcc.dg/ipa/ipa-clone-3.c
@@ -34,8 +34,10 @@ int recur_fn (struct V * __restrict v)
 
 int main ()
 {
-  struct V v = {1, 3};
+  struct V v;
 
+  v.f0 = 1;
+  v.f1 = 3;
   return recur_fn (&v);
 }
 
-- 
2.24.1

Re: [PATCH] Fix a bug that propagation in recursive function uses wrong aggregate lattice (PR ipa/93084)

[Jeff Law]

On Tue, 2020-01-07 at 14:18 +0100, Martin Jambor wrote:
> Hi,
> 
> On Fri, Jan 03 2020, Feng Xue OS wrote:
> > When checking a self-recursively generated value for aggregate jump
> > function, wrong aggregate lattice was used, which will cause infinite
> > constant propagation. This patch is composed to fix this issue. 
> > 
> > 2020-01-03  Feng Xue  <fxue@os.amperecomputing.com>
> > 
> >         PR ipa/93084
> >         * ipa-cp.c (self_recursively_generated_p): Find matched aggregate
> >         lattice for a value to check.
> >         (propagate_vals_across_arith_jfunc): Add an assertion to ensure
> >         finite propagation in self-recursive scc.
> 
> as far as I am concerned, the patch looks good.
In that case, OK for the trunk :-)

jeff
>