Re: [PATCH] sra: Avoid verification failure (PR 93516)
On Fri, 14 Feb 2020, Martin Jambor wrote:
> Hi,
>
> get_ref_base_and_extent can return different sizes for COMPONENT_REFs
> and DECLs of the same type, with the latter including (more?) padding.
> When in the IL there is an assignment between such a COMPONENT_REF and a
> DECL, SRA will try to propagate the access from the former as a child of
> the latter, creating an artificial reference that does not match the
> access's declared size, which triggers a verifier assert.
"both" use DECL_SIZE, but those can differ dependent on alignment
for example.
> Fixed by teaching the propagation functions about this special situation
> so that they don't do it. The condition is the same that
> build_user_friendly_ref_for_offset uses so the artificial reference
> causing the verifier is guaranteed not to be created.
>
> Bootstrapped and tested on x86_64-linux. OK for trunk?
OK.
Thanks,
Richard.
> Thanks,
>
> Martin
>
>
> 2020-02-10 Martin Jambor
> PR tree-optimization/93516
> * tree-sra.c (propagate_subaccesses_from_rhs): Do not create
> access of the same type as the parent.
> (propagate_subaccesses_from_lhs): Likewise.
>
> gcc/testsuite/
> * g++.dg/tree-ssa/pr93516.C: New test.
> ---
> gcc/ChangeLog | 7 ++
> gcc/testsuite/ChangeLog | 5
> gcc/testsuite/g++.dg/tree-ssa/pr93516.C | 24 +++
> gcc/tree-sra.c | 31 +++--
> 4 files changed, 60 insertions(+), 7 deletions(-)
> create mode 100644 gcc/testsuite/g++.dg/tree-ssa/pr93516.C
>
> diff --git a/gcc/ChangeLog b/gcc/ChangeLog
> index fa3cbb895e9..daba335cf79 100644
> --- a/gcc/ChangeLog
> +++ b/gcc/ChangeLog
> @@ -1,3 +1,10 @@
> +2020-02-10 Martin Jambor
> +
> + PR tree-optimization/93516
> + * tree-sra.c (propagate_subaccesses_from_rhs): Do not create
> + access of the same type as the parent.
> + (propagate_subaccesses_from_lhs): Likewise.
> +
> 2020-02-10 Feng Xue
>
> PR ipa/93203
> diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
> index 0ede9604611..eada2ebed63 100644
> --- a/gcc/testsuite/ChangeLog
> +++ b/gcc/testsuite/ChangeLog
> @@ -1,3 +1,8 @@
> +2020-02-10 Martin Jambor
> +
> + PR tree-optimization/93516
> + * g++.dg/tree-ssa/pr93516.C: New test.
> +
> 2020-02-10 Feng Xue
>
> PR ipa/93203
> diff --git a/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
> b/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
> new file mode 100644
> index 000..2bba37c1386
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
> @@ -0,0 +1,24 @@
> +// { dg-do compile }
> +// { dg-options "-O2" } */
> +
> +struct b;
> +struct c {
> + b *operator->();
> +};
> +class e {
> + void *f;
> + int d;
> +
> +public:
> + template a g() { return *static_cast(this); }
> +};
> +struct h : e {};
> +struct b {
> + void i(e);
> + e j();
> +};
> +void m() {
> + c k;
> + h l = k->j().g();
> + k->i(l);
> +}
> diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
> index ea8594db193..a3bcd5f3e03 100644
> --- a/gcc/tree-sra.c
> +++ b/gcc/tree-sra.c
> @@ -2785,9 +2785,17 @@ propagate_subaccesses_from_rhs (struct access *lacc,
> struct access *racc)
> }
>
>rchild->grp_hint = 1;
> - new_acc = create_artificial_child_access (lacc, rchild, norm_offset,
> - false, (lacc->grp_write
> - || rchild->grp_write));
> + /* Because get_ref_base_and_extent always includes padding in size for
> + accesses to DECLs but not necessarily for COMPONENT_REFs of the same
> + type, we might be actually attempting to here to create a child of the
> + same type as the parent. */
> + if (!types_compatible_p (lacc->type, rchild->type))
> + new_acc = create_artificial_child_access (lacc, rchild, norm_offset,
> + false,
> + (lacc->grp_write
> +|| rchild->grp_write));
> + else
> + new_acc = lacc;
>gcc_checking_assert (new_acc);
>if (racc->first_child)
> propagate_subaccesses_from_rhs (new_acc, rchild);
> @@ -2834,10 +2842,19 @@ propagate_subaccesses_from_lhs (struct access *lacc,
> struct access *racc)
> continue;
> }
>
> - struct access *new_acc
> - = create_artificial_child_access (racc, lchild, norm_offset,
> -true, false);
> - propagate_subaccesses_from_lhs (lchild, new_acc);
> + /* Because get_ref_base_and_extent always includes padding in size for
> + accesses to DECLs but not necessarily for COMPONENT_REFs of the same
> + type, we might be actually attempting to here to create a child of the
> + same type as the parent. */
> + if (!types_compatible_p (racc->type, lchild->type))
> +
[PATCH] sra: Avoid verification failure (PR 93516)
Hi,
get_ref_base_and_extent can return different sizes for COMPONENT_REFs
and DECLs of the same type, with the latter including (more?) padding.
When in the IL there is an assignment between such a COMPONENT_REF and a
DECL, SRA will try to propagate the access from the former as a child of
the latter, creating an artificial reference that does not match the
access's declared size, which triggers a verifier assert.
Fixed by teaching the propagation functions about this special situation
so that they don't do it. The condition is the same that
build_user_friendly_ref_for_offset uses so the artificial reference
causing the verifier is guaranteed not to be created.
Bootstrapped and tested on x86_64-linux. OK for trunk?
Thanks,
Martin
2020-02-10 Martin Jambor
PR tree-optimization/93516
* tree-sra.c (propagate_subaccesses_from_rhs): Do not create
access of the same type as the parent.
(propagate_subaccesses_from_lhs): Likewise.
gcc/testsuite/
* g++.dg/tree-ssa/pr93516.C: New test.
---
gcc/ChangeLog | 7 ++
gcc/testsuite/ChangeLog | 5
gcc/testsuite/g++.dg/tree-ssa/pr93516.C | 24 +++
gcc/tree-sra.c | 31 +++--
4 files changed, 60 insertions(+), 7 deletions(-)
create mode 100644 gcc/testsuite/g++.dg/tree-ssa/pr93516.C
diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index fa3cbb895e9..daba335cf79 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2020-02-10 Martin Jambor
+
+ PR tree-optimization/93516
+ * tree-sra.c (propagate_subaccesses_from_rhs): Do not create
+ access of the same type as the parent.
+ (propagate_subaccesses_from_lhs): Likewise.
+
2020-02-10 Feng Xue
PR ipa/93203
diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 0ede9604611..eada2ebed63 100644
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2020-02-10 Martin Jambor
+
+ PR tree-optimization/93516
+ * g++.dg/tree-ssa/pr93516.C: New test.
+
2020-02-10 Feng Xue
PR ipa/93203
diff --git a/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
b/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
new file mode 100644
index 000..2bba37c1386
--- /dev/null
+++ b/gcc/testsuite/g++.dg/tree-ssa/pr93516.C
@@ -0,0 +1,24 @@
+// { dg-do compile }
+// { dg-options "-O2" } */
+
+struct b;
+struct c {
+ b *operator->();
+};
+class e {
+ void *f;
+ int d;
+
+public:
+ template a g() { return *static_cast(this); }
+};
+struct h : e {};
+struct b {
+ void i(e);
+ e j();
+};
+void m() {
+ c k;
+ h l = k->j().g();
+ k->i(l);
+}
diff --git a/gcc/tree-sra.c b/gcc/tree-sra.c
index ea8594db193..a3bcd5f3e03 100644
--- a/gcc/tree-sra.c
+++ b/gcc/tree-sra.c
@@ -2785,9 +2785,17 @@ propagate_subaccesses_from_rhs (struct access *lacc,
struct access *racc)
}
rchild->grp_hint = 1;
- new_acc = create_artificial_child_access (lacc, rchild, norm_offset,
- false, (lacc->grp_write
- || rchild->grp_write));
+ /* Because get_ref_base_and_extent always includes padding in size for
+accesses to DECLs but not necessarily for COMPONENT_REFs of the same
+type, we might be actually attempting to here to create a child of the
+same type as the parent. */
+ if (!types_compatible_p (lacc->type, rchild->type))
+ new_acc = create_artificial_child_access (lacc, rchild, norm_offset,
+ false,
+ (lacc->grp_write
+ || rchild->grp_write));
+ else
+ new_acc = lacc;
gcc_checking_assert (new_acc);
if (racc->first_child)
propagate_subaccesses_from_rhs (new_acc, rchild);
@@ -2834,10 +2842,19 @@ propagate_subaccesses_from_lhs (struct access *lacc,
struct access *racc)
continue;
}
- struct access *new_acc
- = create_artificial_child_access (racc, lchild, norm_offset,
- true, false);
- propagate_subaccesses_from_lhs (lchild, new_acc);
+ /* Because get_ref_base_and_extent always includes padding in size for
+accesses to DECLs but not necessarily for COMPONENT_REFs of the same
+type, we might be actually attempting to here to create a child of the
+same type as the parent. */
+ if (!types_compatible_p (racc->type, lchild->type))
+ {
+ struct access *new_acc
+ = create_artificial_child_access (racc, lchild, norm_offset,
+ true, false);
+ propagate_subaccesses_from_lhs (lchild, new_acc);
+ }
+ else
+ propagate_subaccesses_from_lhs (lchild, racc);
ret = true;
}
return
