[PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[Kees Cook]

With -fno-strict-aliasing, IPA ICF merges functions that copy the same
fields between two struct types of different sizes, and the merged body,
inlined at the other function's call sites, can trip diagnostics like
-Warray-bounds:

    struct small { char pad[16]; int a; unsigned int b, c; };
    struct big   { char pad[16]; int a; unsigned int b, c; int extra; };

    void small_to_big (struct big *dst, const struct small *src)
    {
      dst->a = src->a;
      ...
    }

    void big_to_small (struct small *dst, const struct big *src)
    {
      dst->a = src->a;
      ...
    }

    void f (struct big *p)
    {
      struct small tmp;

      big_to_small (&tmp, p);
      ...
    }

    In function 'small_to_big', inlined from 'f' at t.c:22:3:
    warning: array subscript 'struct big[0]' is partly outside array bounds of 'struct small[1]' [-Warray-bounds=]

Without TBAA, compare those sizes at each level of the path, so we don't
see mismatched sizes on a merged body.  Require the same tree codes at
each level too, as the TBAA comparison does, and array members of the
same kind: -Warray-bounds treats an array member that ends its object as
possibly flexible, so "int arr[1]" at the end of one struct and in the
middle of another, at the same offset in structs of the same size, must
not merge either.  The sizes are compared with ICF's operand_equal_p, so
those of variably modified types, which refer to each function's own
decls, still match.  Functions accessing distinct types of the same
layout still merge, as gcc.dg/lto/pr113359-*.c tests expect.

The Linux kernel (which uses -fno-strict-aliasing) hit this with GCC 16 in
net/ipv6/ip6_tunnel.c: ip6_tnl_parm_from_user() and ip6_tnl_parm_to_user()
copy the same fields between struct ip6_tnl_parm and the larger struct
__ip6_tnl_parm in opposite directions. This was only seen on m68k,
parisc, arc, xtensa, and openrisc.  Those functions also memcpy a char
array member, and since r17-4004 (PR119006) ICF compares the fields in
those addresses and keeps the two apart, but functions that copy only
the scalar fields (as Linux is) still merged.

The stricter comparison doesn't cost merges in the kernel code I
measured: across the 636 objects of net/ipv6, fs/ext4, kernel, and mm in
an x86 allmodconfig build, none of the 1569 ICF merges is between
accesses whose sizes or array kinds differ, so the count and the text
size are the same with and without this change.

This turned out to match PR105438, but on trunk (with -fstrict-aliasing)
the access path comparison refused the merge. But with -fno-strict-aliasing
it was exposed.  (Its testcase also now needed -fno-early-inlining, or both
instantiations were inlined before ICF runs.)

Assisted-by: LLM [tests]

gcc/ChangeLog:

	PR ipa/105438
	* tree-ssa-alias.cc (ao_compare::compare_ao_refs): Without TBAA,
	compare the tree codes, the sizes of the types, and the kinds of
	array members along the access paths.

gcc/testsuite/ChangeLog:

	PR ipa/105438
	* g++.dg/ipa/pr105438.C: New test.
	* gcc.dg/ipa/ipa-icf-41.c: New test.
	* gcc.dg/ipa/ipa-icf-42.c: New test.

Signed-off-by: Kees Cook <kees@kernel.org>
---
I tried to copy neighboring style. I don't like "while (true)", but that
seemed to be idiomatic?
---
 gcc/testsuite/g++.dg/ipa/pr105438.C   | 33 ++++++++++++
 gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c | 73 ++++++++++++++++++++++++++
 gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c | 75 +++++++++++++++++++++++++++
 gcc/tree-ssa-alias.cc                 | 34 +++++++++++-
 4 files changed, 213 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/ipa/pr105438.C
 create mode 100644 gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
 create mode 100644 gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c

diff --git a/gcc/testsuite/g++.dg/ipa/pr105438.C b/gcc/testsuite/g++.dg/ipa/pr105438.C
new file mode 100644
index 000000000000..45daece51abf
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ipa/pr105438.C
@@ -0,0 +1,33 @@
+// PR ipa/105438
+// { dg-do compile }
+// { dg-options "-O2 -fno-strict-aliasing -fno-early-inlining -Warray-bounds -fdump-ipa-icf-optimized" }
+
+// With -fno-strict-aliasing, IPA ICF merged configure<7> and configure<2>,
+// whose accesses through IN differ only in the bound of the array type,
+// and the merged body, inlined for the call with SHORTER, read it as a
+// const int[7].  (-fno-early-inlining keeps the instantiations from being
+// inlined into main before ICF runs.)
+
+int longer[7] = {};
+int shorter[2] = {};
+int out[10] = {};
+
+template <int N>
+void configure(const int(&in)[N], const int nrows = N)
+{
+    if (nrows <= 10)
+    {
+        for (int i = 0; i < nrows; i++)
+        {
+            out[i] = in[i];	// { dg-bogus "outside array bounds" }
+        }
+    }
+}
+
+int main()
+{
+  configure(longer);
+  configure(shorter);
+}
+
+// { dg-final { scan-ipa-dump-not "Semantic equality hit" "icf" } }
diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c b/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
new file mode 100644
index 000000000000..64bf871de0f8
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
@@ -0,0 +1,73 @@
+/* Verify that with -fno-strict-aliasing, IPA ICF doesn't merge functions
+   whose memory accesses go through types of different sizes, which would
+   put one function's types into the other's inlined copies, and that it
+   still merges functions whose accesses go through the same types, or
+   through distinct types of the same layout.
+   { dg-do compile }
+   { dg-options "-O2 -fno-strict-aliasing -fno-early-inlining -Warray-bounds -fdump-ipa-icf-optimized" } */
+
+struct small { char pad[16]; int a; unsigned int b, c; };
+struct big { char pad[16]; int a; unsigned int b, c; int extra; };
+struct small_twin { char pad[16]; int a; unsigned int b, c; };
+
+/* Not merged: the same offsets, but through types of different sizes.
+   Merged, the call to big_to_small in f would get small_to_big's body,
+   which stores to TMP through a "struct big *".  */
+
+void small_to_big (struct big *dst, const struct small *src)
+{
+  dst->a = src->a;	/* { dg-bogus "outside array bounds" } */
+  dst->b = src->b;	/* { dg-bogus "outside array bounds" } */
+  dst->c = src->c;	/* { dg-bogus "outside array bounds" } */
+}
+
+void big_to_small (struct small *dst, const struct big *src)
+{
+  dst->a = src->a;
+  dst->b = src->b;
+  dst->c = src->c;
+}
+
+void f (struct big *p)
+{
+  struct small tmp;
+
+  big_to_small (&tmp, p);
+  small_to_big (p, &tmp);
+}
+
+/* Merged: the same types.  */
+
+void big_to_big_1 (struct big *dst, const struct big *src)
+{
+  dst->a = src->a;
+  dst->b = src->b;
+  dst->c = src->c;
+}
+
+void big_to_big_2 (struct big *dst, const struct big *src)
+{
+  dst->a = src->a;
+  dst->b = src->b;
+  dst->c = src->c;
+}
+
+/* Merged: distinct types of the same layout.  */
+
+void small_to_twin (struct small_twin *dst, const struct small *src)
+{
+  dst->a = src->a;
+  dst->b = src->b;
+  dst->c = src->c;
+}
+
+void twin_to_small (struct small *dst, const struct small_twin *src)
+{
+  dst->a = src->a;
+  dst->b = src->b;
+  dst->c = src->c;
+}
+
+/* { dg-final { scan-ipa-dump-not {Semantic equality hit:[^\n]*(small_to_big|big_to_small)/} "icf" } } */
+/* { dg-final { scan-ipa-dump {Semantic equality hit:big_to_big_[12]/[0-9]+->big_to_big_[12]/[0-9]+} "icf" } } */
+/* { dg-final { scan-ipa-dump {Semantic equality hit:(small_to_twin|twin_to_small)/[0-9]+->(small_to_twin|twin_to_small)/[0-9]+} "icf" } } */
diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c b/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c
new file mode 100644
index 000000000000..cb6a0ab9fdc3
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c
@@ -0,0 +1,75 @@
+/* Verify that with -fno-strict-aliasing, IPA ICF compares each level of
+   an access path, not just its base: an array element against a scalar
+   member, a nested size mismatch, an array member that ends its object
+   against one that doesn't, and a flexible array member against a
+   zero-length one keep functions apart, while
+   flexible array members of the same layout, and identical functions
+   accessing variably modified types, still merge.
+   { dg-do compile }
+   { dg-options "-O2 -fno-strict-aliasing -fdump-ipa-icf-details" } */
+
+/* ICF hashes memory accesses with their alias sets even with
+   -fno-strict-aliasing, so functions accessing distinct structs would
+   not get the same hash and would never be compared.  may_alias gives
+   each struct alias set 0, so each pair below is compared, which the
+   "Equals called for" scans check, along with the reason for each
+   refusal.  */
+
+#define MA __attribute__ ((may_alias))
+
+/* Not merged: the stores are to the same offset and size, but one is to
+   an array element and the other to a scalar member, so the access
+   paths have different tree codes.  (The tree code of a load is part of
+   ICF's hash, so a load would not be compared.)  */
+
+struct MA c1 { int tag[2]; int x[1]; int y; };
+struct MA c2 { int tag[2]; int x; int y; };
+
+void codes_1 (struct c1 *dst, int v) { dst->x[0] = v; }
+void codes_2 (struct c2 *dst, int v) { dst->x = v; }
+
+/* Not merged: the outer structs are the same size, but the structs
+   holding A are 16 and 8 bytes.  */
+
+struct MA n1 { struct { int a; int pad[3]; } in; int t[1]; };
+struct MA n2 { struct { int a; int pad; } in; int t[3]; };
+
+void nested_1 (struct n1 *dst, const struct n1 *src) { dst->in.a = src->in.a; }
+void nested_2 (struct n2 *dst, const struct n2 *src) { dst->in.a = src->in.a; }
+
+/* Not merged: both structs are 16 bytes, but ARR ends struct t1, so it
+   may be a flexible array, and not struct t2.  */
+
+struct MA __attribute__ ((aligned (16))) t1 { int tag; int arr[1]; };
+struct MA __attribute__ ((aligned (16))) t2 { int tag; int arr[1]; int x; };
+
+void trailing_1 (struct t1 *dst, const struct t1 *src) { dst->arr[0] = src->arr[0]; }
+void trailing_2 (struct t2 *dst, const struct t2 *src) { dst->arr[0] = src->arr[0]; }
+
+/* Merged: flexible array members, whose types have no size.  Not merged:
+   either of those and a zero-length array.  */
+
+struct MA f1 { int tag; int arr[]; };
+struct MA f2 { int tag; int arr[]; };
+struct MA z { int tag; int arr[0]; };
+
+void flex_1 (struct f1 *dst, const struct f1 *src, int i) { dst->arr[i] = src->arr[i]; }
+void flex_2 (struct f2 *dst, const struct f2 *src, int i) { dst->arr[i] = src->arr[i]; }
+void zero_len (struct z *dst, const struct z *src, int i) { dst->arr[i] = src->arr[i]; }
+
+/* Merged: the sizes of variably modified types refer to each function's
+   own parameters.  */
+
+void vla_1 (int n, int (*p)[n], int i) { p[i][1] = 1; p[i][2] = 2; }
+void vla_2 (int n, int (*p)[n], int i) { p[i][1] = 1; p[i][2] = 2; }
+
+/* Each refusal is logged as its reason, a few more lines, and then the
+   pair compared; match the reason and the pair with no other comparison
+   between them.  */
+/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: codes_[12]/[0-9]+:codes_[12]/[0-9]+ with result: false} "icf" } } */
+/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: nested_[12]/[0-9]+:nested_[12]/[0-9]+ with result: false} "icf" } } */
+/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: trailing_[12]/[0-9]+:trailing_[12]/[0-9]+ with result: false} "icf" } } */
+/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: (zero_len/[0-9]+:flex_[12]|flex_[12]/[0-9]+:zero_len)/[0-9]+ with result: false} "icf" } } */
+/* { dg-final { scan-ipa-dump {Semantic equality hit:flex_[12]/[0-9]+->flex_[12]/[0-9]+} "icf" } } */
+/* { dg-final { scan-ipa-dump {Semantic equality hit:vla_[12]/[0-9]+->vla_[12]/[0-9]+} "icf" } } */
+/* { dg-final { scan-ipa-dump-not {Semantic equality hit:[^\n]*(codes_|nested_|trailing_|zero_len)} "icf" } } */
diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
index a8196d116355..2f3d7f523a02 100644
--- a/gcc/tree-ssa-alias.cc
+++ b/gcc/tree-ssa-alias.cc
@@ -4262,7 +4262,9 @@ types_equal_for_same_type_for_tbaa_p (tree type1, tree type2)
 /* Compare REF1 and REF2 and return flags specifying their differences.
    If LTO_STREAMING_SAFE is true do not use alias sets and canonical
    types that are going to be recomputed.
-   If TBAA is true also compare TBAA metadata.  */
+   If TBAA is true also compare TBAA metadata, and otherwise only the
+   shape of the access paths: the tree codes along them, the sizes of
+   their types, and the kind of each array member.  */
 
 int
 ao_compare::compare_ao_refs (ao_ref *ref1, ao_ref *ref2,
@@ -4379,8 +4381,36 @@ ao_compare::compare_ao_refs (ao_ref *ref1, ao_ref *ref2,
 	  || (MR_DEPENDENCE_BASE (rbase1) != MR_DEPENDENCE_BASE (rbase2))))
     flags |= DEPENDENCE_CLIQUE;
 
+  /* Without TBAA, the types along the access paths don't matter for
+     aliasing, but other consumers read their sizes (e.g. ranger
+     diagnostics like -Warray-bounds).  Require the same tree codes at
+     each level, with sizes that match and array members of the same
+     kind, so the references of a merged function are valid for the
+     callers of both.  The types themselves may differ, so functions
+     accessing distinct but identically laid out types still merge.  */
   if (!tbaa)
-    return flags;
+    {
+      tree c1 = ref1->ref, c2 = ref2->ref;
+      while (true)
+	{
+	  if (TREE_CODE (c1) != TREE_CODE (c2))
+	    return flags | ACCESS_PATH;
+	  tree size1 = TYPE_SIZE (TREE_TYPE (c1));
+	  tree size2 = TYPE_SIZE (TREE_TYPE (c2));
+	  if (size1 && size2
+	      ? !operand_equal_p (size1, size2, 0)
+	      : size1 != size2)
+	    return flags | ACCESS_PATH;
+	  if (TREE_CODE (c1) == COMPONENT_REF
+	      && component_ref_sam_type (c1) != component_ref_sam_type (c2))
+	    return flags | ACCESS_PATH;
+	  if (!handled_component_p (c1))
+	    break;
+	  c1 = TREE_OPERAND (c1, 0);
+	  c2 = TREE_OPERAND (c2, 0);
+	}
+      return flags;
+    }
 
   /* Alias sets are not stable across LTO sreaming; be conservative here
      and compare types the alias sets are ultimately based on.  */
-- 
2.55.0

Re: [PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[Richard Biener]

On Sat, Oct 10, 2026 at 7:06 AM Kees Cook <kees@kernel.org> wrote:
>
> With -fno-strict-aliasing, IPA ICF merges functions that copy the same
> fields between two struct types of different sizes, and the merged body,
> inlined at the other function's call sites, can trip diagnostics like
> -Warray-bounds:
>
>     struct small { char pad[16]; int a; unsigned int b, c; };
>     struct big   { char pad[16]; int a; unsigned int b, c; int extra; };
>
>     void small_to_big (struct big *dst, const struct small *src)
>     {
>       dst->a = src->a;
>       ...
>     }
>
>     void big_to_small (struct small *dst, const struct big *src)
>     {
>       dst->a = src->a;
>       ...
>     }
>
>     void f (struct big *p)
>     {
>       struct small tmp;
>
>       big_to_small (&tmp, p);
>       ...
>     }
>
>     In function 'small_to_big', inlined from 'f' at t.c:22:3:
>     warning: array subscript 'struct big[0]' is partly outside array bounds of 'struct small[1]' [-Warray-bounds=]

I think these diagnostics are dubious with -fno-strict-aliasing which
essentially
says you can access every object via any other objects type.

> Without TBAA, compare those sizes at each level of the path, so we don't
> see mismatched sizes on a merged body.  Require the same tree codes at
> each level too, as the TBAA comparison does, and array members of the
> same kind: -Warray-bounds treats an array member that ends its object as
> possibly flexible, so "int arr[1]" at the end of one struct and in the
> middle of another, at the same offset in structs of the same size, must
> not merge either.  The sizes are compared with ICF's operand_equal_p, so
> those of variably modified types, which refer to each function's own
> decls, still match.  Functions accessing distinct types of the same
> layout still merge, as gcc.dg/lto/pr113359-*.c tests expect.
>
> The Linux kernel (which uses -fno-strict-aliasing) hit this with GCC 16 in
> net/ipv6/ip6_tunnel.c: ip6_tnl_parm_from_user() and ip6_tnl_parm_to_user()
> copy the same fields between struct ip6_tnl_parm and the larger struct
> __ip6_tnl_parm in opposite directions. This was only seen on m68k,
> parisc, arc, xtensa, and openrisc.  Those functions also memcpy a char
> array member, and since r17-4004 (PR119006) ICF compares the fields in
> those addresses and keeps the two apart, but functions that copy only
> the scalar fields (as Linux is) still merged.
>
> The stricter comparison doesn't cost merges in the kernel code I
> measured: across the 636 objects of net/ipv6, fs/ext4, kernel, and mm in
> an x86 allmodconfig build, none of the 1569 ICF merges is between
> accesses whose sizes or array kinds differ, so the count and the text
> size are the same with and without this change.
>
> This turned out to match PR105438, but on trunk (with -fstrict-aliasing)
> the access path comparison refused the merge. But with -fno-strict-aliasing
> it was exposed.  (Its testcase also now needed -fno-early-inlining, or both
> instantiations were inlined before ICF runs.)

So instead of this I'd rather work towards inlining the original
body rather than the ICF prevailing one.

And force-disable these kind of diagnostics with -fno-strict-aliasing?

Richard.

>
> Assisted-by: LLM [tests]
>
> gcc/ChangeLog:
>
>         PR ipa/105438
>         * tree-ssa-alias.cc (ao_compare::compare_ao_refs): Without TBAA,
>         compare the tree codes, the sizes of the types, and the kinds of
>         array members along the access paths.
>
> gcc/testsuite/ChangeLog:
>
>         PR ipa/105438
>         * g++.dg/ipa/pr105438.C: New test.
>         * gcc.dg/ipa/ipa-icf-41.c: New test.
>         * gcc.dg/ipa/ipa-icf-42.c: New test.
>
> Signed-off-by: Kees Cook <kees@kernel.org>
> ---
> I tried to copy neighboring style. I don't like "while (true)", but that
> seemed to be idiomatic?
> ---
>  gcc/testsuite/g++.dg/ipa/pr105438.C   | 33 ++++++++++++
>  gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c | 73 ++++++++++++++++++++++++++
>  gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c | 75 +++++++++++++++++++++++++++
>  gcc/tree-ssa-alias.cc                 | 34 +++++++++++-
>  4 files changed, 213 insertions(+), 2 deletions(-)
>  create mode 100644 gcc/testsuite/g++.dg/ipa/pr105438.C
>  create mode 100644 gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
>  create mode 100644 gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c
>
> diff --git a/gcc/testsuite/g++.dg/ipa/pr105438.C b/gcc/testsuite/g++.dg/ipa/pr105438.C
> new file mode 100644
> index 000000000000..45daece51abf
> --- /dev/null
> +++ b/gcc/testsuite/g++.dg/ipa/pr105438.C
> @@ -0,0 +1,33 @@
> +// PR ipa/105438
> +// { dg-do compile }
> +// { dg-options "-O2 -fno-strict-aliasing -fno-early-inlining -Warray-bounds -fdump-ipa-icf-optimized" }
> +
> +// With -fno-strict-aliasing, IPA ICF merged configure<7> and configure<2>,
> +// whose accesses through IN differ only in the bound of the array type,
> +// and the merged body, inlined for the call with SHORTER, read it as a
> +// const int[7].  (-fno-early-inlining keeps the instantiations from being
> +// inlined into main before ICF runs.)
> +
> +int longer[7] = {};
> +int shorter[2] = {};
> +int out[10] = {};
> +
> +template <int N>
> +void configure(const int(&in)[N], const int nrows = N)
> +{
> +    if (nrows <= 10)
> +    {
> +        for (int i = 0; i < nrows; i++)
> +        {
> +            out[i] = in[i];    // { dg-bogus "outside array bounds" }
> +        }
> +    }
> +}
> +
> +int main()
> +{
> +  configure(longer);
> +  configure(shorter);
> +}
> +
> +// { dg-final { scan-ipa-dump-not "Semantic equality hit" "icf" } }
> diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c b/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
> new file mode 100644
> index 000000000000..64bf871de0f8
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/ipa/ipa-icf-41.c
> @@ -0,0 +1,73 @@
> +/* Verify that with -fno-strict-aliasing, IPA ICF doesn't merge functions
> +   whose memory accesses go through types of different sizes, which would
> +   put one function's types into the other's inlined copies, and that it
> +   still merges functions whose accesses go through the same types, or
> +   through distinct types of the same layout.
> +   { dg-do compile }
> +   { dg-options "-O2 -fno-strict-aliasing -fno-early-inlining -Warray-bounds -fdump-ipa-icf-optimized" } */
> +
> +struct small { char pad[16]; int a; unsigned int b, c; };
> +struct big { char pad[16]; int a; unsigned int b, c; int extra; };
> +struct small_twin { char pad[16]; int a; unsigned int b, c; };
> +
> +/* Not merged: the same offsets, but through types of different sizes.
> +   Merged, the call to big_to_small in f would get small_to_big's body,
> +   which stores to TMP through a "struct big *".  */
> +
> +void small_to_big (struct big *dst, const struct small *src)
> +{
> +  dst->a = src->a;     /* { dg-bogus "outside array bounds" } */
> +  dst->b = src->b;     /* { dg-bogus "outside array bounds" } */
> +  dst->c = src->c;     /* { dg-bogus "outside array bounds" } */
> +}
> +
> +void big_to_small (struct small *dst, const struct big *src)
> +{
> +  dst->a = src->a;
> +  dst->b = src->b;
> +  dst->c = src->c;
> +}
> +
> +void f (struct big *p)
> +{
> +  struct small tmp;
> +
> +  big_to_small (&tmp, p);
> +  small_to_big (p, &tmp);
> +}
> +
> +/* Merged: the same types.  */
> +
> +void big_to_big_1 (struct big *dst, const struct big *src)
> +{
> +  dst->a = src->a;
> +  dst->b = src->b;
> +  dst->c = src->c;
> +}
> +
> +void big_to_big_2 (struct big *dst, const struct big *src)
> +{
> +  dst->a = src->a;
> +  dst->b = src->b;
> +  dst->c = src->c;
> +}
> +
> +/* Merged: distinct types of the same layout.  */
> +
> +void small_to_twin (struct small_twin *dst, const struct small *src)
> +{
> +  dst->a = src->a;
> +  dst->b = src->b;
> +  dst->c = src->c;
> +}
> +
> +void twin_to_small (struct small *dst, const struct small_twin *src)
> +{
> +  dst->a = src->a;
> +  dst->b = src->b;
> +  dst->c = src->c;
> +}
> +
> +/* { dg-final { scan-ipa-dump-not {Semantic equality hit:[^\n]*(small_to_big|big_to_small)/} "icf" } } */
> +/* { dg-final { scan-ipa-dump {Semantic equality hit:big_to_big_[12]/[0-9]+->big_to_big_[12]/[0-9]+} "icf" } } */
> +/* { dg-final { scan-ipa-dump {Semantic equality hit:(small_to_twin|twin_to_small)/[0-9]+->(small_to_twin|twin_to_small)/[0-9]+} "icf" } } */
> diff --git a/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c b/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c
> new file mode 100644
> index 000000000000..cb6a0ab9fdc3
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/ipa/ipa-icf-42.c
> @@ -0,0 +1,75 @@
> +/* Verify that with -fno-strict-aliasing, IPA ICF compares each level of
> +   an access path, not just its base: an array element against a scalar
> +   member, a nested size mismatch, an array member that ends its object
> +   against one that doesn't, and a flexible array member against a
> +   zero-length one keep functions apart, while
> +   flexible array members of the same layout, and identical functions
> +   accessing variably modified types, still merge.
> +   { dg-do compile }
> +   { dg-options "-O2 -fno-strict-aliasing -fdump-ipa-icf-details" } */
> +
> +/* ICF hashes memory accesses with their alias sets even with
> +   -fno-strict-aliasing, so functions accessing distinct structs would
> +   not get the same hash and would never be compared.  may_alias gives
> +   each struct alias set 0, so each pair below is compared, which the
> +   "Equals called for" scans check, along with the reason for each
> +   refusal.  */
> +
> +#define MA __attribute__ ((may_alias))
> +
> +/* Not merged: the stores are to the same offset and size, but one is to
> +   an array element and the other to a scalar member, so the access
> +   paths have different tree codes.  (The tree code of a load is part of
> +   ICF's hash, so a load would not be compared.)  */
> +
> +struct MA c1 { int tag[2]; int x[1]; int y; };
> +struct MA c2 { int tag[2]; int x; int y; };
> +
> +void codes_1 (struct c1 *dst, int v) { dst->x[0] = v; }
> +void codes_2 (struct c2 *dst, int v) { dst->x = v; }
> +
> +/* Not merged: the outer structs are the same size, but the structs
> +   holding A are 16 and 8 bytes.  */
> +
> +struct MA n1 { struct { int a; int pad[3]; } in; int t[1]; };
> +struct MA n2 { struct { int a; int pad; } in; int t[3]; };
> +
> +void nested_1 (struct n1 *dst, const struct n1 *src) { dst->in.a = src->in.a; }
> +void nested_2 (struct n2 *dst, const struct n2 *src) { dst->in.a = src->in.a; }
> +
> +/* Not merged: both structs are 16 bytes, but ARR ends struct t1, so it
> +   may be a flexible array, and not struct t2.  */
> +
> +struct MA __attribute__ ((aligned (16))) t1 { int tag; int arr[1]; };
> +struct MA __attribute__ ((aligned (16))) t2 { int tag; int arr[1]; int x; };
> +
> +void trailing_1 (struct t1 *dst, const struct t1 *src) { dst->arr[0] = src->arr[0]; }
> +void trailing_2 (struct t2 *dst, const struct t2 *src) { dst->arr[0] = src->arr[0]; }
> +
> +/* Merged: flexible array members, whose types have no size.  Not merged:
> +   either of those and a zero-length array.  */
> +
> +struct MA f1 { int tag; int arr[]; };
> +struct MA f2 { int tag; int arr[]; };
> +struct MA z { int tag; int arr[0]; };
> +
> +void flex_1 (struct f1 *dst, const struct f1 *src, int i) { dst->arr[i] = src->arr[i]; }
> +void flex_2 (struct f2 *dst, const struct f2 *src, int i) { dst->arr[i] = src->arr[i]; }
> +void zero_len (struct z *dst, const struct z *src, int i) { dst->arr[i] = src->arr[i]; }
> +
> +/* Merged: the sizes of variably modified types refer to each function's
> +   own parameters.  */
> +
> +void vla_1 (int n, int (*p)[n], int i) { p[i][1] = 1; p[i][2] = 2; }
> +void vla_2 (int n, int (*p)[n], int i) { p[i][1] = 1; p[i][2] = 2; }
> +
> +/* Each refusal is logged as its reason, a few more lines, and then the
> +   pair compared; match the reason and the pair with no other comparison
> +   between them.  */
> +/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: codes_[12]/[0-9]+:codes_[12]/[0-9]+ with result: false} "icf" } } */
> +/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: nested_[12]/[0-9]+:nested_[12]/[0-9]+ with result: false} "icf" } } */
> +/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: trailing_[12]/[0-9]+:trailing_[12]/[0-9]+ with result: false} "icf" } } */
> +/* { dg-final { scan-ipa-dump {access path difference\)'[^\n]*\n(?:(?!Equals called)[^\n]*\n)*Equals called for: (zero_len/[0-9]+:flex_[12]|flex_[12]/[0-9]+:zero_len)/[0-9]+ with result: false} "icf" } } */
> +/* { dg-final { scan-ipa-dump {Semantic equality hit:flex_[12]/[0-9]+->flex_[12]/[0-9]+} "icf" } } */
> +/* { dg-final { scan-ipa-dump {Semantic equality hit:vla_[12]/[0-9]+->vla_[12]/[0-9]+} "icf" } } */
> +/* { dg-final { scan-ipa-dump-not {Semantic equality hit:[^\n]*(codes_|nested_|trailing_|zero_len)} "icf" } } */
> diff --git a/gcc/tree-ssa-alias.cc b/gcc/tree-ssa-alias.cc
> index a8196d116355..2f3d7f523a02 100644
> --- a/gcc/tree-ssa-alias.cc
> +++ b/gcc/tree-ssa-alias.cc
> @@ -4262,7 +4262,9 @@ types_equal_for_same_type_for_tbaa_p (tree type1, tree type2)
>  /* Compare REF1 and REF2 and return flags specifying their differences.
>     If LTO_STREAMING_SAFE is true do not use alias sets and canonical
>     types that are going to be recomputed.
> -   If TBAA is true also compare TBAA metadata.  */
> +   If TBAA is true also compare TBAA metadata, and otherwise only the
> +   shape of the access paths: the tree codes along them, the sizes of
> +   their types, and the kind of each array member.  */
>
>  int
>  ao_compare::compare_ao_refs (ao_ref *ref1, ao_ref *ref2,
> @@ -4379,8 +4381,36 @@ ao_compare::compare_ao_refs (ao_ref *ref1, ao_ref *ref2,
>           || (MR_DEPENDENCE_BASE (rbase1) != MR_DEPENDENCE_BASE (rbase2))))
>      flags |= DEPENDENCE_CLIQUE;
>
> +  /* Without TBAA, the types along the access paths don't matter for
> +     aliasing, but other consumers read their sizes (e.g. ranger
> +     diagnostics like -Warray-bounds).  Require the same tree codes at
> +     each level, with sizes that match and array members of the same
> +     kind, so the references of a merged function are valid for the
> +     callers of both.  The types themselves may differ, so functions
> +     accessing distinct but identically laid out types still merge.  */
>    if (!tbaa)
> -    return flags;
> +    {
> +      tree c1 = ref1->ref, c2 = ref2->ref;
> +      while (true)
> +       {
> +         if (TREE_CODE (c1) != TREE_CODE (c2))
> +           return flags | ACCESS_PATH;
> +         tree size1 = TYPE_SIZE (TREE_TYPE (c1));
> +         tree size2 = TYPE_SIZE (TREE_TYPE (c2));
> +         if (size1 && size2
> +             ? !operand_equal_p (size1, size2, 0)
> +             : size1 != size2)
> +           return flags | ACCESS_PATH;
> +         if (TREE_CODE (c1) == COMPONENT_REF
> +             && component_ref_sam_type (c1) != component_ref_sam_type (c2))
> +           return flags | ACCESS_PATH;
> +         if (!handled_component_p (c1))
> +           break;
> +         c1 = TREE_OPERAND (c1, 0);
> +         c2 = TREE_OPERAND (c2, 0);
> +       }
> +      return flags;
> +    }
>
>    /* Alias sets are not stable across LTO sreaming; be conservative here
>       and compare types the alias sets are ultimately based on.  */
> --
> 2.55.0
>

Re: [PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[Kees Cook]

On Sun, Oct 11, 2026 at 04:50:41PM +0200, Richard Biener wrote:
> >     In function 'small_to_big', inlined from 'f' at t.c:22:3:
> >     warning: array subscript 'struct big[0]' is partly outside array bounds of 'struct small[1]' [-Warray-bounds=]
>
> I think these diagnostics are dubious with -fno-strict-aliasing which
> essentially says you can access every object via any other objects type.

The source has no such access, though: each function reads and writes
members through its own types, and every byte touched is inside the
object. The "struct big" access to a "struct small" exists only in the
body ICF picked, so the warning is about code GCC wrote rather than
code the programmer wrote.

> So instead of this I'd rather work towards inlining the original
> body rather than the ICF prevailing one.

That does seem like it could be better, but it's not clear to me how
large a project that would be to accomplish.

This patch seems to DTRT, though. It doesn't appear to regress merging
in a meaningful way for the Linux at least.

> And force-disable these kind of diagnostics with -fno-strict-aliasing?

I'd really rather not. Linux builds everything with -fno-strict-aliasing,
so that would remove these checks from the kernel entirely. GCC's
-Warray-bounds is still off by default for the kernel because of false
positives like this one (which is why I'm trying to solve it), and I've
been working toward enabling it, since it finds real bugs. (I have a
stack of a dozen real issues that this found while doing a recent pass.)

For example (since I don't yet have links to the above stack), here
are some random examples from a few years ago: a response struct passed
where a bigger error struct was expected, whose "error" field was past
the end of the object[1]:

  drivers/thunderbolt/xdomain.c:226:22: error: array subscript 'const struct tb_xdp_error_response[0]' is partly outside array bounds of 'struct tb_xdp_properties_changed_response[1]'

and a u64 remainder stored into an int on the stack[2]:

  include/linux/math64.h:56:20: error: array subscript 'u64 {aka long long unsigned int}[0]' is partly outside array bounds of 'int[1]'

But if you did disable diagnostics, which did you have in mind? If it's
only warnings about an access whose type is bigger than the object while
the bytes it touches are in bounds, that's narrower, but the first
example above would still need to be caught, so I'd much rather just
get ICF behaving nicely. :)

-Kees

[1] https://git.kernel.org/linus/19813551701d0
[2] https://git.kernel.org/linus/9f3d543a9f737

-- 
Kees Cook

Re: [PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[Richard Biener]

On Wed, 14 Oct 2026, Kees Cook wrote:

> On Sun, Oct 11, 2026 at 04:50:41PM +0200, Richard Biener wrote:
> > >     In function 'small_to_big', inlined from 'f' at t.c:22:3:
> > >     warning: array subscript 'struct big[0]' is partly outside array bounds of 'struct small[1]' [-Warray-bounds=]
> >
> > I think these diagnostics are dubious with -fno-strict-aliasing which
> > essentially says you can access every object via any other objects type.
> 
> The source has no such access, though: each function reads and writes
> members through its own types, and every byte touched is inside the
> object. The "struct big" access to a "struct small" exists only in the
> body ICF picked, so the warning is about code GCC wrote rather than
> code the programmer wrote.
> 
> > So instead of this I'd rather work towards inlining the original
> > body rather than the ICF prevailing one.
> 
> That does seem like it could be better, but it's not clear to me how
> large a project that would be to accomplish.

I think for the case of ICF and non-LTO it should be possible to
"hack" this in;  it should be just to make sure the inline edges
point to the original body and hopefully the ICF "merge" is somehow
encoded clearly.  When the inline edge is set up correctly it should
possibly even work with LTO.

> This patch seems to DTRT, though. It doesn't appear to regress merging
> in a meaningful way for the Linux at least.

I don't dispute this, but it feels like a hack.  Either ICF should
work as if -fstrict-aliasing were given, but consistently so, and
for a good reason, or you could as well do -fno-ipa-icf when
building the kernel as you are doing so many we-don't-quite-want-C
things already.

> > And force-disable these kind of diagnostics with -fno-strict-aliasing?
> 
> I'd really rather not. Linux builds everything with -fno-strict-aliasing,
> so that would remove these checks from the kernel entirely. GCC's
> -Warray-bounds is still off by default for the kernel because of false
> positives like this one (which is why I'm trying to solve it), and I've
> been working toward enabling it, since it finds real bugs. (I have a
> stack of a dozen real issues that this found while doing a recent pass.)
> 
> For example (since I don't yet have links to the above stack), here
> are some random examples from a few years ago: a response struct passed
> where a bigger error struct was expected, whose "error" field was past
> the end of the object[1]:
> 
>   drivers/thunderbolt/xdomain.c:226:22: error: array subscript 'const struct tb_xdp_error_response[0]' is partly outside array bounds of 'struct tb_xdp_properties_changed_response[1]'
> 
> and a u64 remainder stored into an int on the stack[2]:
> 
>   include/linux/math64.h:56:20: error: array subscript 'u64 {aka long long unsigned int}[0]' is partly outside array bounds of 'int[1]'
> 
> But if you did disable diagnostics, which did you have in mind? If it's
> only warnings about an access whose type is bigger than the object while
> the bytes it touches are in bounds, that's narrower, but the first
> example above would still need to be caught, so I'd much rather just
> get ICF behaving nicely. :)
> 
> -Kees
> 
> [1] https://git.kernel.org/linus/19813551701d0
> [2] https://git.kernel.org/linus/9f3d543a9f737
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)

Re: [PATCH] ipa-icf: Compare access path type sizes without TBAA [PR105438]

[Richard Biener]

[-- Attachment #1: Type: text/plain, Size: 3998 bytes --]

On Thu, 15 Oct 2026, Richard Biener wrote:

> On Wed, 14 Oct 2026, Kees Cook wrote:
> 
> > On Sun, Oct 11, 2026 at 04:50:41PM +0200, Richard Biener wrote:
> > > >     In function 'small_to_big', inlined from 'f' at t.c:22:3:
> > > >     warning: array subscript 'struct big[0]' is partly outside array bounds of 'struct small[1]' [-Warray-bounds=]
> > >
> > > I think these diagnostics are dubious with -fno-strict-aliasing which
> > > essentially says you can access every object via any other objects type.
> > 
> > The source has no such access, though: each function reads and writes
> > members through its own types, and every byte touched is inside the
> > object. The "struct big" access to a "struct small" exists only in the
> > body ICF picked, so the warning is about code GCC wrote rather than
> > code the programmer wrote.
> > 
> > > So instead of this I'd rather work towards inlining the original
> > > body rather than the ICF prevailing one.
> > 
> > That does seem like it could be better, but it's not clear to me how
> > large a project that would be to accomplish.
> 
> I think for the case of ICF and non-LTO it should be possible to
> "hack" this in;  it should be just to make sure the inline edges
> point to the original body and hopefully the ICF "merge" is somehow
> encoded clearly.  When the inline edge is set up correctly it should
> possibly even work with LTO.

I one-shotted the attached.  It definitely fails to do this in a
proper IPA way - it should have used IPA function/variable transform
stages and stream out the "defered aliases" in 
write/read_optimization_summary.

I'm sure this can be completed with some interaction, but our AI
policy makes this questionable.

Richard.

> > This patch seems to DTRT, though. It doesn't appear to regress merging
> > in a meaningful way for the Linux at least.
> 
> I don't dispute this, but it feels like a hack.  Either ICF should
> work as if -fstrict-aliasing were given, but consistently so, and
> for a good reason, or you could as well do -fno-ipa-icf when
> building the kernel as you are doing so many we-don't-quite-want-C
> things already.
> 
> > > And force-disable these kind of diagnostics with -fno-strict-aliasing?
> > 
> > I'd really rather not. Linux builds everything with -fno-strict-aliasing,
> > so that would remove these checks from the kernel entirely. GCC's
> > -Warray-bounds is still off by default for the kernel because of false
> > positives like this one (which is why I'm trying to solve it), and I've
> > been working toward enabling it, since it finds real bugs. (I have a
> > stack of a dozen real issues that this found while doing a recent pass.)
> > 
> > For example (since I don't yet have links to the above stack), here
> > are some random examples from a few years ago: a response struct passed
> > where a bigger error struct was expected, whose "error" field was past
> > the end of the object[1]:
> > 
> >   drivers/thunderbolt/xdomain.c:226:22: error: array subscript 'const struct tb_xdp_error_response[0]' is partly outside array bounds of 'struct tb_xdp_properties_changed_response[1]'
> > 
> > and a u64 remainder stored into an int on the stack[2]:
> > 
> >   include/linux/math64.h:56:20: error: array subscript 'u64 {aka long long unsigned int}[0]' is partly outside array bounds of 'int[1]'
> > 
> > But if you did disable diagnostics, which did you have in mind? If it's
> > only warnings about an access whose type is bigger than the object while
> > the bytes it touches are in bounds, that's narrower, but the first
> > example above would still need to be caught, so I'd much rather just
> > get ICF behaving nicely. :)
> > 
> > -Kees
> > 
> > [1] https://git.kernel.org/linus/19813551701d0
> > [2] https://git.kernel.org/linus/9f3d543a9f737
> > 
> > 
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Stefan Gaiser, Jochen Jaser, Abhinav Puri; (HRB 36809, AG Nuernberg)

[-- Attachment #2: Type: text/plain, Size: 8757 bytes --]

From 60418b7da3a4b88a47336488b1d1c2656786c5c4 Mon Sep 17 00:00:00 2001
From: Richard Biener <rguenther@suse.de>
Date: Thu, 15 Oct 2026 10:42:08 +0200
Subject: [PATCH] ipa-icf: Defer creating aliases until after inlining
 [PR105438]

When IPA-ICF merges two functions by creating an alias and releasing
the alias's body before inlining, inlining calls to the alias inlines
the ICF-chosen prevailing body instead of the original body.  When the
functions operate on different types (such as different array bounds
or structure layouts), this mixes up types and declarations in the
inlined body, triggering bogus diagnostics such as -Warray-bounds.

This patch defers the creation of ICF symbol aliases and body removal
until after inlining.  If an alias function was inlined into all of its
callers, its original body was inlined with its original types and declarations
and no alias needs to be created.  If calls remain or the function is
needed offline, the alias to the prevailing function is created after inlining.

	PR ipa/105438

gcc/ChangeLog:

	* ipa-icf.h (ipa_icf_create_deferred_aliases): Declare.
	* tree-pass.h (ipa_icf_create_deferred_aliases): Declare.
	* ipa-icf.cc (deferred_merge_aliases): New static variable.
	(sem_function::merge): Defer alias creation by pushing to
	deferred_merge_aliases.
	(ipa_icf_create_deferred_aliases): New function.
	(ipa_icf_cc_finalize): Release deferred_merge_aliases.
	* ipa-inline.cc (ipa_inline): Call ipa_icf_create_deferred_aliases.

gcc/testsuite/ChangeLog:

	* g++.dg/ipa/pr105438.C: New test.
	* g++.dg/ipa/pr105438-2.C: New test.
---
 gcc/ipa-icf.cc                        | 74 +++++++++++++++++++++------
 gcc/ipa-icf.h                         |  2 +
 gcc/ipa-inline.cc                     |  5 ++
 gcc/testsuite/g++.dg/ipa/pr105438-2.C | 25 +++++++++
 gcc/testsuite/g++.dg/ipa/pr105438.C   | 25 +++++++++
 gcc/tree-pass.h                       |  1 +
 6 files changed, 117 insertions(+), 15 deletions(-)
 create mode 100644 gcc/testsuite/g++.dg/ipa/pr105438-2.C
 create mode 100644 gcc/testsuite/g++.dg/ipa/pr105438.C

diff --git a/gcc/ipa-icf.cc b/gcc/ipa-icf.cc
index 64cc722bd7a..5e13b267b7e 100644
--- a/gcc/ipa-icf.cc
+++ b/gcc/ipa-icf.cc
@@ -100,6 +100,7 @@ namespace ipa_icf {
    are propagated later on.  */
 
 static sem_item_optimizer *optimizer = NULL;
+static vec<std::pair<cgraph_node *, cgraph_node *> > deferred_merge_aliases;
 
 /* Constructor.  */
 
@@ -1285,22 +1286,9 @@ sem_function::merge (sem_item *alias_item)
   else if (create_alias)
     {
       alias->icf_merged = true;
+      original->icf_merged = true;
 
-      /* Remove the function's body.  */
-      ipa_merge_profiles (original, alias);
-      symtab->call_cgraph_removal_hooks (alias);
-      alias->release_body (true);
-      alias->reset ();
-      /* Notice global symbol possibly produced RTL.  */
-      ((symtab_node *)alias)->call_for_symbol_and_aliases (clear_decl_rtl,
-							   NULL, true);
-
-      /* Create the alias.  */
-      cgraph_node::create_alias (alias_func->decl, decl);
-      alias->resolve_alias (original);
-
-      original->call_for_symbol_thunks_and_aliases
-	 (set_local, (void *)(size_t) original->local_p (), true);
+      deferred_merge_aliases.safe_push (std::make_pair (original, alias));
 
       if (dump_enabled_p ())
 	dump_printf (MSG_OPTIMIZED_LOCATIONS,
@@ -3722,12 +3710,67 @@ public:
 
 } // ipa_icf namespace
 
+using namespace ipa_icf;
+
 ipa_opt_pass_d *
 make_pass_ipa_icf (gcc::context *ctxt)
 {
   return new ipa_icf::pass_ipa_icf (ctxt);
 }
 
+/* Create deferred ICF aliases after inlining.  */
+
+void
+ipa_icf_create_deferred_aliases (void)
+{
+  for (unsigned i = 0; i < deferred_merge_aliases.length (); i++)
+    {
+      cgraph_node *original
+	= deferred_merge_aliases[i].first->ultimate_alias_target ();
+      cgraph_node *alias = deferred_merge_aliases[i].second;
+
+      if (alias->body_removed || alias->inlined_to)
+	continue;
+
+      /* If the alias function was inlined into all callers and can be
+	 removed, its original body was inlined and no offline copy is
+	 needed.  */
+      if (!alias->callers && alias->can_remove_if_no_direct_calls_p ())
+	continue;
+
+      if (original->inlined_to)
+	continue;
+
+      if (alias->clones)
+	{
+	  /* If alias has inline clones, we cannot release its body.
+	     Redirect direct callers to original instead.  */
+	  redirect_all_callers (alias, original);
+	  continue;
+	}
+
+      cgraph_edge *e, *next;
+      for (e = alias->callees; e; e = next)
+	{
+	  next = e->next_callee;
+	  if (!e->inline_failed)
+	    e->callee->remove_symbol_and_inline_clones ();
+	}
+
+      ipa_merge_profiles (original, alias);
+      symtab->call_cgraph_removal_hooks (alias);
+      alias->release_body (true);
+      alias->reset ();
+      ((symtab_node *)alias)->call_for_symbol_and_aliases (clear_decl_rtl,
+							   NULL, true);
+      cgraph_node::create_alias (alias->decl, original->decl);
+      alias->resolve_alias (original);
+      original->call_for_symbol_thunks_and_aliases
+	(set_local, (void *)(size_t) original->local_p (), true);
+    }
+  deferred_merge_aliases.release ();
+}
+
 /* Reset all state within ipa-icf.cc so that we can rerun the compiler
    within the same process.  For use by toplev::finalize.  */
 
@@ -3735,4 +3778,5 @@ void
 ipa_icf_cc_finalize (void)
 {
   ipa_icf::optimizer = NULL;
+  deferred_merge_aliases.release ();
 }
diff --git a/gcc/ipa-icf.h b/gcc/ipa-icf.h
index 4135ad3eb9c..daaa15e9d55 100644
--- a/gcc/ipa-icf.h
+++ b/gcc/ipa-icf.h
@@ -673,3 +673,5 @@ private:
 }; // class sem_item_optimizer
 
 } // ipa_icf namespace
+
+void ipa_icf_create_deferred_aliases (void);
diff --git a/gcc/ipa-inline.cc b/gcc/ipa-inline.cc
index ba8bf11b66e..767068e85ea 100644
--- a/gcc/ipa-inline.cc
+++ b/gcc/ipa-inline.cc
@@ -124,6 +124,8 @@ along with GCC; see the file COPYING3.  If not see
 #include "ipa-modref-tree.h"
 #include "ipa-modref.h"
 
+extern void ipa_icf_create_deferred_aliases (void);
+
 /* Inliner uses greedy algorithm to inline calls in a priority order.
    Badness is used as the key in a Fibonacci heap which roughly corresponds
    to negation of benefit to cost ratios.
@@ -3004,6 +3006,9 @@ ipa_inline (void)
 
   if (dump_file)
     ipa_dump_fn_summaries (dump_file);
+
+  ipa_icf_create_deferred_aliases ();
+
   return remove_functions ? TODO_remove_functions : 0;
 }
 
diff --git a/gcc/testsuite/g++.dg/ipa/pr105438-2.C b/gcc/testsuite/g++.dg/ipa/pr105438-2.C
new file mode 100644
index 00000000000..64aa30c588a
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ipa/pr105438-2.C
@@ -0,0 +1,25 @@
+// PR ipa/105438
+// { dg-do compile }
+// { dg-options "-O2 -fno-strict-aliasing -fno-early-inlining -Warray-bounds" }
+
+int longer[7] = {};
+int shorter[2] = {};
+int out[10] = {};
+
+template <int N>
+void configure(const int(&in)[N], const int nrows = N)
+{
+    if (nrows <= 10)
+    {
+        for (int i = 0; i < nrows; i++)
+        {
+            out[i] = in[i]; // { dg-bogus "outside array bounds" }
+        }
+    }
+}
+
+int main()
+{
+  configure(longer);
+  configure(shorter);
+}
diff --git a/gcc/testsuite/g++.dg/ipa/pr105438.C b/gcc/testsuite/g++.dg/ipa/pr105438.C
new file mode 100644
index 00000000000..c84fcb67b9f
--- /dev/null
+++ b/gcc/testsuite/g++.dg/ipa/pr105438.C
@@ -0,0 +1,25 @@
+// PR ipa/105438
+// { dg-do compile }
+// { dg-options "-O2 -fno-strict-aliasing -Warray-bounds" }
+
+int longer[7] = {};
+int shorter[2] = {};
+int out[10] = {};
+
+template <int N>
+void configure(const int(&in)[N], const int nrows = N)
+{
+    if (nrows <= 10)
+    {
+        for (int i = 0; i < nrows; i++)
+        {
+            out[i] = in[i]; // { dg-bogus "outside array bounds" }
+        }
+    }
+}
+
+int main()
+{
+  configure(longer);
+  configure(shorter);
+}
diff --git a/gcc/tree-pass.h b/gcc/tree-pass.h
index a3b35e009e0..bdd60f72d81 100644
--- a/gcc/tree-pass.h
+++ b/gcc/tree-pass.h
@@ -539,6 +539,7 @@ extern simple_ipa_opt_pass *make_pass_ipa_free_fn_summary (gcc::context *ctxt);
 extern ipa_opt_pass_d *make_pass_ipa_cp (gcc::context *ctxt);
 extern ipa_opt_pass_d *make_pass_ipa_sra (gcc::context *ctxt);
 extern ipa_opt_pass_d *make_pass_ipa_icf (gcc::context *ctxt);
+extern void ipa_icf_create_deferred_aliases (void);
 extern ipa_opt_pass_d *make_pass_ipa_devirt (gcc::context *ctxt);
 extern ipa_opt_pass_d *make_pass_ipa_odr (gcc::context *ctxt);
 extern ipa_opt_pass_d *make_pass_ipa_reference (gcc::context *ctxt);
-- 
2.51.0