ipa: do not match bit-field predicate conditions against byte offsets

Message ID prr43p0s-s333-sq3p-6rp3-s3061r7p7938@fhfr.qr
State New
Headers
Series ipa: do not match bit-field predicate conditions against byte offsets |

Checks

Context Check Description
linaro-tcwg-bot/tcwg_simplebootstrap_build--master-arm-bootstrap success Build passed

Commit Message

Richard Biener Sept. 2, 2026, 8:05 a.m. UTC
  Conditions of IPA predicates record the position of an aggregate load as
a bit offset (ipa_load_from_parm_agg derives it from
get_ref_base_and_extent_hwi and add_condition stores it verbatim), while
the aggregate values of jump functions are indexed by byte offsets
(ipa_argagg_value::unit_offset).

evaluate_conditions_for_known_args bridged the two by simply dividing by
BITS_PER_UNIT, which for a bit-field silently drops its sub-byte
position and matches the constant recorded for the containing byte.  The
subsequent compatibility check only compares TYPE_SIZE, which is the
mode size and therefore equal for a narrow bit-field type and a char, so
the whole byte is then reinterpreted as the field with a
VIEW_CONVERT_EXPR.

In the testcase a 3-bit signed bit-field holding 1 sits at bits 2..4 of
a byte whose value is 4; the byte is reinterpreted as -4, the guard
((int) p.f3.f7) >= 0 folds to false, the guarded call edge gets a false
predicate and edge_set_predicate turns it into __builtin_unreachable.
Everything dominated by the guard is then removed, main loses its
return statement and falls through into _start.

Since ipa_argagg_value_list cannot represent a sub-byte position, skip
the lookup altogether when the condition is not byte aligned.

Boostrap and regtest running on x86_64-unknown-linux-gnu.

OK if that succeeds?

Thanks,
Richard.

	PR ipa/126153
	* ipa-fnsummary.cc (evaluate_conditions_for_known_args): Do not
	look up an aggregate value for a condition whose offset is not
	byte aligned.

	* gcc.dg/torture/pr126153.c: New test.

Assisted-By: Claude Opus 5
---
 gcc/ipa-fnsummary.cc                    |  6 ++-
 gcc/testsuite/gcc.dg/torture/pr126153.c | 52 +++++++++++++++++++++++++
 2 files changed, 57 insertions(+), 1 deletion(-)
 create mode 100644 gcc/testsuite/gcc.dg/torture/pr126153.c
  

Comments

Martin Jambor Sept. 2, 2026, 10:49 a.m. UTC | #1
Hello,

On Wed, Sep 02 2026, Richard Biener wrote:
> Conditions of IPA predicates record the position of an aggregate load as
> a bit offset (ipa_load_from_parm_agg derives it from
> get_ref_base_and_extent_hwi and add_condition stores it verbatim), while
> the aggregate values of jump functions are indexed by byte offsets
> (ipa_argagg_value::unit_offset).
>
> evaluate_conditions_for_known_args bridged the two by simply dividing by
> BITS_PER_UNIT, which for a bit-field silently drops its sub-byte
> position and matches the constant recorded for the containing byte.  The
> subsequent compatibility check only compares TYPE_SIZE, which is the
> mode size and therefore equal for a narrow bit-field type and a char, so
> the whole byte is then reinterpreted as the field with a
> VIEW_CONVERT_EXPR.
>
> In the testcase a 3-bit signed bit-field holding 1 sits at bits 2..4 of
> a byte whose value is 4; the byte is reinterpreted as -4, the guard
> ((int) p.f3.f7) >= 0 folds to false, the guarded call edge gets a false
> predicate and edge_set_predicate turns it into __builtin_unreachable.
> Everything dominated by the guard is then removed, main loses its
> return statement and falls through into _start.
>
> Since ipa_argagg_value_list cannot represent a sub-byte position, skip
> the lookup altogether when the condition is not byte aligned.
>
> Boostrap and regtest running on x86_64-unknown-linux-gnu.
>
> OK if that succeeds?

Yes.  Thank you very much for looking into this.

Martin


>
> Thanks,
> Richard.
>
> 	PR ipa/126153
> 	* ipa-fnsummary.cc (evaluate_conditions_for_known_args): Do not
> 	look up an aggregate value for a condition whose offset is not
> 	byte aligned.
>
> 	* gcc.dg/torture/pr126153.c: New test.
>
> Assisted-By: Claude Opus 5
> ---
>  gcc/ipa-fnsummary.cc                    |  6 ++-
>  gcc/testsuite/gcc.dg/torture/pr126153.c | 52 +++++++++++++++++++++++++
>  2 files changed, 57 insertions(+), 1 deletion(-)
>  create mode 100644 gcc/testsuite/gcc.dg/torture/pr126153.c
>
> diff --git a/gcc/ipa-fnsummary.cc b/gcc/ipa-fnsummary.cc
> index a6ccdb1852a..5ce4fb8bdc5 100644
> --- a/gcc/ipa-fnsummary.cc
> +++ b/gcc/ipa-fnsummary.cc
> @@ -417,7 +417,11 @@ evaluate_conditions_for_known_args (struct cgraph_node *node,
>  
>  	  if (tree sval = avals->safe_sval_at (c->operand_num))
>  	    val = ipa_find_agg_cst_from_init (sval, c->offset, c->by_ref);
> -	  if (!val)
> +	  /* ipa_argagg_value_list is indexed by byte offsets, so a condition
> +	     which does not start at a byte boundary (a bit-field) cannot be
> +	     looked up in it; the containing byte would be reinterpreted as
> +	     the whole field below.  */
> +	  if (!val && (c->offset % BITS_PER_UNIT) == 0)
>  	    {
>  	      ipa_argagg_value_list avs (avals);
>  	      val = avs.get_value (c->operand_num, c->offset / BITS_PER_UNIT,
> diff --git a/gcc/testsuite/gcc.dg/torture/pr126153.c b/gcc/testsuite/gcc.dg/torture/pr126153.c
> new file mode 100644
> index 00000000000..af37f47eef5
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/torture/pr126153.c
> @@ -0,0 +1,52 @@
> +/* { dg-do run { target le } } */
> +/* { dg-require-effective-target int32plus } */
> +/* { dg-options "-O2" } */
> +
> +/* Conditions of IPA predicates record the position of an aggregate load as
> +   a bit offset, while the aggregate values of jump functions are indexed by
> +   byte offsets.  Truncating the former used to match the constant recorded
> +   for the byte containing the bit-field, which was then reinterpreted with
> +   a VIEW_CONVERT_EXPR as if it were the whole field.  Here that turned the
> +   value 1 of f7 into -4, the guard below into a false predicate and the
> +   call to shifter() into __builtin_unreachable.  */
> +
> +struct S0 { unsigned f5 : 2; signed f7 : 3; unsigned f6 : 3; };
> +struct S1 { int a; struct S0 f3; char pad; short s; };
> +
> +int g;
> +
> +__attribute__((noipa)) void ext (int x) { g += x; }
> +
> +__attribute__((noinline, noclone)) static long
> +shifter (long l, int r)
> +{
> +  if (l < 0 || r < 0 || r >= 32 || l > (0x7fffffffffffffffL >> r))
> +    return l;
> +  return l << r;
> +}
> +
> +static void
> +callee (struct S1 p, int n)
> +{
> +  ext (n);
> +  if ((int) p.f3.f7 >= 0)
> +    ext ((int) shifter (0x350631DD6B880108LL, (int) p.f3.f7));
> +  ext (n);
> +}
> +
> +/* A second caller, so that callee is not inlined before IPA.  */
> +void other (struct S1 q, int n) { callee (q, n); }
> +
> +int
> +main (void)
> +{
> +  struct S1 l;
> +  l.a = 5;
> +  l.pad = 7;
> +  l.s = 9;
> +  *(char *) &l.f3 = 4;		/* f5 = 0, f7 = 1, f6 = 0 */
> +  callee (l, 3);
> +  if (g != 3 + 3 + (int) (0x350631DD6B880108LL << 1))
> +    __builtin_abort ();
> +  return 0;
> +}
> -- 
> 2.51.0
  

Patch

diff --git a/gcc/ipa-fnsummary.cc b/gcc/ipa-fnsummary.cc
index a6ccdb1852a..5ce4fb8bdc5 100644
--- a/gcc/ipa-fnsummary.cc
+++ b/gcc/ipa-fnsummary.cc
@@ -417,7 +417,11 @@  evaluate_conditions_for_known_args (struct cgraph_node *node,
 
 	  if (tree sval = avals->safe_sval_at (c->operand_num))
 	    val = ipa_find_agg_cst_from_init (sval, c->offset, c->by_ref);
-	  if (!val)
+	  /* ipa_argagg_value_list is indexed by byte offsets, so a condition
+	     which does not start at a byte boundary (a bit-field) cannot be
+	     looked up in it; the containing byte would be reinterpreted as
+	     the whole field below.  */
+	  if (!val && (c->offset % BITS_PER_UNIT) == 0)
 	    {
 	      ipa_argagg_value_list avs (avals);
 	      val = avs.get_value (c->operand_num, c->offset / BITS_PER_UNIT,
diff --git a/gcc/testsuite/gcc.dg/torture/pr126153.c b/gcc/testsuite/gcc.dg/torture/pr126153.c
new file mode 100644
index 00000000000..af37f47eef5
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr126153.c
@@ -0,0 +1,52 @@ 
+/* { dg-do run { target le } } */
+/* { dg-require-effective-target int32plus } */
+/* { dg-options "-O2" } */
+
+/* Conditions of IPA predicates record the position of an aggregate load as
+   a bit offset, while the aggregate values of jump functions are indexed by
+   byte offsets.  Truncating the former used to match the constant recorded
+   for the byte containing the bit-field, which was then reinterpreted with
+   a VIEW_CONVERT_EXPR as if it were the whole field.  Here that turned the
+   value 1 of f7 into -4, the guard below into a false predicate and the
+   call to shifter() into __builtin_unreachable.  */
+
+struct S0 { unsigned f5 : 2; signed f7 : 3; unsigned f6 : 3; };
+struct S1 { int a; struct S0 f3; char pad; short s; };
+
+int g;
+
+__attribute__((noipa)) void ext (int x) { g += x; }
+
+__attribute__((noinline, noclone)) static long
+shifter (long l, int r)
+{
+  if (l < 0 || r < 0 || r >= 32 || l > (0x7fffffffffffffffL >> r))
+    return l;
+  return l << r;
+}
+
+static void
+callee (struct S1 p, int n)
+{
+  ext (n);
+  if ((int) p.f3.f7 >= 0)
+    ext ((int) shifter (0x350631DD6B880108LL, (int) p.f3.f7));
+  ext (n);
+}
+
+/* A second caller, so that callee is not inlined before IPA.  */
+void other (struct S1 q, int n) { callee (q, n); }
+
+int
+main (void)
+{
+  struct S1 l;
+  l.a = 5;
+  l.pad = 7;
+  l.s = 9;
+  *(char *) &l.f3 = 4;		/* f5 = 0, f7 = 1, f6 = 0 */
+  callee (l, 3);
+  if (g != 3 + 3 + (int) (0x350631DD6B880108LL << 1))
+    __builtin_abort ();
+  return 0;
+}