gimple-ssa-warn-access: Cast huge params to sizetype before using them in maybe_check_access_sizes [PR113410]

Message ID ZaesBsXuRvdIwLH9@tucnak
State Unresolved
Headers
Series gimple-ssa-warn-access: Cast huge params to sizetype before using them in maybe_check_access_sizes [PR113410] |

Checks

Context Check Description
snail/gcc-patch-check warning Git am fail log

Commit Message

Jakub Jelinek Jan. 17, 2024, 10:29 a.m. UTC
  Hi!

WHen a VLA is created with some very high precision size expression
(say __int128, or _BitInt(65535) etc.), we cast it to sizetype, because
we can't have arrays longer than what can be expressed in sizetype.

But the maybe_check_access_sizes code when trying to determine ranges
wasn't doing this but was using fixed buffers for the sizes.  While
__int128 could still be handled (fit into the buffers), obviously
arbitrary _BitInt parameter ranges can't, they can be in the range of
up to almost 20KB per number.  It doesn't make sense to print such
ranges though, no array can be larger than sizetype precision, and
ranger's range_of_expr can handle NOP_EXPRs/CONVERT_EXPRs wrapping a
PARM_DECL just fine, so the following patch just casts the excessively
large counters for the range determination purposes to sizetype.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2024-01-17  Jakub Jelinek  <jakub@redhat.com>

	PR middle-end/113410
	* gimple-ssa-warn-access.cc (pass_waccess::maybe_check_access_sizes):
	If access_nelts is integral with larger precision than sizetype,
	fold_convert it to sizetype.

	* gcc.dg/bitint-72.c: New test.


	Jakub
  

Comments

Richard Biener Jan. 17, 2024, 12:31 p.m. UTC | #1
On Wed, 17 Jan 2024, Jakub Jelinek wrote:

> Hi!
> 
> WHen a VLA is created with some very high precision size expression
> (say __int128, or _BitInt(65535) etc.), we cast it to sizetype, because
> we can't have arrays longer than what can be expressed in sizetype.
> 
> But the maybe_check_access_sizes code when trying to determine ranges
> wasn't doing this but was using fixed buffers for the sizes.  While
> __int128 could still be handled (fit into the buffers), obviously
> arbitrary _BitInt parameter ranges can't, they can be in the range of
> up to almost 20KB per number.  It doesn't make sense to print such
> ranges though, no array can be larger than sizetype precision, and
> ranger's range_of_expr can handle NOP_EXPRs/CONVERT_EXPRs wrapping a
> PARM_DECL just fine, so the following patch just casts the excessively
> large counters for the range determination purposes to sizetype.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK

> 2024-01-17  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR middle-end/113410
> 	* gimple-ssa-warn-access.cc (pass_waccess::maybe_check_access_sizes):
> 	If access_nelts is integral with larger precision than sizetype,
> 	fold_convert it to sizetype.
> 
> 	* gcc.dg/bitint-72.c: New test.
> 
> --- gcc/gimple-ssa-warn-access.cc.jj	2024-01-03 11:51:30.087751231 +0100
> +++ gcc/gimple-ssa-warn-access.cc	2024-01-16 19:25:35.408958088 +0100
> @@ -3406,6 +3406,15 @@ pass_waccess::maybe_check_access_sizes (
>        else
>  	access_nelts = rwm->get (sizidx)->size;
>  
> +      /* If access_nelts is e.g. a PARM_DECL with larger precision than
> +	 sizetype, such as __int128 or _BitInt(34123) parameters,
> +	 cast it to sizetype.  */
> +      if (access_nelts
> +	  && INTEGRAL_TYPE_P (TREE_TYPE (access_nelts))
> +	  && (TYPE_PRECISION (TREE_TYPE (access_nelts))
> +	      > TYPE_PRECISION (sizetype)))
> +	access_nelts = fold_convert (sizetype, access_nelts);
> +
>        /* Format the value or range to avoid an explosion of messages.  */
>        char sizstr[80];
>        tree sizrng[2] = { size_zero_node, build_all_ones_cst (sizetype) };
> --- gcc/testsuite/gcc.dg/bitint-72.c.jj	2024-01-16 19:31:33.839938120 +0100
> +++ gcc/testsuite/gcc.dg/bitint-72.c	2024-01-16 19:31:06.000328741 +0100
> @@ -0,0 +1,16 @@
> +/* PR middle-end/113410 */
> +/* { dg-do compile { target bitint } } */
> +/* { dg-options "-std=c23" } */
> +
> +#if __BITINT_MAXWIDTH__ >= 905
> +void bar (_BitInt(905) n, int[n]);
> +#else
> +void bar (int n, int[n]);
> +#endif
> +
> +void
> +foo (int n)
> +{
> +  int buf[n];
> +  bar (n, buf);
> +}
> 
> 	Jakub
> 
>
  

Patch

--- gcc/gimple-ssa-warn-access.cc.jj	2024-01-03 11:51:30.087751231 +0100
+++ gcc/gimple-ssa-warn-access.cc	2024-01-16 19:25:35.408958088 +0100
@@ -3406,6 +3406,15 @@  pass_waccess::maybe_check_access_sizes (
       else
 	access_nelts = rwm->get (sizidx)->size;
 
+      /* If access_nelts is e.g. a PARM_DECL with larger precision than
+	 sizetype, such as __int128 or _BitInt(34123) parameters,
+	 cast it to sizetype.  */
+      if (access_nelts
+	  && INTEGRAL_TYPE_P (TREE_TYPE (access_nelts))
+	  && (TYPE_PRECISION (TREE_TYPE (access_nelts))
+	      > TYPE_PRECISION (sizetype)))
+	access_nelts = fold_convert (sizetype, access_nelts);
+
       /* Format the value or range to avoid an explosion of messages.  */
       char sizstr[80];
       tree sizrng[2] = { size_zero_node, build_all_ones_cst (sizetype) };
--- gcc/testsuite/gcc.dg/bitint-72.c.jj	2024-01-16 19:31:33.839938120 +0100
+++ gcc/testsuite/gcc.dg/bitint-72.c	2024-01-16 19:31:06.000328741 +0100
@@ -0,0 +1,16 @@ 
+/* PR middle-end/113410 */
+/* { dg-do compile { target bitint } } */
+/* { dg-options "-std=c23" } */
+
+#if __BITINT_MAXWIDTH__ >= 905
+void bar (_BitInt(905) n, int[n]);
+#else
+void bar (int n, int[n]);
+#endif
+
+void
+foo (int n)
+{
+  int buf[n];
+  bar (n, buf);
+}