[PATCH] cse: Fix up CSE const_anchor handling [PR108193]

[PATCH] cse: Fix up CSE const_anchor handling [PR108193]

[Jakub Jelinek]

Hi!

The following testcase ICEs on aarch64, because insert_const_anchor
inserts invalid CONST_INT into the CSE tables - 0x80000000 for SImode.
The second hunk of the patch fixes that, the first one is to avoid
triggering undefined behavior at compile time during compute_const_anchors
computations - performing those additions and subtractions in
HOST_WIDE_INT means it can overflow for certain constants.

Bootstrapped/regtested on aarch64-linux (which does have nonzero
target.const_anchor) and x86_64-linux and i686-linux (which have it
zero).  Ok for trunk?

2022-12-22  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/108193
	* cse.cc (compute_const_anchors): Change n type to
	unsigned HOST_WIDE_INT, adjust comparison against it to avoid
	warnings.  Formatting fix.
	(insert_const_anchor): Use gen_int_mode instead of GEN_INT.

	* gfortran.dg/pr108193.f90: New test.

--- gcc/cse.cc.jj	2022-06-28 13:03:30.699692752 +0200
+++ gcc/cse.cc	2022-12-21 12:58:24.277945065 +0100
@@ -1169,14 +1169,14 @@ compute_const_anchors (rtx cst,
 		       HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
 		       HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
 {
-  HOST_WIDE_INT n = INTVAL (cst);
+  unsigned HOST_WIDE_INT n = INTVAL (cst);
 
   *lower_base = n & ~(targetm.const_anchor - 1);
-  if (*lower_base == n)
+  if ((unsigned HOST_WIDE_INT) *lower_base == n)
     return false;
 
-  *upper_base =
-    (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
+  *upper_base = ((n + (targetm.const_anchor - 1))
+		 & ~(targetm.const_anchor - 1));
   *upper_offs = n - *upper_base;
   *lower_offs = n - *lower_base;
   return true;
@@ -1193,7 +1193,7 @@ insert_const_anchor (HOST_WIDE_INT ancho
   rtx anchor_exp;
   rtx exp;
 
-  anchor_exp = GEN_INT (anchor);
+  anchor_exp = gen_int_mode (anchor, mode);
   hash = HASH (anchor_exp, mode);
   elt = lookup (anchor_exp, hash, mode);
   if (!elt)
--- gcc/testsuite/gfortran.dg/pr108193.f90.jj	2022-12-21 13:02:21.925513332 +0100
+++ gcc/testsuite/gfortran.dg/pr108193.f90	2022-12-21 13:01:38.595139040 +0100
@@ -0,0 +1,24 @@
+! PR rtl-optimization/108193
+! { dg-do compile { target pthread } }
+! { dg-options "-O2 -fsplit-loops -ftree-parallelize-loops=2 -fno-tree-dominator-opts" }
+
+subroutine foo (n, r)
+  implicit none
+  integer :: i, j, n
+  real :: s, r(*)
+
+  s = 0.0
+
+  do j = 1, 2
+     do i = j, n
+        s = r(i)
+     end do
+  end do
+
+  do i = 1, n
+     do j = i, n
+        s = s + 1
+     end do
+     r(i) = s
+  end do
+end subroutine foo

	Jakub

Re: [PATCH] cse: Fix up CSE const_anchor handling [PR108193]

[Richard Biener]

On Thu, 22 Dec 2022, Jakub Jelinek wrote:

> Hi!
> 
> The following testcase ICEs on aarch64, because insert_const_anchor
> inserts invalid CONST_INT into the CSE tables - 0x80000000 for SImode.
> The second hunk of the patch fixes that, the first one is to avoid
> triggering undefined behavior at compile time during compute_const_anchors
> computations - performing those additions and subtractions in
> HOST_WIDE_INT means it can overflow for certain constants.
> 
> Bootstrapped/regtested on aarch64-linux (which does have nonzero
> target.const_anchor) and x86_64-linux and i686-linux (which have it
> zero).  Ok for trunk?
> 
> 2022-12-22  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR rtl-optimization/108193
> 	* cse.cc (compute_const_anchors): Change n type to
> 	unsigned HOST_WIDE_INT, adjust comparison against it to avoid
> 	warnings.  Formatting fix.
> 	(insert_const_anchor): Use gen_int_mode instead of GEN_INT.
> 
> 	* gfortran.dg/pr108193.f90: New test.
> 
> --- gcc/cse.cc.jj	2022-06-28 13:03:30.699692752 +0200
> +++ gcc/cse.cc	2022-12-21 12:58:24.277945065 +0100
> @@ -1169,14 +1169,14 @@ compute_const_anchors (rtx cst,
>  		       HOST_WIDE_INT *lower_base, HOST_WIDE_INT *lower_offs,
>  		       HOST_WIDE_INT *upper_base, HOST_WIDE_INT *upper_offs)
>  {
> -  HOST_WIDE_INT n = INTVAL (cst);
> +  unsigned HOST_WIDE_INT n = INTVAL (cst);

UINTVAL?

Otherwise OK.

Thanks,
Richard.

>  
>    *lower_base = n & ~(targetm.const_anchor - 1);
> -  if (*lower_base == n)
> +  if ((unsigned HOST_WIDE_INT) *lower_base == n)
>      return false;
>  
> -  *upper_base =
> -    (n + (targetm.const_anchor - 1)) & ~(targetm.const_anchor - 1);
> +  *upper_base = ((n + (targetm.const_anchor - 1))
> +		 & ~(targetm.const_anchor - 1));
>    *upper_offs = n - *upper_base;
>    *lower_offs = n - *lower_base;
>    return true;
> @@ -1193,7 +1193,7 @@ insert_const_anchor (HOST_WIDE_INT ancho
>    rtx anchor_exp;
>    rtx exp;
>  
> -  anchor_exp = GEN_INT (anchor);
> +  anchor_exp = gen_int_mode (anchor, mode);
>    hash = HASH (anchor_exp, mode);
>    elt = lookup (anchor_exp, hash, mode);
>    if (!elt)
> --- gcc/testsuite/gfortran.dg/pr108193.f90.jj	2022-12-21 13:02:21.925513332 +0100
> +++ gcc/testsuite/gfortran.dg/pr108193.f90	2022-12-21 13:01:38.595139040 +0100
> @@ -0,0 +1,24 @@
> +! PR rtl-optimization/108193
> +! { dg-do compile { target pthread } }
> +! { dg-options "-O2 -fsplit-loops -ftree-parallelize-loops=2 -fno-tree-dominator-opts" }
> +
> +subroutine foo (n, r)
> +  implicit none
> +  integer :: i, j, n
> +  real :: s, r(*)
> +
> +  s = 0.0
> +
> +  do j = 1, 2
> +     do i = j, n
> +        s = r(i)
> +     end do
> +  end do
> +
> +  do i = 1, n
> +     do j = i, n
> +        s = s + 1
> +     end do
> +     r(i) = s
> +  end do
> +end subroutine foo
> 
> 	Jakub
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)