This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] fix arm neon ICE by widening tree_type's precision field
- From: Nathan Froyd <froydnj at codesourcery dot com>
- To: gcc-patches at gcc dot gnu dot org
- Date: Mon, 8 Jun 2009 08:32:04 -0700
- Subject: [PATCH] fix arm neon ICE by widening tree_type's precision field
ARM's NEON support in GCC includes some rather wide types: XImode, the
biggest, is a 64-byte integer type. Compiling the simple testcase in
the patch below, which uses XImode, results in an ICE.
The reason behind this is because XImode is defined as a
FRACTIONAL_INT_MODE of 511 bits, rather than an INT_MODE of 64 bytes.
This hack was done because the 'precision' field of tree_type is only 9
bits wide...and therefore can't store XImode's actual precision of 512
bits. So when we ask mode_for_size for a MODE_INT of 512 bits, it
thinks that the only reasonable choice is BLKmode--since XImode is not
a MODE_INT mode. We eventually wind up calling simplify_subreg with the
input mode of BLKmode, which triggers an ICE.
The fix is to widen the 'precision' field of tree_type to 10 bits.
Unfortunately, there's not much space in tree_type to widen that field;
the structure is nicely packed for 32-bit and 64-bit hosts. The
approach I've taken in the patch below is to move packed_flag into
tree_base and rearrange the bitfields in tree_type. I realize this is
slightly gross, but I don't see a better way--suggestions welcome.
Tested with cross to arm-none-eabi. OK to commit? (Need middle-end
approval for the tree.h changes and ARM approval for arm-modes.def.)
-Nathan
2009-06-08 Nathan Froyd <froydnj@codesourcery.com>
* tree.h (tree_base): Add packed_flag, moved from...
(tree_type): ...here. Widen precision field to 10 bits and
rearrange fields to pack nicely.
(TYPE_PACKED): Update for new location of packed_flag.
* config/arm/arm-modes.def (XImode): Define as a INT_MODE.
Index: config/arm/arm-modes.def
===================================================================
--- config/arm/arm-modes.def (revision 148220)
+++ config/arm/arm-modes.def (working copy)
@@ -62,6 +62,4 @@ VECTOR_MODES (FLOAT, 16); /* V
INT_MODE (EI, 24);
INT_MODE (OI, 32);
INT_MODE (CI, 48);
-/* ??? This should actually have 512 bits but the precision only has 9
- bits. */
-FRACTIONAL_INT_MODE (XI, 511, 64);
+INT_MODE (XI, 64);
Index: testsuite/ChangeLog
===================================================================
--- testsuite/ChangeLog (revision 148220)
+++ testsuite/ChangeLog (working copy)
@@ -1,3 +1,7 @@
+2009-06-08 Nathan Froyd <froydnj@codesourcery.com>
+
+ * gcc.target/arm/neon-dse-1.c: New test.
+
2009-06-05 Jakub Jelinek <jakub@redhat.com>
PR middle-end/40340
Index: testsuite/gcc.target/arm/neon-dse-1.c
===================================================================
--- testsuite/gcc.target/arm/neon-dse-1.c (revision 0)
+++ testsuite/gcc.target/arm/neon-dse-1.c (revision 0)
@@ -0,0 +1,14 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_neon_ok } */
+/* { dg-options "-O1" } */
+/* { dg-add-options arm_neon } */
+
+#include <arm_neon.h>
+
+void neon_internal_error(int *dst, int *src)
+{
+ uint16x8x4_t sval;
+
+ sval = vld4q_u16((void *)src);
+ vst4q_u16((void *)dst,sval);
+}
Index: tree.h
===================================================================
--- tree.h (revision 148220)
+++ tree.h (working copy)
@@ -373,8 +373,10 @@ struct GTY(()) tree_base {
unsigned lang_flag_6 : 1;
unsigned visited : 1;
+ /* For tree_type. */
+ unsigned packed_flag : 1;
- unsigned spare : 23;
+ unsigned spare : 22;
union tree_ann_d *ann;
};
@@ -2219,7 +2221,7 @@ extern enum machine_mode vector_type_mod
/* Indicated that objects of this type should be laid out in as
compact a way as possible. */
-#define TYPE_PACKED(NODE) (TYPE_CHECK (NODE)->type.packed_flag)
+#define TYPE_PACKED(NODE) (TYPE_CHECK (NODE)->common.base.packed_flag)
/* Used by type_contains_placeholder_p to avoid recomputation.
Values are: 0 (unknown), 1 (false), 2 (true). Never access
@@ -2237,17 +2239,16 @@ struct GTY(()) tree_type {
tree attributes;
unsigned int uid;
- unsigned int precision : 9;
- ENUM_BITFIELD(machine_mode) mode : 7;
-
- unsigned string_flag : 1;
+ unsigned int precision : 10;
unsigned no_force_blk_flag : 1;
unsigned needs_constructing_flag : 1;
unsigned transparent_union_flag : 1;
- unsigned packed_flag : 1;
unsigned restrict_flag : 1;
unsigned contains_placeholder_bits : 2;
+ ENUM_BITFIELD(machine_mode) mode : 7;
+ unsigned string_flag : 1;
+
unsigned lang_flag_0 : 1;
unsigned lang_flag_1 : 1;
unsigned lang_flag_2 : 1;