This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[RFC, rs6000, v2] Fix alignment of non-Altivec vector struct fields
- From: "Ulrich Weigand" <uweigand at de dot ibm dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: dje dot gcc at gcc dot gnu dot org
- Date: Mon, 14 Jul 2014 20:55:44 +0200 (CEST)
- Subject: [RFC, rs6000, v2] Fix alignment of non-Altivec vector struct fields
- Authentication-results: sourceware.org; auth=none
Hello,
this is an attempt to update the prior patch:
https://gcc.gnu.org/ml/gcc-patches/2014-07/msg00635.html
to add a -Wpsabi note as discussed.
However, this is causing a bit of difficulties. First of all, the warning
triggers in a larger number of tests -- which was probably to be expected
since a number of tests in the suite explicitly work on GCC vectors defined
using various vector_size values, and any size except 16 will result in the
warning. Note that the warning gets issued at the site of definition of
the structure type -- it does not have to be used in a function call.
More problematic is that this new warning causes four tests to fail
(it's actually 38 new FAILS since each of the tests fails in multiple
optimization levels):
FAIL: gcc.c-torture/execute/20050316-1.c compilation
gcc.c-torture/execute/20050316-1.c:53:9: note: the layout of aggregates containing vectors with 8-byte alignment has changed in GCC 4.10
FAIL: gcc.c-torture/execute/20050316-3.c compilation
gcc.c-torture/execute/20050316-3.c:26:9: note: the layout of aggregates containing vectors with 8-byte alignment has changed in GCC 4.10
FAIL: gcc.c-torture/execute/20050604-1.c compilation
gcc.c-torture/execute/20050604-1.c:12:1: note: the layout of aggregates containing vectors with 8-byte alignment has changed in GCC 4.10
FAIL: gcc.c-torture/execute/pr23135.c compilation
gcc.c-torture/execute/pr23135.c:20:1: note: the layout of aggregates containing vectors with 8-byte alignment has changed in GCC 4.10
Unfortunately, while most of the tests in the suite use the dg-test
framework which ignores "note" messages, the gcc.c-torture/execute
tests still use the old "torture" test framework, which does *not*.
(Also, in the old framework you cannot even add dg-... commands to
ignore those messages for a certain test, or add a -Wno-psabi option.)
Therefore, I'm not proposing to commit this patch as-is, but would
like to ask for feedback on how to best proceed with this.
One option might be to move the affected tests to the dg-test framework.
Or else we might decide we don't actually need a warning for this change,
since it only affects GCC synthetic vectors, where we might not have made
strict ABI guarantees anyway ...
Thoughts?
Bye,
Ulrich
gcc/ChangeLog:
* config/rs6000/rs6000-protos.h (rs6000_special_adjust_field_align_p):
Add prototype.
* config/rs6000/rs6000.c (rs6000_special_adjust_field_align_p): New
function.
* config/rs6000/sysv4.h (ADJUST_FIELD_ALIGN): Call it.
* config/rs6000/linux64.h (ADJUST_FIELD_ALIGN): Likewise.
* config/rs6000/freebsd64.h (ADJUST_FIELD_ALIGN): Likewise.
gcc/testsuite/ChangeLog:
* gcc.target/powerpc/ppc64-abi-warn-3.c: New test.
Index: gcc/config/rs6000/rs6000-protos.h
===================================================================
--- gcc/config/rs6000/rs6000-protos.h (revision 212147)
+++ gcc/config/rs6000/rs6000-protos.h (working copy)
@@ -155,6 +155,7 @@
#ifdef TREE_CODE
extern unsigned int rs6000_data_alignment (tree, unsigned int, enum data_align);
+extern bool rs6000_special_adjust_field_align_p (tree, unsigned int);
extern unsigned int rs6000_special_round_type_align (tree, unsigned int,
unsigned int);
extern unsigned int darwin_rs6000_special_round_type_align (tree, unsigned int,
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 212147)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -5878,6 +5878,32 @@
return align;
}
+/* Previous GCC releases forced all vector types to have 16-byte alignment. */
+
+bool
+rs6000_special_adjust_field_align_p (tree field, unsigned int computed)
+{
+ if (TARGET_ALTIVEC && TREE_CODE (TREE_TYPE (field)) == VECTOR_TYPE)
+ {
+ if (computed != 128)
+ {
+ static bool warned;
+ if (!warned && warn_psabi)
+ {
+ warned = true;
+ inform (input_location,
+ "the layout of aggregates containing vectors with"
+ " %d-byte alignment has changed in GCC 4.10",
+ computed / BITS_PER_UNIT);
+ }
+ }
+ /* In current GCC there is no special case. */
+ return false;
+ }
+
+ return false;
+}
+
/* AIX increases natural record alignment to doubleword if the first
field is an FP double while the FP fields remain word aligned. */
Index: gcc/config/rs6000/sysv4.h
===================================================================
--- gcc/config/rs6000/sysv4.h (revision 212147)
+++ gcc/config/rs6000/sysv4.h (working copy)
@@ -292,7 +292,7 @@
/* An expression for the alignment of a structure field FIELD if the
alignment computed in the usual way is COMPUTED. */
#define ADJUST_FIELD_ALIGN(FIELD, COMPUTED) \
- ((TARGET_ALTIVEC && TREE_CODE (TREE_TYPE (FIELD)) == VECTOR_TYPE) \
+ (rs6000_special_adjust_field_align_p ((FIELD), (COMPUTED)) \
? 128 : COMPUTED)
#undef BIGGEST_FIELD_ALIGNMENT
Index: gcc/config/rs6000/linux64.h
===================================================================
--- gcc/config/rs6000/linux64.h (revision 212147)
+++ gcc/config/rs6000/linux64.h (working copy)
@@ -246,7 +246,7 @@
/* PowerPC64 Linux word-aligns FP doubles when -malign-power is given. */
#undef ADJUST_FIELD_ALIGN
#define ADJUST_FIELD_ALIGN(FIELD, COMPUTED) \
- ((TARGET_ALTIVEC && TREE_CODE (TREE_TYPE (FIELD)) == VECTOR_TYPE) \
+ (rs6000_special_adjust_field_align_p ((FIELD), (COMPUTED)) \
? 128 \
: (TARGET_64BIT \
&& TARGET_ALIGN_NATURAL == 0 \
Index: gcc/config/rs6000/freebsd64.h
===================================================================
--- gcc/config/rs6000/freebsd64.h (revision 212147)
+++ gcc/config/rs6000/freebsd64.h (working copy)
@@ -367,7 +367,7 @@
/* PowerPC64 Linux word-aligns FP doubles when -malign-power is given. */
#undef ADJUST_FIELD_ALIGN
#define ADJUST_FIELD_ALIGN(FIELD, COMPUTED) \
- ((TARGET_ALTIVEC && TREE_CODE (TREE_TYPE (FIELD)) == VECTOR_TYPE) \
+ (rs6000_special_adjust_field_align_p ((FIELD), (COMPUTED)) \
? 128 \
: (TARGET_64BIT \
&& TARGET_ALIGN_NATURAL == 0 \
Index: gcc/testsuite/gcc.target/powerpc/ppc64-abi-warn-3.c
===================================================================
--- /dev/null
+++ gcc/testsuite/gcc.target/powerpc/ppc64-abi-warn-3.c (revision 0)
@@ -0,0 +1,9 @@
+/* { dg-do compile { target { powerpc*-*-linux* && lp64 } } } */
+/* { dg-require-effective-target powerpc_altivec_ok } */
+/* { dg-options "-maltivec" } */
+
+struct test
+ {
+ int a __attribute__((vector_size (8)));
+ }; /* { dg-message "note: the layout of aggregates containing vectors with 8-byte alignment has changed" } */
+
--
Dr. Ulrich Weigand
GNU/Linux compilers and toolchain
Ulrich.Weigand@de.ibm.com