* [Ada] Alignment promotion vs packedness
@ 2008-03-08 11:31 Eric Botcazou
0 siblings, 0 replies; only message in thread
From: Eric Botcazou @ 2008-03-08 11:31 UTC (permalink / raw)
To: gcc-patches
[-- Attachment #1: Type: text/plain, Size: 718 bytes --]
The Ada compiler needs to promote the alignment of packed records if this is
deemed profitable, for example
package P is
type R1 is record
I : Integer;
end record;
pragma Pack (R1);
end P;
in order to avoid creating superfluous temporaries around function calls.
Fixed thusly, tested on i586-suse-linux, applied on the mainline.
2008-03-08 Eric Botcazou <ebotcazou@adacore.com>
* utils.c (finish_record_type): Clear DECL_BIT_FIELD on sufficiently
aligned bit-fields, bumping the alignment of the record type if deemed
profitable.
(value_factor_p): Return false instead of 0.
2008-03-08 Eric Botcazou <ebotcazou@adacore.com>
* gnat.dg/pack4.adb: New test.
--
Eric Botcazou
[-- Attachment #2: p.diff --]
[-- Type: text/x-diff, Size: 3542 bytes --]
Index: utils.c
===================================================================
--- utils.c (revision 133014)
+++ utils.c (working copy)
@@ -752,6 +752,7 @@ finish_record_type (tree record_type, tr
tree size = bitsize_zero_node;
bool had_size = TYPE_SIZE (record_type) != 0;
bool had_size_unit = TYPE_SIZE_UNIT (record_type) != 0;
+ bool had_align = TYPE_ALIGN (record_type) != 0;
tree field;
if (name && TREE_CODE (name) == TYPE_DECL)
@@ -804,24 +805,55 @@ finish_record_type (tree record_type, tr
for (field = fieldlist; field; field = TREE_CHAIN (field))
{
- tree pos = bit_position (field);
-
tree type = TREE_TYPE (field);
+ tree pos = bit_position (field);
tree this_size = DECL_SIZE (field);
- tree this_ada_size = DECL_SIZE (field);
+ tree this_ada_size;
- if ((TREE_CODE (type) == RECORD_TYPE || TREE_CODE (type) == UNION_TYPE
- || TREE_CODE (type) == QUAL_UNION_TYPE)
+ if ((TREE_CODE (type) == RECORD_TYPE
+ || TREE_CODE (type) == UNION_TYPE
+ || TREE_CODE (type) == QUAL_UNION_TYPE)
&& !TYPE_IS_FAT_POINTER_P (type)
&& !TYPE_CONTAINS_TEMPLATE_P (type)
&& TYPE_ADA_SIZE (type))
this_ada_size = TYPE_ADA_SIZE (type);
+ else
+ this_ada_size = this_size;
/* Clear DECL_BIT_FIELD for the cases layout_decl does not handle. */
- if (DECL_BIT_FIELD (field) && !STRICT_ALIGNMENT
- && value_factor_p (pos, BITS_PER_UNIT)
+ if (DECL_BIT_FIELD (field)
&& operand_equal_p (this_size, TYPE_SIZE (type), 0))
- DECL_BIT_FIELD (field) = 0;
+ {
+ unsigned int align = TYPE_ALIGN (type);
+
+ /* In the general case, type alignment is required. */
+ if (value_factor_p (pos, align))
+ {
+ /* The enclosing record type must be sufficiently aligned.
+ Otherwise, if no alignment was specified for it and it
+ has been laid out already, bump its alignment to the
+ desired one if this is compatible with its size. */
+ if (TYPE_ALIGN (record_type) >= align)
+ {
+ DECL_ALIGN (field) = MAX (DECL_ALIGN (field), align);
+ DECL_BIT_FIELD (field) = 0;
+ }
+ else if (!had_align
+ && rep_level == 0
+ && value_factor_p (TYPE_SIZE (record_type), align))
+ {
+ TYPE_ALIGN (record_type) = align;
+ DECL_ALIGN (field) = MAX (DECL_ALIGN (field), align);
+ DECL_BIT_FIELD (field) = 0;
+ }
+ }
+
+ /* In the non-strict alignment case, only byte alignment is. */
+ if (!STRICT_ALIGNMENT
+ && DECL_BIT_FIELD (field)
+ && value_factor_p (pos, BITS_PER_UNIT))
+ DECL_BIT_FIELD (field) = 0;
+ }
/* If we still have DECL_BIT_FIELD set at this point, we know the field
is technically not addressable. Except that it can actually be
@@ -830,7 +862,9 @@ finish_record_type (tree record_type, tr
DECL_NONADDRESSABLE_P (field)
|= DECL_BIT_FIELD (field) && DECL_MODE (field) != BLKmode;
- if ((rep_level > 0) && !DECL_BIT_FIELD (field))
+ /* A type must be as aligned as its most aligned field that is not
+ a bit-field. But this is already enforced by layout_type. */
+ if (rep_level > 0 && !DECL_BIT_FIELD (field))
TYPE_ALIGN (record_type)
= MAX (TYPE_ALIGN (record_type), DECL_ALIGN (field));
@@ -1800,7 +1834,7 @@ value_factor_p (tree value, HOST_WIDE_IN
return (value_factor_p (TREE_OPERAND (value, 0), factor)
|| value_factor_p (TREE_OPERAND (value, 1), factor));
- return 0;
+ return false;
}
/* Given 2 consecutive field decls PREV_FIELD and CURR_FIELD, return true
[-- Attachment #3: pack4.adb --]
[-- Type: text/x-adasrc, Size: 590 bytes --]
-- { dg-do run }
procedure Pack4 is
type Time_T is record
Hour : Integer;
end record;
type Date_And_Time_T is record
Date : Integer;
Time : Time_T;
end record;
pragma Pack(Date_And_Time_T);
procedure
Assign_Hour_Of (T : out Time_T)
is
begin
T.Hour := 44;
end;
procedure
Clobber_Hour_Of (DT: out Date_And_Time_T)
is
begin
Assign_Hour_Of (Dt.Time);
end;
DT : Date_And_Time_T;
begin
DT.Time.Hour := 22;
Clobber_Hour_Of (DT);
if DT.Time.Hour /= 44 then
raise Program_Error;
end if;
end;
^ permalink raw reply [flat|nested] only message in thread
only message in thread, other threads:[~2008-03-08 11:31 UTC | newest]
Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2008-03-08 11:31 [Ada] Alignment promotion vs packedness Eric Botcazou
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).