RE: [PATCH, i386] Handle extended family cpuid info for AMD
Uros!
I would like to have a check for a family at the beginning, something like:
if (name == signature_NSC_ebx)
processor = PROCESSOR_GEODE;
else if (family == 22)
{
if (has_movbe)
I get your idea of having the family checked first and then differentiating
with cpuid info later.
But, this case is getting interesting because, BTVER1 and BTVER2 are two
variants but don't really have same family numbers.
BTVER1 is family 14h and BTVER2 is family 16h. I don't see near term plans for
any additional cpus to either 14h or 16h.
Given that fact, this particular check is applicable only for BTVER2.
In that case, having
else if (family == 22)
if (has_movbe)
processor = PROCESSOR_BTVER2;
looks odd.
Regards
Ganesh
Re: [AArch64] Implement movmem for the benefit of inline memcpy
On Fri, Jun 6, 2014 at 1:50 AM, James Greenhalgh
james.greenha...@arm.com wrote:
Hi,
The move_by_pieces infrastructure performs a copy by repeatedly trying
the largest safe copy it can make. So for a 15-byte copy we might see:
offset amount bytes copied
08 0-7
84 8-11
12 2 12-13
14 1 14
However, we can implement a 15-byte copy as so:
offset amount bytes copied
08 0-7
78 7-14
Which can prove more efficient for both space and speed.
In this patch we set MOVE_RATIO low to avoid using move_by_pieces, and
implement the movmem pattern name to expand small block copy cases. Note, this
optimization does not apply for -mstrict-align targets, which must continue
copying byte-by-byte.
Setting MOVE_RATIO low in this way causes a few tests to begin failing,
both of these are documented in the test-case as expected to fail for
low MOVE_RATIO targets, which do not allow certain Tree-Level optimizations.
I think you should reevaluate setting MOVE_RATIO this low. It is used
for SRA and IPA-SRA which both are very useful; more useful than
memmove optimizations can do.
In fact this optimization is not even valid for volatile variables.
Here is a testcase for the volatile issue:
struct __attribute__((packed)) t15{
long long t8;
int t4;
short t2;
unsigned char t1;
};
volatile struct t15 t15;
int f(struct t15 *a)
{
t15 = *a;
}
Notice how we are writing to byte 7 twice to t15 in the outputted code.
Thanks,
Andrew Pinski
Bootstrapped on aarch64-unknown-linux-gnu with no issues.
OK for trunk?
Thanks,
James
---
gcc/
2014-06-06 James Greenhalgh james.greenha...@arm.com
* config/aarch64/aarch64-protos.h (aarch64_expand_movmem): New.
* config/aarch64/aarch64.c (aarch64_move_pointer): New.
(aarch64_progress_pointer): Likewise.
(aarch64_copy_one_part_and_move_pointers): Likewise.
(aarch64_expand_movmen): Likewise.
* config/aarch64/aarch64.h (MOVE_RATIO): Set low.
* config/aarch64/aarch64.md (movmemmode): New.
gcc/testsuite/
2014-06-06 James Greenhalgh james.greenha...@arm.com
* gcc.dg/tree-ssa/pr42585.c: Skip for AArch64.
* gcc.dg/tree-ssa/sra-12.c: Likewise.
Re: [PATCH] Keep patch file permissions in mklog
On 01-08-14 06:21, Yury Gribov wrote: Cc Diego. --- From: Tom de Vries tom_devr...@mentor.com Sent: Saturday, May 31, 2014 11:47AM To: Geoff Keating geo...@geoffk.org Cc: GCC Patches Subject: [PATCH] Keep patch file permissions in mklog On 05/31/2014 11:47 AM, Tom de Vries wrote: Geoff, Currently, mklog resets the permissions on a patch file: Yuri, Diego, The permissions problem does not occur anymore with the latests versions of mklog, since it doesn't modify the patch file anymore. This patch adds a script contrib/mklog-in-patch, which uses mklog to generate the skeleton log, but generates the log at the start of the patch as mklog did before (which is how I like to use it). OK for trunk? I can also try to add an --inline option to mklog instead. Thanks, - Tom 2014-08-01 Tom de Vries t...@codesourcery.com * mklog-in-patch: New script. diff --git a/contrib/mklog-in-patch b/contrib/mklog-in-patch new file mode 100755 index 000..6d041eb --- /dev/null +++ b/contrib/mklog-in-patch @@ -0,0 +1,61 @@ +#!/bin/sh + +# Copyright (C) 2014 Free Software Foundation, Inc. +# +# This file is part of GCC. +# +# GCC is free software; you can redistribute it and/or modify +# it under the terms of the GNU General Public License as published by +# the Free Software Foundation; either version 3, or (at your option) +# any later version. +# +# GCC is distributed in the hope that it will be useful, +# but WITHOUT ANY WARRANTY; without even the implied warranty of +# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the +# GNU General Public License for more details. +# +# You should have received a copy of the GNU General Public License +# along with GCC; see the file COPYING. If not, write to +# the Free Software Foundation, 51 Franklin Street, Fifth Floor, +# Boston, MA 02110-1301, USA. + + +# Check for errors +set -e + +scriptdir=$(dirname $0) +prog=$(basename $0) + +diff=$1 + +if [ $diff = ]; then + usage +echo EOF +usage: $prog file.diff + +Invokes mklog, and prepends the generated log to file.diff. +EOF +elif [ ! -r $diff ]; then +echo Could not open file $diff for reading +exit 1 +fi + +temp=$(mktemp) + +# Generate log in $temp +if ! $scriptdir/mklog $diff $temp; then +echo Could not write log file to $temp +exit 1 +fi + +# Append patch to $temp, and move $temp to $patch. For the latter, we use cat +# rather than mv to keep $diff permissions the same. +if ! ( cat $diff $temp cat $temp $diff ); then +echo Could not add the ChangeLog entry to $diff +exit 1 +fi + +if ! rm $temp; then +echo Could not remove $temp +exit 1 +fi
Re: [PATCH] Fix PR61375: cancel bswap optimization when value doesn't fit in a HOST_WIDE_INT
On Thu, Jul 31, 2014 at 02:57:34PM +0800, Thomas Preud'homme wrote: Now that GCC 4.9 branch is opened again and GCC 4.8 branch still open, is the following backported patch ok for both branches? Ok. Ok, what about the following patch and associated ChangeLog entries? 2014-06-24 Thomas Preud'homme thomas.preudho...@arm.com PR tree-optimization/61375 * tree-ssa-math-opts.c (find_bswap_or_nop_1): Cancel optimization if symbolic number cannot be represented in an unsigned HOST_WIDE_INT. (execute_optimize_bswap): Cancel optimization if CHAR_BIT != 8. 2014-06-24 Thomas Preud'homme thomas.preudho...@arm.com PR tree-optimization/61375 * gcc.c-torture/execute/pr61375-1.c: New test. Jakub
Re: [PATCH] Move Asan instrumentation to sanopt pass
On Mon, Jul 28, 2014 at 06:14:08PM +0400, Yury Gribov wrote:
On 07/24/2014 11:48 AM, Jakub Jelinek wrote:
So, either support for just EAF*, or perhaps support for DECL_ATTRIBUTES
for internal-fns, say by having some tree array where you'd store what you
stick into DECL_ATTRIBUTES normally.
I'd prefer to avoid attributes. Would something like this be enough?
(not yet fully tested, just ran asan.exp tests).
I think I'd prefer if you'd specify the fnspec as either NULL (most internal
functions) or normal string and reuse the string parsing in
gimple_call_arg_flags.
So, you'd just replace the:
tree attr = gimple_call_fnspec (stmt);
if (!attr || 1 + arg = (unsigned) TREE_STRING_LENGTH (attr))
return 0;
switch (TREE_STRING_POINTER (attr)[1 + arg])
with:
const char *fnspec;
if (gimple_call_arg_flags (stmt))
{
fnspec = ...;
if (!fnspec || 1 + arg = strlen (fnspec))
return 0;
}
else
{
tree attr = gimple_call_fnspec (stmt);
if (!attr || 1 + arg = (unsigned) TREE_STRING_LENGTH (attr))
return 0;
fnspec = TREE_STRING_POINTER (attr);
}
switch (fnspec[1 + arg])
Jakub
Re: [PATCH] Keep patch file permissions in mklog
On 08/01/2014 10:52 AM, Tom de Vries wrote: This patch adds a script contrib/mklog-in-patch, which uses mklog to generate the skeleton log, but generates the log at the start of the patch as mklog did before (which is how I like to use it). Yeah, we had some argument about this but kind of agreed that separate log is preferred. I can also try to add an --inline option to mklog instead. I'd prefer this. -Y
Re: FWD: Re: OpenACC subarray specifications in the GCC Fortran front end
On Wed, Jul 23, 2014 at 05:42:32PM -0700, Cesar Philippidis wrote:
Jakub, before your Fortran OpenMP 4 target changes, Ilmir had written the
test case gcc/testsuite/gfortran.dg/gomp/map-1.f90 (based on his
interpretation and implementation of OpenMP 4 target), which I have now
amended with XFAILs and changed error messages -- anything in there that
you'd like to see addressed for Fortran OpenMP 4 target?
+ !$omp target map(j(5:4)) ! { dg-error Lower bound of OpenMP array
section in greater than upper { xfail *-*-* } }
+ !$omp end target
I think this isn't an error in Fortran, if low bound is above upper bound,
then it is considered a zero size array section. Though supposedly for
depend clause we might want to diagnose that.
+ !$omp target map(aas) ! { dg-error The upper bound in the last
dimension must appear { xfail *-*-* } }
+ !$omp end target
Assumed-size in map without array section would be indeed nice thing to
diagnose.
+ !$omp target map(tt%i) ! { dg-error Syntax error in OpenMP variable
list }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
Right now the parsing of !$omp directives in case of parsing error rejects
the whole directive, perhaps it should be reconsidered unless it is a fatal
error from which there is no easy way out.
+ !$omp target map(tt%j(1)) ! { dg-bogus Syntax error in OpenMP variable
list { xfail *-*-* } }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
+
+ !$omp target map(tt%j(1:)) ! { dg-bogus Syntax error in OpenMP
variable list { xfail *-*-* } }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
These two are pending resolution on omp-lang, I had exchanged a few mails
about it, I think we shouldn't support those for consistency with the C/C++
support, where tt.j[1] or tt.j[1:] and similar is explicitly invalid.
Jakub, should I drop the map-1.f90 test?
Not the whole testcase, just the problematic parts (or, just remove the
dg-error/xfail or replace dg-bogus xfail with dg-error), for the j(5:4)
and tt%j(1)/tt%j(1:) cases?
Jakub
Re: [PATCH, i386] Handle extended family cpuid info for AMD
On Fri, Aug 1, 2014 at 8:18 AM, Gopalasubramanian, Ganesh
ganesh.gopalasubraman...@amd.com wrote:
I would like to have a check for a family at the beginning, something like:
if (name == signature_NSC_ebx)
processor = PROCESSOR_GEODE;
else if (family == 22)
{
if (has_movbe)
I get your idea of having the family checked first and then differentiating
with cpuid info later.
But, this case is getting interesting because, BTVER1 and BTVER2 are two
variants but don't really have same family numbers.
BTVER1 is family 14h and BTVER2 is family 16h. I don't see near term plans
for any additional cpus to either 14h or 16h.
Given that fact, this particular check is applicable only for BTVER2.
In that case, having
else if (family == 22)
if (has_movbe)
processor = PROCESSOR_BTVER2;
looks odd.
In this case, having only check for family ID should be enough. If
BTVER1 and BTVER2 can be uniquely determined by their family IDs ,
then:
... (signature checks)
else if (family == 0x16h)
processor = PROCESSOR_BTVER2;
else if (family == 0x14h)
processor = PROCESSOR_BTVER1;
else
... (detect processor in a generic way by its cpuid features.)
IMO, this would be the most future-proof approach. Signature checks
will override family id checks which will override cpuid checks.
Uros.
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
On Thu, Jul 24, 2014 at 04:39:28PM +0400, Maxim Ostapenko wrote: Ping. Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time 2) recently a bug has been reported that some ICEs which print RTL to stderr are considered non-reproduceable, because the addresses e.g. in call_insn fndecl change through address space randomization. For the retries the driver should supposedly append -fdump-noaddr. Jakub
Re: [Patch] PR 61692 - Fix for inline asm ICE
On 7/30/2014 9:58 PM, Jeff Law wrote: On 07/28/14 16:39, David Wohlferd wrote: On 7/28/2014 12:42 PM, Jeff Law wrote: On 07/27/14 01:26, David Wohlferd wrote: I'm not sure which maintainer to cc for inline asm stuff? I have a release on file with the FSF, but don't have SVN write access. Problem: extract_insn() in recog.c will ICE if (noperands MAX_RECOG_OPERANDS). Normally this isn't a problem since expand_asm_operands() in cfgexpand.c catches and reports a proper error for this condition. However, expand_asm_operands() only checks (ninputs + noutputs) instead of (ninputs + noutputs + nlabels), so you can get the ICE when using asm goto. See the bugzilla entry for sample code. ChangeLog: 2014-07-27 David Wohlferd d...@limegreensocks.com PR target/61692 * cfgexpand.c (expand_asm_operands): Count all inline asm parameters. You should also include 'nclobbers'. Reading thru asm_noperands (which is what extract_insn uses to count operands), I would have thought you were right. But while making this fail with nLabels was easy, I wasn't able to get this to ICE at all using clobbers (30 labels + 11 clobbers still didn't ICE). And I'm reluctant to propose that change unless I can see it fail. I understand, but I'm still quite confident it's the right thing to do. Running that 30 label + 11 clobber testcase under valgrind might show the problem, if you can stand waiting that long... I'd love to. Unfortunately, my platform doesn't support valgrind. Also, please include the testcase you had nlabels part. I have created the testcase for the 31 labels problem. However, not so much for the nclobbers part. And if I'm going to patch both, I should have testcases for both. I also worry about potentially breaking existing code. While the 31 labels thing I proposed won't break existing code (it would have been ICE-ing), adding nclobbers to the count of parameters could do so. When Jeff Law says that something is so about gcc, I tend to believe that it is so. But, unless I can see an example of the actual problem here, I have no idea how to create a test case for it. So I spent the last several hours hunting for it. If valgrind was going to show the problem, presumably we were expecting some type of array overrun. And there are a LOT of places that do foo[MAX_RECOG_OPERANDS]. Absent valgrind, I resorted to using printfs sprinkled throughout the code to check indices. However, despite my best efforts, I was unable to see any out of range accesses, gcc_asserts, segment faults or any other indications of a problem. All the places I tried seem to treat clobbers separately. That doesn't prove anything: I could easily have missed something. But I did make a sincere effort. Given how confident you are that there's a problem, could you point out a place+condition I should focus on? That way I can create the additional test case, satisfy my curiosity, and sleep soundly at night knowing I haven't (unnecessarily) broken people's code. Thanks, dw
[Ada] Remove VMS handling in most tools
Continuation of VMS removal in the GNAT front-end.
Tested on x86_64-pc-linux-gnu, committed on trunk
2014-08-01 Arnaud Charlet char...@adacore.com
* binde.adb, bindgen.adb, butil.adb, clean.adb, gnatbind.adb,
gnatchop.adb, gnatcmd.adb, gnatls.adb, gnatname.adb, krunch.adb,
make.adb, makeutl.adb, memtrack.adb, mlib-prj.adb, mlib.adb,
mlib.ads, tempdir.adb: Remove VMS handling.
Index: binde.adb
===
--- binde.adb (revision 213263)
+++ binde.adb (working copy)
@@ -6,7 +6,7 @@
-- --
-- B o d y --
-- --
--- Copyright (C) 1992-2013, Free Software Foundation, Inc. --
+-- Copyright (C) 1992-2014, Free Software Foundation, Inc. --
-- --
-- GNAT is free software; you can redistribute it and/or modify it under --
-- terms of the GNU General Public License as published by the Free Soft- --
@@ -31,7 +31,6 @@
with Opt; use Opt;
with Osint;
with Output; use Output;
-with Targparm; use Targparm;
with System.Case_Util; use System.Case_Util;
@@ -1089,12 +1088,7 @@
if Pessimistic_Elab_Order
and not Dynamic_Elaboration_Checks_Specified
then
- if OpenVMS_On_Target then
-Error_Msg (?use of /PESSIMISTIC_ELABORATION questionable);
- else
-Error_Msg (?use of -p switch questionable);
- end if;
-
+ Error_Msg (?use of -p switch questionable);
Error_Msg (?since all units compiled with static elaboration model);
end if;
Index: bindgen.adb
===
--- bindgen.adb (revision 213329)
+++ bindgen.adb (working copy)
@@ -52,10 +52,6 @@
Last : Natural := 0;
-- Last location in Statement_Buffer currently set
- With_DECGNAT : Boolean := False;
- -- Flag which indicates whether the program uses the DECGNAT library
- -- (presence of the unit DEC).
-
With_GNARL : Boolean := False;
-- Flag which indicates whether the program uses the GNARL library
-- (presence of the unit System.OS_Interface)
@@ -325,9 +321,7 @@
-- Move routine for sorting linker options
procedure Resolve_Binder_Options;
- -- Set the value of With_GNARL and With_DECGNAT. The latter only on VMS
- -- since it tests for a package named dec which might cause a conflict
- -- on non-VMS systems.
+ -- Set the value of With_GNARL.
procedure Set_Char (C : Character);
-- Set given character in Statement_Buffer at the Last + 1 position
@@ -659,36 +653,6 @@
__gnat_finalize_library_objects););
end if;
- -- Import entry point for environment feature enable/disable
- -- routine, and indication that it's been called previously.
-
- if OpenVMS_On_Target then
-WBI ();
-WBI ( procedure Set_Features;);
-WBI ( pragma Import (C, Set_Features,
- __gnat_set_features););
-WBI ();
-WBI ( Features_Set : Integer;);
-WBI ( pragma Import (C, Features_Set,
- __gnat_features_set););
-
-if Opt.Heap_Size /= 0 then
- WBI ();
- WBI ( Heap_Size : Integer;);
- WBI ( pragma Import (C, Heap_Size,
-__gl_heap_size););
-
- Write_Statement_Buffer;
-end if;
-
-WBI ();
-WBI ( Float_Format : Character;);
-WBI ( pragma Import (C, Float_Format,
-__gl_float_format););
-
-Write_Statement_Buffer;
- end if;
-
-- Initialize stack limit variable of the environment task if the
-- stack check method is stack limit and stack check is enabled.
@@ -886,44 +850,6 @@
WBI ( Install_Handler;);
WBI ( end if;);
end if;
-
- -- Generate call to Set_Features
-
- if OpenVMS_On_Target then
-
--- Set_Features will call IEEE$SET_FP_CONTROL appropriately
--- depending on the setting of Float_Format.
-
-WBI ();
-Set_String ( Float_Format := ');
-
-if Float_Format_Specified = 'G'
- or else
- Float_Format_Specified = 'D'
-then
- Set_Char ('V');
-else
- Set_Char ('I');
-end if;
-
-Set_String (';);
-Write_Statement_Buffer;
-
-WBI ();
-WBI ( if Features_Set = 0 then);
-WBI ( Set_Features;);
-
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
On 08/01/2014 12:00 PM, Jakub Jelinek wrote: Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time I think we should use David Malcolm's approach i.e. add some --report-bug flag to driver. This could be enabled by default at configure time via --with-spec. 2) recently a bug has been reported that some ICEs which print RTL to stderr are considered non-reproduceable, because the addresses e.g. in call_insn fndecl change through address space randomization. For the etries the driver should supposedly append -fdump-noaddr. +1 -Y
[Ada] More VMS clean ups
In various parts of the front-end. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Robert Dewar de...@adacore.com * inline.adb, inline.ads, fe.h, einfo.adb, einfo.ads, sem_util.adb, sem_util.ads, exp_ch4.adb, exp_ch11.adb, exp_ch6.adb, cstand.adb, sem_mech.adb, sem_ch6.adb, sem_ch8.adb, sem_ch11.adb, snames.ads-tmpl: Remove VMS-specific code. Index: inline.adb === --- inline.adb (revision 213373) +++ inline.adb (working copy) @@ -165,10 +165,10 @@ function Has_Single_Return (N : Node_Id) return Boolean; -- In general we cannot inline functions that return unconstrained type. - -- However, we can handle such functions if all return statements return - -- a local variable that is the only declaration in the body of the - -- function. In that case the call can be replaced by that local - -- variable as is done for other inlined calls. + -- However, we can handle such functions if all return statements return a + -- local variable that is the only declaration in the body of the function. + -- In that case the call can be replaced by that local variable as is done + -- for other inlined calls. function In_Main_Unit_Or_Subunit (E : Entity_Id) return Boolean; -- Return True if E is in the main unit or its spec or in a subunit @@ -429,7 +429,7 @@ procedure Register_Backend_Not_Inlined_Subprogram (Subp : Entity_Id); -- Append Subp to the list of subprograms that cannot be inlined by - -- the backend + -- the backend. -- Back_End_Cannot_Inline -- @@ -3332,7 +3332,7 @@ -- expanded into a procedure call which must be added after the -- object declaration. -if Is_Unc_Decl and then Back_End_Inlining then +if Is_Unc_Decl and Back_End_Inlining then Insert_Action_After (Parent (N), Blk); else Set_Expression (Parent (N), Empty); @@ -4329,9 +4329,9 @@ return False; end Has_Initialized_Type; - - -- Has_Single_Return -- - + --- + -- Has_Single_Return -- + --- function Has_Single_Return (N : Node_Id) return Boolean is Return_Statement : Node_Id := Empty; @@ -4376,8 +4376,8 @@ return Abandon; end if; - -- We can only inline a build-in-place function if - -- it has a single extended return. + -- We can only inline a build-in-place function if it has a single + -- extended return. elsif Nkind (N) = N_Extended_Return_Statement then if No (Return_Statement) then @@ -4572,6 +4572,8 @@ -- Number_Of_Statements -- -- + -- Why not List_Length??? + function Number_Of_Statements (Stats : List_Id) return Natural is Stat_Count : Integer := 0; Stmt : Node_Id; Index: inline.ads === --- inline.ads (revision 213373) +++ inline.ads (working copy) @@ -131,6 +131,9 @@ Table_Increment = Alloc.Pending_Instantiations_Increment, Table_Name = Pending_Descriptor); + -- The following should be initialized in an init call in Frontend, we + -- have thoughts of making the frontend reusable in future ??? + Inlined_Calls : Elist_Id := No_Elist; -- List of frontend inlined calls @@ -242,13 +245,14 @@ function Has_Excluded_Declaration (Subp : Entity_Id; Decls : List_Id) return Boolean; - -- Check for declarations that make inlining not worthwhile inlining Subp + -- Check a list of declarations, Decls, that make the inlining of Subp not + -- worthwhile function Has_Excluded_Statement (Subp : Entity_Id; Stats : List_Id) return Boolean; - -- Check for statements that make inlining not worthwhile: any tasking - -- statement, nested at any level. + -- Check a list of statements, Stats, that make inlining of Subp not + -- worthwhile, including any tasking statement, nested at any level. procedure Register_Backend_Call (N : Node_Id); -- Append N to the list Backend_Calls Index: fe.h === --- fe.h(revision 213353) +++ fe.h(working copy) @@ -154,11 +154,6 @@ extern Boolean Is_Fully_Repped_Tagged_Type (Entity_Id); -/* exp_vfpt: */ - -#define Get_Vax_Real_Literal_As_Signed exp_vfpt__get_vax_real_literal_as_signed -extern Ureal Get_Vax_Real_Literal_As_Signed (Node_Id); - /* lib: */ #define Cunit lib__cunit Index: einfo.adb === --- einfo.adb (revision 213371) +++ einfo.adb (working copy) @@ -195,7 +195,6 @@
[PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
currently, aarch64 LEGITIMIZE_ADDRESS_P hook will reject all reg + offset
address given
offset is beyond supported range.
while this may be too strict. we should honor the strict_p parameter in the
hook. before
reload, we accept all offset if it's a frame access, because the offset may
change during
later register elimination.
the early reject of reg + offset may cause extra registers created, and if
that register
live range is across function invoking then callee saved reg needed, thus
introduce extra
reg save/restore also.
give a simple example as:
int
test15 (void)
{
unsigned char a[480];
initialize_array (a, 480);
if (a[0] == 0x10)
return 1;
return 0;
}
.S before the patch
(-O2 -fPIC)
===
test15:
sub sp, sp, #480
mov w1, 480
stp x29, x30, [sp, -32]!
add x29, sp, 0
str x19, [sp, 16]
add x19, x29, 32
mov x0, x19
bl initialize_array
ldrbw0, [x19]
ldr x19, [sp, 16]
ldp x29, x30, [sp], 32
cmp w0, 16
csetw0, eq
add sp, sp, 480
ret
.S after the patch
===
test15:
stp x29, x30, [sp, -496]!
mov w1, 480
add x29, sp, 0
add x0, x29, 16
bl initialize_array
ldrbw0, [x29, 16]
ldp x29, x30, [sp], 496
cmp w0, 16
csetw0, eq
ret
test done
=
no regression on aarch64-none-elf bare metal.
bootstrap OK on aarch64.
OK for trunk?
thanks.
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset for
frame access
when strict_p is false.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index ed80269..c8e5808 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -3248,6 +3248,21 @@ aarch64_classify_address (struct aarch64_address_info *info,
case PLUS:
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
+
+ if (! strict_p
+ GET_CODE (op0) == REG
+ (op0 == virtual_stack_vars_rtx
+ || op0 == frame_pointer_rtx
+ || op0 == arg_pointer_rtx)
+ GET_CODE (op1) == CONST_INT)
+ {
+ info-type = ADDRESS_REG_IMM;
+ info-base = op0;
+ info-offset = op1;
+
+ return true;
+ }
+
if (GET_MODE_SIZE (mode) != 0
CONST_INT_P (op1)
aarch64_base_register_rtx_p (op0, strict_p))
diff --git a/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c b/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c
new file mode 100644
index 000..e5f0bd1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c
@@ -0,0 +1,21 @@
+/* { dg-do compile } */
+/* { dg-options -O2 -fdump-rtl-expand } */
+
+extern void initialize_array (unsigned char *, int);
+
+int
+test15 (void)
+{
+ unsigned char a[480];
+
+ initialize_array (a, 480);
+
+ if (a[0] == 0x10)
+return 1;
+
+ return 0;
+}
+
+/* { dg-final { scan-rtl-dump mem/j/c:QI \\(plus:DI expand } } */
+
+/* { dg-final { cleanup-rtl-dump expand } } */
Re: [PATCH][1-3] New configure options that make the compiler use -fPIE and -pie as default option
Hi Magnus,
a couple of comments, mostly nits.
2014-07-31 Magnus Granberg zo...@gentoo.org
/gcc
* config/gnu-user.h: Define PIE_DRIVER_SELF_SPECS for PIE
as default and GNU_DRIVER_SELF_SPECS.
* config/i386/gnu-user-common.h: Define DRIVER_SELF_SPECS
* configure.ac: Add new option that enable PIE as default.
* configure, config.in: Rebuild.
* Makefile.in: Disable PIE when building the compiler.
* doc/install.texi: Add the new configure option default PIE.
* doc/invoke.texi: Add note for the new configure option default PIE.
Many of those entries are mis-formatted. See other examples and the GNU
Coding Standards for details. E.g. the first would be
* config/gnu-user.h (PIE_DRIVER_SELF_SPECS): Define.
In general, you need to mention which macro, variable, manual section
you change. Emacs' add-change-log-entry does the basics for you.
Besides, you only state what changed, not why.
Apart from that, I don't think defining PIE_DRIVER_SELF_SPECS in
gnu-user.h is a good idea. This way, every other target supporting the
option would have to duplicate that stuff.
* testsuite/gcc/default-pie.c: New test for new configure option
--enale-default-pie
gcc/testsuite has its own ChangeLog file. Typo for --enale-...
* testsuite/gcc.dg/other/anon5.C: Add skip test as it fail to link
on effective_target default_pie.
should be
* g++.dg/other/anon5.C: Skip if default_pie.
No explanations in ChangeLog entries; they belong into the code.
Besides, you had the first dir component wrong. Again, Emacs does this
for you.
* testsuite/lib/target-supports.exp (check_profiling_available):
We can't use profiling on effective target default_pie.
(check_effective_target_pie): Add check_effective_target_default_pie.
Wrong: should be
* lib/target-supports.exp (check_effective_target_default_pie):
New proc.
The new default_pic effective-target keyword needs to be documented in
doc/sourcebuild.texi.
--- a/gcc/testsuite/gcc.dg/default-pie.c2013-11-09 21:07:16.741479728
+0100
+++ b/gcc/testsuite/gcc.dg/default-pie.c2013-11-09 21:05:07.801479218
+0100
@@ -0,0 +1,12 @@
+/* { dg-do compile { target *-*-linux* *-*-gnu* } } */
+/* { dg-require-effective-target default_pie } */
Why restrict to Linux, GNU? default_pie should be enough once other
targets add this.
--- a/gcc/testsuite/gcc.dg/tree-ssa/ssa-store-ccp-3.c 2012-03-14
17:33:37.0 +0100
+++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-store-ccp-3.c 2014-07-29
00:55:17.421086416 +0200
@@ -2,6 +2,9 @@
/* Skipped on MIPS GNU/Linux target because __PIC__ can be
defined for executables as well as shared libraries. */
/* { dg-skip-if { *-*-darwin* hppa*64*-*-* mips*-*-linux* *-*-mingw* } {
* } { } } */
+/* Skipped on default_pie targets because __PIC__ is
+ defined for executables. */
+/* { dg-skip-if { default_pie } { * } { } } */
Emit those default args, they're unnecessary. Also in g++.dg/other/anon5.C.
--- a/gcc/testsuite/g++.dg/other/anon5.C2012-11-10 15:34:42.0
+0100
+++ b/gcc/testsuite/g++.dg/other/anon5.C2013-11-09 14:49:52.281390127
+0100
@@ -1,5 +1,6 @@
// PR c++/34094
// { dg-do link { target { ! { *-*-darwin* *-*-hpux* *-*-solaris2.* } } } }
+// { dg-skip-if { default_pie } { * } { } }
The first arg to dg-skip-if should explain why you're skipping the test.
--- a/gcc/testsuite/lib/target-supports.exp 2013-10-01 11:18:30.0
+0200
+++ b/gcc/testsuite/lib/target-supports.exp 2013-10-25 22:01:46.743388469
+0200
@@ -474,6 +474,11 @@ proc check_profiling_available { test_wh
}
}
+# Profiling don't work with default -fPIE -pie.
Grammar: doesn't work.
+# Return 1 if -pie, -fPIE are default enable, 0 otherwise.
+
+proc check_effective_target_default_pie { } {
Hard to understand, perhaps
# Return 1 if -pie -fPIE are enabled by default, 0 otherwise.
--- a/gcc/doc/invoke.texi 2013-10-03 19:13:50.0 +0200
+++ b/gcc/doc/invoke.texi 2013-11-17 21:30:02.784220111 +0100
@@ -10535,6 +10535,12 @@ For predictable results, you must also s
used for compilation (@option{-fpie}, @option{-fPIE},
or model suboptions) when you specify this linker option.
+NOTE: With configure --enable-default-pie this option is enabled by default
With the @option{--enable-default-pie} configure option, ...
+for C, C++, ObjC, ObjC++, if none of @option{-fno-PIE}, @option{-fno-pie},
+@option{-fPIC}, @option{-fpic}, @option{-fno-PIC}, @option{-fno-pic},
+@option{-nostdlib}, @option{-nostartfiles}, @option{-shared},
+@option{-nodefaultlibs}, nor @option{static} are found.
@option{-static}.
Rainer
--
-
Rainer Orth, Center for Biotechnology, Bielefeld University
Re: [PATCH, Cilk+] Fix for PR61962
On Thu, Jul 31, 2014 at 9:51 PM, Zamyatin, Igor igor.zamya...@intel.com wrote:
Hi!
This patch fixes endless compilation for the case of array notation for an
array which is a structure member
Ok for trunk/4.9 once testing finished?
Thanks,
Igor
Changelog:
gcc/c-family:
2014-07-31 Igor Zamyatin igor.zamya...@intel.com
PR other/61962
* array-notation-common.c (find_rank): Added handling for other
types of references.
gcc/testsuite:
2014-07-31 Igor Zamyatin igor.zamya...@intel.com
PR other/61962
* c-c++-common/cilk-plus/AN/pr61962.c: New test.
diff --git a/gcc/c-family/array-notation-common.c
b/gcc/c-family/array-notation-common.c
index c010039..5db14c6 100644
--- a/gcc/c-family/array-notation-common.c
+++ b/gcc/c-family/array-notation-common.c
@@ -221,7 +221,9 @@ find_rank (location_t loc, tree orig_expr, tree expr,
bool ignore_builtin_fn,
current_rank++;
ii_tree = ARRAY_NOTATION_ARRAY (ii_tree);
}
- else if (TREE_CODE (ii_tree) == ARRAY_REF)
+ else if (TREE_CODE (ii_tree) == ARRAY_REF
+ || TREE_CODE (ii_tree) == INDIRECT_REF
+ || TREE_CODE (ii_tree) == COMPONENT_REF)
Maybe you want handled_component_p (ii_tree)? The above misses
REALPART_EXPR, IMAGPART_EXRP, VIEW_CONVERT_EXPR
and BIT_FIELD_REF.
And I'm missing a
else
gcc_unreachable ();
Richard.
ii_tree = TREE_OPERAND (ii_tree, 0);
else if (TREE_CODE (ii_tree) == PARM_DECL
|| TREE_CODE (ii_tree) == VAR_DECL)
diff --git a/gcc/testsuite/c-c++-common/cilk-plus/AN/pr61962.c
b/gcc/testsuite/c-c++-common/cilk-plus/AN/pr61962.
new file mode 100644
index 000..08d4fe2
--- /dev/null
+++ b/gcc/testsuite/c-c++-common/cilk-plus/AN/pr61962.c
@@ -0,0 +1,14 @@
+/* PR other/61962 */
+/* { dg-do compile } */
+/* { dg-options -fcilkplus } */
+
+struct FloatStruct
+{
+float *f;
+};
+
+/* Either SRC or DST must be a struct, otherwise the bug does not occur. */
+void f (struct FloatStruct* dst, float *src, unsigned int length)
+{
+dst-f[0:length] = src[0:length];
+}
RE: [PATCH, Cilk+] Fix for PR61962
diff --git a/gcc/c-family/array-notation-common.c b/gcc/c-family/array-notation-common.c index c010039..5db14c6 100644 --- a/gcc/c-family/array-notation-common.c +++ b/gcc/c-family/array-notation-common.c @@ -221,7 +221,9 @@ find_rank (location_t loc, tree orig_expr, tree expr, bool ignore_builtin_fn, current_rank++; ii_tree = ARRAY_NOTATION_ARRAY (ii_tree); } - else if (TREE_CODE (ii_tree) == ARRAY_REF) + else if (TREE_CODE (ii_tree) == ARRAY_REF + || TREE_CODE (ii_tree) == INDIRECT_REF + || TREE_CODE (ii_tree) == COMPONENT_REF) Maybe you want handled_component_p (ii_tree)? The above misses REALPART_EXPR, IMAGPART_EXRP, VIEW_CONVERT_EXPR and BIT_FIELD_REF. And I'm missing a else gcc_unreachable (); Thanks, will look at it. Igor Richard. ii_tree = TREE_OPERAND (ii_tree, 0); else if (TREE_CODE (ii_tree) == PARM_DECL || TREE_CODE (ii_tree) == VAR_DECL) diff --git
Re: [AArch64] Implement movmem for the benefit of inline memcpy
On Fri, Aug 1, 2014 at 8:38 AM, Andrew Pinski pins...@gmail.com wrote:
On Fri, Jun 6, 2014 at 1:50 AM, James Greenhalgh
james.greenha...@arm.com wrote:
Hi,
The move_by_pieces infrastructure performs a copy by repeatedly trying
the largest safe copy it can make. So for a 15-byte copy we might see:
offset amount bytes copied
08 0-7
84 8-11
12 2 12-13
14 1 14
However, we can implement a 15-byte copy as so:
offset amount bytes copied
08 0-7
78 7-14
Which can prove more efficient for both space and speed.
In this patch we set MOVE_RATIO low to avoid using move_by_pieces, and
implement the movmem pattern name to expand small block copy cases. Note,
this
optimization does not apply for -mstrict-align targets, which must continue
copying byte-by-byte.
Setting MOVE_RATIO low in this way causes a few tests to begin failing,
both of these are documented in the test-case as expected to fail for
low MOVE_RATIO targets, which do not allow certain Tree-Level optimizations.
I think you should reevaluate setting MOVE_RATIO this low. It is used
for SRA and IPA-SRA which both are very useful; more useful than
memmove optimizations can do.
Maybe we should finally decouple SRA and IPA-SRA from MOVE_RATIO
and have a --param to control the heuristic a target can adjust separately.
(we can still default to MOVE_RATIO here). Ok, maybe we need
two params, one for size and one for speed optimization (though
the size effects are hard to estimate).
Richard.
In fact this optimization is not even valid for volatile variables.
Here is a testcase for the volatile issue:
struct __attribute__((packed)) t15{
long long t8;
int t4;
short t2;
unsigned char t1;
};
volatile struct t15 t15;
int f(struct t15 *a)
{
t15 = *a;
}
Notice how we are writing to byte 7 twice to t15 in the outputted code.
Thanks,
Andrew Pinski
Bootstrapped on aarch64-unknown-linux-gnu with no issues.
OK for trunk?
Thanks,
James
---
gcc/
2014-06-06 James Greenhalgh james.greenha...@arm.com
* config/aarch64/aarch64-protos.h (aarch64_expand_movmem): New.
* config/aarch64/aarch64.c (aarch64_move_pointer): New.
(aarch64_progress_pointer): Likewise.
(aarch64_copy_one_part_and_move_pointers): Likewise.
(aarch64_expand_movmen): Likewise.
* config/aarch64/aarch64.h (MOVE_RATIO): Set low.
* config/aarch64/aarch64.md (movmemmode): New.
gcc/testsuite/
2014-06-06 James Greenhalgh james.greenha...@arm.com
* gcc.dg/tree-ssa/pr42585.c: Skip for AArch64.
* gcc.dg/tree-ssa/sra-12.c: Likewise.
Re: [PATCH] Move Asan instrumentation to sanopt pass
On Fri, Aug 1, 2014 at 9:17 AM, Jakub Jelinek ja...@redhat.com wrote:
On Mon, Jul 28, 2014 at 06:14:08PM +0400, Yury Gribov wrote:
On 07/24/2014 11:48 AM, Jakub Jelinek wrote:
So, either support for just EAF*, or perhaps support for DECL_ATTRIBUTES
for internal-fns, say by having some tree array where you'd store what you
stick into DECL_ATTRIBUTES normally.
I'd prefer to avoid attributes. Would something like this be enough?
(not yet fully tested, just ran asan.exp tests).
I think I'd prefer if you'd specify the fnspec as either NULL (most internal
functions) or normal string and reuse the string parsing in
gimple_call_arg_flags.
Yeah, me too.
Richard.
So, you'd just replace the:
tree attr = gimple_call_fnspec (stmt);
if (!attr || 1 + arg = (unsigned) TREE_STRING_LENGTH (attr))
return 0;
switch (TREE_STRING_POINTER (attr)[1 + arg])
with:
const char *fnspec;
if (gimple_call_arg_flags (stmt))
{
fnspec = ...;
if (!fnspec || 1 + arg = strlen (fnspec))
return 0;
}
else
{
tree attr = gimple_call_fnspec (stmt);
if (!attr || 1 + arg = (unsigned) TREE_STRING_LENGTH (attr))
return 0;
fnspec = TREE_STRING_POINTER (attr);
}
switch (fnspec[1 + arg])
Jakub
Re: [AArch64] Implement movmem for the benefit of inline memcpy
On Jun 6, 2014, at 1:50 AM, James Greenhalgh james.greenha...@arm.com wrote: Hi, The move_by_pieces infrastructure performs a copy by repeatedly trying the largest safe copy it can make. So for a 15-byte copy we might see: offset amount bytes copied 08 0-7 84 8-11 12 2 12-13 14 1 14 However, we can implement a 15-byte copy as so: offset amount bytes copied 08 0-7 78 7-14 Which can prove more efficient for both space and speed. In this patch we set MOVE_RATIO low to avoid using move_by_pieces, and implement the movmem pattern name to expand small block copy cases. Note, this optimization does not apply for -mstrict-align targets, which must continue copying byte-by-byte. Why not change move_by_pieces instead of having a target specific code? This seems like a better option. You can check is unaligned slow target macro to see if you want to do this optimization too. As I mentioned in the other email make sure you check the volatile ness of the from and to before doing this optimization. Thanks, Andrew Setting MOVE_RATIO low in this way causes a few tests to begin failing, both of these are documented in the test-case as expected to fail for low MOVE_RATIO targets, which do not allow certain Tree-Level optimizations. Bootstrapped on aarch64-unknown-linux-gnu with no issues. OK for trunk? Thanks, James --- gcc/ 2014-06-06 James Greenhalgh james.greenha...@arm.com * config/aarch64/aarch64-protos.h (aarch64_expand_movmem): New. * config/aarch64/aarch64.c (aarch64_move_pointer): New. (aarch64_progress_pointer): Likewise. (aarch64_copy_one_part_and_move_pointers): Likewise. (aarch64_expand_movmen): Likewise. * config/aarch64/aarch64.h (MOVE_RATIO): Set low. * config/aarch64/aarch64.md (movmemmode): New. gcc/testsuite/ 2014-06-06 James Greenhalgh james.greenha...@arm.com * gcc.dg/tree-ssa/pr42585.c: Skip for AArch64. * gcc.dg/tree-ssa/sra-12.c: Likewise. 0001-AArch64-Implement-movmem-for-the-benefit-of-inline-m.patch
[Ada] Remove xgnatugn in doc generation
[ Note: the previous related change was rejected due to a too large patch file attached. The mail was: This is another step in removing VMS code, and also a step in converting the documentation to the rest/sphinx format. The xgnatugn preprocessing step will also soon be removed. 2014-08-01 Ben Brosgol bros...@adacore.com * gnat_ugn.texi, projects.texi, xgnatugn.adb: Removed all VMS conditionalization from gnat_ugn.texi and projects.texi, and updated (considerably simplified) xgnatugn.adb, to be removed soon. ] Following up from previous doc clean ups (see above), this change removes the preprocessing step needed to generate gnat_ugn.info. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Arnaud Charlet char...@adacore.com * ug_words, xgnatugn.adb, gcc-interface/Make-lang.in: Remove xgnatugn.adb and ug_words. Index: ug_words === --- ug_words(revision 213263) +++ ug_words(working copy) @@ -1,271 +0,0 @@ -b_ ^ B_ -b~ ^ B__ -cc1 ^ CC1 -Cc1 ^ CC1 -emacs ^ EMACS -Emacs ^ EMACS -gdb ^ GDB -Gdb ^ GDB -gnat1 ^ GNAT1 -Gnat1 ^ GNAT1 -gnatbind^ GNAT BIND -Gnatbind^ GNAT BIND -gnatcheck ^ GNAT CHECK -Gnatcheck ^ GNAT CHECK -gnatchop^ GNAT CHOP -Gnatchop^ GNAT CHOP -gnatclean ^ GNAT CLEAN -Gnatclean ^ GNAT CLEAN -gnatelim^ GNAT ELIM -Gnatelim^ GNAT ELIM -gnatf ^ GNAT XREF -Gnatf ^ GNAT XREF -gnatfind^ GNAT FIND -Gnatfind^ GNAT FIND -gnatkr ^ GNAT KRUNCH -Gnatkr ^ GNAT KRUNCH -gnatlink^ GNAT LINK -Gnatlink^ GNAT LINK -gnatls ^ GNAT LIST -Gnatls ^ GNAT LIST -gnatmake^ GNAT MAKE -Gnatmake^ GNAT MAKE -gnatmetric ^ GNAT METRIC -Gnatmetric ^ GNAT METRIC -gnatname^ GNAT NAME -Gnatname^ GNAT NAME -gnatpp ^ GNAT PRETTY -Gnatpp ^ GNAT PRETTY -gnatprep^ GNAT PREPROCESS -Gnatprep^ GNAT PREPROCESS -gnatstub^ GNAT STUB -Gnatstub^ GNAT STUB -gnatxref^ GNAT XREF -Gnatxref^ GNAT XREF -gcc ^ GNAT COMPILE -gcc -c ^ GNAT COMPILE --fno-inline ^ /INLINE=SUPPRESS --fstack-check ^ /CHECKS=STACK --fno-strict-aliasing ^ /OPTIMIZE=NO_STRICT_ALIASING --gnata ^ /CHECKS=ASSERTIONS --gnatA ^ /NO_GNAT_ADC --gnatb ^ /REPORT_ERRORS=BRIEF --gnatB ^ /ASSUME_VALID --gnatc ^ /NOLOAD --gnatct ^ /NOLOAD /TREE_OUTPUT --gnatdc ^ /TRACE_UNITS --gnatdO ^ /REPORT_ERRORS=IMMEDIATE --gnatC ^ /COMPRESS_NAMES --gnatDG ^ /XDEBUG /EXPAND_SOURCEA --gnatD ^ /XDEBUG --gnateA ^ /ALIASING_CHECK --gnatec ^ /CONFIGURATION_PRAGMAS_FILE --gnated ^ /DISABLE_ATOMIC_SYNCHRONIZATION --gnateD ^ /SYMBOL_PREPROCESSING --gnateE ^ /EXTRA_EXCEPTION_INFORMATION --gnatef ^ /FULL_PATH_IN_BRIEF_MESSAGES --gnateF ^ /FLOAT_OVERFLOW_CHECK --gnateG ^ /GENERATE_PROCESSED_SOURCE --gnatei ^ /MAX_INSTANTIATIONS= --gnateI ^ /MULTI_UNIT_INDEX= --gnatel ^ /ELABORATION_INFO_MESSAGES --gnateL ^ /NOELABORATION_INFO_MESSAGES --gnatem ^ /MAPPING_FILE --gnatep ^ /DATA_PREPROCESSING --gnateP ^ /CATEGORIZATION_WARNINGS --gnateS ^ /SCO_OUTPUT --gnatet ^ /WRITE_TARGET_DEPENDENT_INFO --gnateT ^ /READ_TARGET_DEPENDENT_INFO --gnateu ^ /IGNORE_UNRECOGNIZED --gnateV ^ /PARAMETER_VALIDITY_CHECK --gnateY ^ /IGNORE_STYLE_CHECKS_PRAGMAS --gnatE ^ /CHECKS=ELABORATION --gnatf ^ /REPORT_ERRORS=FULL --gnatF ^ /UPPERCASE_EXTERNALS --gnatg ^ /STYLE_CHECKS=GNAT --gnatG ^ /EXPAND_SOURCE --gnatk ^ /FILE_NAME_MAX_LENGTH --gnatl ^ /LIST --gnatL ^ /LONGJMP_SETJMP --gnatj ^ /JUSTIFY_MESSAGES --gnatj0 ^ /NO_JUSTIFY_MESSAGES --gnatjnn^ /JUSTIFY_MESSAGES=nn --gnatL ^ /INTERSPERSE_SOURCE --gnatm ^ /ERROR_LIMIT --gnatm2 ^ /ERROR_LIMIT=2 --gnatn ^ /INLINE=PRAGMA --gnatn1 ^ /INLINE=PRAGMA_LEVEL_1 --gnatn2 ^ /INLINE=PRAGMA_LEVEL_2 --gnatN ^ /INLINE=FULL --gnato ^ /CHECKS=OVERFLOW --gnato? ^ /OVERFLOW_CHECKS=? --gnato??^ /OVERFLOW_CHECKS=?? --gnatp ^ /CHECKS=SUPPRESS_ALL --gnat-p ^ /CHECKS=UNSUPPRESS_ALL --gnatP ^ /POLLING --gnatR ^ /REPRESENTATION_INFO --gnatR0 ^ /REPRESENTATION_INFO=NONE --gnatR1 ^ /REPRESENTATION_INFO=ARRAYS --gnatR2 ^ /REPRESENTATION_INFO=OBJECTS --gnatR3 ^ /REPRESENTATION_INFO=SYMBOLIC --gnatq ^ /TRY_SEMANTICS --gnatQ ^ /FORCE_ALI --gnatr ^
[Ada] Generate Machine, Model and Rounding FP attributes in line
This change makes it so that the Machine, Model and Rounding FP attributes are generated in line by the compiler (in conjunction with a conversion to an integer type for the third) and optimized on architectures that do not make use of internal extended precision in the FPU. The following function must be optimized into a return on PowerPC and SPARC: function Machine (F: Float) return Float is begin return Float'Machine (F); end; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Eric Botcazou ebotca...@adacore.com * exp_attr.adb (Expand_N_Attribute_Reference): Check whether expansion can be avoid for Machine, Model and Rounding. (Is_Inline_Floating_Point_Attribute): Return true for Machine Model, as well as Rounding if wrapped in a conversion to an integer type. * sem_res.adb (Simplify_Type_Conversion): Deal with Rounding as well. * gcc-interface/gigi.h (fp_arith_may_widen): Declare. * gcc-interface/utils.c (fp_arith_may_widen): New global variable. * gcc-interface/misc.c (enumerate_modes): Compute it. * gcc-interface/trans.c (FP_ARITH_MAY_WIDEN): Delete. (lvalue_required_for_attribute_p): Deal with Descriptor_Size, Machine and Model. (Attribute_to_gnu) Attr_Model: New case. Attr_Machine): Likewise. (convert_with_check): Test fp_arith_may_widen variable. Index: exp_attr.adb === --- exp_attr.adb(revision 213371) +++ exp_attr.adb(working copy) @@ -4068,10 +4068,13 @@ - -- Transforms 'Machine into a call to the floating-point attribute - -- function Machine in Fat_xxx (where xxx is the root type) + -- function Machine in Fat_xxx (where xxx is the root type). + -- Expansion is avoided for cases the back end can handle directly. when Attribute_Machine = - Expand_Fpt_Attribute_R (N); + if not Is_Inline_Floating_Point_Attribute (N) then +Expand_Fpt_Attribute_R (N); + end if; -- -- Machine_Rounding -- @@ -4335,10 +4338,13 @@ --- -- Transforms 'Model into a call to the floating-point attribute - -- function Model in Fat_xxx (where xxx is the root type) + -- function Model in Fat_xxx (where xxx is the root type). + -- Expansion is avoided for cases the back end can handle directly. when Attribute_Model = - Expand_Fpt_Attribute_R (N); + if not Is_Inline_Floating_Point_Attribute (N) then +Expand_Fpt_Attribute_R (N); + end if; - -- Object_Size -- @@ -5411,9 +5417,12 @@ -- Transforms 'Rounding into a call to the floating-point attribute -- function Rounding in Fat_xxx (where xxx is the root type) + -- Expansion is avoided for cases the back end can handle directly. when Attribute_Rounding = - Expand_Fpt_Attribute_R (N); + if not Is_Inline_Floating_Point_Attribute (N) then +Expand_Fpt_Attribute_R (N); + end if; - -- Scaling -- @@ -7946,7 +7955,10 @@ Id : constant Attribute_Id := Get_Attribute_Id (Attribute_Name (N)); begin - if Nkind (Parent (N)) /= N_Type_Conversion + if Id = Attribute_Machine or else Id = Attribute_Model then + return True; + + elsif Nkind (Parent (N)) /= N_Type_Conversion or else not Is_Integer_Type (Etype (Parent (N))) then return False; @@ -7955,7 +7967,7 @@ -- Should also support 'Machine_Rounding and 'Unbiased_Rounding, but -- required back end support has not been implemented yet ??? - return Id = Attribute_Truncation; + return Id = Attribute_Rounding or else Id = Attribute_Truncation; end Is_Inline_Floating_Point_Attribute; end Exp_Attr; Index: sem_res.adb === --- sem_res.adb (revision 213373) +++ sem_res.adb (working copy) @@ -259,7 +259,7 @@ procedure Simplify_Type_Conversion (N : Node_Id); -- Called after N has been resolved and evaluated, but before range checks -- have been applied. Currently simplifies a combination of floating-point - -- to integer conversion and Truncation attribute. + -- to integer conversion and Rounding or Truncation attribute. function Unique_Fixed_Point_Type (N : Node_Id) return Entity_Id; -- A universal_fixed expression in an universal context is unambiguous if @@ -11082,29 +11082,36 @@ Opnd_Typ : constant Entity_Id := Etype (Operand); begin -if Is_Floating_Point_Type (Opnd_Typ) - and then -(Is_Integer_Type (Target_Typ) - or else (Is_Fixed_Point_Type (Target_Typ) - and then Conversion_OK (N))) -
Re: [PATCH] Extended if-conversion for loops marked with pragma omp simd.
On Wed, Jun 25, 2014 at 4:06 PM, Yuri Rumyantsev ysrum...@gmail.com wrote:
Hi All,
We implemented additional support for pragma omp simd in part of
extended if-conversion loops with such pragma. These extensions
include:
1. All extensions are performed only if considered loop or its outer
loop was marked with pragma omp simd (force_vectorize); For ordinary
loops behavior was not changed.
2. Took off cfg restriction on basic block which can have more than 2
predecessors.
3. Put additional restriction on phi nodes which was missed in current design:
all phi nodes must be in non-predicated basic block to conform
semantic of COND_EXPR which is used for transformation.
How is that so? If the PHI is predicated then its result will be used
in a PHI node again and thus we'd create a sequence of COND_EXPRs.
No?
4. Extend predication of phi nodes: phi may have more than 2 arguments
with some limitations:
- for phi nodes which have more than 2 arguments, but only two
arguments are different and one of them has the only occurence,
transformation to single COND_EXPR can be done.
- if phi node has more different arguments and all edge predicates
correspondent to phi-arguments are disjoint, a chain of COND_EXPR
will be generated for it. In current design very simple check is used:
check starting from end that two edges correspondent to neighbor
arguments have common predecessor which is used for further check
with next edge.
These guarantee that phi predication will produce the correct result.
Btw, you can think of these extensions as unfactoring a PHI node by
inserting forwarder blocks. Thus
x = PHI 1(2), 1(3), 2(4)
becomes
bb 5: forwarder-from(2)-and(3)
x = PHI 1(5), 2(4)
and
x = PHI 1(2), 2(3), 3(4)
becomes
bb 5:
x' = PHI 1(2), 2(3)
b = PHIx'(5), 3(4)
which means that 3) has to work. Note that we want this kind of
PHI transforms for out-of-SSA as well to reduce the number of
copies we need to insert on edges.
Thus it would be nice if you implemented 4) in terms of a pre-pass
over the force_vect loops PHI nodes, applying that CFG transform.
And make 3) work properly if it doesn't already.
It looks like you introduce a negate predicate to work around the
critical edge limitation? Please instead change if-conversion to
work with edge predicates (as opposed to BB predicates).
Thanks,
Richard.
Here is example of such extended predication (compile with -march=core-avx2):
#pragma omp simd safelen(8)
for (i=0; i512; i++)
{
float t = a[i];
if (t 0 t 1.0e+17f)
if (c[i] != 0)
res += 1;
}
bb 4:
# res_15 = PHI res_1(5), 0(3)
# i_16 = PHI i_11(5), 0(3)
# ivtmp_17 = PHI ivtmp_14(5), 512(3)
t_5 = a[i_16];
_6 = t_5 0.0;
_7 = t_5 9.998430674944e+16;
_8 = _7 _6;
_ifc__28 = (unsigned int) _8;
_10 = c[i_16];
_ifc__36 = _ifc__28 != 0 ? 4294967295 : 0;
_9 = MASK_LOAD (_10, 0B, _ifc__36);
_ifc__29 = _ifc__28 != 0 ? 1 : 0;
_ifc__30 = (int) _ifc__29;
_ifc__31 = _9 != 0 ? _ifc__30 : 0;
_ifc__32 = _ifc__28 != 0 ? 1 : 0;
_ifc__33 = (int) _ifc__32;
_ifc__34 = _9 == 0 ? _ifc__33 : 0;
_ifc__35 = _ifc__31 != 0 ? 1 : 0;
res_1 = res_15 + _ifc__35;
i_11 = i_16 + 1;
ivtmp_14 = ivtmp_17 - 1;
if (ivtmp_14 != 0)
goto bb 4;
Bootstrap and regression testing did not show any new failures.
gcc/ChageLog
2014-06-25 Yuri Rumyantsev ysrum...@gmail.com
* tree-if-conv.c (flag_force_vectorize): New variable.
(struct bb_predicate_s): Add negate_predicate field.
(bb_negate_predicate): New function.
(set_bb_negate_predicate): New function.
(bb_copy_predicate): New function.
(add_stmt_to_bb_predicate_gimplified_stmts): New function.
(init_bb_predicate): Add initialization of negate_predicate field.
(reset_bb_predicate): Reset negate_predicate to NULL_TREE.
(convert_name_to_cmp): New function.
(get_type_for_cond): New function.
(convert_bool_predicate): New function.
(predicate_disjunction): New function.
(predicate_conjunction): New function.
(add_to_predicate_list): Add convert_bool argument.
Add call of predicate_disjunction if convert_bool argument is true.
(add_to_dst_predicate_list): Add convert_bool argument.
Add early function exit if edge target block is always executed.
Add call of predicate_conjunction if convert_bool argument is true.
Pass convert_bool argument for add_to_predicate_list.
(equal_phi_args): New function.
(phi_has_two_different_args): New function.
(phi_args_disjoint): New function.
(if_convertible_phi_p): Accept phi nodes with more than two args
for loops marked with pragma omp simd. Add check that phi nodes are
in non-predicated basic blocks.
(ifcvt_can_use_mask_load_store): Use flag_force_vectorize.
(all_edges_are_critical): New function.
(if_convertible_bb_p): Allow bb has more than two predecessors if
flag_force_vectorize was setup. Use call of all_edges_are_critical
to reject block
[C++ Patch, committed] Use DECL_TYPE_TEMPLATE_P
Hi,
committed as obvious the below.
Thanks,
Paolo.
///
2014-08-01 Paolo Carlini paolo.carl...@oracle.com
* pt.c (lookup_template_class_1): Use DECL_TYPE_TEMPLATE_P.
Index: pt.c
===
--- pt.c(revision 213428)
+++ pt.c(working copy)
@@ -7414,9 +7414,7 @@ lookup_template_class_1 (tree d1, tree arglist, tr
templ = TYPE_TI_TEMPLATE (d1);
d1 = DECL_NAME (templ);
}
- else if (TREE_CODE (d1) == TEMPLATE_DECL
-DECL_TEMPLATE_RESULT (d1)
- TREE_CODE (DECL_TEMPLATE_RESULT (d1)) == TYPE_DECL)
+ else if (DECL_TYPE_TEMPLATE_P (d1))
{
templ = d1;
d1 = DECL_NAME (templ);
[Ada] Improve messages for lack of inlining in GNATprove mode
Expression functions with a separate declarations should never be inlined in GNATprove mode, hence no message should be issued for these. Correct a problem in call resolution which lead to such a message. Also change the message for lack of inlining in GNATprove mode, so that an info message is issued instead of a warning. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Yannick Moy m...@adacore.com * inline.adb (Cannot_Inline): Issue info message instead of warning for subprograms not inlined in GNATprove mode. * sem_res.adb (Resolve_Call): Take body into account for deciding whether subprogram can be inlined in GNATprove mode or not. Index: inline.adb === --- inline.adb (revision 213437) +++ inline.adb (working copy) @@ -1239,11 +1239,12 @@ and then Msg (Msg'First .. Msg'First + 12) = cannot inline then declare -Len1 : constant Positive := 13; -- cannot inline -Len2 : constant Positive := 25; -- no contextual analysis of +Len1 : constant Positive := 13; -- length of cannot inline +Len2 : constant Positive := 31; +-- lenth of info: no contextual analysis of New_Msg : String (1 .. Msg'Length + Len2 - Len1); begin -New_Msg (1 .. Len2) := no contextual analysis of; +New_Msg (1 .. Len2) := info: no contextual analysis of; New_Msg (Len2 + 1 .. Msg'Length + Len2 - Len1) := Msg (Msg'First + Len1 .. Msg'Last); Cannot_Inline (New_Msg, N, Subp, Is_Serious); Index: sem_res.adb === --- sem_res.adb (revision 213441) +++ sem_res.adb (working copy) @@ -6217,8 +6217,9 @@ -- being inlined. declare -Nam_UA : constant Entity_Id := Ultimate_Alias (Nam); -Decl : constant Node_Id := Unit_Declaration_Node (Nam_UA); +Nam_UA : constant Entity_Id := Ultimate_Alias (Nam); +Decl: constant Node_Id := Unit_Declaration_Node (Nam_UA); +Body_Id : constant Entity_Id := Corresponding_Body (Decl); begin -- If the subprogram is not eligible for inlining in GNATprove @@ -6226,7 +6227,7 @@ if Nkind (Decl) /= N_Subprogram_Declaration or else not Is_Inlined_Always (Nam_UA) - or else not Can_Be_Inlined_In_GNATprove_Mode (Nam_UA, Empty) + or else not Can_Be_Inlined_In_GNATprove_Mode (Nam_UA, Body_Id) then null; @@ -6245,7 +6246,7 @@ -- With the one-pass inlining technique, a call cannot be -- inlined if the corresponding body has not been seen yet. - if No (Corresponding_Body (Decl)) then + if No (Body_Id) then Error_Msg_NE (?no contextual analysis of (body not seen yet), N, Nam);
gcov tidy up
In developing the next patch I found some places to tidy up first. Committed.
nathan
2014-08-01 Nathan Sidwell nat...@acm.org
* Makefile.in (LIBGCOV_MERGE, LIBGCOV_PROFILER,
LIBGCOV_INTERFACE): Reformat.
* libgcov-driver.c (gcov_exit, __gcov_init): Disable when
IN_GCOV_TOOL.
* libgcov-interface.c: Reformat some comments.
(__gcov_flush_mx): Add declaration. Tidy up definition.
Index: libgcc/Makefile.in
===
--- libgcc/Makefile.in (revision 213384)
+++ libgcc/Makefile.in (working copy)
@@ -852,13 +852,15 @@ include $(iterator)
# Build libgcov components.
-LIBGCOV_MERGE = _gcov_merge_add _gcov_merge_single _gcov_merge_delta
_gcov_merge_ior \
- _gcov_merge_time_profile
-LIBGCOV_PROFILER = _gcov_interval_profiler _gcov_pow2_profiler
_gcov_one_value_profiler \
-_gcov_indirect_call_profiler _gcov_average_profiler _gcov_ior_profiler \
-_gcov_indirect_call_profiler_v2 _gcov_time_profiler
-LIBGCOV_INTERFACE = _gcov_flush _gcov_fork _gcov_execl _gcov_execlp
_gcov_execle \
-_gcov_execv _gcov_execvp _gcov_execve _gcov_reset _gcov_dump
+LIBGCOV_MERGE = _gcov_merge_add _gcov_merge_single _gcov_merge_delta \
+ _gcov_merge_ior _gcov_merge_time_profile
+LIBGCOV_PROFILER = _gcov_interval_profiler _gcov_pow2_profiler \
+ _gcov_one_value_profiler _gcov_indirect_call_profiler \
+ _gcov_average_profiler _gcov_ior_profiler \
+ _gcov_indirect_call_profiler_v2 _gcov_time_profiler
+LIBGCOV_INTERFACE = _gcov_flush _gcov_fork _gcov_execl _gcov_execlp\
+ _gcov_execle _gcov_execv _gcov_execvp _gcov_execve _gcov_reset \
+ _gcov_dump
LIBGCOV_DRIVER = _gcov
libgcov-merge-objects = $(patsubst %,%$(objext),$(LIBGCOV_MERGE))
Index: libgcc/libgcov-driver.c
===
--- libgcc/libgcov-driver.c (revision 213384)
+++ libgcc/libgcov-driver.c (working copy)
@@ -775,6 +775,7 @@ gcov_do_dump (struct gcov_info *list, in
free (gf.filename);
}
+#if !IN_GCOV_TOOL
void
gcov_exit (void)
{
@@ -840,6 +841,6 @@ __gcov_init (struct gcov_info *info)
}
info-version = 0;
}
-
+#endif /* !IN_GCOV_TOOL */
#endif /* L_gcov */
#endif /* inhibit_libc */
Index: libgcc/libgcov-interface.c
===
--- libgcc/libgcov-interface.c (revision 213384)
+++ libgcc/libgcov-interface.c (working copy)
@@ -44,20 +44,21 @@ void __gcov_dump (void) {}
extern void gcov_clear (void) ATTRIBUTE_HIDDEN;
extern void gcov_exit (void) ATTRIBUTE_HIDDEN;
+extern __gthread_mutex_t __gcov_flush_mx ATTRIBUTE_HIDDEN;
#ifdef L_gcov_flush
-
#ifdef __GTHREAD_MUTEX_INIT
-ATTRIBUTE_HIDDEN __gthread_mutex_t __gcov_flush_mx = __GTHREAD_MUTEX_INIT;
+__gthread_mutex_t __gcov_flush_mx = __GTHREAD_MUTEX_INIT;
#define init_mx_once()
#else
-__gthread_mutex_t __gcov_flush_mx ATTRIBUTE_HIDDEN;
+__gthread_mutex_t __gcov_flush_mx;
static void
init_mx (void)
{
__GTHREAD_MUTEX_INIT_FUNCTION (__gcov_flush_mx);
}
+
static void
init_mx_once (void)
{
@@ -160,8 +161,8 @@ __gcov_execl (const char *path, char *ar
#endif
#ifdef L_gcov_execlp
-/* A wrapper for the execlp function. Flushes the accumulated profiling data,
so
- that they are not lost. */
+/* A wrapper for the execlp function. Flushes the accumulated
+ profiling data, so that they are not lost. */
int
__gcov_execlp (const char *path, char *arg, ...)
@@ -191,8 +192,8 @@ __gcov_execlp (const char *path, char *a
#endif
#ifdef L_gcov_execle
-/* A wrapper for the execle function. Flushes the accumulated profiling data,
so
- that they are not lost. */
+/* A wrapper for the execle function. Flushes the accumulated
+ profiling data, so that they are not lost. */
int
__gcov_execle (const char *path, char *arg, ...)
@@ -224,8 +225,8 @@ __gcov_execle (const char *path, char *a
#endif
#ifdef L_gcov_execv
-/* A wrapper for the execv function. Flushes the accumulated profiling data,
so
- that they are not lost. */
+/* A wrapper for the execv function. Flushes the accumulated
+ profiling data, so that they are not lost. */
int
__gcov_execv (const char *path, char *const argv[])
@@ -236,8 +237,8 @@ __gcov_execv (const char *path, char *co
#endif
#ifdef L_gcov_execvp
-/* A wrapper for the execvp function. Flushes the accumulated profiling data,
so
- that they are not lost. */
+/* A wrapper for the execvp function. Flushes the accumulated
+ profiling data, so that they are not lost. */
int
__gcov_execvp (const char *path, char *const argv[])
@@ -248,8 +249,8 @@ __gcov_execvp (const char *path, char *c
#endif
#ifdef L_gcov_execve
-/* A wrapper for the execve function. Flushes the accumulated profiling data,
so
- that they are not lost. */
+/* A wrapper for the execve function. Flushes the accumulated
+
[Ada] Legality checks involving aspect Import
This patch adds some missing legality checks on uses of aspect Import. These checks were performed properly in the presence of pragmas, but were incomplete when using aspects. Compiling q.ads must yield: q.ads:2:28: imported entities cannot have explicit initialization (RM 8.1 (24)) q.ads:6:24: no initialization allowed for declaration of J at line 5 q.ads:6:24: imported entities cannot be initialized (RM B.1(24)) q.ads:9:24: no initialization allowed for declaration of K at line 8 q.ads:9:24: imported entities cannot be initialized (RM B.1(24)) --- package Q is R : constant Integer := 42 with Import, Volatile; Q1 : constant := R*R + R*R; J : constant Integer := -999 with Volatile; pragma Import (Ada, J); K : Integer := 1123; pragma Import (Ada, K); end; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Ed Schonberg schonb...@adacore.com * sem_ch13.adb (Analyze_Aspect_Specifications, case Aspect_Import): Set Is_Imported flag at once, to simplify subsequent legality checks. Reject the aspect on an object whose declaration has an explicit initial value. * sem_prag.adb (Process_Import_Or_Interface): Use original node to check legality of an initial value for an imported entity. Set Is_Imported flag in case of error to prevent cascaded errors. Do not set the Is_Imported flag if the pragma comes from an aspect, because it is already done when analyzing the aspect. Index: sem_prag.adb === --- sem_prag.adb(revision 213440) +++ sem_prag.adb(working copy) @@ -7838,8 +7838,14 @@ -- the code generator making an implicit initialization explicit. elsif Present (Expression (Parent (Def_Id))) - and then Comes_From_Source (Expression (Parent (Def_Id))) + and then Comes_From_Source + (Original_Node (Expression (Parent (Def_Id then + + -- Set imported flag to prevent cascaded errors. + + Set_Is_Imported (Def_Id); + Error_Msg_Sloc := Sloc (Def_Id); Error_Pragma_Arg (no initialization allowed for declaration of #, @@ -7847,7 +7853,13 @@ Arg2); else - Set_Imported (Def_Id); + -- If the pragma comes from an aspect specification the + -- Is_Imported flag has already been set. + + if not From_Aspect_Specification (N) then + Set_Imported (Def_Id); + end if; + Process_Interface_Name (Def_Id, Arg3, Arg4); -- Note that we do not set Is_Public here. That's because we @@ -7922,8 +7934,13 @@ exit; else - Set_Imported (Def_Id); + -- If the pragma comes from an aspect specification the + -- Is_Imported flag has already been set. + if not From_Aspect_Specification (N) then + Set_Imported (Def_Id); + end if; + -- Reject an Import applied to an abstract subprogram if Is_Subprogram (Def_Id) Index: sem_ch13.adb === --- sem_ch13.adb(revision 213446) +++ sem_ch13.adb(working copy) @@ -2915,6 +2915,21 @@ -- that verifed that there was a matching convention -- is now obsolete. + if A_Id = Aspect_Import then +Set_Is_Imported (E); + +-- An imported entity cannot have an explicit +-- initialization. + +if Nkind (N) = N_Object_Declaration + and then Present (Expression (N)) +then + Error_Msg_N + (imported entities cannot be initialized + (RM B.1(24)), Expression (N)); +end if; + end if; + goto Continue; end if; @@ -2930,7 +2945,7 @@ and then Nkind (Parent (N)) /= N_Compilation_Unit then Error_Msg_N -(incorrect context for library unit aspect, Id); + (incorrect context for library unit aspect, Id); goto Continue; end if;
[Ada] Add support for non-capturing parenthesis in GNAT.Regpat
It is now possible to use the (?:...) syntax to group elements in a regular expression without making their matched substring available in the Match_Array. The following test must output: Matched (0)= 1.. 6 Matched (1)= 5.. 5 Matched (2)= 0.. 0 with GNAT.Regpat; use GNAT.Regpat; with Ada.Text_IO; use Ada.Text_IO; procedure Main is P : constant Pattern_Matcher := Compile (ab(?:a*)(a+)b); M : Match_Array (0 .. 3); begin Match (P, abaaab, M); if M (0) /= No_Match then Put_Line (Matched (0)= M (0).First'Img .. M (0).Last'Img); Put_Line (Matched (1)= M (1).First'Img .. M (1).Last'Img); Put_Line (Matched (2)= M (2).First'Img .. M (2).Last'Img); end if; end Main; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Emmanuel Briot br...@adacore.com * g-regpat.adb (Parse): Add support for non-capturing parenthesis. Index: s-regpat.adb === --- s-regpat.adb(revision 213263) +++ s-regpat.adb(working copy) @@ -7,7 +7,7 @@ -- B o d y -- -- -- -- Copyright (C) 1986 by University of Toronto. -- --- Copyright (C) 1999-2013, AdaCore-- +-- Copyright (C) 1999-2014, AdaCore-- -- -- -- GNAT is free software; you can redistribute it and/or modify it under -- -- terms of the GNU General Public License as published by the Free Soft- -- @@ -410,10 +410,13 @@ procedure Parse (Parenthesized : Boolean; + Capturing : Boolean; Flags : out Expression_Flags; IP: out Pointer); -- Parse regular expression, i.e. main body or parenthesized thing -- Caller must absorb opening parenthesis. + -- Capturing should be set to True when we have an open parenthesis + -- from which we want the user to extra text. procedure Parse_Branch (Flags : out Expression_Flags; @@ -831,9 +834,10 @@ -- the branches to what follows makes it hard to avoid. procedure Parse - (Parenthesized : Boolean; - Flags : out Expression_Flags; - IP : out Pointer) + (Parenthesized : Boolean; + Capturing : Boolean; + Flags : out Expression_Flags; + IP: out Pointer) is E : String renames Expression; Br, Br2 : Pointer; @@ -847,7 +851,7 @@ -- Make an OPEN node, if parenthesized - if Parenthesized then + if Parenthesized and then Capturing then if Matcher.Paren_Count Max_Paren_Count then Fail (too many ()); end if; @@ -856,7 +860,6 @@ Matcher.Paren_Count := Matcher.Paren_Count + 1; IP := Emit_Node (OPEN); Emit (Character'Val (Par_No)); - else IP := 0; Par_No := 0; @@ -913,14 +916,19 @@ -- Make a closing node, and hook it on the end if Parenthesized then -Ender := Emit_Node (CLOSE); -Emit (Character'Val (Par_No)); +if Capturing then + Ender := Emit_Node (CLOSE); + Emit (Character'Val (Par_No)); + Link_Tail (IP, Ender); +else + -- need to keep looking after the closing parenthesis + null; +end if; else Ender := Emit_Node (EOP); +Link_Tail (IP, Ender); end if; - Link_Tail (IP, Ender); - if Have_Branch and then Emit_Ptr = PM.Size + 1 then -- Hook the tails of the branches to the closing node @@ -945,7 +953,7 @@ elsif Parse_Pos = Parse_End then if E (Parse_Pos) = ')' then - Fail (unmatched ()); + Fail (unmatched ')'); else Fail (junk on end); -- Can't happen end if; @@ -1003,16 +1011,24 @@ New_Flags : Expression_Flags; begin - Parse (True, New_Flags, IP); - - if IP = 0 then - return; + if Parse_Pos = Parse_End - 1 + and then Expression (Parse_Pos) = '?' + and then Expression (Parse_Pos + 1) = ':' + then + Parse_Pos := Parse_Pos + 2; + -- non-capturing parenthesis + Parse (True, False, New_Flags, IP); + else + -- capturing parenthesis +
[RFC][Vectorizer, AArch64] Fix PR/61114 by redefining REDUC_xxx_EXPR tree codes to return scalars
This fixes PR/61114 by redefining the REDUC_{MIN,MAX,PLUS}_EXPR tree codes.
These are presently documented as producing a vector with the result in element
0, and this is inconsistent with their use in tree-vect-loop.c (which on
bigendian targets pulls the bits out of the other end of the vector result).
This leads to bugs on bigendian targets - see
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61114 for a small testcase.
I discounted fixing the vectorizer (to always read from element 0) and then
making the optab reverse the result on bigendian targets (whose architectural
insn produces the result in lane N-1), as optimization of vectors in RTL seems
unlikely to remove such a reverse/permute and so this would lead to a
performance regression (specifically on PowerPC).
Instead it seems more natural for the tree code to produce a scalar result
(producing a vector with the result in lane 0 has already caused confusion, e.g.
https://gcc.gnu.org/ml/gcc-patches/2012-10/msg01100.html).
This patch preserves the meaning of the existing optab (producing a result in
lane 0 on little-endian architectures or N-1 on bigendian), thus generally
avoiding the need to change backends. Hence, expr.c extracts an
endianness-dependent element from the optab result to give the result expected
for the tree code.
Significant complication in the AArch64 backend stems from the existence of
builtins for reduction operations, which are gimple_fold'd to the tree code.
Hence, I introduce new define_expands, and map the existing
__builtin_aarch64_reduc_s{plus,min,max}_mode functions to those, with scalar
result types, matching the result of the tree code to which these are still
gimple_folded.
If the above/proposed solution is acceptable, I'd make a longer patch series,
including some cleanup to tree-vect-loop.c (vect_create_epilog_for_reduction now
has only one case where extract_scalar_result == true), and separating out
AArch64 changes. Further, I'd like to propose creating a new optab that directly
outputs a scalar, as a migration path away from the existing optab whose meaning
is endianness-dependent, i.e. such that expand_unop falls back to the existing
optab only if the new one is not defined.
Patch as it stands has been bootstrapped on x86_64 and regression tested on
aarch64 and aarch64_be without regressions. On x86_64 there is a regression in
gcc.target/i386/pr51235.c, where it seems my check in tree-cfg.c is too strict -
we end up with a reduction from vector (4) unsigned long int to void *.
(Even if I modify tree-vect-loop.c to build the REDUC_..._EXPR as returning the
element type of the input vector, its return type is later changed.) It seems I
can get away with a less-strict check in tree-cfg.c, i.e. allowing the case
where the modes of the expected and actual result types match (rather than
useless_type_conversion_p holding between said types), but if anyone can
suggest an alternative/better check then it'd be great to hear it...
--Alandiff --git a/gcc/config/aarch64/aarch64-builtins.c b/gcc/config/aarch64/aarch64-builtins.c
index fee17ecf637436c8704f565be2eb9ef23891209a..77ed36ecc4cade4c2c6cafd16070198dacb0b869 100644
--- a/gcc/config/aarch64/aarch64-builtins.c
+++ b/gcc/config/aarch64/aarch64-builtins.c
@@ -1396,21 +1396,21 @@ aarch64_gimple_fold_builtin (gimple_stmt_iterator *gsi)
switch (fcode)
{
- BUILTIN_VALL (UNOP, reduc_splus_, 10)
+ BUILTIN_VALL (UNOP, reduc_splus_, 0)
new_stmt = gimple_build_assign_with_ops (
REDUC_PLUS_EXPR,
gimple_call_lhs (stmt),
args[0],
NULL_TREE);
break;
- BUILTIN_VDQIF (UNOP, reduc_smax_, 10)
+ BUILTIN_VDQIF (UNOP, reduc_smax_, 0)
new_stmt = gimple_build_assign_with_ops (
REDUC_MAX_EXPR,
gimple_call_lhs (stmt),
args[0],
NULL_TREE);
break;
- BUILTIN_VDQIF (UNOP, reduc_smin_, 10)
+ BUILTIN_VDQIF (UNOP, reduc_smin_, 0)
new_stmt = gimple_build_assign_with_ops (
REDUC_MIN_EXPR,
gimple_call_lhs (stmt),
diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def b/gcc/config/aarch64/aarch64-simd-builtins.def
index 268432cc117b7027ee9472fc5a4f9b1ea13bea0f..ef363c676bf5bb05cfdaf0f111324d42c3f3d992 100644
--- a/gcc/config/aarch64/aarch64-simd-builtins.def
+++ b/gcc/config/aarch64/aarch64-simd-builtins.def
@@ -251,13 +251,19 @@
BUILTIN_VSDQ_I_DI (BINOP, cmgtu, 0)
BUILTIN_VSDQ_I_DI (BINOP, cmtst, 0)
+ /* Implemented by aarch64_reduc_splus_mode. */
+ BUILTIN_VALL (UNOP, reduc_splus_, 0)
+
/* Implemented by reduc_surplus_mode. */
- BUILTIN_VALL (UNOP, reduc_splus_, 10)
BUILTIN_VDQ (UNOP, reduc_uplus_, 10)
+ /* Implemented by aarch64_reduc_smax_mode. */
+ BUILTIN_VDQIF (UNOP, reduc_smax_, 0)
+
+ /* Implemented by aarch64_reduc_smin_mode. */
+ BUILTIN_VDQIF (UNOP, reduc_smin_, 0)
+
/* Implemented by reduc_maxmin_uns_mode. */
- BUILTIN_VDQIF (UNOP, reduc_smax_, 10)
- BUILTIN_VDQIF (UNOP,
Re: Patch for constexpr variable templates
On 07/31/2014 08:53 AM, Jason Merrill wrote:
In the original mail you mentioned doing it in cp_parser_template_id;
I agree that we probably don't want to do it there because of the
complications you are seeing. I was thinking about doing it when we
normally resolve an id-expression that refers to a variable, once we
know we're looking at an expression and not a declarator-id. But I
think my suggestion of cp_parser_id_expression was wrong, and it would
be better in finish_id_expression.
Jason
Ahh sorry, I misread the function name you gave. finish_id_expression
works just fine and doesn't require any extra work. How does this patch
look?
- Braden Obrzut
2014-08-01 Braden Obrzut ad...@maniacsvault.net
* decl.c (grokvardecl): Handle specializations of variable templates.
(grokdeclarator): Handle variable template id expressions and NULL_TREE
return from grokvardecl.
* decl2.c (check_member_template): Allow declaration of template member
variables.
* parser.c (cp_parser_template_id): Build a TEMPLATE_ID_EXPR for
variable templates.
* pt.c (determine_specialization): Accept variable templates.
(check_template_variable): Fixed wanted template header count and
change non static data member error to variable template warning.
(check_explicit_specialization): Handle and check for malformed
variable template specializations.
(lookup_template_variable): New.
(tsubst_decl): Handle variable template specializations.
(do_decl_instantiation): Handle template variables.
(instantiate_decl): Handle template variables.
* semantics.c (finish_template_variable): New.
(finish_id_expression): Instantiate variable templates.
2014-08-01 Braden Obrzut ad...@maniacsvault.net
* g++.dg/cpp1y/var-templ1.C: New.
* g++.dg/cpp1y/var-templ2.C: New.
* g++.dg/cpp1y/var-templ3.C: New.
* g++.dg/cpp1y/var-templ4.C: New.
* g++.dg/cpp1y/var-templ5.C: New.
* g++.dg/cpp1y/pr59638.C: Marked xfail for template variables as
function pointers taking auto are considered templates at the moment.
* g++.dg/parse/error50.C: Malformed statements look like variable
templ.
* g++.dg/template/crash71.C: Looks like variable templ.
* g++.old-deja/g++.oliva/template10.C: Looks like variable templ.
* g++.old-deja/g++.pt/var1.C: Non-constexpr variable templ.
2013-03-29 Gabriel Dos Reis g...@integrable-solutions.net
* cp-tree.h (variable_template_p): Do not check scope.
* pt.c (check_template_variable): Fix thinko from previous change.
(push_template_decl_real): Fix formatting.
2013-03-29 Gabriel Dos Reis g...@integrable-solutions.net
* cp-tree.h (variable_template_p): New.
* pt.c (check_template_variable): Accept variable temploids at
non-class scope.
(push_template_decl_real): The current instantiation of a template
can be a VAR_DECL.
diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h
index 4a5cb98..c6c 100644
--- a/gcc/cp/cp-tree.h
+++ b/gcc/cp/cp-tree.h
@@ -5027,6 +5027,17 @@ class_of_this_parm (const_tree fntype)
return TREE_TYPE (type_of_this_parm (fntype));
}
+/* True if T designates a variable template declaration. */
+inline bool
+variable_template_p (tree t)
+{
+ if (TREE_CODE (t) != TEMPLATE_DECL)
+return false;
+ if (tree r = DECL_TEMPLATE_RESULT (t))
+return VAR_P (r);
+ return false;
+}
+
/* A parameter list indicating for a function with no parameters,
e.g int f(void). */
extern cp_parameter_declarator *no_parameters;
@@ -5554,6 +5565,7 @@ extern bool redeclare_class_template (tree, tree);
extern tree lookup_template_class (tree, tree, tree, tree,
int, tsubst_flags_t);
extern tree lookup_template_function (tree, tree);
+extern tree lookup_template_variable (tree, tree);
extern int uses_template_parms (tree);
extern int uses_template_parms_level (tree, int);
extern bool in_template_function (void);
@@ -5816,6 +5828,7 @@ extern tree perform_koenig_lookup (tree, vectree, va_gc *,
tsubst_flags_t);
extern tree finish_call_expr (tree, vectree, va_gc **, bool,
bool, tsubst_flags_t);
+extern tree finish_template_variable (tree);
extern tree finish_increment_expr (tree, enum tree_code);
extern tree finish_this_expr (void);
extern tree finish_pseudo_destructor_expr (tree, tree, tree, location_t);
diff --git a/gcc/cp/decl.c b/gcc/cp/decl.c
index dae85c2..2ac0472 100644
--- a/gcc/cp/decl.c
+++ b/gcc/cp/decl.c
@@ -80,8 +80,8 @@ static int ambi_op_p (enum tree_code);
static int unary_op_p (enum tree_code);
static void push_local_name (tree);
static tree grok_reference_init (tree, tree, tree, int);
-static tree grokvardecl (tree, tree, const cp_decl_specifier_seq *,
- int, int, tree);
+static tree grokvardecl (tree, tree, tree, const cp_decl_specifier_seq *,
+ int, int, int, tree);
static int check_static_variable_definition (tree, tree);
static void record_unknown_type (tree, const char *);
[Ada] Spurious warning on generated with_clause for renamed unit
This patch suppresses spurious warnings that may be generated when the prefix of a with_clause for a child unit denotes the renaming of a parent unit. Such renamings generate internal with_clauses that are not marked Implicit_With because they must be properly installed as part of the context, but they are never redundant with explicit with_clauses. The following must compile quietly: gcc -c -gnatwa pkg_b2.adb --- with Renamed_Pkg_A; package Pkg_B2 is procedure Proc_1( Val_1 : in out Renamed_Pkg_A.Type_A ); end Pkg_B2; --- package Pkg_A is type Type_A is ( A1, A2 ); end Pkg_A; --- with Pkg_A; package Renamed_Pkg_A renames Pkg_A; --- package Pkg_A.Child_A is type Type_B is ( B1, B2 ); end Pkg_A.Child_A; --- with Renamed_Pkg_A.Child_A; package body Pkg_B2 is use type Renamed_Pkg_A.Child_A.Type_B; Val_B : Renamed_Pkg_A.Child_A.Type_B := Renamed_Pkg_A.Child_A.Type_B'First; procedure Proc_1( Val_1 : in out Renamed_Pkg_A.Type_A ) is begin if Val_B = Renamed_Pkg_A.Child_A.Type_B'First then Val_1 := Renamed_Pkg_A.Type_A'Last; Val_B := Renamed_Pkg_A.Child_A.Type_B'Last; else Val_B := Renamed_Pkg_A.Child_A.Type_B'First; end if; end Proc_1; end Pkg_B2; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Ed Schonberg schonb...@adacore.com * sem_ch10.adb (Analyze_Compilation_Unit): Do not place a warning on a with_clause created for the renaming of a parent unit in an explicit with_clause. Index: sem_ch10.adb === --- sem_ch10.adb(revision 213440) +++ sem_ch10.adb(working copy) @@ -242,7 +242,7 @@ -- on the context. Note that in contrast with the handling of private -- types, the limited view and the non-limited view of a type are treated -- as separate entities, and no entity exchange needs to take place, which - -- makes the implementation must simpler than could be feared. + -- makes the implementation much simpler than could be feared. -- -- Analyze_Compilation_Unit -- @@ -507,11 +507,15 @@ -- Avoid checking implicitly generated with clauses, limited with -- clauses or withs that have pragma Elaborate or Elaborate_All. +-- With_clauses introduced for renamings of parent clauses are not +-- marked implicit because they need to be properly installed, but +-- they do not come from source and do not require warnings. if Nkind (Clause) = N_With_Clause and then not Implicit_With (Clause) and then not Limited_Present (Clause) and then not Elaborate_Present (Clause) + and then Comes_From_Source (Clause) then -- Package body-to-spec check
[Ada] Conformance checking in instantiations
Within an instance, a subprogram declaration and the corresponding body may be fully conformant, but one may use in its profile a formal type while the other uses a declared subtype of that formal. The routine Same_Generic_Actual is used to recognize this case and prevent a spurious conformance error or the creation of two different subprograms, which leads to errors at link time. This patch makes the check symmetric, so that either the spec or the body may use a declared subtype of the generic formal. The following must execute quietly: gnatmake -q main --- with specific_vector; procedure main is sv : specific_vector.Object := (1.0, 2.0, 3.0); begin sv := specific_vector.-(Right = sv); end main; --- with Vector_Types; generic type Vector_Element is digits ; type Vector is array (Vector_Types.Coordinate) of Vector_Element; Package Generic_Vector is subtype Object is Vector; function - (Right : Object) return Object; end Generic_Vector; --- with Vector_Types; package body Generic_Vector is function - (Right : Vector) return Vector is Negated_Vector : Vector := (vector_types.X = -Right(vector_types.X), vector_types.Y = -Right(vector_types.Y), vector_types.Z = -Right(vector_types.Z) ); begin return Negated_Vector; end -; end Generic_Vector; --- with interfaces; with Vector_Types; with Generic_Vector; package Specific_Vector is new Generic_Vector(Vector_Element = Interfaces.IEEE_Float_64, Vector = Vector_Types.Float_Vector_64); --- with interfaces; package Vector_Types is type Coordinate is (X, Y, Z); type Float_Vector_64 is array (Coordinate) of Interfaces.IEEE_Float_64; end Vector_Types; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Ed Schonberg schonb...@adacore.com * sem_ch6.adb (Same_Generic_Actual): Make function symmetric, because either type may be a subtype of the other. Index: sem_ch6.adb === --- sem_ch6.adb (revision 213434) +++ sem_ch6.adb (working copy) @@ -7274,21 +7274,38 @@ -- Check that the types of corresponding formals have the same -- generic actual if any. We have to account for subtypes of a -- generic formal, declared between a spec and a body, which may - -- appear distinct in an instance but matched in the generic. + -- appear distinct in an instance but matched in the generic, and + -- the subtype may be used either in the spec or the body of the + -- subprogram being checked. - -- Same_Generic_Actual -- - function Same_Generic_Actual (T1, T2 : Entity_Id) return Boolean is + +function Is_Declared_Subtype (S1, S2 : Entity_Id) return Boolean; +-- Predicate to check whether S1 is a subtype of S2 in the source +-- of the instance. + +- +-- Is_Declared_Subtype -- +- + +function Is_Declared_Subtype (S1, S2 : Entity_Id) return Boolean is +begin + return Comes_From_Source (Parent (S1)) + and then Nkind (Parent (S1)) = N_Subtype_Declaration + and then Is_Entity_Name (Subtype_Indication (Parent (S1))) + and then Entity (Subtype_Indication (Parent (S1))) = S2; +end Is_Declared_Subtype; + + -- Start of processing for Same_Generic_Actual + begin return Is_Generic_Actual_Type (T1) = Is_Generic_Actual_Type (T2) - or else -(Present (Parent (T1)) - and then Comes_From_Source (Parent (T1)) - and then Nkind (Parent (T1)) = N_Subtype_Declaration - and then Is_Entity_Name (Subtype_Indication (Parent (T1))) - and then Entity (Subtype_Indication (Parent (T1))) = T2); + or else Is_Declared_Subtype (T1, T2) + or else Is_Declared_Subtype (T2, T1); end Same_Generic_Actual; -- Start of processing for Different_Generic_Profile
Re: [PATCH][testsuite] Don't run cproj-fails-with-broken-glibc.c for broken glibc
Hi Tom, The test-case cproj-fails-with-broken-glibc.c does not work with broken glibcs, as the header comment mentions: ... Check the runtime behavior of the C library's cproj() function and whether it follows the standard. Versions of GLIBC through 2.11.1 had an incorrect implementation which will conflict with GCC's builtin cproj(). GLIBC 2.12+ should be okay. ... This patch skips the test for known broken glibcs. OK for trunk? I'm not at all happy with this patch: unless absolutely necessary, we shouldn't check for version numbers when there's any way to check for working/broken features instead. I'm all for keeping testsuite results clean, but this seems like a total corner case to me. Which distributions do still use affected glibc versions? That test, even if we go the glibc version route, needs to be XFAILed instead of requiring the working version. Apart from that, new effective-target keywords need documenting in doc/sourcebuild.texi. Thanks. Rainer -- - Rainer Orth, Center for Biotechnology, Bielefeld University
[PATCH] convert many pointer_map to hash_map
From: Trevor Saunders tsaund...@mozilla.com
Hi,
This patch replaces a bunch of usage of pointer_map with hash_map. It also
adds an overload to hash_map::traverse that allows modifying the value, and a
remove method.
bootstrapped + regtested on x86_64-unknown-linux-gnu, ok?
Trev
c-family/
* cilk.c: Use hash_map instead of pointer_map.
c/
* c-typeck.c: Use hash_map instead of pointer_map.
cp/
* optimize.c, semantics.c: Use hash_map instead of pointer_map.
gcc/
* hash-map.h (default_hashmap_traits::mark_key_deleted):
Fix cast.
(hash_map::remove): New method.
(hash_map::traverse): New method.
* cgraph.h, except.c, except.h, gimple-ssa-strength-reduction.c,
ipa-utils.c, lto-cgraph.c, lto-streamer.h, omp-low.c, predict.c,
tree-cfg.c, tree-cfgcleanup.c, tree-eh.c, tree-eh.h, tree-inline.c,
tree-inline.h, tree-nested.c, tree-sra.c, tree-ssa-loop-im.c,
tree-ssa-loop-ivopts.c, tree-ssa-reassoc.c, tree-ssa-structalias.c,
tree-ssa.c, tree-ssa.h, var-tracking.c: Use hash_map instead of
pointer_map.
diff --git a/gcc/c-family/cilk.c b/gcc/c-family/cilk.c
index b864bb1..e0d1141 100644
--- a/gcc/c-family/cilk.c
+++ b/gcc/c-family/cilk.c
@@ -64,7 +64,7 @@ struct wrapper_data
/* Containing function. */
tree context;
/* Disposition of all variables in the inner statement. */
- struct pointer_map_t *decl_map;
+ hash_maptree, tree *decl_map;
/* True if this function needs a static chain. */
bool nested;
/* Arguments to be passed to wrapper function, currently a list. */
@@ -335,12 +335,11 @@ create_cilk_helper_decl (struct wrapper_data *wd)
/* A function used by walk tree to find wrapper parms. */
-static bool
-wrapper_parm_cb (const void *key0, void **val0, void *data)
+bool
+wrapper_parm_cb (tree const key0, tree *val0, wrapper_data *wd)
{
- struct wrapper_data *wd = (struct wrapper_data *) data;
- tree arg = * (tree *)key0;
- tree val = (tree)*val0;
+ tree arg = key0;
+ tree val = *val0;
tree parm;
if (val == error_mark_node || val == arg)
@@ -387,7 +386,7 @@ build_wrapper_type (struct wrapper_data *wd)
wd-parms = NULL_TREE;
wd-argtypes = void_list_node;
- pointer_map_traverse (wd-decl_map, wrapper_parm_cb, wd);
+ wd-decl_map-traversewrapper_data *, wrapper_parm_cb (wd);
gcc_assert (wd-type != CILK_BLOCK_FOR);
/* Now build a function.
@@ -452,25 +451,22 @@ copy_decl_for_cilk (tree decl, copy_body_data *id)
/* Copy all local variables. */
-static bool
-for_local_cb (const void *k_v, void **vp, void *p)
+bool
+for_local_cb (tree const k, tree *vp, copy_body_data *id)
{
- tree k = *(tree *) k_v;
- tree v = (tree) *vp;
+ tree v = *vp;
if (v == error_mark_node)
-*vp = copy_decl_no_change (k, (copy_body_data *) p);
+*vp = copy_decl_no_change (k, id);
return true;
}
/* Copy all local declarations from a _Cilk_spawned function's body. */
-static bool
-wrapper_local_cb (const void *k_v, void **vp, void *data)
+bool
+wrapper_local_cb (tree const key, tree *vp, copy_body_data *id)
{
- copy_body_data *id = (copy_body_data *) data;
- tree key = *(tree *) k_v;
- tree val = (tree) *vp;
+ tree val = *vp;
if (val == error_mark_node)
*vp = copy_decl_for_cilk (key, id);
@@ -514,8 +510,11 @@ cilk_outline (tree inner_fn, tree *stmt_p, void *w)
insert_decl_map (id, wd-block, DECL_INITIAL (inner_fn));
/* We don't want the private variables any more. */
- pointer_map_traverse (wd-decl_map, nested ? for_local_cb : wrapper_local_cb,
- id);
+ if (nested)
+wd-decl_map-traversecopy_body_data *, for_local_cb (id);
+ else
+wd-decl_map-traversecopy_body_data *, wrapper_local_cb (id);
+
walk_tree (stmt_p, copy_tree_body_r, (void *) id, NULL);
/* See if this function can throw or calls something that should
@@ -576,7 +575,7 @@ init_wd (struct wrapper_data *wd, enum cilk_block_type type)
wd-type = type;
wd-fntype = NULL_TREE;
wd-context = current_function_decl;
- wd-decl_map = pointer_map_create ();
+ wd-decl_map = new hash_maptree, tree;
/* _Cilk_for bodies are always nested. Others start off as
normal functions. */
wd-nested = (type == CILK_BLOCK_FOR);
@@ -590,7 +589,7 @@ init_wd (struct wrapper_data *wd, enum cilk_block_type type)
static void
free_wd (struct wrapper_data *wd)
{
- pointer_map_destroy (wd-decl_map);
+ delete wd-decl_map;
wd-nested = false;
wd-arglist = NULL_TREE;
wd-argtypes = NULL_TREE;
@@ -618,12 +617,11 @@ free_wd (struct wrapper_data *wd)
(var, ???) -- Pure output argument, handled similarly to above.
*/
-static bool
-declare_one_free_variable (const void *var0, void **map0,
- void *data ATTRIBUTE_UNUSED)
+bool
+declare_one_free_variable (tree const var0, tree *map0, wrapper_data )
{
- const_tree var = (const_tree) var0;
- tree map = (tree)*map0;
+ const_tree var = var0;
+ tree map = *map0;
Re: [RFC][Vectorizer, AArch64] Fix PR/61114 by redefining REDUC_xxx_EXPR tree codes to return scalars
On Fri, Aug 1, 2014 at 12:19 PM, Alan Lawrence alan.lawre...@arm.com wrote:
This fixes PR/61114 by redefining the REDUC_{MIN,MAX,PLUS}_EXPR tree codes.
These are presently documented as producing a vector with the result in
element 0, and this is inconsistent with their use in tree-vect-loop.c
(which on bigendian targets pulls the bits out of the other end of the
vector result). This leads to bugs on bigendian targets - see
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61114 for a small testcase.
I discounted fixing the vectorizer (to always read from element 0) and
then making the optab reverse the result on bigendian targets (whose
architectural insn produces the result in lane N-1), as optimization of
vectors in RTL seems unlikely to remove such a reverse/permute and so this
would lead to a performance regression (specifically on PowerPC).
Instead it seems more natural for the tree code to produce a scalar result
(producing a vector with the result in lane 0 has already caused confusion,
e.g. https://gcc.gnu.org/ml/gcc-patches/2012-10/msg01100.html).
This patch preserves the meaning of the existing optab (producing a result
in lane 0 on little-endian architectures or N-1 on bigendian), thus
generally avoiding the need to change backends. Hence, expr.c extracts an
endianness-dependent element from the optab result to give the result
expected for the tree code.
Significant complication in the AArch64 backend stems from the existence of
builtins for reduction operations, which are gimple_fold'd to the tree code.
Hence, I introduce new define_expands, and map the existing
__builtin_aarch64_reduc_s{plus,min,max}_mode functions to those, with
scalar result types, matching the result of the tree code to which these are
still gimple_folded.
If the above/proposed solution is acceptable, I'd make a longer patch
series, including some cleanup to tree-vect-loop.c
(vect_create_epilog_for_reduction now has only one case where
extract_scalar_result == true), and separating out AArch64 changes. Further,
I'd like to propose creating a new optab that directly outputs a scalar, as
a migration path away from the existing optab whose meaning is
endianness-dependent, i.e. such that expand_unop falls back to the existing
optab only if the new one is not defined.
Patch as it stands has been bootstrapped on x86_64 and regression tested on
aarch64 and aarch64_be without regressions. On x86_64 there is a regression
in gcc.target/i386/pr51235.c, where it seems my check in tree-cfg.c is too
strict - we end up with a reduction from vector (4) unsigned long int to
void *. (Even if I modify tree-vect-loop.c to build the REDUC_..._EXPR as
returning the element type of the input vector, its return type is later
changed.) It seems I can get away with a less-strict check in tree-cfg.c,
i.e. allowing the case where the modes of the expected and actual result
types match (rather than useless_type_conversion_p holding between said
types), but if anyone can suggest an alternative/better check then it'd be
great to hear it...
We should fix the vectorizer code-generation instead.
Makes sense to me - non-aarch64 parts of the patch are ok. The optab
migration strategy is as well.
Thanks,
Richard.
--Alan
Re: [PATCH] convert many pointer_map to hash_map
On Fri, Aug 1, 2014 at 12:34 PM, tsaund...@mozilla.com wrote:
From: Trevor Saunders tsaund...@mozilla.com
Hi,
This patch replaces a bunch of usage of pointer_map with hash_map. It also
adds an overload to hash_map::traverse that allows modifying the value, and a
remove method.
bootstrapped + regtested on x86_64-unknown-linux-gnu, ok?
Ok.
Thanks,
Richard.
Trev
c-family/
* cilk.c: Use hash_map instead of pointer_map.
c/
* c-typeck.c: Use hash_map instead of pointer_map.
cp/
* optimize.c, semantics.c: Use hash_map instead of pointer_map.
gcc/
* hash-map.h (default_hashmap_traits::mark_key_deleted):
Fix cast.
(hash_map::remove): New method.
(hash_map::traverse): New method.
* cgraph.h, except.c, except.h, gimple-ssa-strength-reduction.c,
ipa-utils.c, lto-cgraph.c, lto-streamer.h, omp-low.c, predict.c,
tree-cfg.c, tree-cfgcleanup.c, tree-eh.c, tree-eh.h, tree-inline.c,
tree-inline.h, tree-nested.c, tree-sra.c, tree-ssa-loop-im.c,
tree-ssa-loop-ivopts.c, tree-ssa-reassoc.c, tree-ssa-structalias.c,
tree-ssa.c, tree-ssa.h, var-tracking.c: Use hash_map instead of
pointer_map.
diff --git a/gcc/c-family/cilk.c b/gcc/c-family/cilk.c
index b864bb1..e0d1141 100644
--- a/gcc/c-family/cilk.c
+++ b/gcc/c-family/cilk.c
@@ -64,7 +64,7 @@ struct wrapper_data
/* Containing function. */
tree context;
/* Disposition of all variables in the inner statement. */
- struct pointer_map_t *decl_map;
+ hash_maptree, tree *decl_map;
/* True if this function needs a static chain. */
bool nested;
/* Arguments to be passed to wrapper function, currently a list. */
@@ -335,12 +335,11 @@ create_cilk_helper_decl (struct wrapper_data *wd)
/* A function used by walk tree to find wrapper parms. */
-static bool
-wrapper_parm_cb (const void *key0, void **val0, void *data)
+bool
+wrapper_parm_cb (tree const key0, tree *val0, wrapper_data *wd)
{
- struct wrapper_data *wd = (struct wrapper_data *) data;
- tree arg = * (tree *)key0;
- tree val = (tree)*val0;
+ tree arg = key0;
+ tree val = *val0;
tree parm;
if (val == error_mark_node || val == arg)
@@ -387,7 +386,7 @@ build_wrapper_type (struct wrapper_data *wd)
wd-parms = NULL_TREE;
wd-argtypes = void_list_node;
- pointer_map_traverse (wd-decl_map, wrapper_parm_cb, wd);
+ wd-decl_map-traversewrapper_data *, wrapper_parm_cb (wd);
gcc_assert (wd-type != CILK_BLOCK_FOR);
/* Now build a function.
@@ -452,25 +451,22 @@ copy_decl_for_cilk (tree decl, copy_body_data *id)
/* Copy all local variables. */
-static bool
-for_local_cb (const void *k_v, void **vp, void *p)
+bool
+for_local_cb (tree const k, tree *vp, copy_body_data *id)
{
- tree k = *(tree *) k_v;
- tree v = (tree) *vp;
+ tree v = *vp;
if (v == error_mark_node)
-*vp = copy_decl_no_change (k, (copy_body_data *) p);
+*vp = copy_decl_no_change (k, id);
return true;
}
/* Copy all local declarations from a _Cilk_spawned function's body. */
-static bool
-wrapper_local_cb (const void *k_v, void **vp, void *data)
+bool
+wrapper_local_cb (tree const key, tree *vp, copy_body_data *id)
{
- copy_body_data *id = (copy_body_data *) data;
- tree key = *(tree *) k_v;
- tree val = (tree) *vp;
+ tree val = *vp;
if (val == error_mark_node)
*vp = copy_decl_for_cilk (key, id);
@@ -514,8 +510,11 @@ cilk_outline (tree inner_fn, tree *stmt_p, void *w)
insert_decl_map (id, wd-block, DECL_INITIAL (inner_fn));
/* We don't want the private variables any more. */
- pointer_map_traverse (wd-decl_map, nested ? for_local_cb :
wrapper_local_cb,
- id);
+ if (nested)
+wd-decl_map-traversecopy_body_data *, for_local_cb (id);
+ else
+wd-decl_map-traversecopy_body_data *, wrapper_local_cb (id);
+
walk_tree (stmt_p, copy_tree_body_r, (void *) id, NULL);
/* See if this function can throw or calls something that should
@@ -576,7 +575,7 @@ init_wd (struct wrapper_data *wd, enum cilk_block_type
type)
wd-type = type;
wd-fntype = NULL_TREE;
wd-context = current_function_decl;
- wd-decl_map = pointer_map_create ();
+ wd-decl_map = new hash_maptree, tree;
/* _Cilk_for bodies are always nested. Others start off as
normal functions. */
wd-nested = (type == CILK_BLOCK_FOR);
@@ -590,7 +589,7 @@ init_wd (struct wrapper_data *wd, enum cilk_block_type
type)
static void
free_wd (struct wrapper_data *wd)
{
- pointer_map_destroy (wd-decl_map);
+ delete wd-decl_map;
wd-nested = false;
wd-arglist = NULL_TREE;
wd-argtypes = NULL_TREE;
@@ -618,12 +617,11 @@ free_wd (struct wrapper_data *wd)
(var, ???) -- Pure output argument, handled similarly to above.
*/
-static bool
-declare_one_free_variable (const void *var0, void **map0,
- void *data ATTRIBUTE_UNUSED)
Re: [PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
Hi Jiong,
(Snip)
+ (op0 == virtual_stack_vars_rtx
+ || op0 == frame_pointer_rtx
+ || op0 == arg_pointer_rtx)
(Snip)
The above check is means that these are the ways to access the frame.
is it possible to have stack_pointer_rtx has op0?
On 1 August 2014 14:01, Jiong Wang jiong.w...@arm.com wrote:
currently, aarch64 LEGITIMIZE_ADDRESS_P hook will reject all reg + offset
address given
offset is beyond supported range.
while this may be too strict. we should honor the strict_p parameter in
the hook. before
reload, we accept all offset if it's a frame access, because the offset may
change during
later register elimination.
the early reject of reg + offset may cause extra registers created, and if
that register
live range is across function invoking then callee saved reg needed, thus
introduce extra
reg save/restore also.
give a simple example as:
int
test15 (void)
{
unsigned char a[480];
initialize_array (a, 480);
if (a[0] == 0x10)
return 1;
return 0;
}
.S before the patch
(-O2 -fPIC)
===
test15:
sub sp, sp, #480
mov w1, 480
stp x29, x30, [sp, -32]!
add x29, sp, 0
str x19, [sp, 16]
add x19, x29, 32
mov x0, x19
bl initialize_array
ldrbw0, [x19]
ldr x19, [sp, 16]
ldp x29, x30, [sp], 32
cmp w0, 16
csetw0, eq
add sp, sp, 480
ret
.S after the patch
===
test15:
stp x29, x30, [sp, -496]!
mov w1, 480
add x29, sp, 0
add x0, x29, 16
bl initialize_array
ldrbw0, [x29, 16]
ldp x29, x30, [sp], 496
cmp w0, 16
csetw0, eq
ret
test done
=
no regression on aarch64-none-elf bare metal.
bootstrap OK on aarch64.
OK for trunk?
thanks.
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset
for frame access
when strict_p is false.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
Re: [PATCH 2/2] Enable elimination of zext/sext
On Fri, Aug 1, 2014 at 6:51 AM, Kugan kugan.vivekanandara...@linaro.org wrote:
+ lhs_type = lang_hooks.types.type_for_mode (lhs_mode, lhs_uns);
...
+ ((!lhs_uns !wi::neg_p (min, TYPE_SIGN (lhs_type)))
...
+ type_min = wide_int::from (TYPE_MIN_VALUE (lhs_type), prec,
+TYPE_SIGN (TREE_TYPE (ssa)));
+ type_max = wide_int::from (TYPE_MAX_VALUE (lhs_type), prec,
+TYPE_SIGN (TREE_TYPE (ssa)));
you shouldn't try getting at lhs_type. Btw, do you want to constrain
lhs_mode to MODE_INTs somewhere?
Is this in addition to !INTEGRAL_TYPE_P (TREE_TYPE (ssa)))? Do you mean
that I should check lhs_mode as well?
No, that's probably enough.
For TYPE_SIGN use lhs_uns instead, for the min/max value you
should use wi::min_value () and wi::max_value () instead.
You are still using TYPE_SIGN (TREE_TYPE (ssa)) here and later,
but we computed rhs_uns properly using PROMOTE_MODE.
I think the code with re-setting lhs_uns if rhs_uns != lhs_uns
and later using TYPE_SIGN again is pretty hard to follow.
Btw, it seems you need to conditionalize the call to PROMOTE_MODE
on its availability.
Isn't it simply about choosing a proper range we need to restrict
ssa to? That is, dependent on rhs_uns computed by PROMOTE_MODE,
simply:
+ mode = TYPE_MODE (TREE_TYPE (ssa));
+ rhs_uns = TYPE_UNSIGNED (TREE_TYPE (ssa));
#ifdef PROMOTE_MODE
+ PROMOTE_MODE (mode, rhs_uns, TREE_TYPE (ssa));
#endif
if (rhs_uns)
return wi::ge_p (min, 0); // if min = 0 then range contains positive
values
else
return wi::le_p (max, wi::max_value (TYPE_PRECISION (TREE_TYPE
(ssa)), SIGNED); // if max = signed-max-of-type then range doesn't
need sign-extension
I think we will have to check that ssa has necessary sign/zero extension
when assigned to lhs_type. If PROMOTE_MODE tells us that ssa's type will
be interpreted differently, the value range of ssa also will have
corresponding range. In this cases, shouldn’t we have to check for
upper and lower limit for both min and max?
Hmm? That's exactly what the check is testing... we know that
min = max thus if min = 0 then max = 0.
zero_extension will never do anything on [0, INF]
If max MAX-SIGNED then sign-extension will not do anything. Ok,
sign-extension will do sth for negative values still. So rather
if (rhs_uns)
return wi::geu_p (min, 0);
else
return wi::ges_p (min, 0) wi::les_p (max, wi::max_value
(TYPE_PRECISION (TREE_TYPE (ssa)), SIGNED));
?
I don't like the use of int_fits_type_p you propose.
Richard.
How about this?
bool
promoted_for_type_p (tree ssa, enum machine_mode lhs_mode, bool lhs_uns)
{
wide_int min, max;
tree lhs_type, rhs_type;
bool rhs_uns;
enum machine_mode rhs_mode;
tree min_tree, max_tree;
if (ssa == NULL_TREE
|| TREE_CODE (ssa) != SSA_NAME
|| !INTEGRAL_TYPE_P (TREE_TYPE (ssa)))
return false;
/* Return FALSE if value_range is not recorded for SSA. */
if (get_range_info (ssa, min, max) != VR_RANGE)
return false;
rhs_uns = TYPE_UNSIGNED (TREE_TYPE (ssa));
if (rhs_uns != lhs_uns)
{
/* Signedness of LHS and RHS differs and values also cannot be
represented in LHS range. */
unsigned int prec = min.get_precision ();
if ((lhs_uns wi::neg_p (min, rhs_uns ? UNSIGNED : SIGNED))
|| (!lhs_uns !wi::le_p (max,
wi::max_value (prec, SIGNED),
rhs_uns ? UNSIGNED : SIGNED)))
return false;
}
/* In some architectures, modes are promoted and sign changed with
target defined PROMOTE_MODE macro. If PROMOTE_MODE tells you to
promote _not_ according to ssa's sign then honour that. */
rhs_mode = TYPE_MODE (TREE_TYPE (ssa));
#ifdef PROMOTE_MODE
PROMOTE_MODE (rhs_mode, rhs_uns, TREE_TYPE (ssa));
#endif
rhs_type = lang_hooks.types.type_for_mode (rhs_mode, rhs_uns);
lhs_type = lang_hooks.types.type_for_mode (lhs_mode, lhs_uns);
min_tree = wide_int_to_tree (rhs_type, min);
max_tree = wide_int_to_tree (rhs_type, max);
/* Check if values lies in-between the type range. */
if (int_fits_type_p (min_tree, lhs_type)
int_fits_type_p (max_tree, lhs_type))
return true;
else
return false;
}
Thanks,
Kugan
Re: [RFC, PATCH 3/n] IPA C++ refactoring
On 08/01/2014 02:34 AM, Trevor Saunders wrote:
+++ b/gcc/cgraph.h
@@ -27,6 +27,7 @@ along with GCC; see the file COPYING3. If not see
#include basic-block.h
#include function.h
#include ipa-ref.h
+#include indexed-set.h
you don't actually use that here so why are you including it?
+++ b/gcc/cgraphunit.c
@@ -211,11 +211,12 @@ along with GCC; see the file COPYING3. If not see
#include tree-nested.h
#include gimplify.h
#include dbgcnt.h
+#include indexed-set.h
afaict you don't use it here either.
@@ -529,9 +529,9 @@ cgraph_node::add_new_function (tree fndecl, bool lowered)
}
if (lowered)
node-lowered = true;
-if (!cgraph_new_nodes)
- cgraph_new_nodes = cgraph_node_set_new ();
-cgraph_node_set_add (cgraph_new_nodes, node);
+if (!cgraph_new_nodes.exists ())
+ cgraph_new_nodes.create (4);
I don't think you need this, if the vector doesn't exist it will point
at null, and safe_push knows how to handle that.
+++ b/gcc/indexed-set.h
+templatetypename T
+class indexed_set
per below I don't think you actually need this class after all.
+{
+public:
+ class iterator
+ {
+ public:
+iterator (const vecT *v, unsigned index) : m_index (index), m_vector
(v) {};
+
+inline T operator * ()
+{
+ gcc_assert (m_index m_vector-length ());
+
+ return (*m_vector)[m_index];
+}
+
+inline iterator operator ++ ()
+{
+ m_index++;
+ return *this;
+}
+
+bool operator != (const iterator other) const
+{
+ return m_vector != other.m_vector || m_index != other.m_index;
+}
+
+inline int
+get_index (void) const
+{
+ return m_index;
+}
+
+ private:
+unsigned m_index;
+const vec T *m_vector;
+ };
+
+ /* Iterator to the start of data structure. */
+
+ iterator begin () const
+ {
+gcc_assert (!is_empty ());
it seems to me it would be better to just return the iterator for end ()
if there's no elements.
+
+iterator iter (m_vector, 0);
+return iter;
+ }
+
+ /* Iterator to the end of data structure. */
+
+ iterator end () const
+ {
+return iterator (m_vector, m_vector.length ());
+ }
+
+ /* Add item V to data structure. */
+
+ inline void
afaik defining within the class implies inline.
+ add (T v)
You are passing by value here, which you probably don't want to do if T
is big.
+ {
+m_map.put (v, m_vector.length ());
+m_vector.safe_push (v);
I think this blows up if v was already in the map, so assert it isn't?
+ }
+
+ /* Return iterator for an item V. */
+
+ inline iterator
+ find (T v)
+ {
+int *item = m_map.get (v);
+gcc_unreachable ();
+
+return item ? iterator (m_vector, *item) : end ();
returning an iterator here seems silly, the only information you get is
the index and if the element exists in the map, so just return an int?
+ }
+
+ /* Return if the data structure is empty. */
+
+ inline bool
+ is_empty (void) const
+ {
+return m_vector.is_empty ();
+ }
+
+ /* Return number of items stored in data structure. */
+
+ inline unsigned
+ length (void)
+ {
+return m_vector.length ();
+ }
+
+
+private:
+ hash_mapT, int m_map;
+ auto_vecT m_vector;
I know you don't use this with anything but pointers now, but this will
do weird things if T is some sort of object.
+};
+
+#endif
diff --git a/gcc/tree-emutls.c b/gcc/tree-emutls.c
index 89197c7..9c06aec 100644
--- a/gcc/tree-emutls.c
+++ b/gcc/tree-emutls.c
@@ -42,6 +42,7 @@ along with GCC; see the file COPYING3. If not see
#include target.h
#include targhooks.h
#include tree-iterator.h
+#include indexed-set.h
/* Whenever a target does not support thread-local storage (TLS) natively,
@@ -70,7 +71,7 @@ along with GCC; see the file COPYING3. If not see
/* These two vectors, once fully populated, are kept in lock-step so that
the index of a TLS variable equals the index of its control variable in
the other vector. */
-static varpool_node_set tls_vars;
+static indexed_setvarpool_node * tls_vars;
Note that will create a static initializer that calls malloc, and tls
emulation probably isn't needed on many targets, so I'd guess your
better off using a hash_map * with new and delete as annoying as that is.
static vecvarpool_node * control_vars;
so, looking at this thing more closely it looks like this is just a
weirdo implementation of a hash map from varpool_node * to (varpool_node
*, tree) so I think you could just do
struct tls_var_data
{
varpool_node *control_var;
tree access;
};
hash_mapvar_pool_node *, tls_var_data
Thanks!
Trev
Hello,
I followed Trevor's advice and indexed_set won't be needed.
There's a new patch that utilizes just sent hash_map modification
(hash_map::traverse can modify values).
Regtested on x86_64-pc-linux-gnu with --enable-tls=no.
Thanks,
Martin
Re: [RFC, PATCH 3/n] IPA C++ refactoring
Hi! Just a random comment: On Fri, Aug 01, 2014 at 01:35:47PM +0200, Martin Liška wrote: @@ -357,8 +356,8 @@ cgraph_process_new_functions (void) break; } } - free_cgraph_node_set (cgraph_new_nodes); - cgraph_new_nodes = NULL; + +cgraph_new_nodes.release (); } Please watch out formatting. This used to be indented by 2 spaces, but you changed it to 4, that doesn't look right. Jakub
Re: [RFC, PATCH 3/n] IPA C++ refactoring
Hi,
thank you for your comment. Fixed.
Martin
On 08/01/2014 01:46 PM, Jakub Jelinek wrote:
Hi!
Just a random comment:
On Fri, Aug 01, 2014 at 01:35:47PM +0200, Martin Liška wrote:
@@ -357,8 +356,8 @@ cgraph_process_new_functions (void)
break;
}
}
- free_cgraph_node_set (cgraph_new_nodes);
- cgraph_new_nodes = NULL;
+
+cgraph_new_nodes.release ();
}
Please watch out formatting. This used to be indented by 2 spaces, but
you changed it to 4, that doesn't look right.
Jakub
diff --git a/gcc/cgraph.h b/gcc/cgraph.h
index d8651e2..edf3f7f 100644
--- a/gcc/cgraph.h
+++ b/gcc/cgraph.h
@@ -1200,41 +1200,6 @@ public:
unsigned calls_comdat_local : 1;
};
-/* A cgraph node set is a collection of cgraph nodes. A cgraph node
- can appear in multiple sets. */
-struct cgraph_node_set_def
-{
- struct pointer_map_t *map;
- veccgraph_node * nodes;
-};
-
-typedef cgraph_node_set_def *cgraph_node_set;
-typedef struct varpool_node_set_def *varpool_node_set;
-
-class varpool_node;
-
-/* A varpool node set is a collection of varpool nodes. A varpool node
- can appear in multiple sets. */
-struct varpool_node_set_def
-{
- struct pointer_map_t * map;
- vecvarpool_node * nodes;
-};
-
-/* Iterator structure for cgraph node sets. */
-struct cgraph_node_set_iterator
-{
- cgraph_node_set set;
- unsigned index;
-};
-
-/* Iterator structure for varpool node sets. */
-struct varpool_node_set_iterator
-{
- varpool_node_set set;
- unsigned index;
-};
-
/* Structure containing additional information about an indirect call. */
struct GTY(()) cgraph_indirect_call_info
@@ -1509,7 +1474,7 @@ enum cgraph_state
};
extern enum cgraph_state cgraph_state;
extern bool cgraph_function_flags_ready;
-extern cgraph_node_set cgraph_new_nodes;
+extern veccgraph_node * cgraph_new_nodes;
extern GTY(()) struct asm_node *asm_nodes;
extern GTY(()) int symtab_order;
@@ -1615,24 +1580,7 @@ void record_references_in_initializer (tree, bool);
/* In ipa.c */
bool symtab_remove_unreachable_nodes (bool, FILE *);
-cgraph_node_set cgraph_node_set_new (void);
-cgraph_node_set_iterator cgraph_node_set_find (cgraph_node_set,
- cgraph_node *);
-void cgraph_node_set_add (cgraph_node_set, cgraph_node *);
-void cgraph_node_set_remove (cgraph_node_set, cgraph_node *);
-void dump_cgraph_node_set (FILE *, cgraph_node_set);
-void debug_cgraph_node_set (cgraph_node_set);
-void free_cgraph_node_set (cgraph_node_set);
void cgraph_build_static_cdtor (char which, tree body, int priority);
-
-varpool_node_set varpool_node_set_new (void);
-varpool_node_set_iterator varpool_node_set_find (varpool_node_set,
- varpool_node *);
-void varpool_node_set_add (varpool_node_set, varpool_node *);
-void varpool_node_set_remove (varpool_node_set, varpool_node *);
-void dump_varpool_node_set (FILE *, varpool_node_set);
-void debug_varpool_node_set (varpool_node_set);
-void free_varpool_node_set (varpool_node_set);
void ipa_discover_readonly_nonaddressable_vars (void);
/* In predict.c */
@@ -1948,93 +1896,6 @@ cgraph_next_function_with_gimple_body (cgraph_node *node)
/* Create a new static variable of type TYPE. */
tree add_new_static_var (tree type);
-/* Return true if iterator CSI points to nothing. */
-static inline bool
-csi_end_p (cgraph_node_set_iterator csi)
-{
- return csi.index = csi.set-nodes.length ();
-}
-
-/* Advance iterator CSI. */
-static inline void
-csi_next (cgraph_node_set_iterator *csi)
-{
- csi-index++;
-}
-
-/* Return the node pointed to by CSI. */
-static inline cgraph_node *
-csi_node (cgraph_node_set_iterator csi)
-{
- return csi.set-nodes[csi.index];
-}
-
-/* Return an iterator to the first node in SET. */
-static inline cgraph_node_set_iterator
-csi_start (cgraph_node_set set)
-{
- cgraph_node_set_iterator csi;
-
- csi.set = set;
- csi.index = 0;
- return csi;
-}
-
-/* Return true if SET contains NODE. */
-static inline bool
-cgraph_node_in_set_p (cgraph_node *node, cgraph_node_set set)
-{
- cgraph_node_set_iterator csi;
- csi = cgraph_node_set_find (set, node);
- return !csi_end_p (csi);
-}
-
-/* Return number of nodes in SET. */
-static inline size_t
-cgraph_node_set_size (cgraph_node_set set)
-{
- return set-nodes.length ();
-}
-
-/* Return true if iterator VSI points to nothing. */
-static inline bool
-vsi_end_p (varpool_node_set_iterator vsi)
-{
- return vsi.index = vsi.set-nodes.length ();
-}
-
-/* Advance iterator VSI. */
-static inline void
-vsi_next (varpool_node_set_iterator *vsi)
-{
- vsi-index++;
-}
-
-/* Return the node pointed to by VSI. */
-static inline varpool_node *
-vsi_node (varpool_node_set_iterator vsi)
-{
- return vsi.set-nodes[vsi.index];
-}
-
-/* Return an iterator to the first node in SET. */
-static inline varpool_node_set_iterator
-vsi_start (varpool_node_set set)
-{
- varpool_node_set_iterator vsi;
-
- vsi.set = set;
- vsi.index = 0;
- return vsi;
-}
-
-/* Return number of nodes in SET. */
-static
[C++ PATCH] Improve location info [1/many]
This is the first patch in a series that strives to improve C++ location info. The following is more like a base for another improvements, thus it contains only one small test. It adds a location_t parameter and a vector of location_t of arguments to finish_call_expr - so functions such as build_over_call and build_new_method_call can have this info as well. Next patch will pass these locations down to build_cxx_call/build_call_a/... functions - I hope we'll then get rid of that pesky issue that the caret points to final ) of the function, which is a pain in gdb if the function declarator spans multiple lines. I'd prefer to do smaller (1000 lines) incremental patches, where every patch improves something concrete, rather than one huge megapatch. Maybe you'll want to preapprove similar patches (add location_t to get better diagnostics) in due course - up to you. Bootstrapped/regtested on x86_64-linux, ok for trunk? 2014-08-01 Marek Polacek pola...@redhat.com * call.c (convert_like, convert_like_with_context): Add location_t parameter. (build_user_type_conversion): Adjust convert_like call. (build_integral_nontype_arg_conv): Likewise. (build_new_function_call): Add location_t and veclocation_t parameters. Use them. (build_operator_new_call): Adjust build_over_call call. (build_op_call_1): Adjust build_over_call and convert_like_with_context calls. (build_new_op_1): Likewise. (build_conditional_expr_1): Adjust convert_like call. (convert_like_real): Add location_t parameter. Use it. (build_over_call): Add location_t and veclocation_t parameters. Use them. (build_special_member_call): Adjust build_new_method_call call. (build_new_method_call_1): Add location_t and veclocation_t parameters. Use them. (build_new_method_call): Likewise. (perform_implicit_conversion_flags): Adjust convert_like call. (perform_direct_initialization_if_possible): Adjust convert_like_real call. (initialize_reference): Adjust convert_like calls. * cp-array-notation.c (expand_sec_reduce_builtin): Adjust finish_call_expr call. * cp-tree.h: Adjust declarations. * init.c (build_new_1): Adjust build_new_method_call call. (build_dtor_call): Likewise. * method.c (locate_fn_flags): Likewise. * parser.c (cp_parser_userdef_char_literal): Adjust finish_call_expr call. (cp_parser_userdef_numeric_literal): Likewise. (cp_parser_userdef_string_literal): Likewise. (cp_parser_postfix_expression): New vector of locations. Adjust cp_parser_parenthesized_expression_list call. Make it fill the vector of locations of arguments. Adjust build_new_method_call and finish_call_expr calls. Release the vector of locations. (cp_parser_parenthesized_expression_list): Fill the vector of locations of arguments. (cp_parser_new_placement): Adjust cp_parser_parenthesized_expression_list call. (cp_parser_new_initializer): Likewise. (cp_parser_mem_initializer): Likewise. (cp_parser_initializer): Likewise. (cp_parser_gnu_attribute_list): Likewise. (cp_parser_std_attribute): Likewise. (cp_parser_functional_cast): Likewise. (cp_parser_omp_declare_reduction_exprs): Likewise. (cp_parser_perform_range_for_lookup): Adjust finish_call_expr calls. (cp_parser_range_for_member_function): Likewise. * pt.c (tsubst_copy_and_build): Adjust build_new_method_call and finish_call_expr calls. * semantics.c (finish_call_expr): Add location_t and veclocation_t parameters. Use them. (finish_omp_barrier): Update finish_call_expr call. (finish_omp_flush): Likewise. (finish_omp_taskwait): Likewise. (finish_omp_taskyield): Likewise. (finish_omp_cancel): Likewise. (finish_omp_cancellation_point): Likewise. * g++.dg/diagnostic/location-1.C: New test. diff --git gcc/cp/call.c gcc/cp/call.c index 4d37c65..b9db4b9 100644 --- gcc/cp/call.c +++ gcc/cp/call.c @@ -148,18 +148,19 @@ static int equal_functions (tree, tree); static int joust (struct z_candidate *, struct z_candidate *, bool, tsubst_flags_t); static int compare_ics (conversion *, conversion *); -static tree build_over_call (struct z_candidate *, int, tsubst_flags_t); +static tree build_over_call (location_t, veclocation_t, +struct z_candidate *, int, tsubst_flags_t); static tree build_java_interface_fn_ref (tree, tree); -#define convert_like(CONV, EXPR, COMPLAIN) \ - convert_like_real ((CONV), (EXPR), NULL_TREE, 0, 0, \ -/*issue_conversion_warnings=*/true,\ -/*c_cast_p=*/false, (COMPLAIN)) -#define convert_like_with_context(CONV,
Re: [PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
Hi Venkataramanan,
thanks for review.
On 01/08/14 11:59, Venkataramanan Kumar wrote:
Hi Jiong,
(Snip)
+ (op0 == virtual_stack_vars_rtx
+ || op0 == frame_pointer_rtx
+ || op0 == arg_pointer_rtx)
(Snip)
The above check is means that these are the ways to access the frame.
is it possible to have stack_pointer_rtx has op0?
I think there is no necessary to add stack_pointer_rtx check here.
because normally all local variables will be addressed using
virtual_stack_vars_rtx,
and then rebased to frame_pointer_rtx in instantiate_new_reg, and finally
eliminated
to fp/sp during reload depending on whether frame pointer is needed. so
+ (op0 == virtual_stack_vars_rtx
+ || op0 == frame_pointer_rtx
could actually catch almost all we want.
-- Jiong
On 1 August 2014 14:01, Jiong Wang jiong.w...@arm.com wrote:
currently, aarch64 LEGITIMIZE_ADDRESS_P hook will reject all reg + offset
address given
offset is beyond supported range.
while this may be too strict. we should honor the strict_p parameter in
the hook. before
reload, we accept all offset if it's a frame access, because the offset may
change during
later register elimination.
the early reject of reg + offset may cause extra registers created, and if
that register
live range is across function invoking then callee saved reg needed, thus
introduce extra
reg save/restore also.
give a simple example as:
int
test15 (void)
{
unsigned char a[480];
initialize_array (a, 480);
if (a[0] == 0x10)
return 1;
return 0;
}
.S before the patch
(-O2 -fPIC)
===
test15:
sub sp, sp, #480
mov w1, 480
stp x29, x30, [sp, -32]!
add x29, sp, 0
str x19, [sp, 16]
add x19, x29, 32
mov x0, x19
bl initialize_array
ldrbw0, [x19]
ldr x19, [sp, 16]
ldp x29, x30, [sp], 32
cmp w0, 16
csetw0, eq
add sp, sp, 480
ret
.S after the patch
===
test15:
stp x29, x30, [sp, -496]!
mov w1, 480
add x29, sp, 0
add x0, x29, 16
bl initialize_array
ldrbw0, [x29, 16]
ldp x29, x30, [sp], 496
cmp w0, 16
csetw0, eq
ret
test done
=
no regression on aarch64-none-elf bare metal.
bootstrap OK on aarch64.
OK for trunk?
thanks.
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset
for frame access
when strict_p is false.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
Re: [PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
On 1 August 2014 09:31, Jiong Wang jiong.w...@arm.com wrote:
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset
for frame access
when strict_p is false.
This part is OK.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
This test case relies on:
+/* { dg-final { scan-rtl-dump mem/j/c:QI \\(plus:DI expand } } */
... which is not very specific. Can we tighten this up to look for
something directly related to the relevant frame access?
/Marcus
[PATCH] Fix PR61762 testcase
After trying to fix the missed optimization on strict-align targets
in a fancy way for some time I gave up and this fixes the testcase
instead.
Tested on a sparc cross and on x86_64-unknown-linux-gnu, applied.
Richard.
2014-08-01 Richard Biener rguent...@suse.de
PR middle-end/61762
* gcc.dg/pr61762.c: Align the string to make the testcase work
on strict-align targets.
Index: gcc/testsuite/gcc.dg/pr61762.c
===
--- gcc/testsuite/gcc.dg/pr61762.c (revision 213342)
+++ gcc/testsuite/gcc.dg/pr61762.c (working copy)
@@ -3,7 +3,7 @@
unsigned int f()
{
- static const char string[] = Private;
+ static const char string[] __attribute__((aligned(sizeof(int = Private;
unsigned int priv;
__builtin_memcpy(priv, string[0], sizeof(priv));
[PATCH] Move memory builtin foldings to GIMPLE
The following patch removes folding of bzero, memset, bcopy,
memcpy, mempcpy and memmove from builtins.c and re-implements
them in gimple-fold.c. This means those foldings are no longer
accessible from GENERIC folding.
Bootstrapped on x86_64-unknown-linux-gnu - testing seems to reveal
some testsuite fallout I still have to look into.
Does this look sensible? In particular, does any Frontend maintainer
object to no longer getting these simplifications? Note they
will still apply during gimplification and given they now work
on GENERIC we can simplify them quite a bit (but as followup).
Thanks,
Richard.
2014-08-01 Richard Biener rguent...@suse.de
* builtins.c (fold_builtin_memory_op, var_decl_component_p,
fold_builtin_memset): Move to gimple-fold.c.
(fold_builtin_2): Remove handling of bzero, memset, bcopy,
memcpy, mempcpy and memmove.
* gimple-fold.h (gimple_fold_builtin): Adjust signature.
* gimple-fold.c: Include output.h.
(readonly_data_expr): New function copied from builtins.c.
(var_decl_component_p): Moved from builtins.c.
(gimple_fold_builtin_memory_op): Moved from builtins.c
fold_builtin_memory_op and rewritten to GIMPLE.
(gimple_fold_builtin_memset): Likewise.
(gimple_fold_builtin_1): Split out from gimple_fold_builtin.
(gimple_fold_builtin): Change signature and handle
bzero, memset, bcopy, memcpy, mempcpy and memmove folding
here.
(gimple_fold_call): Adjust.
* tree-ssa-ccp.c (pass_fold_builtins::execute): Adjust
for gimple_fold_builtin change.
Index: gcc/builtins.c
===
*** gcc/builtins.c.orig 2014-07-31 15:10:04.127659292 +0200
--- gcc/builtins.c 2014-08-01 13:59:34.686001948 +0200
*** static tree fold_builtin_ceil (location_
*** 164,170
static tree fold_builtin_round (location_t, tree, tree);
static tree fold_builtin_int_roundingfn (location_t, tree, tree);
static tree fold_builtin_bitop (tree, tree);
- static tree fold_builtin_memory_op (location_t, tree, tree, tree, tree, bool,
int);
static tree fold_builtin_strchr (location_t, tree, tree, tree);
static tree fold_builtin_memchr (location_t, tree, tree, tree, tree);
static tree fold_builtin_memcmp (location_t, tree, tree, tree);
--- 164,169
*** fold_builtin_exponent (location_t loc, t
*** 8496,8961
return NULL_TREE;
}
- /* Return true if VAR is a VAR_DECL or a component thereof. */
-
- static bool
- var_decl_component_p (tree var)
- {
- tree inner = var;
- while (handled_component_p (inner))
- inner = TREE_OPERAND (inner, 0);
- return SSA_VAR_P (inner);
- }
-
- /* Fold function call to builtin memset. Return
-NULL_TREE if no simplification can be made. */
-
- static tree
- fold_builtin_memset (location_t loc, tree dest, tree c, tree len,
-tree type, bool ignore)
- {
- tree var, ret, etype;
- unsigned HOST_WIDE_INT length, cval;
-
- if (! validate_arg (dest, POINTER_TYPE)
- || ! validate_arg (c, INTEGER_TYPE)
- || ! validate_arg (len, INTEGER_TYPE))
- return NULL_TREE;
-
- if (! tree_fits_uhwi_p (len))
- return NULL_TREE;
-
- /* If the LEN parameter is zero, return DEST. */
- if (integer_zerop (len))
- return omit_one_operand_loc (loc, type, dest, c);
-
- if (TREE_CODE (c) != INTEGER_CST || TREE_SIDE_EFFECTS (dest))
- return NULL_TREE;
-
- var = dest;
- STRIP_NOPS (var);
- if (TREE_CODE (var) != ADDR_EXPR)
- return NULL_TREE;
-
- var = TREE_OPERAND (var, 0);
- if (TREE_THIS_VOLATILE (var))
- return NULL_TREE;
-
- etype = TREE_TYPE (var);
- if (TREE_CODE (etype) == ARRAY_TYPE)
- etype = TREE_TYPE (etype);
-
- if (!INTEGRAL_TYPE_P (etype)
-!POINTER_TYPE_P (etype))
- return NULL_TREE;
-
- if (! var_decl_component_p (var))
- return NULL_TREE;
-
- length = tree_to_uhwi (len);
- if (GET_MODE_SIZE (TYPE_MODE (etype)) != length
- || get_pointer_alignment (dest) / BITS_PER_UNIT length)
- return NULL_TREE;
-
- if (length HOST_BITS_PER_WIDE_INT / BITS_PER_UNIT)
- return NULL_TREE;
-
- if (integer_zerop (c))
- cval = 0;
- else
- {
- if (CHAR_BIT != 8 || BITS_PER_UNIT != 8 || HOST_BITS_PER_WIDE_INT 64)
- return NULL_TREE;
-
- cval = TREE_INT_CST_LOW (c);
- cval = 0xff;
- cval |= cval 8;
- cval |= cval 16;
- cval |= (cval 31) 1;
- }
-
- ret = build_int_cst_type (etype, cval);
- var = build_fold_indirect_ref_loc (loc,
-fold_convert_loc (loc,
- build_pointer_type (etype),
- dest));
- ret = build2 (MODIFY_EXPR, etype, var, ret);
- if (ignore)
- return ret;
-
- return omit_one_operand_loc (loc, type, dest, ret);
- }
Re: [PATCH, libfortran] Backport xmallocarray to 4.8/4.9 (CVE-2014-5044)
On Thu, Jul 31, 2014 at 11:32:12PM +0300, Janne Blomqvist wrote:
a while ago I committed a patch to trunk adding a function
xmallocarray to libgfortran, which is a malloc wrapper like xmalloc
but has two arguments and does an overflow check before multiplying
them together.
That seems to be unnecessarily expensive for the common case where both
nmemb and size are small.
calloc in glibc uses something like following to avoid the division most of
the time, if both nmemb and size are small enough then nmemb * size can't
overflow. At least for 64-bit architectures small is smaller than 4GB
for both numbers.
2014-08-01 Jakub Jelinek ja...@redhat.com
* runtime/memory.c (xmallocarray): Avoid division for the common case.
--- libgfortran/runtime/memory.c.jj 2014-06-18 08:50:33.0 +0200
+++ libgfortran/runtime/memory.c2014-08-01 14:41:08.385856116 +0200
@@ -56,7 +56,9 @@ xmallocarray (size_t nmemb, size_t size)
if (!nmemb || !size)
size = nmemb = 1;
- else if (nmemb SIZE_MAX / size)
+#define HALF_SIZE_T (((size_t) 1) (__CHAR_BIT__ * sizeof (size_t) / 2))
+ else if (__builtin_expect ((nmemb | size) = HALF_SIZE_T, 0)
+ nmemb SIZE_MAX / size)
{
errno = ENOMEM;
os_error (Integer overflow in xmallocarray);
Jakub
Re: [PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
On 01/08/14 13:20, Marcus Shawcroft wrote:
On 1 August 2014 09:31, Jiong Wang jiong.w...@arm.com wrote:
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset
for frame access
when strict_p is false.
This part is OK.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
This test case relies on:
+/* { dg-final { scan-rtl-dump mem/j/c:QI \\(plus:DI expand } } */
... which is not very specific. Can we tighten this up to look for
something directly related to the relevant frame access?
Hmm... thanks for pointing this out.
updated the patch to match three keyword. mem, plus and virtual-stack-vars
ok for trunk?
thanks.
-- Jiong
gcc/
* config/aarch64/aarch64.c (aarch64_classify_address): Accept all offset for
frame access
when strict_p is false.
gcc/testsuite
* gcc.target/aarch64/legitimize_stack_var_before_reload_1.c: New testcase.
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index ed80269..c8e5808 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -3248,6 +3248,21 @@ aarch64_classify_address (struct aarch64_address_info *info,
case PLUS:
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
+
+ if (! strict_p
+ GET_CODE (op0) == REG
+ (op0 == virtual_stack_vars_rtx
+ || op0 == frame_pointer_rtx
+ || op0 == arg_pointer_rtx)
+ GET_CODE (op1) == CONST_INT)
+ {
+ info-type = ADDRESS_REG_IMM;
+ info-base = op0;
+ info-offset = op1;
+
+ return true;
+ }
+
if (GET_MODE_SIZE (mode) != 0
CONST_INT_P (op1)
aarch64_base_register_rtx_p (op0, strict_p))
diff --git a/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c b/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c
new file mode 100644
index 000..e5f0bd1
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/legitimize_stack_var_before_reload_1.c
@@ -0,0 +1,21 @@
+/* { dg-do compile } */
+/* { dg-options -O2 -fdump-rtl-expand } */
+
+extern void initialize_array (unsigned char *, int);
+
+int
+test15 (void)
+{
+ unsigned char a[480];
+
+ initialize_array (a, 480);
+
+ if (a[0] == 0x10)
+return 1;
+
+ return 0;
+}
+
+/* { dg-final { scan-rtl-dump \\(mem\[^\\n\]*\\(plus\[^\\n\]*virtual-stack-vars expand } } */
+
+/* { dg-final { cleanup-rtl-dump expand } } */
[Ada] Don't unconditionally define _LARGEFILE_SOURCE
This patch 2014-08-01 Pascal Obry o...@adacore.com * cstreams.c: Only enable large file support on know supported platforms. Add missing defines/includes. broke Ada bootstrap on Solaris: /vol/gcc/src/hg/trunk/local/gcc/ada/cstreams.c:34:0: error: _LARGEFILE_SOURCE redefined [-Werror] #define _LARGEFILE_SOURCE ^ built-in: note: this is the location of the previous definition Fixed as follows, i386-pc-solaris2.10 bootstrap is past the failure now. Ok for mainline? Rainer 2014-08-01 Rainer Orth r...@cebitec.uni-bielefeld.de gcc/ada: * cstreams.c (_LARGEFILE_SOURCE): Guard definition. diff --git a/gcc/ada/cstreams.c b/gcc/ada/cstreams.c --- a/gcc/ada/cstreams.c +++ b/gcc/ada/cstreams.c @@ -31,7 +31,9 @@ /* Routines required for implementing routines in Interfaces.C.Streams. */ +#ifndef _LARGEFILE_SOURCE #define _LARGEFILE_SOURCE +#endif #define _FILE_OFFSET_BITS 64 /* the define above will make off_t a 64bit type on GNU/Linux */ -- - Rainer Orth, Center for Biotechnology, Bielefeld University
Re: [Ada] Don't unconditionally define _LARGEFILE_SOURCE
2014-08-01 Pascal Obry o...@adacore.com * cstreams.c: Only enable large file support on know supported platforms. Add missing defines/includes. broke Ada bootstrap on Solaris: /vol/gcc/src/hg/trunk/local/gcc/ada/cstreams.c:34:0: error: _LARGEFILE_SOURCE redefined [-Werror] #define _LARGEFILE_SOURCE ^ built-in: note: this is the location of the previous definition Fixed as follows, i386-pc-solaris2.10 bootstrap is past the failure now. Ok for mainline? Yes, thanks.
Re: [PATCH][AArch64]Fix offset glitch in load reg pair pattern
On 29 July 2014 16:25, Jiong Wang jiong.w...@arm.com wrote: gcc/ * config/aarch64/aarch64.md (loadwb_pairGPI:mode_P:mode): Fix offset. (loadwb_pairGPI:mode_P:mode): Likewise. * config/aarch64/aarch64.c (aarch64_gen_loadwb_pair): Likewise. OK /Marcus
Re: [PATCH][AArch64] Add constrain to address offset in storewb_pair/loadwb_pair insns
On 29 July 2014 16:31, Jiong Wang jiong.w...@arm.com wrote: gcc/ * config/aarch64/aarch64.c (offset_7bit_signed_scaled_p): Rename to aarch64_offset_7bit_signed_scaled_p, remove static and use it . * config/aarch64/aarch64-protos.h (aarch64_offset_7bit_signed_scaled_p): Declaration. * config/aarch64/predicates.md (aarch64_mem_pair_offset): Define new predicate. * config/aarch64/aarch64.md (loadwb_pair, storewb_pair): Use aarch64_mem_pair_offset. OK /Marcus
Re: [PATCH][AArch64] Improve TARGET_LEGITIMIZE_ADDRESS_P hook
On 1 August 2014 13:49, Jiong Wang jiong.w...@arm.com wrote: Hmm... thanks for pointing this out. updated the patch to match three keyword. mem, plus and virtual-stack-vars ok for trunk? OK /Marcus
[Ada] Give missing entity messages in -gnatc mode
This patch forces expansion on (but not actual code generation) for configurable run time mode or no run time mode in -gnatc mode. This means that we get error messages about unsupported features even in -gnatc mode, which was not true before. The following is compiled in -gnatc mode with a gnat.adc file containing pragma No_Run_Time: 1. function CRT_GNATC (X : Integer) return Integer is 2. begin 3.return X ** X; | construct not allowed in no run time mode entity System.Exp_Int.Exp_Integer not defined 4. end; prior to this patch, no message was given in -gnatc mode. A debug flag -gnatd.Z can be used to disable this behavior in case we run into any trouble from doing the expansion in the future. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Robert Dewar de...@adacore.com * debug.adb: Document debug switch -gnatd.Z. * sem.adb (Semantics): Force expansion on in no or configurable run time mode. Index: debug.adb === --- debug.adb (revision 213453) +++ debug.adb (working copy) @@ -143,7 +143,7 @@ -- d.W Print out debugging information for Walk_Library_Items -- d.X Old treatment of indexing aspects -- d.Y - -- d.Z + -- d.Z Do not enable expansion in configurable run-time mode -- d1 Error msgs have node numbers where possible -- d2 Eliminate error flags in verbose form error messages @@ -686,6 +686,12 @@ -- is preserved temporarily for use by the modelling project under -- debug flag d.X. + -- d.Z Normally we always enable expansion in configurable run-time mode + -- to make sure we get error messages about unsupported features even + -- when compiling in -gnatc mode. But expansion is turned off in this + -- case if debug flag -gnatd.Z is used. This is to deal with the case + -- where we discover difficulties in this new processing. + -- d1 Error messages have node numbers where possible. Normally error -- messages have only source locations. This option is useful when -- debugging errors caused by expanded code, where the source location Index: sem.adb === --- sem.adb (revision 213437) +++ sem.adb (working copy) @@ -1410,11 +1410,33 @@ GNAT_Mode := True; end if; + -- For generic main, never do expansion + if Generic_Main then Expander_Mode_Save_And_Set (False); + + -- Non generic case + else Expander_Mode_Save_And_Set - (Operating_Mode = Generate_Code or Debug_Flag_X); + + -- Turn on expansion if generating code + + (Operating_Mode = Generate_Code + + -- or if special debug flag -gnatdx is set + + or else Debug_Flag_X + + -- Or if in configuration run-time mode. We do this so we get + -- error messages about missing entities in the run-time even + -- if we are compiling in -gnatc (no code generation) mode. + -- Similar processing applies to No_Run_Time_Mode. However, + -- don't do this if debug flag -gnatd.Z is set (this is to handle + -- a situation where this new processing causes trouble). + + or else ((Configurable_Run_Time_Mode or No_Run_Time_Mode) + and not Debug_Flag_Dot_ZZ)); end if; Full_Analysis := True;
[Ada] Improved handling of info messages
The count of info messages is now separated from the count of warnings and -gnatwe never turns info messages into errors. The following compiles as shown, and generates object code since no error is found when compilation options -gnatwe -gnatld7 -gnatj55 are used. Compiling: infoerr.adb 1. with Text_IO; use Text_IO; 2. procedure InfoErr is 3.J : Natural; 4. begin 5.J := 1; 6.ENTER | info: code between label and backwards goto rewritten as loop 7.Put_Line (J'Img); 8.J := J + 1; 9.if J 3 then 10. goto DONE; 11.end if; 12.goto ENTER; 13.DONE 14.null; 15. end; 15 lines: No errors, 1 info message Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Robert Dewar de...@adacore.com * atree.ads (Info_Messages): New counter. * err_vars.ads: Minor comment update. * errout.adb (Delete_Warning_And_Continuations): Deal with new Info_Messages counter. (Error_Msg_Internal): ditto. (Delete_Warning): ditto. (Initialize): ditto. (Write_Error_Summary): ditto. (Output_Messages): ditto. (To_Be_Removed): ditto. * erroutc.adb (Delete_Msg): Deal with Info_Messages counter. (Compilation_Errors): ditto. * errutil.adb (Error_Msg): Deal with Info_Messages counter. (Finalize): ditto. (Initialize): ditto. * sem_prag.adb (Analyze_Pragma): Minor comment addition. * gnat_ugn.texi: Document that -gnatwe does not affect info messages. Index: atree.ads === --- atree.ads (revision 213263) +++ atree.ads (working copy) @@ -313,8 +313,13 @@ Warnings_Detected : Nat := 0; -- Number of warnings detected. Initialized to zero at the start of - -- compilation. Initialized for -gnatVa use, see comment above. + -- compilation. Initialized for -gnatVa use, see comment above. This + -- count includes the count of style and info messages. + Info_Messages : Nat := 0; + -- Number of info messages generated. Info messages are neved treated as + -- errors (whether from use of the pragma, or the compiler switch -gnatwe). + Warnings_Treated_As_Errors : Nat := 0; -- Number of warnings changed into errors as a result of matching a pattern -- given in a Warning_As_Error configuration pragma. Index: err_vars.ads === --- err_vars.ads(revision 213431) +++ err_vars.ads(working copy) @@ -39,10 +39,10 @@ -- from invalid values in such cases. -- Note on error counts (Serious_Errors_Detected, Total_Errors_Detected, - -- Warnings_Detected). These counts might more logically appear in this - -- unit, but we place them in atree.ads, because of licensing issues. We - -- need to be able to access these counts from units that have the more - -- general licensing conditions. + -- Warnings_Detected, Info_Messages). These counts might more logically + -- appear in this unit, but we place them instead in atree.ads, because of + -- licensing issues. We need to be able to access these counts from units + -- that have the more general licensing conditions. -- -- Error Message Mode Variables -- Index: errout.adb === --- errout.adb (revision 213369) +++ errout.adb (working copy) @@ -261,8 +261,12 @@ M.Deleted := True; Warnings_Detected := Warnings_Detected - 1; + if M.Info then + Info_Messages := Info_Messages - 1; + end if; + if M.Warn_Err then - Warnings_Treated_As_Errors := Warnings_Treated_As_Errors + 1; + Warnings_Treated_As_Errors := Warnings_Treated_As_Errors - 1; end if; end if; @@ -1132,6 +1136,10 @@ if Errors.Table (Cur_Msg).Warn or else Errors.Table (Cur_Msg).Style then Warnings_Detected := Warnings_Detected + 1; + if Errors.Table (Cur_Msg).Info then +Info_Messages := Info_Messages + 1; + end if; + else Total_Errors_Detected := Total_Errors_Detected + 1; @@ -1340,8 +1348,12 @@ Errors.Table (E).Deleted := True; Warnings_Detected := Warnings_Detected - 1; +if Errors.Table (E).Info then + Info_Messages := Info_Messages - 1; +end if; + if Errors.Table (E).Warn_Err then - Warnings_Treated_As_Errors := Warnings_Treated_As_Errors + 1; + Warnings_Treated_As_Errors := Warnings_Treated_As_Errors - 1; end if; end if; end Delete_Warning; @@ -1566,6 +1578,7 @@ Total_Errors_Detected := 0;
[Ada] Parameterless calls to protected operations on selected components.
A multi-level selected component can designate a call to a parameterless protected operation whose target object is itself given by a selected component. When the node is rewritten as a function call, it is necessary to preserve the tree structure of the name, so that overload information and subsequent action insertions work properly. THe following must compile quietly: gcc -c ia-ib-ic-d.adb --- WITH Ada.Unchecked_Deallocation; PACKAGE BODY ia.ib.ic.d IS -- -- Protected declaration -- PROTECTED TYPE network_state_controller IS PROCEDURE set_send_success (success : IN Boolean); PROCEDURE set_receive_success (success : IN Boolean); PROCEDURE set_connected; PROCEDURE set_disconnected; PROCEDURE set_open_connection; PROCEDURE set_close_connection; FUNCTION is_connected RETURN Boolean; FUNCTION is_opened RETURN Boolean; ENTRY wait_disconnected; ENTRY wait_connected; PRIVATE connected : Boolean := False; open_connection : Boolean := False; send_success: Boolean := True; receive_success : Boolean := True; END network_state_controller; TYPE internals IS RECORD -- object management -- - is_initialized : Boolean := False; -- Network -- --- network_state : network_state_controller; -- values that may be sent to the gun -- -- -- values reported back by the gun -- --- -- raw telegram buffers -- -- comms house keeping and control -- --- -- fire zone comms -- --- fire_zone_request_timeout: Natural := 0; -- 10 ms cycle count fire_zone_request_started_at : Natural := 0;-- 10 ms cycle count -- Trace -- - use_tracer_position_entry : Boolean := False; tracer_dusc : Boolean := False; tracer_dusc_overwrite : Boolean := False; tracer_zoom_overwrite : Boolean := False; print_counter : Positive := 1; -- for debugging information call_counter : Natural := 0; tick_cycle : Natural := 0; proceed_counter : Natural := 0; -- for test purpose self: reference := NULL; trace_alarm : Boolean := False; trace_fire : Boolean := False; END RECORD; FUNCTION create RETURN reference IS ref : reference := NEW object; BEGIN ref.hidden := NEW internals; ref.hidden.self := ref; RETURN ref; EXCEPTION WHEN error : OTHERS = RETURN NULL; END create; PROCEDURE initialize (obj : IN OUT object; success : OUT Boolean) IS BEGIN IF NOT obj.hidden.is_initialized THEN obj.hidden.is_initialized := True; success := True; END IF; EXCEPTION WHEN error : OTHERS = success := False; END initialize; -- PROCEDURE free IS NEW Ada.Unchecked_Deallocation (object'class, reference); PROCEDURE free IS NEW Ada.Unchecked_Deallocation (internals, access_internals); -- PROCEDURE finalize (obj : IN OUT reference) IS BEGIN IF obj /= NULL THEN IF obj.hidden /= NULL THEN free (obj.hidden); END IF; free (obj); END IF; EXCEPTION WHEN error : OTHERS = NULL; END finalize; FUNCTION is_connected (obj : IN object) RETURN Boolean IS BEGIN IF obj.hidden.network_state.is_opened AND THEN obj.hidden.network_state.is_connected THEN RETURN True; ELSE RETURN False; END IF; EXCEPTION WHEN error : OTHERS = RETURN False; END is_connected; max_before_reconnect : CONSTANT Positive := 100; PROTECTED BODY network_state_controller IS PROCEDURE set_send_success (success : IN Boolean) IS BEGIN send_success := success; END set_send_success; PROCEDURE set_receive_success (success : IN Boolean) IS BEGIN receive_success := success; END set_receive_success; PROCEDURE set_connected IS BEGIN connected := True; END set_connected; PROCEDURE set_disconnected IS BEGIN connected := False; END set_disconnected; FUNCTION is_connected RETURN Boolean IS BEGIN RETURN open_connection AND THEN connected; END is_connected; FUNCTION is_opened RETURN Boolean IS BEGIN RETURN open_connection; END is_opened; PROCEDURE set_close_connection IS BEGIN connected :=
[PATCH] Add target macro to override DWARF2 frame register size
This patch adds a target macro to allow a backend to override the size
of a dwarf frame register. This is already supported to some extent
by allowing the HARD_REGNO_CALL_PART_CLOBBERED to reduce the size of
a specific register to only that which is saved. However, for the
new MIPS O32 FP64 ABI extension I need more control.
The standard MIPS O32 ABI has 32-bit floating point-registers which
are saved as pairs to form 64-bit values. The FP64 extension widens
all floating-point registers to 64-bit but only saves the even numbered
registers thus not increasing the callee-saved state. In order to
support interlinking between the standard O32 ABI and the extension I
must keep the layout of the frame information the same. The dwarf
frame registers are therefore re-interpreted to refer to the low and
high parts of even numbered 64-bit registers rather than 32 32-bit
registers. To achieve this I must artificially reduce the size of the
frame registers for the widened 64-bit registers down to 32-bit but
cannot use the HARD_REGNO_CALL_PART_CLOBBERED macro to achieve this.
Thanks,
Matthew
gcc/
* dwarf2cfi (DWARF_REG_MODE): Define with default implementation.
(expand_builtin_init_dwarf_reg_sizes): Use DWARF_REG_MODE.
* doc/tm.texi.in (DWARF_REG_MODE): Document.
* doc/tm.texi: Regenerate.
---
gcc/doc/tm.texi.in | 9 +
gcc/dwarf2cfi.c| 5 +
2 files changed, 14 insertions(+)
diff --git a/gcc/doc/tm.texi.in b/gcc/doc/tm.texi.in
index dd72b98..71799f7 100644
--- a/gcc/doc/tm.texi.in
+++ b/gcc/doc/tm.texi.in
@@ -2809,6 +2809,12 @@ in DWARF 2 debug information. The default is zero. A
different value
may reduce the size of debug information on some ports.
@end defmac
+@defmac DWARF_REG_MODE (@var{N}, @var{mode})
+If defined, a C expression whose value is a mode that should be used
+to represent the @var{N}th DWARF frame register. By default this is
+the same @var{mode} as the hard register it relates to.
+@end defmac
+
@node Exception Handling
@subsection Exception Handling Support
@cindex exception handling
diff --git a/gcc/dwarf2cfi.c b/gcc/dwarf2cfi.c
index 85cfb60..163db5d 100644
--- a/gcc/dwarf2cfi.c
+++ b/gcc/dwarf2cfi.c
@@ -252,6 +252,10 @@ init_return_column_size (enum machine_mode mode, rtx mem,
unsigned int c)
gen_int_mode (size, mode));
}
+#ifndef DWARF_REG_MODE
+#define DWARF_REG_MODE(REGNO, MODE) (MODE)
+#endif
+
/* Generate code to initialize the register size table. */
void
@@ -276,6 +280,7 @@ expand_builtin_init_dwarf_reg_sizes (tree address)
if (HARD_REGNO_CALL_PART_CLOBBERED (i, save_mode))
save_mode = choose_hard_reg_mode (i, 1, true);
+ save_mode = DWARF_REG_MODE (i, save_mode);
if (dnum == DWARF_FRAME_RETURN_COLUMN)
{
if (save_mode == VOIDmode)
--
1.9.4
[Ada] Crash on entry call with preconditions and access parameters
When an entry has preconditions, the entry call is wrapped in a procedure call that incorporates the precondition checks. To prevent a double expansion, with possible duplication of extra formals, that procedure call must only be pre- nalyzed and resolved. Expansion takes place upon return to the caller Resolve_Call. No short example available. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Ed Schonberg schonb...@adacore.com * sem_res.adb (Resolve_Entry_Call): When an entry has preconditions, the entry call is wrapped in a procedure call that incorporates the precondition checks. To prevent a double expansion, with possible duplication of extra formals, that procedure call must only be pre-analyzed and resolved. Expansion takes place upon return to the caller Resolve_Call. Index: sem_res.adb === --- sem_res.adb (revision 213460) +++ sem_res.adb (working copy) @@ -7176,7 +7176,11 @@ New_Occurrence_Of (PPC_Wrapper (Nam), Loc), Parameter_Associations = New_Actuals); Rewrite (N, New_Call); -Analyze_And_Resolve (N); + +-- Preanalyze and resolve new call. Current procedure is called +-- from Resolve_Call, after which expansion will take place. + +Preanalyze_And_Resolve (N); return; end; end if;
[Ada] Access types that designate limited views of types with tasks.
This patch fixes the handling of access types whose designated types are limited views of untagged types with tasks. executing the following: gnatmake -q main.adb main must yield: Task started Task started --- with Package_Common; use Package_Common; with Package_A; use Package_A; procedure Main is pClassA : Class_A_Ptr := new Class_A; aClassA : access Class_A := new Class_A; pTypeA : Type_A_Ptr := new Type_A; aTypeA : access Type_A := new Type_A; pTaskA : Task_A_Ptr := new Task_A; aTaskA : access Task_A := new Task_A; begin -- OK pClassA.Start; pClassA.Start; Start (pTypeA); Start (aTypeA); pTaskA.Start; aTaskA.Start; end Main; --- package Package_A is type Class_A is tagged null record; procedure Start (self : in out Class_A) is null; type Type_A is null record; procedure Start (obj : access Type_A) is null; task type Task_A is entry Start; end Task_A; end Package_A; --- limited with Package_A; package Package_Common is type Class_A_Ptr is access all Package_A.Class_A; type Type_A_Ptr is access all Package_A.Type_A; type Task_A_Ptr is access all Package_A.Task_A; end Package_Common; --- with text_IO; use text_IO; package body Package_A is task body Task_A is begin accept Start do Put_Line (Task started); end; end Task_A; end Package_A; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Ed Schonberg schonb...@adacore.com * sem_ch3.adb (Access_Type_Declaration): If designated type is a limited view, create a master entity (as is already done for class-wide types) in case the full view designates a type that contains tasks. * sem_ch8.adb (Find_Selected_Component): If prefix is a dereference and the designated type is a limited view, use the non-limited view if available. Index: sem_ch3.adb === --- sem_ch3.adb (revision 213451) +++ sem_ch3.adb (working copy) @@ -1331,9 +1331,23 @@ if Ekind (Root_Type (Entity (S))) = E_Incomplete_Type then Set_Directly_Designated_Type (T, Entity (S)); + +-- If the designated type is a limited view, we cannot tell if +-- the full view contains tasks, and there is no way to handle +-- that full view in a client. We create a master entity for the +-- scope, which will be used when a client determines that one +-- is needed. + +if From_Limited_With (Entity (S)) + and then not Is_Class_Wide_Type (Entity (S)) +then + Set_Ekind (T, E_Access_Type); + Build_Master_Entity (T); + Build_Master_Renaming (T); +end if; + else -Set_Directly_Designated_Type (T, - Process_Subtype (S, P, T, 'P')); +Set_Directly_Designated_Type (T, Process_Subtype (S, P, T, 'P')); end if; -- If the access definition is of the form: ACCESS NOT NULL .. Index: sem_ch8.adb === --- sem_ch8.adb (revision 213440) +++ sem_ch8.adb (working copy) @@ -6236,6 +6236,25 @@ Write_Entity_Info (P_Type, ); Write_Eol; end if; + -- The designated type may be a limited view with no components. + -- Check whether the non-limited view is available, because in some + -- cases this will not be set when instlling the context. + + if Is_Access_Type (P_Type) then +declare + D : constant Entity_Id := Directly_Designated_Type (P_Type); +begin + if Is_Incomplete_Type (D) + and then not Is_Class_Wide_Type (D) + and then From_Limited_With (D) + and then Present (Non_Limited_View (D)) + and then not Is_Class_Wide_Type (Non_Limited_View (D)) + then + Set_Directly_Designated_Type (P_Type, Non_Limited_View (D)); + end if; +end; + end if; + -- First check for components of a record object (not the -- result of a call, which is handled below).
[Ada] Minor internal cleanup
Underlying_Type now recursively goes down the chain of private derivations, including the Underlying_Full_View if any, so we call it instead of making the descent manually in Build_Initialization_Call. No functional changes. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Eric Botcazou ebotca...@adacore.com * exp_ch3.adb (Build_Initialization_Call): Call Underlying_Type to go down the chain of private derivations. * freeze.adb (Freeze_Entity): Fix typo in comment. Index: freeze.adb === --- freeze.adb (revision 213463) +++ freeze.adb (working copy) @@ -5034,7 +5034,7 @@ -- Otherwise freeze full view and patch the pointers so that -- the freeze node will elaborate both views in the back end. -- However, if full view is itself private, freeze underlying - -- full view instead and patch the pointer so that the freeze + -- full view instead and patch the pointers so that the freeze -- node will elaborate the three views in the back end. else Index: exp_ch3.adb === --- exp_ch3.adb (revision 213439) +++ exp_ch3.adb (working copy) @@ -1491,15 +1491,21 @@ return Empty_List; end if; - -- Go to full view if private type. In the case of successive - -- private derivations, this can require more than one step. + -- Go to full view or underlying full view if private type. In the case + -- of successive private derivations, this can require two steps. - while Is_Private_Type (Full_Type) + if Is_Private_Type (Full_Type) and then Present (Full_View (Full_Type)) - loop + then Full_Type := Full_View (Full_Type); - end loop; + end if; + if Is_Private_Type (Full_Type) +and then Present (Underlying_Full_View (Full_Type)) + then + Full_Type := Underlying_Full_View (Full_Type); + end if; + -- If Typ is derived, the procedure is the initialization procedure for -- the root type. Wrap the argument in an conversion to make it type -- honest. Actually it isn't quite type honest, because there can be @@ -1583,12 +1589,6 @@ begin if Is_Protected_Type (T) then T := Corresponding_Record_Type (T); - - elsif Is_Private_Type (T) - and then Present (Underlying_Full_View (T)) - and then Is_Protected_Type (Underlying_Full_View (T)) - then - T := Corresponding_Record_Type (Underlying_Full_View (T)); end if; Arg :=
[Ada] Crash on generic instance with class-wide actual
This patch reimplements part of the support for AI05-0071 which deals with generic/instance scenarios involving a formal type with unknown discriminants, a generic primitive operation of the formal type declared with a box and an actual class-wide type. -- Source -- -- types.ads package Types is type T1 is tagged null record; procedure Prim_Op (Param : T1); type T2 is tagged null record; procedure Prim_Op (Param : T2'Class); type T3 is tagged null record; procedure Prim_Op (Param : T3); procedure Prim_Op (Param : T3'Class); type T4 is tagged null record; end Types; -- gen.ads generic type Formal_Typ () is private; with procedure Prim_Op (Param : Formal_Typ) is ; package Gen is end Gen; -- instances.ads with Gen; with Types; use Types; package Instances is package Inst1 is new Gen (T1'Class); -- OK package Inst2 is new Gen (T2'Class); -- OK package Inst3 is new Gen (T3'Class); -- ERROR, two primitives visible package Inst4 is new Gen (T4'Class); -- ERROR, no primitives visible end Instances; -- Compilation and output -- $ gcc -c instances.ads instances.ads:7:04: instantiation error at gen.ads:3 instances.ads:7:04: ambiguous actual for generic subprogram Prim_Op instances.ads:7:04: possible interpretation: Prim_Op defined at types.ads:10 instances.ads:7:04: possible interpretation: Prim_Op defined at types.ads:9 instances.ads:8:04: instantiation error at gen.ads:3 instances.ads:8:04: no visible subprogram matches the specification for Prim_Op Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Hristian Kirtchev kirtc...@adacore.com * sem_ch8.adb (Analyze_Subprogram_Renaming): Alphabetize globals and move certain variables to the local variable section. Call Build_Class_Wide_Wrapper when renaming a default actual subprogram with a class-wide actual. (Build_Class_Wide_Wrapper): New routine. (Check_Class_Wide_Actual): Removed. (Find_Renamed_Entity): Code reformatting. (Has_Class_Wide_Actual): Alphabetize. Change the logic of the predicate as the renamed name may not necessarely denote the correct subprogram. Index: sem_ch8.adb === --- sem_ch8.adb (revision 213464) +++ sem_ch8.adb (working copy) @@ -1812,18 +1812,51 @@ - procedure Analyze_Subprogram_Renaming (N : Node_Id) is - Formal_Spec : constant Node_Id := Corresponding_Formal_Spec (N); - Is_Actual : constant Boolean := Present (Formal_Spec); - Inst_Node : Node_Id := Empty; + Formal_Spec : constant Entity_Id:= Corresponding_Formal_Spec (N); + Is_Actual : constant Boolean := Present (Formal_Spec); Nam : constant Node_Id := Name (N); - New_S : Entity_Id; - Old_S : Entity_Id := Empty; - Rename_Spec : Entity_Id; Save_AV : constant Ada_Version_Type := Ada_Version; Save_AVP: constant Node_Id := Ada_Version_Pragma; Save_AV_Exp : constant Ada_Version_Type := Ada_Version_Explicit; Spec: constant Node_Id := Specification (N); + Old_S : Entity_Id := Empty; + Rename_Spec : Entity_Id; + + procedure Build_Class_Wide_Wrapper +(Ren_Id : out Entity_Id; + Wrap_Id : out Entity_Id); + -- Ada 2012 (AI05-0071): A generic/instance scenario involving a formal + -- type with unknown discriminants and a generic primitive operation of + -- the said type with a box require special processing when the actual + -- is a class-wide type: + + --generic + -- type Formal_Typ () is private; + -- with procedure Prim_Op (Param : Formal_Typ) is ; + --package Gen is ... + + --package Inst is new Gen (Actual_Typ'Class); + + -- In this case the general renaming mechanism used in the prologue of + -- an instance no longer applies: + + --procedure Prim_Op (Param : Formal_Typ) renames Prim_Op; + + -- The above is replaced the following wrapper/renaming combination: + + --procedure Prim_Op (Param : Formal_Typ) is -- wrapper + --begin + -- Prim_Op (Param);-- primitive + --end Wrapper; + + --procedure Dummy (Param : Formal_Typ) renames Prim_Op; + + -- This transformation applies only if there is no explicit visible + -- class-wide operation at the point of the instantiation. Ren_Id is + -- the entity of the renaming declaration. Wrap_Id is the entity of + -- the generated class-wide wrapper (or Any_Id). + procedure Check_Null_Exclusion (Ren : Entity_Id; Sub : Entity_Id); @@ -1845,6 +1878,11 @@ --
Re: [RFC: Patch, PR 60102] [4.9/4.10 Regression] powerpc fp-bit ices@dwf_regno
Rohit, #define DWARF_REG_TO_UNWIND_COLUMN(r) \ - ((r) 1200 ? ((r) - 1200 + (DWARF_FRAME_REGISTERS - 32)) : (r)) + ((r) = 1200 ? ((r) - 1200 + (DWARF_FRAME_REGISTERS - 32)) : (r)) OK, makes sense. /* Use gcc hard register numbering for eh_frame. */ -#define DWARF_FRAME_REGNUM(REGNO) (REGNO) +#define DWARF_FRAME_REGNUM(REGNO) \ + ((REGNO) = FIRST_SPE_HIGH_REGNO ? ((REGNO) - FIRST_SPE_HIGH_REGNO + 1200) : (REGNO)) Any reason for not using SPE_HIGH_REGNO_P here, just in case we do get other hard registers at some point? Otherwise this now looks good to me. (Of course, I cannot approve the patch myself.) Thanks, Ulrich -- Dr. Ulrich Weigand GNU/Linux compilers and toolchain ulrich.weig...@de.ibm.com
[Ada] Fix breach of privacy with type derived from private discriminated
This fixes a breach of privacy for types derived from private discriminated record types whose full view has no discriminants, as well as factors out the code doing the full derivation in the various cases into a single child procedure of Build_Derived_Private_Type. As a consequence, this removes obsolete code that was catching the privacy breach in some specific cases. The following specs p1.ads, p2.ads, p3.ads and p4.ads must be rejected: p1.ads:8:18: invalid prefix in selected component A p2.ads:12:18: invalid prefix in selected component A p3.ads:8:18: invalid prefix in selected component A p4.ads:12:18: invalid prefix in selected component A with Q; package P1 is type My_Data is new Q.Data; A : My_Data; B : Natural := A.Length; end P1; with Q; package P2 is type My_Data is private; private type My_Data is new Q.Data; A : My_Data; B : Natural := A.Length; end P2; package Q is type Data is private; private type Small is range 1 .. 10; type Data (D : Small := 1) is record Length : Natural; end record; end Q; with Q2; package P3 is type My_Data is new Q2.Data; A : My_Data; B : Natural := A.Length; end P3; with Q2; package P4 is type My_Data is limited private; private type My_Data is new Q2.Data; A : My_Data; B : Natural := A.Length; end P4; package Q2 is type Data is limited private; private type Small is range 1 .. 10; type Data (D : Small := 1) is tagged limited record Length : Natural; end record; end Q2; Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Eric Botcazou ebotca...@adacore.com * einfo.ads (Has_Private_Ancestor): Remove obsolete usage. * exp_ch4.adb (Expand_Composite_Equality): Add conversion of the actuals in the case of untagged record types too. * sem_ch3.adb (Build_Full_Derivation): New procedure to create the full derivation of a derived private type, extracted from... (Copy_And_Build): In the case of record types and most enumeration types, copy the original declaration. Build the full derivation according to the approach extracted from... (Build_Derived_Private_Type): ...here. Call Build_Full_Derivation to create the full derivation in all existing cases and also create it in the no-discriminants/discriminants case instead of deriving directly from the full view. (Is_Visible_Component): Remove obsolete code. * sem_aggr.adb (Resolve_Record_Aggregate): Likewise. Index: sem_aggr.adb === --- sem_aggr.adb(revision 213458) +++ sem_aggr.adb(working copy) @@ -3984,21 +3984,6 @@ -- Typ is not a derived tagged type else --- A type derived from an untagged private type whose full view --- has discriminants is constructed as a record type but there --- are no legal aggregates for it. - -if Is_Derived_Type (Typ) - and then Has_Private_Ancestor (Typ) - and then Nkind (N) /= N_Extension_Aggregate -then - Error_Msg_Node_2 := Base_Type (Etype (Typ)); - Error_Msg_NE - (no aggregate available for type derived from - private type, N, Typ); - return; -end if; - Record_Def := Type_Definition (Parent (Base_Type (Typ))); if Null_Present (Record_Def) then Index: sem_ch3.adb === --- sem_ch3.adb (revision 213471) +++ sem_ch3.adb (working copy) @@ -6543,40 +6543,143 @@ Loc : constant Source_Ptr := Sloc (N); Der_Base: Entity_Id; Discr : Entity_Id; - Full_Decl : Node_Id := Empty; Full_Der: Entity_Id; Full_P : Entity_Id; Last_Discr : Entity_Id; Par_Scope : constant Entity_Id := Scope (Base_Type (Parent_Type)); - Swapped : Boolean := False; + procedure Build_Full_Derivation; + -- Build full derivation, i.e. derive from the full view + procedure Copy_And_Build; -- Copy derived type declaration, replace parent with its full view, - -- and analyze new declaration. + -- and build derivation + --- + -- Build_Full_Derivation -- + --- + + procedure Build_Full_Derivation is + begin + -- If parent scope is not open, install the declarations + + if not In_Open_Scopes (Par_Scope) then +Install_Private_Declarations (Par_Scope); +Install_Visible_Declarations (Par_Scope); +Copy_And_Build; +Uninstall_Declarations (Par_Scope); + + -- If parent scope is open and in another unit, and parent has a + -- completion, then the derivation is taking place in the visible +
[COMMITTED] Add myself to MAINTAINERS file (Write After Approval)
Hi, This patch adds myself to the MAINTAINERS file (Write After Approval). Commmitted as 213481 Thanks. -- Jiong Index: MAINTAINERS === --- MAINTAINERS (revision 213472) +++ MAINTAINERS (working copy) @@ -564,6 +564,7 @@ Tom de Vries t...@codesourcery.com Nenad Vukicevicne...@intrepid.com Feng Wang fengw...@nudt.edu.cn +Jiong Wang jiong.w...@arm.com Stephen M. Webb stephen.w...@bregmasoft.com John Wehle j...@feith.com Florian Weimer f...@deneb.enyo.de
RE: [PATCH, i386] Handle extended family cpuid info for AMD
In this case, having only check for family ID should be enough. If
BTVER1 and BTVER2 can be uniquely determined by their family IDs ,
IMO, this would be the most future-proof approach. Signature checks will
override family id checks which will override cpuid checks.
Thank you Uros!
I have modified source only for BTVER2.
The way BTVER1 is currently assigned to processor includes more than one
family. So, I am leaving that unmoved.
Bootstrap passes.
Is it OK for trunk and backport to open branches.
Regards
-Ganesh
diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 706fedc..202bd99 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,8 @@
+2014-08-01 Ganesh Gopalasubramanian ganesh.gopalasubraman...@amd.com
+
+ * config/i386/driver-i386.c (host_detect_local_cpu): Handle AMD's
extended family
+ information. Handle BTVER2 cpu with cpuid family value.
+
2014-07-31 James Greenhalgh james.greenha...@arm.com
* config/aarch64/arm_neon.h (vpadd_suf8,16,32,64): Move to
diff --git a/gcc/config/i386/driver-i386.c b/gcc/config/i386/driver-i386.c
index 1c6385f..0402c90 100644
--- a/gcc/config/i386/driver-i386.c
+++ b/gcc/config/i386/driver-i386.c
@@ -432,7 +432,8 @@ const char *host_detect_local_cpu (int argc, const char
**argv)
model = (eax 4) 0x0f;
family = (eax 8) 0x0f;
- if (vendor == signature_INTEL_ebx)
+ if ((vendor == signature_INTEL_ebx) ||
+ (vendor == signature_AMD_ebx))
{
unsigned int extended_model, extended_family;
@@ -576,7 +577,7 @@ const char *host_detect_local_cpu (int argc, const char
**argv)
if (name == signature_NSC_ebx)
processor = PROCESSOR_GEODE;
- else if (has_movbe)
+ else if (family == 22)
processor = PROCESSOR_BTVER2;
else if (has_avx2)
processor = PROCESSOR_BDVER4;
Re: [PATCH] Add target macro to override DWARF2 frame register size
On Fri, Aug 1, 2014 at 3:31 PM, Matthew Fortune wrote: This patch adds a target macro Please don't add target macros. Add a hook if you must, but we're supposed to remove target macros, not add new ones :-) Ciao! Steven
Re: [PATCH, i386] Handle extended family cpuid info for AMD
On Fri, Aug 01, 2014 at 02:52:28PM +, Gopalasubramanian, Ganesh wrote: --- a/gcc/config/i386/driver-i386.c +++ b/gcc/config/i386/driver-i386.c @@ -432,7 +432,8 @@ const char *host_detect_local_cpu (int argc, const char **argv) model = (eax 4) 0x0f; family = (eax 8) 0x0f; - if (vendor == signature_INTEL_ebx) + if ((vendor == signature_INTEL_ebx) || + (vendor == signature_AMD_ebx)) Wrong formatting. No ()s around the comparisons needed, and || should go on the second line, not first. @@ -576,7 +577,7 @@ const char *host_detect_local_cpu (int argc, const char **argv) if (name == signature_NSC_ebx) processor = PROCESSOR_GEODE; - else if (has_movbe) + else if (family == 22) processor = PROCESSOR_BTVER2; Wouldn't it be safer to use has_movbe family == 22? I mean, especially with emulators which choose to provide some architecture, but disable some CPUID flags it is IMHO safer to also check the flags. Jakub
[PATCH AArch64] Removed unused get_lane and dup_lane builtins.
None of the variants of __builtin_aarch64_get_lane or __builtin_aarch64_dup_lane are used in arm_neon.h (vdup_lane uses vget_lane and vdup_n, vget_lane uses be_checked_get_lane to do an endianness swap, vdup_n uses gcc vector extension code). So remove them. Regression tested on aarch64-none-elf. gcc/ChangeLog: * gcc/config/aarch64/aarch64-simd-builtins.def (dup_lane, get_lane): Delete. -- Index: gcc/config/aarch64/aarch64-simd-builtins.def === --- gcc/config/aarch64/aarch64-simd-builtins.def(revision 213452) +++ gcc/config/aarch64/aarch64-simd-builtins.def(working copy) @@ -47,8 +47,6 @@ VAR1 (UNOP, addp, 0, di) BUILTIN_VDQ_BHSI (UNOP, clz, 2) - BUILTIN_VALL (GETLANE, get_lane, 0) - VAR1 (GETLANE, get_lane, 0, di) BUILTIN_VALL (GETLANE, be_checked_get_lane, 0) VAR1 (REINTERP_SS, reinterpretdi, 0, v1df) @@ -74,7 +72,6 @@ VAR1 (REINTERP_PS, reinterpretv2si, 0, v1df) VAR1 (REINTERP_PS, reinterpretv2sf, 0, v1df) - BUILTIN_VDQ_I (BINOP, dup_lane, 0) /* Implemented by aarch64_surqrshlmode. */ BUILTIN_VSDQ_I (BINOP, sqshl, 0) BUILTIN_VSDQ_I (BINOP_UUS, uqshl, 0)
Re: [PATCH] PowerPC: Implement TARGET_ATOMIC_ASSIGN_EXPAND_FENV
On Thu, 31 Jul 2014, David Edelsohn wrote: Thanks for implementing the FENV support. The patch generally looks good to me. My one concern is a detail in the implementation of update. I do not have enough experience with GENERIC to verify the details and it seems like it is missing building an outer COMPOUND_EXPR containing update_mffs and the CALL_EXPR for update mtfsf. I suppose what's actually odd there is that you have + tree update_mffs = build2 (MODIFY_EXPR, void_type_node, old_fenv, call_mffs); + + tree old_llu = build1 (VIEW_CONVERT_EXPR, uint64_type_node, update_mffs); so you build a MODIFY_EXPR in void_type_node but then convert it with a VIEW_CONVERT_EXPR. If you'd built the MODIFY_EXPR in double_type_node then the VIEW_CONVERT_EXPR would be meaningful (the value of an assignment a = b being the new value of a), but reinterpreting a void value doesn't make sense. Or you could probably just use call_mffs directly in the VIEW_CONVERT_EXPR without explicitly creating the old_fenv variable. -- Joseph S. Myers jos...@codesourcery.com
Re: [Ada] Remove xgnatugn in doc generation
On Fri, 1 Aug 2014, Arnaud Charlet wrote: [ Note: the previous related change was rejected due to a too large patch file attached. The mail was: This is another step in removing VMS code, and also a step in converting the documentation to the rest/sphinx format. The xgnatugn preprocessing step will also soon be removed. 2014-08-01 Ben Brosgol bros...@adacore.com * gnat_ugn.texi, projects.texi, xgnatugn.adb: Removed all VMS conditionalization from gnat_ugn.texi and projects.texi, and updated (considerably simplified) xgnatugn.adb, to be removed soon. ] Following up from previous doc clean ups (see above), this change removes the preprocessing step needed to generate gnat_ugn.info. Tested on x86_64-pc-linux-gnu, committed on trunk 2014-08-01 Arnaud Charlet char...@adacore.com * ug_words, xgnatugn.adb, gcc-interface/Make-lang.in: Remove xgnatugn.adb and ug_words. Could you check if this needs any changes to the support in update_web_docs_svn for generating the online manuals, which would previously have run this preprocessing step? -- Joseph S. Myers jos...@codesourcery.com
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
Jakub Jelinek ja...@redhat.com writes: Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time It may also not be desirable for LTO builds, which can take very long by themselves. -Andi -- a...@linux.intel.com -- Speaking for myself only
Re: [Ada] Generate Machine, Model and Rounding FP attributes in line
On Fri, 1 Aug 2014, Arnaud Charlet wrote: This change makes it so that the Machine, Model and Rounding FP attributes are generated in line by the compiler (in conjunction with a conversion to an integer type for the third) and optimized on architectures that do not make use of internal extended precision in the FPU. Using TARGET_FLT_EVAL_METHOD would be better than basing this on examining the machine modes (after all, on x86_64, you have XFmode but use of SSE means there are no excess previous issues; likewise, ia64 has XFmode but proper floating-point operations for all modes, including the formatOf operations that take wider operands and produce a narrower result without double rounding). Except that S/390 defines TARGET_FLT_EVAL_METHOD to 1 for historical reasons (but doesn't, I think, actually have any excess precision issues with the back end - excess precision would only really be used for C with -fexcess-precision=standard), so strictly you'd need a different macro that differs from TARGET_FLT_EVAL_METHOD only on S/390 (i.e., a macro that describes what FLT_EVAL_METHOD is at back-end level rather than what it is at C API level with -fexcess-precision=standard). -- Joseph S. Myers jos...@codesourcery.com
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
On Fri, Aug 01, 2014 at 08:43:27AM -0700, Andi Kleen wrote: Jakub Jelinek ja...@redhat.com writes: Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time It may also not be desirable for LTO builds, which can take very long by themselves. Indeed, and preparing preprocessed sources for all the files is hard in that case anyway. But as it is keyed on cc1 substring being present in the compiler's name and lto uses lto1, it shouldn't trigger for that. Only if you get ICEs with -flto already when compiling from C/C++ source, but then it doesn't take very long usually and there is preprocessed source available. Jakub
Re: [PATCH AArch64] Removed unused get_lane and dup_lane builtins.
On 1 August 2014 16:09, Alan Lawrence alan.lawre...@arm.com wrote: None of the variants of __builtin_aarch64_get_lane or __builtin_aarch64_dup_lane are used in arm_neon.h (vdup_lane uses vget_lane and vdup_n, vget_lane uses be_checked_get_lane to do an endianness swap, vdup_n uses gcc vector extension code). So remove them. Regression tested on aarch64-none-elf. gcc/ChangeLog: * gcc/config/aarch64/aarch64-simd-builtins.def (dup_lane, get_lane): Delete. OK /Marcus
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
On Fri, Aug 01, 2014 at 12:13:18PM +0400, Yury Gribov wrote: On 08/01/2014 12:00 PM, Jakub Jelinek wrote: Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time I think we should use David Malcolm's approach i.e. add some --report-bug flag to driver. This could be enabled by default at configure time via --with-spec. -freport-bug or whatever we call it should not be, at least if it attempts to communicate over the network, should not be required though, lots of people do e.g. package builds in various chroots etc. and you really don't want to perform any network activity from there. Still, having info whether an ICE was reproduceable or not is useful in that case, and having preprocessed source left in /tmp with the path printed to stderr is also useful, you can e.g. grab those /tmp/cc*.out files from the chroot and allow people to report later on. The ICE message can of course suggest an option or script or command line to run to report the bug or navigate the user through bug reporting process. Jakub
Re: [PATCH] Move memory builtin foldings to GIMPLE
On Fri, 1 Aug 2014, Richard Biener wrote: Does this look sensible? In particular, does any Frontend maintainer object to no longer getting these simplifications? Note they will still apply during gimplification and given they now work on GENERIC we can simplify them quite a bit (but as followup). I see no problems with this in C terms. I expect some code may well do things like static size_t len = __builtin_strlen (foo); and expect the static initializer to get folded, but that's only an issue with folding for constant arguments to pure functions - not for the functions involved here, whose whole purpose is to modify memory. -- Joseph S. Myers jos...@codesourcery.com
Re: [PATCH 2/2] Enable elimination of zext/sext
if (rhs_uns)
return wi::ge_p (min, 0); // if min = 0 then range contains positive
values
else
return wi::le_p (max, wi::max_value (TYPE_PRECISION (TREE_TYPE
(ssa)), SIGNED); // if max = signed-max-of-type then range doesn't
need sign-extension
I think we will have to check that ssa has necessary sign/zero extension
when assigned to lhs_type. If PROMOTE_MODE tells us that ssa's type will
be interpreted differently, the value range of ssa also will have
corresponding range. In this cases, shouldn’t we have to check for
upper and lower limit for both min and max?
Hmm? That's exactly what the check is testing... we know that
min = max thus if min = 0 then max = 0.
zero_extension will never do anything on [0, INF]
If max MAX-SIGNED then sign-extension will not do anything. Ok,
sign-extension will do sth for negative values still. So rather
if (rhs_uns)
return wi::geu_p (min, 0);
else
return wi::ges_p (min, 0) wi::les_p (max, wi::max_value
(TYPE_PRECISION (TREE_TYPE (ssa)), SIGNED));
?
Thanks for the explanation. I agree. Don’t we have to however check this
on lhs_uns as this function is checking if ssa is promoted for lhs_sign
and lhs_mode?
Here is an attempt based on this. I ran regression testing with
arm-none-linux-gnueabi on qemu-arm without any new regressions.
Sine I am not comparing value ranges to see if it can be represented in
lhs_sigh, I can now skip the PROMOTED_MODE check.
I am still using wide_int::from (instead of wi::max_value) to get the
limit as I have to match the precision with min, max precision.
otherwise wide_int comparisons will not work. Is there a better way for
this?
/* Return TRUE if value in SSA is already zero/sign extended for lhs type
(type here is the combination of LHS_MODE and LHS_UNS) using value range
information stored. Return FALSE otherwise. */
bool
promoted_for_type_p (tree ssa, enum machine_mode lhs_mode, bool lhs_uns)
{
wide_int min, max, limit;
tree lhs_type;
bool rhs_uns;
signop rhs_signop;
if (ssa == NULL_TREE
|| TREE_CODE (ssa) != SSA_NAME
|| !INTEGRAL_TYPE_P (TREE_TYPE (ssa)))
return false;
/* Return FALSE if value_range is not recorded for SSA. */
if (get_range_info (ssa, min, max) != VR_RANGE)
return false;
rhs_uns = TYPE_UNSIGNED (TREE_TYPE (ssa));
rhs_signop = rhs_uns ? UNSIGNED : SIGNED;
lhs_type = lang_hooks.types.type_for_mode (lhs_mode, lhs_uns);
limit = wide_int::from (TYPE_MAX_VALUE (lhs_type),
TYPE_PRECISION (TREE_TYPE (ssa)), SIGNED);
if (lhs_uns)
/* If min = 0 then range contains positive values and doesnt need
zero-extension. */
return wi::ge_p (min, 0, rhs_signop);
else
/* If min = 0 and max = signed-max-of-type then range doesn't need
sign-extension. */
return wi::ge_p (min, 0, rhs_signop) wi::le_p (max, limit,
rhs_signop);
}
Thanks,
Kugan
Re: [Ada] Legality checks involving aspect Import
On Aug 1, 2014, at 3:16 AM, Arnaud Charlet char...@adacore.com wrote: This patch adds some missing legality checks No, GNU English doesn’t have us using legal/illegal unless the motivation is a law of some specific jurisdiction. q.ads:2:28: imported entities cannot have explicit initialization (RM 8.1 (24)) q.ads:6:24: no initialization allowed for declaration of J at line 5 q.ads:6:24: imported entities cannot be initialized (RM B.1(24)) q.ads:9:24: no initialization allowed for declaration of K at line 8 q.ads:9:24: imported entities cannot be initialized (RM B.1(24)) There are ok, but… * sem_ch13.adb (Analyze_Aspect_Specifications, case Aspect_Import): Set Is_Imported flag at once, to simplify subsequent legality Not this. checks. Reject the aspect on an object whose declaration has an explicit initial value. * sem_prag.adb (Process_Import_Or_Interface): Use original node to check legality of an initial value for an imported entity. Nor this.
Re: [PATCH] Move memory builtin foldings to GIMPLE
On 08/01/14 06:43, Richard Biener wrote: The following patch removes folding of bzero, memset, bcopy, memcpy, mempcpy and memmove from builtins.c and re-implements them in gimple-fold.c. This means those foldings are no longer accessible from GENERIC folding. Bootstrapped on x86_64-unknown-linux-gnu - testing seems to reveal some testsuite fallout I still have to look into. Does this look sensible? In particular, does any Frontend maintainer object to no longer getting these simplifications? Note they will still apply during gimplification and given they now work on GENERIC we can simplify them quite a bit (but as followup). Given the general need to move folding out of the FEs and into the gimple optimizers, I'm all for this. Obviously we have to deal will fallout, but I'm all for the general direction this is going. Thanks, Jeff
a patch was backported to gcc 4.9 branch
The following patch was backported from mainline to gcc-4.9 branch: https://gcc.gnu.org/ml/gcc-patches/2014-07/msg00845.html Committed as rev. 213497. 2014-08-01 Vladimir Makarov vmaka...@redhat.com * lra-constraints.c (remove_inheritance_pseudos): Process destination pseudo too.
RE: [RFC: Patch, PR 60102] [4.9/4.10 Regression] powerpc fp-bit ices@dwf_regno
Hello Ulrich, Thanks. /* Use gcc hard register numbering for eh_frame. */ -#define DWARF_FRAME_REGNUM(REGNO) (REGNO) +#define DWARF_FRAME_REGNUM(REGNO) \ + ((REGNO) = FIRST_SPE_HIGH_REGNO ? ((REGNO) - FIRST_SPE_HIGH_REGNO + +1200) : (REGNO)) Any reason for not using SPE_HIGH_REGNO_P here, just in case we do get other hard registers at some point? Yes, we can use it. I just have to move the definition of SPE_HIGH_REGNO_P macro before DWARF_FRAME_REGNUM macro definition. [Previously, I had defined and placed SPE_HIGH_REGNO_P macro along with similar macros ALTIVEC_REGNO_P etc.] I had updated the patch as required (For this last change, I have checked/tested only the builds: ppc64 trunk, e500v2 v4.9.1 bareboard linux build). PR target/60102 [libgcc] 2014-07-31 Rohit rohitarul...@freescale.com * config/rs6000/linux-unwind.h (ppc_fallback_frame_state): Update based on change in SPE high register numbers and 3 HTM registers. [gcc] 2014-07-31 Rohit rohitarul...@freescale.com * config/rs6000/rs6000.c (rs6000_reg_names) : Add SPE high register names. (alt_reg_names) : Likewise. (rs6000_dwarf_register_span) : For SPE high registers, replace dwarf register numbers with GCC hard register numbers. (rs6000_init_dwarf_reg_sizes_extra) : Likewise. (rs6000_dbx_register_number): For SPE high registers, return dwarf register number for the corresponding GCC hard register number. * config/rs6000/rs6000.h (FIRST_PSEUDO_REGISTER) : Update based on 32 newly added GCC hard register numbers for SPE high registers. (DWARF_FRAME_REGISTERS) : Likewise. (DWARF_REG_TO_UNWIND_COLUMN) : Likewise. (DWARF_FRAME_REGNUM) : Likewise. (FIXED_REGISTERS) : Likewise. (CALL_USED_REGISTERS) : Likewise. (CALL_REALLY_USED_REGISTERS) : Likewise. (REG_ALLOC_ORDER) : Likewise. (enum reg_class) : Likewise. (REG_CLASS_NAMES) : Likewise. (REG_CLASS_CONTENTS) : Likewise. (SPE_HIGH_REGNO_P) : New macro to identify SPE high registers. * gcc.target/powerpc/pr60102.c: New testcase. Regards, Rohit pr60102.patch Description: pr60102.patch
Re: [RFC: Patch, PR 60102] [4.9/4.10 Regression] powerpc fp-bit ices@dwf_regno
On Fri, Aug 01, 2014 at 06:03:56PM +, rohitarul...@freescale.com wrote:
PR target/60102
--- libgcc/config/rs6000/linux-unwind.h (revision 213110)
+++ libgcc/config/rs6000/linux-unwind.h (working copy)
@@ -274,8 +274,8 @@ ppc_fallback_frame_state (struct _Unwind
#ifdef __SPE__
for (i = 14; i 32; i++)
{
- fs-regs.reg[i + FIRST_PSEUDO_REGISTER - 1].how = REG_SAVED_OFFSET;
- fs-regs.reg[i + FIRST_PSEUDO_REGISTER - 1].loc.offset
+ fs-regs.reg[i + FIRST_SPE_HIGH_REGNO - 4].how = REG_SAVED_OFFSET;
+ fs-regs.reg[i + FIRST_SPE_HIGH_REGNO - 4].loc.offset
= (long) regs-vregs - new_cfa + 4 * i;
}
#endif
is a different index, previously i + 116, newly i + 113, is that
intentional?
Jakub
Re: [RFC, PATCH 3/n] IPA C++ refactoring
On 08/01/14 05:35, Martin Liška wrote:
2014-07-31 Martin Liskamli...@suse.cz
* cgraph.h (cgraph_node_set, varpool_node_set): Removed structs and all
related functions that manipulated these data structures.
(cgraph_new_nodes): vec replaces cgraph_node_set.
* cgraphunit.c (cgraph_new_nodes): Likewise.
(cgraph_process_new_functions): Likewise.
(cgraph_node::add_new_function): Likewise.
* ipa-inline-transform.c (can_remove_node_now_p_1): Likewise.
* ipa-utils.c: Implementations of legacy data structures removed,
data structure was replaced by a hash_map.
* tree-emutls.c (tls_vars): hash_map replaces varpool_node_set.
(emutls_index): New functions are used.
(ipa_lower_emutls): Likewise.
Looks like a good cleanup. Just aone nit left with the most recent
version (the one where you fixed the indention issues pointed out by Jakub):
symtab-node-removal.patch
diff --git a/gcc/tree-emutls.c b/gcc/tree-emutls.c
index 89197c7..2c9cfe4 100644
--- a/gcc/tree-emutls.c
+++ b/gcc/tree-emutls.c
@@ -608,13 +581,20 @@ lower_emutls_phi_arg (gimple phi, unsigned int i, struct
lower_emutls_data *d)
}
}
+bool
+reset_access (varpool_node * const , tls_var_data *data, void *)
+{
+ data-access = NULL;
+
+ return true;
+}
Need a comment for this function.
Approved once you add that comment. Please post the final patch for
archival purposes.
Thanks,
Jeff
RE: [RFC: Patch, PR 60102] [4.9/4.10 Regression] powerpc fp-bit ices@dwf_regno
Jakub,
On Fri, Aug 01, 2014 at 06:03:56PM +, rohitarul...@freescale.com wrote:
PR target/60102
--- libgcc/config/rs6000/linux-unwind.h (revision 213110)
+++ libgcc/config/rs6000/linux-unwind.h (working copy)
@@ -274,8 +274,8 @@ ppc_fallback_frame_state (struct _Unwind
#ifdef __SPE__
for (i = 14; i 32; i++)
{
- fs-regs.reg[i + FIRST_PSEUDO_REGISTER - 1].how = REG_SAVED_OFFSET;
- fs-regs.reg[i + FIRST_PSEUDO_REGISTER - 1].loc.offset
+ fs-regs.reg[i + FIRST_SPE_HIGH_REGNO - 4].how = REG_SAVED_OFFSET;
+ fs-regs.reg[i + FIRST_SPE_HIGH_REGNO - 4].loc.offset
= (long) regs-vregs - new_cfa + 4 * i;
}
#endif
is a different index, previously i + 116, newly i + 113, is that intentional?
Yes, it is intentional.
This part of code wasn't updated after the introduction of three HTM registers.
Regards,
Rohit
Re: [PATCH, i386] Handle extended family cpuid info for AMD
On Fri, Aug 1, 2014 at 5:00 PM, Jakub Jelinek ja...@redhat.com wrote: --- a/gcc/config/i386/driver-i386.c +++ b/gcc/config/i386/driver-i386.c @@ -432,7 +432,8 @@ const char *host_detect_local_cpu (int argc, const char **argv) model = (eax 4) 0x0f; family = (eax 8) 0x0f; - if (vendor == signature_INTEL_ebx) + if ((vendor == signature_INTEL_ebx) || + (vendor == signature_AMD_ebx)) Wrong formatting. No ()s around the comparisons needed, and || should go on the second line, not first. @@ -576,7 +577,7 @@ const char *host_detect_local_cpu (int argc, const char **argv) if (name == signature_NSC_ebx) processor = PROCESSOR_GEODE; - else if (has_movbe) + else if (family == 22) processor = PROCESSOR_BTVER2; Wouldn't it be safer to use has_movbe family == 22? I mean, especially with emulators which choose to provide some architecture, but disable some CPUID flags it is IMHO safer to also check the flags. OK, after rethinking this, let's go with the safer original patch (sorry for a bit of confusion, but we cleared a lot of open questions during the discussion). The original patch is OK for mainline and branches. OTOH, it looks we will have to redesign -march=native someday to only pass options, derived from cpuid and eventually pass -mtune based on detected family. Thanks, Uros.
Re: [PATCH][testsuite] Don't run cproj-fails-with-broken-glibc.c for broken glibc
On Aug 1, 2014, at 3:35 AM, Rainer Orth r...@cebitec.uni-bielefeld.de wrote: I'm not at all happy with this patch That test, even if we go the glibc version route, needs to be XFAILed instead of requiring the working version. Apart from that, new effective-target keywords need documenting in doc/sourcebuild.texi. Some background folks: https://gcc.gnu.org/ml/gcc-patches/2010-04/msg00490.html So I too agree that skipping the test case on known broken systems when the entire design of the test case is to fail on exactly those broken systems is borked. So, what should we do with it instead. Either xfailing it on known bad systems, or removing it entirely as misguided seem to be the two choices that make sense. What are people in favor of? Clearly the original author wanted at least the xfail. I lean toward removal I think. Why? The gcc test suite, while it could be a filesystem test suite, an OS test suite, a libc test suite, generally speaking that is beyond the scope of this project. From the test case: Check the runtime behavior of the C library's cproj() function and whether it follows the standard. Versions of GLIBC through 2.11.1 had an incorrect implementation which will conflict with GCC's builtin cproj(). GLIBC 2.12+ should be okay.
Re: [Patch] PR 61692 - Fix for inline asm ICE
On 08/01/14 02:07, David Wohlferd wrote: I'd love to. Unfortunately, my platform doesn't support valgrind. Ah. Also, please include the testcase you had nlabels part. I have created the testcase for the 31 labels problem. However, not so much for the nclobbers part. And if I'm going to patch both, I should have testcases for both. Tell you what, pass along what you've got and I'll run it under valgrind here. I'm quite confident both need to be changed -- though it is possible nothing will trigger with the nclobbers stuff if it is indeed handled separately throughout the guts of GCC. Jeff
Re: [C++ Patch] DR 217 follow up (and more)
On 07/31/2014 01:36 PM, Paolo Carlini wrote: The problem is that when grokfndecl calls duplicate_decls in such member cases it looks through TEMPLATE_DECLs and then telling apart the two cases above is tough, both are FUNCTION_DECLs Ideas about the best way to handle that? Could you just condition it on DECL_TEMPLATE_INFO? tree t2 = TYPE_ARG_TYPES (TREE_TYPE (newdecl)); int i = 1; - if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE) + if (DECL_NONSTATIC_MEMBER_FUNCTION_P (newdecl)) t1 = TREE_CHAIN (t1), t2 = TREE_CHAIN (t2); Let's use FUNCTION_FIRST_USER_PARMTYPE instead. - if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE + if (DECL_FUNCTION_MEMBER_P (newdecl) This change is OK. and in the predicate patch: - || TREE_CODE (TREE_TYPE (target_type)) != METHOD_TYPE); + || !DECL_NONSTATIC_MEMBER_FUNCTION_P (target_type)); This is wrong because target_type is not a FUNCTION_DECL. We should probably either add FUNCTION_DECL_CHECK to DECL_NONSTATIC_MEMBER_FUNCTION_P or make it false if the argument isn't a FUNCTION_DECL. There are a bunch of instances of this in the patch. if (TREE_CODE (TREE_TYPE (expr)) == FUNCTION_TYPE - || TREE_CODE (TREE_TYPE (expr)) == METHOD_TYPE + || DECL_NONSTATIC_MEMBER_FUNCTION_P (expr) Changing this instance and similar ones doesn't make much sense since the condition is looking for any kind of function, not just a non-static member function. Jason
Re: [Ada] Remove xgnatugn in doc generation
2014-08-01 Arnaud Charlet char...@adacore.com
* ug_words, xgnatugn.adb, gcc-interface/Make-lang.in: Remove
xgnatugn.adb and ug_words.
Could you check if this needs any changes to the support in
update_web_docs_svn for generating the online manuals, which would
previously have run this preprocessing step?
Thanks for the note.
Updated as follows, this should do the trick as far as I can tell:
* update_web_docs_svn: Simplify build of gnat_ugn.
--
Index: update_web_docs_svn
===
--- update_web_docs_svn (revision 213156)
+++ update_web_docs_svn (working copy)
@@ -20,11 +20,7 @@ MANUALS=cpp
gcj
gfortran
gfc-internals
- gnat_ug_unx
- gnat_ug_vms
- gnat_ug_vxw
- gnat_ug_wnt
- gnat_ugn_unw
+ gnat_ugn
gnat-style
gnat_rm
libgomp
@@ -118,8 +114,6 @@ fi
find gcc -type f \( -name '*.texi' \
-o -path gcc/gcc/doc/install.texi2html \
-o -path gcc/gcc/doc/include/texinfo.tex \
- -o -path gcc/gcc/ada/xgnatugn.adb \
- -o -path gcc/gcc/ada/ug_words \
-o -path gcc/gcc/BASE-VER \
-o -path gcc/gcc/DEV-PHASE \
-o -print0 \) | xargs -0 rm -f
@@ -131,14 +125,6 @@ tar cf docs-sources.tar gcc
# and fdl.texi.
includedir=gcc/gcc/doc/include
-# Generate gnat_ugn_unw
-
-if [ -f gcc/gcc/ada/xgnatugn.adb ]; then
- gnatmake -q gcc/gcc/ada/xgnatugn
- ./xgnatugn unw gcc/gcc/ada/gnat_ugn.texi \
- gcc/gcc/ada/ug_words gnat_ugn_unw.texi
-fi
-
# Generate gcc-vers.texi.
(
echo @set version-GCC $(cat gcc/gcc/BASE-VER)
@@ -161,7 +147,7 @@ for file in $MANUALS; do
filename=`find . -name ${file}.texi`
if [ ${filename} ]; then
includes=-I ${includedir} -I `dirname ${filename}`
-if [ $file = gnat_ugn_unw ]; then
+if [ $file = gnat_ugn ]; then
includes=$includes -I gcc/gcc/ada
fi
makeinfo --html $includes -o ${file} ${filename}
Re: [PATCH][Ping v5] Add patch for debugging compiler ICEs
On 08/01/14 02:00, Jakub Jelinek wrote: On Thu, Jul 24, 2014 at 04:39:28PM +0400, Maxim Ostapenko wrote: Ping. Don't want to review a patch I wrote partially, so just a few comments: 1) IMHO it should be configure time selectable (not sure about the default, but for non-release branches IMHO it should default to off, for release branches I don't know). The point is that while it is useful for people to report gcc bugs in production compilers, it is not useful for gcc developers on their boxes, having ICEs take 3 times as long is not desirable for people who deal with those all the time We definitely want the ability to select. Waiting on an ICE to trigger 3 times would be painful for the developers. So I'd think that enabled for the release branches and disabled for the trunk would be the right default. 2) recently a bug has been reported that some ICEs which print RTL to stderr are considered non-reproduceable, because the addresses e.g. in call_insn fndecl change through address space randomization. For the retries the driver should supposedly append -fdump-noaddr. Right. I think we discussed this at Cauldron. I think it's conceptually ok with those two changes, then it's just reviewing the details. If you or Maxim could make those changes and repost, I'll cover the review. jeff
Re: [PATCH] libstdc++: add _GLIBCXX_ macro prefix in atexit_thread.cc
On 30/07/14 23:36 +0800, Zifei Tong wrote:
On Wed, Jul 30, 2014 at 11:25 PM, Jonathan Wakely jwak...@redhat.com wrote:
On 30/07/14 22:42 +0800, Zifei Tong wrote:
Hi,
I found an issue that the __cxa_thread_atexit_impl() function never called
by
__cxa_thread_atexit() even with newest glibc which have
__cxa_thread_atexit_impl
implemented.
It turns out that the code tried to use macro
HAVE___CXA_THREAD_ATEXIT_IMPL, but
not _GLIBCXX_HAVE___CXA_THREAD_ATEXIT_IMPL which is defined in
bits/c++config.h
(generated from autoconf scripts).
This patch adds the missing macro prefix.
The patch is correct, thanks very much.
Do you have commit access?
I don't think so, this is my first patch to gcc :)
If not I can do it for you.
Thanks!
I've committed the attached patch to trunk, I'll apply it to the 4.8
and 4.9 branches soon too.
Tested x86_64-linux, with both glibc 2.11 and 2.18 so with and without
__cxa_thread_atexit_impl.
Thanks again for the report and patch.
commit ecaf491ba1a82e858b8a0654efeee74e55c74c05
Author: Jonathan Wakely jwak...@redhat.com
Date: Fri Aug 1 19:53:58 2014 +0100
2014-08-01 Zifei Tong zifeit...@gmail.com
* libsupc++/atexit_thread.cc (HAVE___CXA_THREAD_ATEXIT_IMPL): Add
_GLIBCXX_ prefix to macro.
diff --git a/libstdc++-v3/libsupc++/atexit_thread.cc b/libstdc++-v3/libsupc++/atexit_thread.cc
index db20200..dff08e9 100644
--- a/libstdc++-v3/libsupc++/atexit_thread.cc
+++ b/libstdc++-v3/libsupc++/atexit_thread.cc
@@ -26,7 +26,7 @@
#include new
#include bits/gthr.h
-#if HAVE___CXA_THREAD_ATEXIT_IMPL
+#if _GLIBCXX_HAVE___CXA_THREAD_ATEXIT_IMPL
extern C int __cxa_thread_atexit_impl (void (*func) (void *),
void *arg, void *d);
@@ -38,7 +38,7 @@ __cxxabiv1::__cxa_thread_atexit (void (*dtor)(void *),
return __cxa_thread_atexit_impl (dtor, obj, dso_handle);
}
-#else /* HAVE___CXA_THREAD_ATEXIT_IMPL */
+#else /* _GLIBCXX_HAVE___CXA_THREAD_ATEXIT_IMPL */
namespace {
// One element in a singly-linked stack of cleanups.
@@ -142,4 +142,4 @@ __cxxabiv1::__cxa_thread_atexit (void (*dtor)(void *), void *obj, void */*dso_ha
return 0;
}
-#endif /* HAVE___CXA_THREAD_ATEXIT_IMPL */
+#endif /* _GLIBCXX_HAVE___CXA_THREAD_ATEXIT_IMPL */
Re: [C++ Patch] DR 217 follow up (and more)
Hi,
On 08/01/2014 08:30 PM, Jason Merrill wrote:
On 07/31/2014 01:36 PM, Paolo Carlini wrote:
The problem is that when grokfndecl calls duplicate_decls in such
member cases it looks through TEMPLATE_DECLs
and then telling apart the two cases above is tough, both are
FUNCTION_DECLs
Ideas about the best way to handle that?
Could you just condition it on DECL_TEMPLATE_INFO?
Of course, thanks! I have no idea why I thought I was unable to go back
to the template information :( Then the whole issue - of course, again -
is easy.
tree t2 = TYPE_ARG_TYPES (TREE_TYPE (newdecl));
int i = 1;
- if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE)
+ if (DECL_NONSTATIC_MEMBER_FUNCTION_P (newdecl))
t1 = TREE_CHAIN (t1), t2 = TREE_CHAIN (t2);
Let's use FUNCTION_FIRST_USER_PARMTYPE instead.
- if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE
+ if (DECL_FUNCTION_MEMBER_P (newdecl)
This change is OK.
Thanks, I'm attaching below what I'm going to apply.
I will return to the separate DECL_NONSTATIC_MEMBER_FUNCTION_P mini
project after c++/15339.
Paolo.
Index: decl.c
===
--- decl.c (revision 213439)
+++ decl.c (working copy)
@@ -1706,14 +1706,11 @@ duplicate_decls (tree newdecl, tree olddecl, bool
;
else if (TREE_CODE (olddecl) == FUNCTION_DECL)
{
- tree t1 = TYPE_ARG_TYPES (TREE_TYPE (olddecl));
- tree t2 = TYPE_ARG_TYPES (TREE_TYPE (newdecl));
+ tree t1 = FUNCTION_FIRST_USER_PARMTYPE (olddecl);
+ tree t2 = FUNCTION_FIRST_USER_PARMTYPE (newdecl);
int i = 1;
- if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE)
- t1 = TREE_CHAIN (t1), t2 = TREE_CHAIN (t2);
-
- if (TREE_CODE (TREE_TYPE (newdecl)) == METHOD_TYPE
+ if (DECL_FUNCTION_MEMBER_P (newdecl)
CLASSTYPE_TEMPLATE_INFO (CP_DECL_CONTEXT (newdecl)))
{
/* C++11 8.3.6/6.
Re: [Patch, Fortran] -fcoarray=lib - support CRITICAL, prepare for locking support
Hello, I was implementing lock/unlock on the library side when I found a possible problem in the patch: if (is_lock_type == GFC_CAF_CRITICAL) +reg_type = sym-attr.artificial ? GFC_CAF_CRITICAL : GFC_CAF_LOCK_STATIC; + else +reg_type = GFC_CAF_COARRAY_STATIC; the if statement cannot be true since is_lock_type is a boolean and GFC_CAF_CRITICAL is 4. Using if (is_lock_type) it produces the right result for the lock registration. Regards Alessandro 2014-07-28 14:37 GMT-06:00 Tobias Burnus bur...@net-b.de: This patch implements -fcoarray=lib support for CRITICAL blocks and includes some preparatory work for locking. In particular: * Updated the documentation for locking/critical, minor cleanup. The patch goes on top of the unreviewed patch https://gcc.gnu.org/ml/fortran/2014-07/msg00155.html * Add libcaf_single implementation for lock/unlock * Add lock/unlock calls for CRITICAL * Register static/SAVEd locking variables and locking variables for critical sections. Build and currently regtesting on x86-64-gnu-linux. OK when it regtested successfully? * * * Still to be done as follow up: * Handling the registering of lock-type components in statically allocated derived types * Handling the registering of lock-type variables and components with allocate and with implicit/explicit deallocate * Calling lock/unlock function for those * Test case for locking and critical blocks Other coarray to-do items: * Type-conversion test case missing * Vector subscript library implementation + test cases * Extending the documentation * Issues with striding for coarray components of derived types * Nonallocatable polymophic coarrays and select type/associated * Allocatable/pointer components of coarrays; co_reduce and co_broadcast Tobias
Re: [PATCH, libfortran] Backport xmallocarray to 4.8/4.9 (CVE-2014-5044)
On Fri, Aug 1, 2014 at 3:49 PM, Jakub Jelinek ja...@redhat.com wrote:
On Thu, Jul 31, 2014 at 11:32:12PM +0300, Janne Blomqvist wrote:
a while ago I committed a patch to trunk adding a function
xmallocarray to libgfortran, which is a malloc wrapper like xmalloc
but has two arguments and does an overflow check before multiplying
them together.
That seems to be unnecessarily expensive for the common case where both
nmemb and size are small.
calloc in glibc uses something like following to avoid the division most of
the time, if both nmemb and size are small enough then nmemb * size can't
overflow. At least for 64-bit architectures small is smaller than 4GB
for both numbers.
2014-08-01 Jakub Jelinek ja...@redhat.com
* runtime/memory.c (xmallocarray): Avoid division for the common case.
--- libgfortran/runtime/memory.c.jj 2014-06-18 08:50:33.0 +0200
+++ libgfortran/runtime/memory.c2014-08-01 14:41:08.385856116 +0200
@@ -56,7 +56,9 @@ xmallocarray (size_t nmemb, size_t size)
if (!nmemb || !size)
size = nmemb = 1;
- else if (nmemb SIZE_MAX / size)
+#define HALF_SIZE_T (((size_t) 1) (__CHAR_BIT__ * sizeof (size_t) / 2))
+ else if (__builtin_expect ((nmemb | size) = HALF_SIZE_T, 0)
+ nmemb SIZE_MAX / size)
{
errno = ENOMEM;
os_error (Integer overflow in xmallocarray);
Nice, though as os_error() has the _Noreturn specifier the
__builtin_expect() is not necessary, right? In libgfortran.h we have
in the comment block above the likely/unlikely macros, which are
wrappers around __builtin_expect:
...as __builtin_expect overrides the compiler
heuristic, do not use in conditions where one of the branches ends with a
call to a function with __attribute__((noreturn)): the compiler internal
heuristic will mark this branch as much less likely as unlikely() would
do.
Ok for trunk/4.9/4.8. If you choose to leave the __builtin_expect
there, please explain why?
--
Janne Blomqvist
Re: [PATCH, libfortran] Backport xmallocarray to 4.8/4.9 (CVE-2014-5044)
On Sat, Aug 02, 2014 at 12:09:24AM +0300, Janne Blomqvist wrote:
--- libgfortran/runtime/memory.c.jj 2014-06-18 08:50:33.0 +0200
+++ libgfortran/runtime/memory.c2014-08-01 14:41:08.385856116 +0200
@@ -56,7 +56,9 @@ xmallocarray (size_t nmemb, size_t size)
if (!nmemb || !size)
size = nmemb = 1;
- else if (nmemb SIZE_MAX / size)
+#define HALF_SIZE_T (((size_t) 1) (__CHAR_BIT__ * sizeof (size_t) / 2))
+ else if (__builtin_expect ((nmemb | size) = HALF_SIZE_T, 0)
+ nmemb SIZE_MAX / size)
{
errno = ENOMEM;
os_error (Integer overflow in xmallocarray);
Nice, though as os_error() has the _Noreturn specifier the
__builtin_expect() is not necessary, right? In libgfortran.h we have
The reason for __builtin_expect here was to make already the
nmemb SIZE_MAX / size
computation as unlikely, the noreturn predictor will of course DTRT with the
{} block.
Jakub
[C++ Patch] PR 15339
Hi,
thus this is what I have tested for c++/15339. By the way, I figured out
why yesterday I got stupidly confused about DECL_TEMPLATE_INFO: at this
point, for a member function template, newdecl does *not* have
DECL_TEMPLATE_INFO set, only olddecl does. Consistently, I adjusted all
the code in the condition to work on olddecl. Tested x86_64-linux.
Thanks,
Paolo.
//
/cp
2014-08-01 Paolo Carlini paolo.carl...@oracle.com
PR c++/15339
* decl.c (duplicate_decls): Handle default arguments in function
templates.
/testsuite
2014-08-01 Paolo Carlini paolo.carl...@oracle.com
PR c++/15339
* g++.dg/other/default9.C: New.
* g++.dg/other/default3.C: Remove xfail.
Index: cp/decl.c
===
--- cp/decl.c (revision 213505)
+++ cp/decl.c (working copy)
@@ -1704,19 +1704,23 @@ duplicate_decls (tree newdecl, tree olddecl, bool
if (DECL_LANG_SPECIFIC (olddecl) DECL_USE_TEMPLATE (olddecl))
;
- else if (TREE_CODE (olddecl) == FUNCTION_DECL)
+ else if (DECL_DECLARES_FUNCTION_P (olddecl))
{
tree t1 = FUNCTION_FIRST_USER_PARMTYPE (olddecl);
tree t2 = FUNCTION_FIRST_USER_PARMTYPE (newdecl);
int i = 1;
- if (DECL_FUNCTION_MEMBER_P (newdecl)
- CLASSTYPE_TEMPLATE_INFO (CP_DECL_CONTEXT (newdecl)))
+ if (DECL_FUNCTION_TEMPLATE_P (olddecl)
+ || (DECL_FUNCTION_MEMBER_P (olddecl)
+ (/* grokfndecl passes member function templates too
+as FUNCTION_DECLs. */
+ DECL_TEMPLATE_INFO (olddecl)
+ /* C++11 8.3.6/6.
+Default arguments for a member function of a class
+template shall be specified on the initial declaration
+of the member function within the class template. */
+ || CLASSTYPE_TEMPLATE_INFO (CP_DECL_CONTEXT (olddecl)
{
- /* C++11 8.3.6/6.
-Default arguments for a member function of a class template
-shall be specified on the initial declaration of the member
-function within the class template. */
for (; t2 t2 != void_list_node; t2 = TREE_CHAIN (t2))
if (TREE_PURPOSE (t2))
{
Index: testsuite/g++.dg/other/default3.C
===
--- testsuite/g++.dg/other/default3.C (revision 213504)
+++ testsuite/g++.dg/other/default3.C (working copy)
@@ -25,7 +25,7 @@ templatetypename void g3(int, int);
templatetypename void g3(int = 0, int);// { dg-error default }
templatetypename void g4(int, int);
-templatetypename void g4(int = 0, int) {} // { dg-error default {
xfail *-*-* } }
+templatetypename void g4(int = 0, int) {} // { dg-error default }
templatetypename void g5();
templatetypename void g5(int = 0, int);// { dg-error default }
Index: testsuite/g++.dg/other/default9.C
===
--- testsuite/g++.dg/other/default9.C (revision 0)
+++ testsuite/g++.dg/other/default9.C (working copy)
@@ -0,0 +1,18 @@
+// PR c++/15339
+
+templatetypename void fun(int);
+templatetypename void fun(int = 0); // { dg-error default arguments }
+
+class A
+{
+ templatetypename void fun(int);
+};
+
+templatetypename void A::fun(int = 0) { } // { dg-error default arguments }
+
+class B
+{
+ void fun(int);
+};
+
+void B::fun(int = 0) { }
Re: FWD: Re: OpenACC subarray specifications in the GCC Fortran front end
On 08/01/2014 12:48 AM, Jakub Jelinek wrote:
On Wed, Jul 23, 2014 at 05:42:32PM -0700, Cesar Philippidis wrote:
Jakub, before your Fortran OpenMP 4 target changes, Ilmir had written the
test case gcc/testsuite/gfortran.dg/gomp/map-1.f90 (based on his
interpretation and implementation of OpenMP 4 target), which I have now
amended with XFAILs and changed error messages -- anything in there that
you'd like to see addressed for Fortran OpenMP 4 target?
+ !$omp target map(j(5:4)) ! { dg-error Lower bound of OpenMP array
section in greater than upper { xfail *-*-* } }
+ !$omp end target
I think this isn't an error in Fortran, if low bound is above upper bound,
then it is considered a zero size array section. Though supposedly for
depend clause we might want to diagnose that.
+ !$omp target map(aas) ! { dg-error The upper bound in the last
dimension must appear { xfail *-*-* } }
+ !$omp end target
Assumed-size in map without array section would be indeed nice thing to
diagnose.
+ !$omp target map(tt%i) ! { dg-error Syntax error in OpenMP variable
list }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
Right now the parsing of !$omp directives in case of parsing error rejects
the whole directive, perhaps it should be reconsidered unless it is a fatal
error from which there is no easy way out.
+ !$omp target map(tt%j(1)) ! { dg-bogus Syntax error in OpenMP variable
list { xfail *-*-* } }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
+
+ !$omp target map(tt%j(1:)) ! { dg-bogus Syntax error in OpenMP
variable list { xfail *-*-* } }
+ !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET
statement { xfail *-*-* } }
These two are pending resolution on omp-lang, I had exchanged a few mails
about it, I think we shouldn't support those for consistency with the C/C++
support, where tt.j[1] or tt.j[1:] and similar is explicitly invalid.
Jakub, should I drop the map-1.f90 test?
Not the whole testcase, just the problematic parts (or, just remove the
dg-error/xfail or replace dg-bogus xfail with dg-error), for the j(5:4)
and tt%j(1)/tt%j(1:) cases?
Thank you for the feedback. I've committed those changes to gomp-4_0-branch.
Cesar
2014-08-01 Cesar Philippidis ce...@codesourcery.com
gcc/testsuite/
* gfortran.dg/gomp/map-1.f90 (test): Update error reporting.
diff --git a/gcc/testsuite/gfortran.dg/gomp/map-1.f90 b/gcc/testsuite/gfortran.dg/gomp/map-1.f90
index de96ed2..e4b8b86 100644
--- a/gcc/testsuite/gfortran.dg/gomp/map-1.f90
+++ b/gcc/testsuite/gfortran.dg/gomp/map-1.f90
@@ -45,7 +45,7 @@ subroutine test(aas)
!$omp target map(j(0:)) ! { dg-warning out of bounds }
!$omp end target
- !$omp target map(j(5:4)) ! { dg-error Lower bound of OpenMP array section in greater than upper { xfail *-*-* } }
+ !$omp target map(j(5:4))
!$omp end target
!$omp target map(j(5:))
@@ -93,16 +93,17 @@ subroutine test(aas)
!$omp end target
!$omp target map(tt%i) ! { dg-error Syntax error in OpenMP variable list }
- !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET statement { xfail *-*-* } }
+ !$omp end target ! { dg-error Unexpected !\\\$OMP END TARGET statement }
!$omp target map(tt%j) ! { dg-error Syntax error in OpenMP variable list }
- !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET statement { xfail *-*-* } }
+ !$omp end target ! { dg-error Unexpected !\\\$OMP END TARGET statement }
- !$omp target map(tt%j(1)) ! { dg-bogus Syntax error in OpenMP variable list { xfail *-*-* } }
- !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET statement { xfail *-*-* } }
+ ! broken test
+ !$omp target map(tt%j(1)) ! { dg-error Syntax error in OpenMP variable list }
+ !$omp end target ! { dg-error Unexpected !\\\$OMP END TARGET statement }
- !$omp target map(tt%j(1:)) ! { dg-bogus Syntax error in OpenMP variable list { xfail *-*-* } }
- !$omp end target ! { dg-bogus Unexpected !\\\$OMP END TARGET statement { xfail *-*-* } }
+ !$omp target map(tt%j(1:)) ! { dg-error Syntax error in OpenMP variable list }
+ !$omp end target ! { dg-error Unexpected !\\\$OMP END TARGET statement }
!$omp target map(tp) ! { dg-error THREADPRIVATE object 'tp' in MAP clause }
!$omp end target
Re: C++ PATCH for c++/60417 (explicit ctor vs aggregate init)
On 03/04/2014 05:14 PM, Jason Merrill wrote:
In C++11, copy-list-initialization by explicit constructor is an error,
even for the default constructor. But this combined with the change of
aggregate initialization to use {} for any missing initializers means
that well-formed C++03 code becomes ill-formed in C++11. Until the
committee decides what to do about this, let's call the implicit
initializers direct-initialization.
...and for array initialization as well.
Tested x86_64-pc-linux-gnu, applying to trunk.
commit ef09581840cc49d85599a2a9f2c855d8c8b60d9f
Author: Jason Merrill ja...@redhat.com
Date: Fri Aug 1 16:13:54 2014 -0400
PR c++/60417
* init.c (build_vec_init): Set CONSTRUCTOR_IS_DIRECT_INIT on
init-list for trailing elements.
* typeck2.c (process_init_constructor_array): Likewise.
diff --git a/gcc/cp/init.c b/gcc/cp/init.c
index f8cae28..5bb3 100644
--- a/gcc/cp/init.c
+++ b/gcc/cp/init.c
@@ -3545,19 +3545,11 @@ build_vec_init (tree base, tree maxindex, tree init,
try_block = begin_try_block ();
}
- /* If the initializer is {}, then all elements are initialized from {}.
- But for non-classes, that's the same as value-initialization. */
+ bool empty_list = false;
if (init BRACE_ENCLOSED_INITIALIZER_P (init)
CONSTRUCTOR_NELTS (init) == 0)
-{
- if (CLASS_TYPE_P (type))
- /* Leave init alone. */;
- else
- {
- init = NULL_TREE;
- explicit_value_init_p = true;
- }
-}
+/* Skip over the handling of non-empty init lists. */
+empty_list = true;
/* Maybe pull out constant value when from_array? */
@@ -3677,14 +3669,8 @@ build_vec_init (tree base, tree maxindex, tree init,
vec_free (new_vec);
}
- /* Any elements without explicit initializers get {}. */
- if (cxx_dialect = cxx11 AGGREGATE_TYPE_P (type))
- init = build_constructor (init_list_type_node, NULL);
- else
- {
- init = NULL_TREE;
- explicit_value_init_p = true;
- }
+ /* Any elements without explicit initializers get T{}. */
+ empty_list = true;
}
else if (from_array)
{
@@ -3699,6 +3685,26 @@ build_vec_init (tree base, tree maxindex, tree init,
}
}
+ /* If the initializer is {}, then all elements are initialized from T{}.
+ But for non-classes, that's the same as value-initialization. */
+ if (empty_list)
+{
+ if (cxx_dialect = cxx11 AGGREGATE_TYPE_P (type))
+ {
+ if (BRACE_ENCLOSED_INITIALIZER_P (init)
+ CONSTRUCTOR_NELTS (init) == 0)
+ /* Reuse it. */;
+ else
+ init = build_constructor (init_list_type_node, NULL);
+ CONSTRUCTOR_IS_DIRECT_INIT (init) = true;
+ }
+ else
+ {
+ init = NULL_TREE;
+ explicit_value_init_p = true;
+ }
+}
+
/* Now, default-initialize any remaining elements. We don't need to
do that if a) the type does not need constructing, or b) we've
already initialized all the elements.
diff --git a/gcc/cp/typeck2.c b/gcc/cp/typeck2.c
index 59a4760..20523fa 100644
--- a/gcc/cp/typeck2.c
+++ b/gcc/cp/typeck2.c
@@ -1239,8 +1239,9 @@ process_init_constructor_array (tree type, tree init,
{
/* If this type needs constructors run for default-initialization,
we can't rely on the back end to do it for us, so make the
- initialization explicit by list-initializing from {}. */
+ initialization explicit by list-initializing from T{}. */
next = build_constructor (init_list_type_node, NULL);
+ CONSTRUCTOR_IS_DIRECT_INIT (next) = true;
next = massage_init_elt (TREE_TYPE (type), next, complain);
if (initializer_zerop (next))
/* The default zero-initialization is fine for us; don't
diff --git a/gcc/testsuite/g++.dg/init/explicit2.C b/gcc/testsuite/g++.dg/init/explicit2.C
new file mode 100644
index 000..d1dbb39
--- /dev/null
+++ b/gcc/testsuite/g++.dg/init/explicit2.C
@@ -0,0 +1,8 @@
+// PR c++/60417
+
+struct A { explicit A(int = 0); };
+
+int main()
+{
+ A a[1] = { };
+}
libgo patch committed: Ignore small argv[0] for backtrace
Reportedly in some cases Docker starts processes with argv[0] pointing
to an empty file. That would cause libgo to pass that empty file to
libbacktrace, which would then fail to do any backtraces. Everything
should work fine if libbacktrace falls back to /proc/self/exe.
This patch to libgo works around the problem by ignoring argv[0] if it
is a small file, or if stat fails. This is not a perfect fix but it's
an unusual problem.
This is PR 61895.
Bootstrapped and ran Go testsuite on x86_64-unknown-linux-gnu.
Committed to mainline and 4.9 branch.
Ian
diff -r 3674ebeb1d40 libgo/runtime/go-caller.c
--- a/libgo/runtime/go-caller.c Sun Jul 20 12:24:14 2014 -0700
+++ b/libgo/runtime/go-caller.c Fri Aug 01 17:41:05 2014 -0700
@@ -7,6 +7,9 @@
/* Implement runtime.Caller. */
#include stdint.h
+#include sys/types.h
+#include sys/stat.h
+#include unistd.h
#include backtrace.h
@@ -99,6 +102,7 @@
if (back_state == NULL)
{
const char *filename;
+ struct stat s;
filename = (const char *) runtime_progname ();
@@ -108,6 +112,14 @@
if (__builtin_strchr (filename, '/') == NULL)
filename = NULL;
+ /* If the file is small, then it's not the real executable.
+ This is specifically to deal with Docker, which uses a bogus
+ argv[0] (http://gcc.gnu.org/PR61895). It would be nice to
+ have a better check for whether this file is the real
+ executable. */
+ if (stat (filename, s) 0 || s.st_size 1024)
+ filename = NULL;
+
back_state = backtrace_create_state (filename, 1, error_callback, NULL);
}
runtime_unlock (back_state_lock);
Re: [C++ Patch] PR 15339
... the below seems better to me:
1- FUNCTION_DECLs and TEMPLATE_DECLs are handled in their own places.
2- When I tweaked default3.C I didn't notice at first the real reason
for the existing xfail: in such case we didn't give the
check_default_args diagnostics. Restored in the below.
3- By handling TEMPLATE_DECLs separately, we can give the
check_redeclaration_no_default_args permerror and the check_default_args
error in a consistent order, that is the former before the latter, which
makes most sense to me (also matches the clang order).
Thanks!
Paolo.
//
/cp
2014-08-01 Paolo Carlini paolo.carl...@oracle.com
PR c++/15339
* decl.c (check_redeclaration_no_default_args): New.
(duplicate_decls): Use it, handle default arguments
in redeclarations of function templates.
/testsuite
2014-08-01 Paolo Carlini paolo.carl...@oracle.com
PR c++/15339
* g++.dg/other/default9.C: New.
* g++.dg/other/default10.C: Likewise.
* g++.dg/other/default3.C: Remove xfail.
Index: cp/decl.c
===
--- cp/decl.c (revision 213505)
+++ cp/decl.c (working copy)
@@ -1236,6 +1236,27 @@ validate_constexpr_redeclaration (tree old_decl, t
return true;
}
+/* DECL is a redeclaration of a function or function template. If
+ it does have default arguments issue a diagnostic. Note: this
+ function is used to enforce the requirements in C++11 8.3.6 about
+ no default arguments in redeclarations. */
+
+static void
+check_redeclaration_no_default_args (tree decl)
+{
+ gcc_assert (DECL_DECLARES_FUNCTION_P (decl));
+
+ for (tree t = FUNCTION_FIRST_USER_PARMTYPE (decl);
+ t t != void_list_node; t = TREE_CHAIN (t))
+if (TREE_PURPOSE (t))
+ {
+ permerror (input_location,
+ redeclaration of %q#D may not have default
+ arguments, decl);
+ return;
+ }
+}
+
#define GNU_INLINE_P(fn) (DECL_DECLARED_INLINE_P (fn) \
lookup_attribute (gnu_inline,\
DECL_ATTRIBUTES (fn)))
@@ -1706,28 +1727,23 @@ duplicate_decls (tree newdecl, tree olddecl, bool
;
else if (TREE_CODE (olddecl) == FUNCTION_DECL)
{
- tree t1 = FUNCTION_FIRST_USER_PARMTYPE (olddecl);
- tree t2 = FUNCTION_FIRST_USER_PARMTYPE (newdecl);
- int i = 1;
-
- if (DECL_FUNCTION_MEMBER_P (newdecl)
- CLASSTYPE_TEMPLATE_INFO (CP_DECL_CONTEXT (newdecl)))
- {
- /* C++11 8.3.6/6.
-Default arguments for a member function of a class template
-shall be specified on the initial declaration of the member
-function within the class template. */
- for (; t2 t2 != void_list_node; t2 = TREE_CHAIN (t2))
- if (TREE_PURPOSE (t2))
- {
- permerror (input_location,
- redeclaration of %q#D may not have default
- arguments, newdecl);
- break;
- }
- }
+ /* Note: free functions, as TEMPLATE_DECLs, are handled below. */
+ if (DECL_FUNCTION_MEMBER_P (olddecl)
+ (/* grokfndecl passes member function templates too
+as FUNCTION_DECLs. */
+ DECL_TEMPLATE_INFO (olddecl)
+ /* C++11 8.3.6/6.
+Default arguments for a member function of a class
+template shall be specified on the initial declaration
+of the member function within the class template. */
+ || CLASSTYPE_TEMPLATE_INFO (CP_DECL_CONTEXT (olddecl
+ check_redeclaration_no_default_args (newdecl);
else
{
+ tree t1 = FUNCTION_FIRST_USER_PARMTYPE (olddecl);
+ tree t2 = FUNCTION_FIRST_USER_PARMTYPE (newdecl);
+ int i = 1;
+
for (; t1 t1 != void_list_node;
t1 = TREE_CHAIN (t1), t2 = TREE_CHAIN (t2), i++)
if (TREE_PURPOSE (t1) TREE_PURPOSE (t2))
@@ -1864,6 +1880,12 @@ duplicate_decls (tree newdecl, tree olddecl, bool
if (DECL_FUNCTION_TEMPLATE_P (newdecl))
{
+ /* Per C++11 8.3.6/4, default arguments cannot be added in later
+declarations of a function template. */
+ check_redeclaration_no_default_args (newdecl);
+
+ check_default_args (newdecl);
+
if (GNU_INLINE_P (old_result) != GNU_INLINE_P (new_result)
DECL_INITIAL (new_result))
{
Index: testsuite/g++.dg/other/default10.C
===
--- testsuite/g++.dg/other/default10.C (revision 0)
+++ testsuite/g++.dg/other/default10.C (working copy)
@@ -0,0 +1,4 @@
+// PR c++/15339
+
