[Bug c/45726] New: Thumb2 instruction emitted for incompatible CPU

[Bug c/45726] New: Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

% cat test.c                               
union prop_data_t
{
    unsigned int u;
    struct
    {
        unsigned short s1;
        unsigned short s2;
    };
};

union prop_data_t broken (int a, int b)
{
    union prop_data_t var;

    if (a)
    {
        var.s2 = 0;
        var.s2 = b ? a : 2;
        if (var.s2)
            return var;
    }

    var.u = 0;
    return var;
}
% echo $CC
/usr/local/arm-elf-eabi-cvs/bin/arm-elf-eabi-gcc
% $CC -mcpu=arm9tdmi -c test.c -O -pipe   
{standard input}: Assembler messages:
{standard input}:25: Error: selected processor does not support ARM mode
`movteq r0,2'
% $CC -mcpu=arm9tdmi -c test.c -Ofast -pipe
{standard input}: Assembler messages:
{standard input}:23: Error: selected processor does not support ARM mode
`movteq r3,2'
% $CC -mcpu=arm9tdmi -c test.c -pipe   
% $CC -v                               
Using built-in specs.
COLLECT_GCC=/usr/local/arm-elf-eabi-cvs/bin/arm-elf-eabi-gcc
COLLECT_LTO_WRAPPER=/usr/local/arm-elf-eabi-cvs/libexec/gcc/arm-elf-eabi/4.6.0/lto-wrapper
Target: arm-elf-eabi
Configured with: ../gcc-4.6-20100918/configure
--prefix=/usr/local/arm-elf-eabi-cvs --target=arm-elf-eabi --enable-lto
--enable-languages=c --disable-docs --disable-libssp
Thread model: single
gcc version 4.6.0 20100918 (experimental) (GCC) 
%

Using 20100918 snapshot

reduced from
http://svn.rockbox.org/viewvc.cgi/trunk/apps/plugins/goban/sgf.c?revision=20001&view=markup

I tried to see which optimization broke with -Ox -Q --help=optimizers but the
list printed doesn't seem to be complete.


-- 
           Summary: Thumb2 instruction emitted for incompatible CPU
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: rafael dot carre at gmail dot com
GCC target triplet: arm-elf-eabi


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug c/45726] Thumb2 instruction emitted for incompatible CPU

[mikpe at it dot uu dot se]

------- Comment #1 from mikpe at it dot uu dot se  2010-09-19 11:15 -------
I see the same on arm-linux-gnueabi with 4.6-20100907 and 4.5-20100916.  It
happens regardless of whether I pass -mcpu=arm9tdmi, -mcpu=armv5tel, or no
-mcpu= at all.


-- 

mikpe at it dot uu dot se changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |mikpe at it dot uu dot se


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2010-09-20 04:52 -------
What binutils version are you using?

movteq is a valid ARM v7 instruction.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

------- Comment #3 from rafael dot carre at gmail dot com  2010-09-20 07:24 -------
I made a CVS checkout of binutils yesterday.

You're right, MOVT is supported on ARMv7 because all ARMv7 supports Thumb2,
http://infocenter.arm.com/help/topic/com.arm.doc.qrc0001m/QRC0001_UAL.pdf
(It says "All Thumb-2 versions of ARM v6 and above" so I suppose some ARMv6 CPU
supports it too)

But this instruction is emitted when an ARMv4 (mcpu=arm9tdmi) or ARMv5
(mcpu=armv5tel) CPU is used.

Do you confuse ARM7 and ARMv7 ? ;)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

------- Comment #4 from rafael dot carre at gmail dot com  2010-09-20 07:47 -------
Created an attachment (id=21844)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=21844&action=view)
Only emit MOVT when targetting Thumb2

Tentative patch.

However as suggested by the original TARGET_32BIT, your comment, and the
binutils message; MOVT might be valid as well when targetting ARM?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

------- Comment #5 from rafael dot carre at gmail dot com  2010-09-20 08:17 -------
Created an attachment (id=21845)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=21845&action=view)
Only emit MOVT when the targetted CPU is Thumb2-able


-- 

rafael dot carre at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #21844|0                           |1
        is obsolete|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[mikpe at it dot uu dot se]

------- Comment #6 from mikpe at it dot uu dot se  2010-09-20 10:29 -------
Can you do a bisection to identify the exact commit responsible?  Looking at
the original commit that introduced the movt md pattern (139881) I see a
TARGET_USE_MOVT guard in the C code that _should_ prevent it from being
selected on non Thumb2-capable CPUs.  If these guards are now broken then they
need to be fixed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

------- Comment #7 from rafael dot carre at gmail dot com  2010-09-20 11:08 -------
I didn't bisect.
Did you try r139881? If I download a checkout (I'm using a snapshot) to bisect
I could use this as a working starting point.

TARGET_USE_MOVT uses arm_arch_thumb2 just like my patch so it's not broken.
I double checked by adding an abort() in the 2 points where TARGET_USE_MOVT is
tested in arm.c

BTW I remember seeing this error in 4.5.0 and/or 4.5.1 but I don't have the
builds around to verify.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[mikpe at it dot uu dot se]

------- Comment #8 from mikpe at it dot uu dot se  2010-09-20 12:02 -------
r139881 is good.  I'll start a bisection.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #9 from rearnsha at gcc dot gnu dot org  2010-09-20 15:20 -------
Must also be present (even if latent) on 4.5.


-- 

rearnsha at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rearnsha at gcc dot gnu dot
                   |                            |org
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
           Keywords|                            |assemble-failure, wrong-code
      Known to work|                            |4.5.0
   Last reconfirmed|0000-00-00 00:00:00         |2010-09-20 15:20:37
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

-- 

rearnsha at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |rearnsha at gcc dot gnu dot
                   |dot org                     |org
             Status|NEW                         |ASSIGNED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #10 from rearnsha at gcc dot gnu dot org  2010-09-20 15:25 -------
Subject: Bug 45726

Author: rearnsha
Date: Mon Sep 20 15:25:44 2010
New Revision: 164436

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=164436
Log:
2010-09-20  Rafael Carre   <rafael.carre@gmail.com>

        PR target/45726
        * arm.md (arm_movtas_ze): Only enable on machine with MOVT.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/arm/arm.md


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #11 from rearnsha at gcc dot gnu dot org  2010-09-20 15:27 -------
Subject: Bug 45726

Author: rearnsha
Date: Mon Sep 20 15:27:13 2010
New Revision: 164437

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=164437
Log:
2010-09-20  Rafael Carre   <rafael.carre@gmail.com>

        PR target/45726
        * arm.md (arm_movtas_ze): Only enable on machine with MOVT.

Modified:
    branches/gcc-4_5-branch/gcc/ChangeLog
    branches/gcc-4_5-branch/gcc/config/arm/arm.md


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #12 from rearnsha at gcc dot gnu dot org  2010-09-20 15:36 -------
Fixed in 4.5.3 and trunk.


-- 

rearnsha at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.5.3


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rafael dot carre at gmail dot com]

------- Comment #13 from rafael dot carre at gmail dot com  2010-09-20 15:46 -------
Is there something wrong with the first hunk of the patch (arm_movt) ?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #14 from rearnsha at gcc dot gnu dot org  2010-09-20 16:13 -------
(In reply to comment #13)
> Is there something wrong with the first hunk of the patch (arm_movt) ?
> 

Nothing really.  I missed that bit.

I think in practice the compiler will never end up matching that pattern (as
lo_sum isn't something the compiler spontaneously generates -- it means
different things on each architecture that supports it, so the generic parts of
the compiler will only use it if they are transforming something that already
uses it), but it should probably be on the trunk version as the current code is
clearly too liberal.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[rearnsha at gcc dot gnu dot org]

------- Comment #15 from rearnsha at gcc dot gnu dot org  2010-09-20 16:22 -------
Subject: Bug 45726

Author: rearnsha
Date: Mon Sep 20 16:21:57 2010
New Revision: 164441

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=164441
Log:
2010-09-20  Rafael Carre   <rafael.carre@gmail.com>

        PR target/45726
        * arm.md (arm_movt): Only enable on machines with MOVT.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/arm/arm.md


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726

[Bug target/45726] Thumb2 instruction emitted for incompatible CPU

[mikpe at it dot uu dot se]

------- Comment #16 from mikpe at it dot uu dot se  2010-09-20 16:37 -------
FWIW, exposed on trunk by r160462 (PR44423 fix), backported to 4.5 in r160775. 
But clearly the issue was latent since the movt patterns were added.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45726