[Bug middle-end/45325] New: target attribute doesn't work with -march=i586

[Bug middle-end/45325] New: target attribute doesn't work with -march=i586

[hjl dot tools at gmail dot com]

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #1: Type: text/plain, Size: 1150 bytes --]

[hjl@gnu-36 gcc]$ cat
../../../../src-trunk/gcc/testsuite/gcc.target/i386/pr38240.c
/* { dg-do compile } */

typedef float V
  __attribute__ ((__vector_size__ (16), __may_alias__));

V __attribute__((target("sse"))) f(const V *ptr) { return *ptr; }

V g(const V *ptr) { return *ptr; }
[hjl@gnu-36 gcc]$ ../../xgcc -B../../
../../../../src-trunk/gcc/testsuite/gcc.target/i386/pr38240.c -S -march=i586
-m32
../../../../src-trunk/gcc/testsuite/gcc.target/i386/pr38240.c: In function ‘g’:
../../../../src-trunk/gcc/testsuite/gcc.target/i386/pr38240.c:8:21: internal
compiler error: in convert_move, at expr.c:326
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.
[hjl@gnu-36 gcc]$


-- 
           Summary: target attribute doesn't work with -march=i586
           Product: gcc
           Version: 4.6.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: hjl dot tools at gmail dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] target attribute doesn't work with -march=i586

[rguenth at gcc dot gnu dot org]

------- Comment #1 from rguenth at gcc dot gnu dot org  2010-08-18 20:31 -------
Well - obviously global typedefs keep their BLKmode.  The target attribute
can't work this way - it's broken by design.


-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|                            |WONTFIX


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] target attribute doesn't work with -march=i586

[steven at gcc dot gnu dot org]

------- Comment #2 from steven at gcc dot gnu dot org  2010-08-18 20:58 -------
WONTFIX for an ICE seems a bit odd to me.
Just needs a diagnostic to reject nonsense target attributes, or a sorry. But
not an ICE.


-- 

steven at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|RESOLVED                    |UNCONFIRMED
         Resolution|WONTFIX                     |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] target attribute doesn't work with -march=i586

[rguenth at gcc dot gnu dot org]

------- Comment #3 from rguenth at gcc dot gnu dot org  2010-08-18 21:18 -------
Well, I think we should back out support for that option.  The set of
"nonsensical" options doesn't include "sse" - but all options are included
in the set of broken options.  At least I expect that you can create a
testcase with such ICEs for every one.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[hjl dot tools at gmail dot com]

------- Comment #4 from hjl dot tools at gmail dot com  2010-08-19 19:38 -------
It is caused by revision 162918:

http://gcc.gnu.org/ml/gcc-cvs/2010-08/msg00129.html


-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rguenth at gcc dot gnu dot
                   |                            |org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[rth at gcc dot gnu dot org]

------- Comment #5 from rth at gcc dot gnu dot org  2010-08-19 19:48 -------
(In reply to comment #3)
> Well, I think we should back out support for that option.  The set of
> "nonsensical" options doesn't include "sse" - but all options are included
> in the set of broken options.  At least I expect that you can create a
> testcase with such ICEs for every one.

Not every one, just those that expand the set of valid backend modes.
Thus "sse3" will not cause additional ice's, assuming "sse2" is in
effect globally.

That's not to say this feature isn't fraught with peril.  I think it
should be improved though, not removed.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[rguenth at gcc dot gnu dot org]

-- 

rguenth at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.6.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[davek at gcc dot gnu dot org]

------- Comment #6 from davek at gcc dot gnu dot org  2010-09-12 23:45 -------
This is also present on i686-pc-cygwin:

> FAIL: gcc.target/i386/pr38240.c (internal compiler error)

ICE happens here:

(gdb) bt
#0  0x006065e0 in convert_move (to=0x7fcc26c0, from=0x7fcc26d0, unsignedp=0)
    at /gnu/gcc/gcc-unpatched/gcc/expr.c:2944
#1  0x00609d2f in store_expr (exp=0x7fde2c10, target=0x7fcc26c0,
    call_param_p=0, nontemporal=0 '\0')
    at /gnu/gcc/gcc-unpatched/gcc/expr.c:2944
#2  0x0060f638 in expand_assignment (to=0x7fe20050, from=0x7fde2c10,
    nontemporal=0 '\0') at /gnu/gcc/gcc-unpatched/gcc/expr.c:2944

(...yes, the line number info is wrong, don't know why but it's unrelated.)

  The patch in r162918 tightens up the condition on calling emit_block_move:

--- trunk/gcc/expr.c    2010/08/05 15:39:54     162917
+++ trunk/gcc/expr.c    2010/08/05 16:37:38     162918
@@ -4752,11 +4752,14 @@
        {
          int unsignedp = TYPE_UNSIGNED (TREE_TYPE (exp));
          if (GET_MODE (target) == BLKmode
-                  || GET_MODE (temp) == BLKmode)
+             && GET_MODE (temp) == BLKmode)
            emit_block_move (target, temp, expr_size (exp),
                             (call_param_p
                              ? BLOCK_OP_CALL_PARM
                              : BLOCK_OP_NORMAL));
+         else if (GET_MODE (target) == BLKmode)
+           store_bit_field (target, INTVAL (expr_size (exp)) * BITS_PER_UNIT,
+                            0, GET_MODE (temp), temp);
          else
            convert_move (target, temp, unsignedp);
        }
so that when temp is BLKmode and target is not, as in this case:

(gdb) call debug_rtx (to) # aka target
(mem/c/i:V4SF (reg/f:SI 58 [ D.1753 ]) [0 <retval>+0 S16 A128])
(gdb) call debug_rtx (from) # aka temp
(mem/c/i:BLK (plus:SI (reg/f:SI 54 virtual-stack-vars)
        (const_int -16 [0xfffffff0])) [0 D.1747+0 S16 A128])

... it falls through into the final else clause, and calls convert_move. 
However convert_move is unable to handle any BLKmode data at all: it has these
asserts near the beginning,

  gcc_assert (to_mode != BLKmode);
  gcc_assert (from_mode != BLKmode);

.. and the second one fires.


-- 

davek at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |davek at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[hjl dot tools at gmail dot com]

------- Comment #7 from hjl dot tools at gmail dot com  2010-09-23 00:20 -------
*** Bug 45753 has been marked as a duplicate of this bug. ***


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[ebotcazou at gcc dot gnu dot org]

-- 

ebotcazou at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|0                           |1
   Last reconfirmed|0000-00-00 00:00:00         |2010-09-23 06:10:37
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[rguenth at gcc dot gnu dot org]

------- Comment #8 from rguenth at gcc dot gnu dot org  2010-09-23 09:16 -------
Yes, my patch should be a no-op on this failure, it just avoids calling
convert_move if it would ICE anyway.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325

[Bug middle-end/45325] [4.6 Regression] target attribute doesn't work with -march=i586

[manu at gcc dot gnu dot org]

------- Comment #9 from manu at gcc dot gnu dot org  2010-09-23 15:53 -------
(In reply to comment #1)
> Well - obviously global typedefs keep their BLKmode.  The target attribute
> can't work this way - it's broken by design.

If it is broken by design, why not remove it before people start depending on
it and GCC has to add layer after layer of workarounds?


-- 

manu at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |manu at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45325