[Bug middle-end/37434] New: [4.4 Regression] ICE in extract_insn, at recog.c:2027

[Bug middle-end/37434] New: [4.4 Regression] ICE in extract_insn, at recog.c:2027

[tbm at cyrius dot com]

With current trunk (revision 140117):

(sid)1122:tbm@em64t: ..4.3-2008-09-08-r140117/gcc] ./cc1plus -quiet -m64 -msse3
~/framewave-Set.ii
....
domain/fwImage/include/Set.h: In function 'void SSE2::Set_AC4R_SETUP_I(long
long int __vector__&, XMM128&, const Fw16s*)':
domain/fwImage/include/Set.h:134: error: unrecognizable insn:
(insn 73 72 74 2
/usr/lib/gcc-snapshot/lib/gcc/x86_64-linux-gnu/4.4.0/include/emmintrin.h:596
(set (reg:V8HI 117)
        (vec_merge:V8HI (vec_duplicate:V8HI (const_int 0 [0x0]))
            (reg:V8HI 117)
            (const_int 2 [0x2]))) -1 (nil))
domain/fwImage/include/Set.h:134: internal compiler error: in extract_insn, at
recog.c:2027


-- 
           Summary: [4.4 Regression] ICE in extract_insn, at recog.c:2027
           Product: gcc
           Version: 4.4.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: tbm at cyrius dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug middle-end/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[tbm at cyrius dot com]

------- Comment #1 from tbm at cyrius dot com  2008-09-08 20:02 -------
Created an attachment (id=16257)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=16257&action=view)
Preprocessed code


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[pinskia at gcc dot gnu dot org]

-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pinskia at gcc dot gnu dot
                   |                            |org
          Component|middle-end                  |target
           Keywords|                            |ice-on-valid-code
            Summary|[4.4 Regression] ICE in     |[4.4 Regression] ICE in
                   |extract_insn, at            |extract_insn, at
                   |recog.c:2027                |recog.c:2027
   Target Milestone|---                         |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[pinskia at gcc dot gnu dot org]

------- Comment #2 from pinskia at gcc dot gnu dot org  2008-09-08 20:31 -------
Reducing ...


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[pinskia at gcc dot gnu dot org]

------- Comment #3 from pinskia at gcc dot gnu dot org  2008-09-08 21:35 -------
Reduced testcase:
typedef short __v8hi __attribute__ ((__vector_size__ (16)));
typedef long long __m128i __attribute__ ((__vector_size__ (16)));
__m128i Set_AC4R_SETUP_I( const short *val ) {
  short D2073 = *val;
  short D2076 = *(val + 2);
  short D2079 = *(val + 4);
  __v8hi D2094 = {D2073, D2076, D2079, 0, D2073, D2076, D2079, 0};
  return (__m128i)D2094;
}


-- 

pinskia 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         |2008-09-08 21:35:02
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl dot tools at gmail dot com]

------- Comment #4 from hjl dot tools at gmail dot com  2008-09-09 03:18 -------
It is caused by my patch:

http://gcc.gnu.org/ml/gcc-cvs/2008-05/msg00569.html


-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         AssignedTo|unassigned at gcc dot gnu   |hjl dot tools at gmail dot
                   |dot org                     |com
             Status|NEW                         |ASSIGNED
   Last reconfirmed|2008-09-08 21:35:02         |2008-09-09 03:18:41
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl dot tools at gmail dot com]

------- Comment #5 from hjl dot tools at gmail dot com  2008-09-09 04:23 -------
This patch fixes the crash:

--- i386.c.sse2 2008-09-08 21:17:15.000000000 -0700
+++ i386.c      2008-09-08 21:17:32.000000000 -0700
@@ -26886,7 +26886,7 @@ static void
 ix86_expand_vector_init_interleave (enum machine_mode mode,
                                    rtx target, rtx *ops, int n)
 {
-  enum machine_mode first_imode, second_imode, third_imode;
+  enum machine_mode first_imode, second_imode, third_imode, inner_mode;
   int i, j;
   rtx op0, op1;
   rtx (*gen_load_even) (rtx, rtx, rtx);
@@ -26899,6 +26899,7 @@ ix86_expand_vector_init_interleave (enum
       gen_load_even = gen_vec_setv8hi;
       gen_interleave_first_low = gen_vec_interleave_lowv4si;
       gen_interleave_second_low = gen_vec_interleave_lowv2di;
+      inner_mode = HImode;
       first_imode = V4SImode;
       second_imode = V2DImode;
       third_imode = VOIDmode;
@@ -26907,6 +26908,7 @@ ix86_expand_vector_init_interleave (enum
       gen_load_even = gen_vec_setv16qi;
       gen_interleave_first_low = gen_vec_interleave_lowv8hi;
       gen_interleave_second_low = gen_vec_interleave_lowv4si;
+      inner_mode = QImode;
       first_imode = V8HImode;
       second_imode = V4SImode;
       third_imode = V2DImode;
@@ -26935,7 +26937,9 @@ ix86_expand_vector_init_interleave (enum
       emit_move_insn (op0, gen_lowpart (mode, op1));

       /* Load even elements into the second positon.  */
-      emit_insn ((*gen_load_even) (op0, ops [i + i + 1],
+      emit_insn ((*gen_load_even) (op0,
+                                  force_reg (inner_mode,
+                                             ops [i + i + 1]),
                                   const1_rtx));

       /* Cast vector to FIRST_IMODE vector.  */

However the generated codes don't look good for 2 reasons:

1. By default, gcc won't generate inter-unit move between
GPR and SSE.
2. ix86_expand_vector_init doesn't take advantage of the
fact that the second half is the same as the first half

 __v8hi D2094 = {D2073, D2076, D2079, 0, D2073, D2076, D2079, 0};


-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ubizjak at gmail dot com


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl dot tools at gmail dot com]

------- Comment #6 from hjl dot tools at gmail dot com  2008-09-09 04:39 -------
This patch will disable SSE4 and SSE2 optimization if inter-unit
move is disabled or there are duplicates:

--- i386.c.sse2 2008-09-08 21:17:15.000000000 -0700
+++ i386.c      2008-09-08 21:36:38.000000000 -0700
@@ -26886,7 +26886,7 @@ static void
 ix86_expand_vector_init_interleave (enum machine_mode mode,
                                    rtx target, rtx *ops, int n)
 {
-  enum machine_mode first_imode, second_imode, third_imode;
+  enum machine_mode first_imode, second_imode, third_imode, inner_mode;
   int i, j;
   rtx op0, op1;
   rtx (*gen_load_even) (rtx, rtx, rtx);
@@ -26899,6 +26899,7 @@ ix86_expand_vector_init_interleave (enum
       gen_load_even = gen_vec_setv8hi;
       gen_interleave_first_low = gen_vec_interleave_lowv4si;
       gen_interleave_second_low = gen_vec_interleave_lowv2di;
+      inner_mode = HImode;
       first_imode = V4SImode;
       second_imode = V2DImode;
       third_imode = VOIDmode;
@@ -26907,6 +26908,7 @@ ix86_expand_vector_init_interleave (enum
       gen_load_even = gen_vec_setv16qi;
       gen_interleave_first_low = gen_vec_interleave_lowv8hi;
       gen_interleave_second_low = gen_vec_interleave_lowv4si;
+      inner_mode = QImode;
       first_imode = V8HImode;
       second_imode = V4SImode;
       third_imode = V2DImode;
@@ -26935,7 +26937,9 @@ ix86_expand_vector_init_interleave (enum
       emit_move_insn (op0, gen_lowpart (mode, op1));

       /* Load even elements into the second positon.  */
-      emit_insn ((*gen_load_even) (op0, ops [i + i + 1],
+      emit_insn ((*gen_load_even) (op0,
+                                  force_reg (inner_mode,
+                                             ops [i + i + 1]),
                                   const1_rtx));

       /* Cast vector to FIRST_IMODE vector.  */
@@ -26998,7 +27002,8 @@ ix86_expand_vector_init_general (bool mm
 {
   rtx ops[32], op0, op1;
   enum machine_mode half_mode = VOIDmode;
-  int n, i;
+  int n, i, h;
+  bool duplicated;

   switch (mode)
     {
@@ -27045,18 +27050,27 @@ half:
       return;

     case V16QImode:
-      if (!TARGET_SSE4_1)
+      if (!TARGET_SSE4_1 || !TARGET_INTER_UNIT_MOVES)
        break;
       /* FALLTHRU */

     case V8HImode:
-      if (!TARGET_SSE2)
+      if (!TARGET_SSE2 || !TARGET_INTER_UNIT_MOVES)
        break;

       n = GET_MODE_NUNITS (mode);
+      h = n >> 1;
+      duplicated = true;
       for (i = 0; i < n; i++)
-       ops[i] = XVECEXP (vals, 0, i);
-      ix86_expand_vector_init_interleave (mode, target, ops, n >> 1);
+       {
+         ops[i] = XVECEXP (vals, 0, i);
+         if (i >= h && !rtx_equal_p (ops[i], ops[i - h]))
+           duplicated = false;
+       }
+      if (duplicated)
+       break;
+
+      ix86_expand_vector_init_interleave (mode, target, ops, h);
       return;

     case V4HImode:


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl dot tools at gmail dot com]

------- Comment #7 from hjl dot tools at gmail dot com  2008-09-09 17:23 -------
A patch is posted at

http://gcc.gnu.org/ml/gcc-patches/2008-09/msg00727.html


-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                URL|                            |http://gcc.gnu.org/ml/gcc-
                   |                            |patches/2008-
                   |                            |09/msg00727.html


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[jsm28 at gcc dot gnu dot org]

-- 

jsm28 at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Priority|P3                          |P2


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl at gcc dot gnu dot org]

------- Comment #8 from hjl at gcc dot gnu dot org  2008-09-10 14:15 -------
Subject: Bug 37434

Author: hjl
Date: Wed Sep 10 14:14:28 2008
New Revision: 140231

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=140231
Log:
gcc/

2008-09-10  H.J. Lu  <hongjiu.lu@intel.com>

        PR target/37434:
        * config/i386/i386.c (ix86_expand_vector_init_interleave): Force
        the even element into register.
        (ix86_expand_vector_init_general): Don't use
        ix86_expand_vector_init_interleave on V16QImode and V8HImode
        if we can't move from GPR to SSE register directly.

gcc/testsuite/

2008-09-10  H.J. Lu  <hongjiu.lu@intel.com>

        PR target/37434:
        * gcc.target/i386/pr37434-1.c: New.
        * gcc.target/i386/pr37434-2.c: Likewise.
        * gcc.target/i386/pr37434-3.c: Likewise.
        * gcc.target/i386/pr37434-4.c: Likewise.
        * gcc.target/i386/sse2-set-v8hi-1a.c: Likewise.
        * gcc.target/i386/sse2-set-v8hi-2a.c: Likewise.
        * gcc.target/i386/sse4_1-set-v16qi-1a.c: Likewise.
        * gcc.target/i386/sse4_1-set-v16qi-2a.c: Likewise.
        * gcc.target/i386/sse4_1-set-v16qi-3a.c: Likewise.

Added:
    trunk/gcc/testsuite/gcc.target/i386/pr37434-1.c
    trunk/gcc/testsuite/gcc.target/i386/pr37434-2.c
    trunk/gcc/testsuite/gcc.target/i386/pr37434-3.c
    trunk/gcc/testsuite/gcc.target/i386/pr37434-4.c
    trunk/gcc/testsuite/gcc.target/i386/sse2-set-v8hi-1a.c
    trunk/gcc/testsuite/gcc.target/i386/sse2-set-v8hi-2a.c
    trunk/gcc/testsuite/gcc.target/i386/sse4_1-set-v16qi-1a.c
    trunk/gcc/testsuite/gcc.target/i386/sse4_1-set-v16qi-2a.c
    trunk/gcc/testsuite/gcc.target/i386/sse4_1-set-v16qi-3a.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/i386/i386.c
    trunk/gcc/testsuite/ChangeLog


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434

[Bug target/37434] [4.4 Regression] ICE in extract_insn, at recog.c:2027

[hjl dot tools at gmail dot com]

------- Comment #9 from hjl dot tools at gmail dot com  2008-09-10 14:18 -------
Fixed.


-- 

hjl dot tools at gmail dot com changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37434