Created attachment 34599 [details]
testcase

Confirmed on all release branches.

Doesn't reproduce with a cross from x86_64-linux on trunk.

> ./xgcc -B. t.c -O3 -v -Wall
Reading specs from ./specs
COLLECT_GCC=./xgcc
COLLECT_LTO_WRAPPER=./lto-wrapper
Target: arm-none-linux-gnueabihf
Configured with: /space/rguenther/src/svn/trunk2/configure --target=arm-none-linux-gnueabihf --with-mode=arm --with-cpu=cortex-a9 --with-fpu=neon
...
t.c: In function ?fail?:
t.c:55:19: warning: ?val2? may be used uninitialized in this function [-Wmaybe-uninitialized]
   result += (val2 - val1) * interpFrac2;
                   ^
t.c:55:19: warning: ?val1? may be used uninitialized in this function [-Wmaybe-uninitialized]
t.c:55:27: warning: ?interpFrac2? may be used uninitialized in this function [-Wmaybe-uninitialized]
   result += (val2 - val1) * interpFrac2;
                           ^
t.c:56:19: warning: ?val3? may be used uninitialized in this function [-Wmaybe-uninitialized]
   result += (val3 - val2) * interpFrac3;
                   ^
t.c:56:27: warning: ?interpFrac3? may be used uninitialized in this function [-Wmaybe-uninitialized]
   result += (val3 - val2) * interpFrac3;
                           ^
t.c:54:27: warning: ?interpFrac1? may be used uninitialized in this function [-Wmaybe-uninitialized]
   result += (val1 - val0) * interpFrac1;
                           ^
COLLECT_GCC_OPTIONS='-B' '.' '-O3' '-v' '-Wall' '-mcpu=cortex-a9' '-mfpu=neon' '-marm' '-mtls-dialect=gnu'
 ./as -v -mcpu=cortex-a9 -mfpu=neon -meabi=5 -o /tmp/cciyGoik.o /tmp/cc4vjgVC.s
./as: line 106: exec: -v: invalid option
exec: usage: exec [-cl] [-a name] [command [arguments ...]] [redirection ...]


are you sure it isn't already fixed?

> 
> are you sure it isn't already fixed?

Don't think so. It ICEs for me with r220203.
I'm using the options -mfpu=neon -mfloat-abi=hard  -O3 -march=armv7-a.
The -mfloat-abi=hard is important, it doesn't reproduce without it

Ok, reproduce with

> ./cc1 -quiet t.c -O3 -mfloat-abi=hard -I include -march=armv7-a -mfpu=neon
t.c: In function ?fail?:
t.c:22:5: internal compiler error: in vect_get_vec_def_for_stmt_copy, at tree-vect-stmts.c:1598
 int fail ( const RMColorData * pInColor,
     ^
0xf89822 vect_get_vec_def_for_stmt_copy(vect_def_type, tree_node*)
        /space/rguenther/src/svn/trunk2/gcc/tree-vect-stmts.c:1598
0xf8991c vect_get_vec_defs_for_stmt_copy
        /space/rguenther/src/svn/trunk2/gcc/tree-vect-stmts.c:1624
0xf94dd8 vectorizable_operation
        /space/rguenther/src/svn/trunk2/gcc/tree-vect-stmts.c:4910
0xf9c3b3 vect_transform_stmt(gimple_statement_base*, gimple_stmt_iterator*, bool*, _slp_tree*, _slp_instance*)
        /space/rguenther/src/svn/trunk2/gcc/tree-vect-stmts.c:7273
0xfb0c26 vect_transform_loop(_loop_vec_info*)
        /space/rguenther/src/svn/trunk2/gcc/tree-vect-loop.c:6156
0xfc5daf vectorize_loops()
        /space/rguenther/src/svn/trunk2/gcc/tree-vectorizer.c:497
0xec3f97 execute
        /space/rguenther/src/svn/trunk2/gcc/tree-ssa-loop.c:295
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.

Must be related to

t.c:48:2: note: vect_recog_widen_shift_pattern: detected:
t.c:48:2: note: patt_42 = _41 w<< 4;
t.c:48:2: note: pattern recognized: patt_42 = _41 w<< 4;

t.c:48:2: note: additional pattern stmt: _41 = (unsigned short) _13;
...
t.c:48:2: note: type of def: 3.
t.c:48:2: note: mark relevant 4, live 0.
t.c:48:2: note: last stmt in pattern. don't mark relevant/live.
t.c:48:2: note: worklist: examine stmt: _41 = (unsigned short) _13;
...
t.c:48:2: note: type of def: 3.
t.c:48:2: note: mark relevant 4, live 0.
t.c:48:2: note: last stmt in pattern. don't mark relevant/live.
t.c:48:2: note: worklist: examine stmt: patt_42 = _41 w<< 4;

I will have a closer look.

In fact we recognize a widening shift but it gets dropped on the floor:

t.c:48:2: note: ------>vectorizing statement: result_15 = val0_14 << 4;
                
t.c:48:2: note: ------>vectorizing statement: _17 = val1_16(D) - val0_14;

huh?  The issue that we hit is that we vectorized the def of val0_14
with ncopies == 1 but now require two copies of it when vectorizing
val1_16(D) - val0_14.  So maybe the shift issue is unrelated.

Sofar we have

  vect__13.17_155 = MEM[(const Uint8 *)vectp_pCornerPoint0.15_153];
  _13 = *pCornerPoint0_48;
  vect__41.18_156 = [vec_unpack_lo_expr] vect__13.17_155;
  vect__41.18_157 = [vec_unpack_hi_expr] vect__13.17_155;
  val0_14 = (Sint32) _13;

Thus we widen a char load to an int but the vectorized version widens to
an unsigned short only.  That's probably because of the consumer

  result_15 = val0_14 << 4;

which was pattern detected as widening shift.  But unfortunately that isn't
the only one and we forgot

  _17 = val1_16(D) - val0_14;

which then results in the ICE.

Index: gcc/tree-vect-patterns.c
===================================================================
--- gcc/tree-vect-patterns.c    (revision 220205)
+++ gcc/tree-vect-patterns.c    (working copy)
@@ -1732,9 +1732,11 @@ vect_recog_widen_shift_pattern (vec<gimp
   if (TREE_CODE (oprnd0) != SSA_NAME || TREE_CODE (oprnd1) != INTEGER_CST)
     return NULL;
 
-  /* Check operand 0: it has to be defined by a type promotion.  */
-  if (!type_conversion_p (oprnd0, last_stmt, false, &half_type0, &def_stmt0,
-                          &promotion)
+  /* Check operand 0: it has to be defined by a type promotion and it
+     should be only used by the shift.  */
+  if (!has_single_use (oprnd0)
+      || !type_conversion_p (oprnd0, last_stmt, false, &half_type0, &def_stmt0,
+                            &promotion)
       || !promotion)
      return NULL;

> 
> Index: gcc/tree-vect-patterns.c
> ===================================================================
> --- gcc/tree-vect-patterns.c    (revision 220205)
> +++ gcc/tree-vect-patterns.c    (working copy)
> @@ -1732,9 +1732,11 @@ vect_recog_widen_shift_pattern (vec<gimp
>    if (TREE_CODE (oprnd0) != SSA_NAME || TREE_CODE (oprnd1) != INTEGER_CST)
>      return NULL;
>  
> -  /* Check operand 0: it has to be defined by a type promotion.  */
> -  if (!type_conversion_p (oprnd0, last_stmt, false, &half_type0, &def_stmt0,
> -                          &promotion)
> +  /* Check operand 0: it has to be defined by a type promotion and it
> +     should be only used by the shift.  */
> +  if (!has_single_use (oprnd0)
> +      || !type_conversion_p (oprnd0, last_stmt, false, &half_type0,
> &def_stmt0,
> +                            &promotion)
>        || !promotion)
>       return NULL;

That fixes the ICE in the testcase for me.
I'll give it a test on arm-none-eabi and an arm bootstrap

Author: rguenth
Date: Fri Jan 30 09:22:17 2015
New Revision: 220275

URL: https://gcc.gnu.org/viewcvs?rev=220275&root=gcc&view=rev
Log:
2015-01-30  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/64829
	* tree-vect-patterns.c (vect_handle_widen_op_by_const): Do
	not add a widening conversion pattern but hand off extra
	widenings to callers.
	(vect_recog_widen_mult_pattern): Handle extra widening produced
	by vect_handle_widen_op_by_const.
	(vect_recog_widen_shift_pattern): Likewise.
	(vect_pattern_recog_1): Remove excess vertical space in dumping.
	* tree-vect-stmts.c (vect_mark_stmts_to_be_vectorized): Likewise.
	(vect_init_vector_1): Likewise.
	(vect_get_vec_def_for_operand): Likewise.
	(vect_finish_stmt_generation): Likewise.
	(vectorizable_load): Likewise.
	(vect_analyze_stmt): Likewise.
	(vect_is_simple_use): Likewise.

	* gcc.dg/vect/pr64829.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/vect/pr64829.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-vect-patterns.c
    trunk/gcc/tree-vect-stmts.c

Fixed.